I think personal agents could have some of the strongest lock-in effects we’ve ever seen in software.

Most software gets better as the product features improve. Personal agents have another advantage: they get better as they get to know you. Every conversation, task, preference and connection adds context. The more useful they become, the more you use them, and the more context they accumulate. This is a really powerful feedback loop.

I’ve tried a fair number of personal assistants but OpenAI’s Dots was the first to make this dynamic feel tangible. It’s “time-to-usefulness” metric for me was something like under a minute. What was special was how proactive it was - it found clever solutions to problems I had, sometimes digging through obscure corners of the internet to find a solve. The product experience of the dot isn’t something I’ve seen a lot written about but this is absolutely next level.

But the bigger question is what happens after a year of this. Or five.

An agent that knows your work, relationships, habits and preferences could become extraordinarily difficult to replace. Not because its underlying model is necessarily better, but because a new agent would have to start learning you all over again.

There’s an interesting tension here. The models themselves may become increasingly interchangeable, while the agents built on top become anything but.

My hunch is that you’ll likely become a lifetime customer of one agent, not many. (not unlike the Apple-Android lock in). There may be plenty of specialized tools, perhaps, but one agent that knows you and coordinates the rest.

The battle to become that agent is going to be fierce. The battle to get you to leave may be even harder. This is the window to gain ground.

-Tara

THE DOWNLOAD

SpaceX acquires nationwide spectrum to challenge top mobile carriers

On October 8, SpaceX agreed to acquire nationwide 800 MHz spectrum from Grain Management, moving Starlink closer to becoming a full-fledged mobile operator competing with AT&T, Verizon and T-Mobile.

The announcement came just days after the FCC authorized SpaceX to deploy 15,000 next-generation direct-to-device satellites. Together, the developments could fundamentally change how mobile networks are built.

SpaceX already connects ordinary phones through satellites, largely in partnership with existing carriers. But satellite coverage has limitations, particularly indoors and in dense urban environments. The newly acquired low-band spectrum can penetrate walls and buildings more effectively, complementing Starlink’s existing satellite frequencies. SpaceX’s ambition is a hybrid network that connects phones from both space and the ground.

The implications are significant. Traditional mobile operators have spent decades building networks of cell towers, acquiring spectrum and establishing customer relationships. SpaceX controls something they don’t: its own launch infrastructure and a rapidly expanding satellite constellation. Adding terrestrial spectrum could give it an entirely different cost structure and network architecture.

Wall Street took notice. According to the Financial Times, approximately $62 billion was erased from the combined market value of AT&T, Verizon and T-Mobile following the announcement.

There are substantial hurdles. The acquisition still requires FCC approval, and SpaceX will need terrestrial infrastructure, network capacity and competitive economics to challenge established carriers. Satellite connectivity cannot simply replace dense urban cellular networks overnight.

But SpaceX is no longer positioning Starlink solely as a solution for remote connectivity or cellular dead zones. It wants to compete for the entire mobile customer relationship.

Why it matters: SpaceX could become the first major mobile operator built around an integrated satellite-and-terrestrial network, potentially challenging the economics of an industry dominated by a handful of incumbents.

Further reading: Grain Management · Ars Technica

OpenAI releases hundreds of mathematical manuscripts, then withdraws three

On October 6, OpenAI released 722 mathematical manuscripts generated by an unreleased AI model, spanning number theory, geometry, algebra and mathematical physics. A day later, three were withdrawn after a sign error invalidated a key construction. Another 14 required corrections or clarifications.

The scale is extraordinary. The repository now contains 719 manuscripts across 372 families of results. OpenAI says roughly 42% of the top-line results have been formalized in Lean, a proof assistant that can mechanically check mathematical arguments.

The response from mathematicians has been mixed. Some see potentially important work. Others have raised questions about novelty, attribution and the burden of evaluating hundreds of unfamiliar arguments. Establishing which results are new, correct and significant could take years.

The withdrawals are not, by themselves, evidence that the larger effort has failed. Mathematics has always advanced through conjecture, error and correction. What’s different is the rate at which candidate results can now be produced.

We could soon have more proposed theorems than mathematicians can reasonably evaluate, with verification, interpretation and scientific judgment becoming the scarce resources.

Why it matters: AI may be making mathematical discovery dramatically cheaper while moving the bottleneck to determining what is actually true, new and worth knowing.

Further reading: OpenAI research repository · The Verge

Google agrees to finance 890 MW of additional nuclear power from existing reactors

Google and Constellation Energy announced a 20-year agreement to add 890 megawatts of nuclear generating capacity to the PJM electricity grid. None of it requires building a new reactor.

Instead, Constellation will upgrade equipment and improve efficiency at 11 existing nuclear units across six sites. The program represents more than $4.3 billion in planned investment, with the first additional capacity expected in 2028. A separate 15-year agreement covers 2,700 MW of existing generation.

As AI data centers strain electricity grids, technology companies are increasingly investing directly in energy infrastructure. Much of the attention has gone to next-generation reactors, restarting shuttered plants or building new nuclear facilities

Improving existing reactors offers a different path. The plants already have operating teams, grid connections and established infrastructure. Upgrades still require engineering work and regulatory approvals, but they can potentially deliver new capacity faster than greenfield projects.

Why it matters: In the near term, some of the most valuable new nuclear capacity may come from reactors we’ve already built.

Further reading: Constellation Energy · Reuters

GlobalFoundries signs $2 billion agreement to manufacture silicon interposers for TSMC

On October 8, GlobalFoundries announced a five-year, $2 billion agreement to manufacture silicon interposers for TSMC’s CoWoS advanced packaging ecosystem. Production will take place in Malta, New York, with volume manufacturing expected to ramp in the first half of 2028.

Why is this important? AI chip performance increasingly depends on how efficiently processors communicate with memory and one another, not simply how many transistors can fit on a chip. Advanced packaging has become a critical bottleneck in scaling AI compute. Silicon interposers are a key part of that system.

It’s an interesting turn for GlobalFoundries. The company abandoned the race to manufacture leading-edge logic chips in 2018, leaving TSMC, Samsung and Intel to pursue smaller transistor technologies. Now it’s entering a strategically important part of the AI supply chain without competing at the smallest process nodes.

As AI systems grow more complex, specialized manufacturing capabilities are becoming increasingly valuable. Memory bandwidth, interconnects, thermal management and packaging can be just as consequential as transistor density.

Why it matters: The race to scale AI compute is making parts of the semiconductor industry beyond leading-edge logic newly strategic.

Further reading: GlobalFoundries · Tom’s Hardware

ALSO

DeepSeek reportedly raises at least $12 billion in Tencent-backed financing. DeepSeek is close to securing at least ¥80 billion ($12 billion), with Tencent and battery giant CATL among its largest reported backers. The final round could approach ¥100 billion ($15 billion), making it one of the largest private AI financings globally. (Reuters)

Lumentum says optical-component capacity is nearly sold out through 2029. Demand for lasers and photonic components used to connect AI processors and data centers is stretching supply. CEO Michael Hurlston says the company cannot meet as much as 70% of demand for certain products through next year, while other product lines face significant shortages through 2028.

Researchers demonstrate the first functioning nuclear clocks. Independent teams in Austria and China have built working clocks based on a transition inside the thorium-229 nucleus, rather than the electronic transitions used by atomic clocks. The devices don’t yet outperform the best optical atomic clocks, but they establish a new approach to precision timekeeping that could eventually improve navigation, communications and experiments probing fundamental physics. (Second Nature paper)

Biohub, NIH, and research partners commit $1.8 billion toward AI-ready biological data. Biohub, the Department of Energy, NIH and industry partners are expanding an effort to build large-scale, interoperable biological datasets for predictive AI models. As models become better at biological prediction, the quality and scale of experimental data may matter as much as the models themselves.

Anthropic discloses AI agents taking unintended actions on live govt websites. During testing, Claude submitted a fabricated tip through a Philadelphia police homicide website, circumvented access restrictions and interacted with sensitive government forms.

Anthropic releases Claude Haiku 5.5. Its latest small model targets high-volume, latency-sensitive work, with Anthropic reporting substantially lower running costs than its predecessor.

Reflection and Mistral announce new frontier-scale open-weight models. Reflection’s 501-billion-parameter Beam and Mistral’s trillion-parameter Large 4 challenge the assumption that the strongest models must remain proprietary. Both companies are promising publicly available weights, though neither had completed that release at the time of the announcements.

DEEP DIVE FROM THE REVIEW

— Fred Parietti, CEO, Multiply Labs

Some of the most advanced medicines in the world are still manufactured by hand. In cell therapy, technical operators work through dozens of delicate steps inside cleanrooms: moving bags of cells between instruments, connecting tubing, drawing samples, adding reagents, counting cells. The product is biological, the process is unforgiving, and a mistake can mean losing a patient’s dose.

It is extraordinary science running on surprisingly manual infrastructure.

That is the problem Multiply Labs is solving.

When I first met founder Fred Parietti during his PhD at MIT almost 12 years ago (!), he was building robots with extra limbs designed to extend what the human body could do. The machines were wonderfully complicated: multiple robotic systems moving together, each action dependent on the next.

Years later, his co-founder, pharmaceutical scientist Alice Melocchi, introduced him to the pain of manufacturing medicine.

Fred had imagined pharmaceutical production would look something like semiconductor manufacturing. Think highly automated facilities producing enormously complex products with extraordinary precision. Instead, in advanced biologics, he found skilled scientists still performing painstaking manual work.

Multiply Labs grew out of a deceptively simple question: what if robots could do it instead?

Not by asking pharmaceutical companies to throw away the processes they had spent years developing. Multiply builds its robots around the instruments, consumables and workflows already used inside pharmaceutical manufacturing facilities.

Its robotic clusters bring multiple arms and manufacturing instruments together inside a closed system. The robots move materials between machines, manipulate bags and syringes, take samples, perform transfers and execute the sequence of steps that would otherwise require skilled human operators. Multiply says its systems can automate more than 50 unit operations and increase throughput per square foot by as much as 100x.

A single manual cell therapy line produces an average annual throughput of roughly 10,000 doses. The waitlist for the medicine is orders of magnitude more.

In pharmaceuticals, changing the way a therapy is manufactured is not like swapping out a piece of software. Manufacturing processes are validated, regulated and deeply embedded into how a drug is produced. The winning automation platform therefore may not be the one that asks the industry to redesign everything around it. It may be the one capable of adapting robotics to the messy reality that already exists.

The opportunity is much larger than replacing labor.

Our thesis was that as medicine becomes more complex, manufacturing itself becomes a massive bottleneck.

And if that is true, the factory starts to matter almost as much as the drug.

Multiply has spent the past several years turning that thesis into real infrastructure. Its announced pharmaceutical collaborators include AstraZeneca, Legend Biotech and Kyverna Therapeutics, and its systems are now being deployed inside pharmaceutical manufacturing environments.

Step into Multiply Labs’ workshop in Dogpatch and it feels familiar to me: part hardware lab, part robotics shop. Engineers work around robotic arms and test fixtures; dexterous mechanical hands sit on workbenches beside half-built systems.

But Multiply is going industrial.

You can see it in the mix of people on the floor: young robotics engineers working alongside seasoned operators like Enrico Antonio Racca, previously Head of Supply Chain and Manufacturing at Ferrari’s Formula One, thinking about supply chains, assembly, and reliability.

Formula One to biopharma sounds like an unusual jump. But the problem is surprisingly similar: taking an extraordinarily complex, high-precision machine and figuring out how to build it repeatedly, reliably, and at scale.

Today, Multiply announced a $75 million Series B, led by serial biotech entrepreneur Dr. Patrick Soon-Shiong and NantWorks, with AstraZeneca, Lingotto, Teradyne and Strange Ventures among the new investors, alongside returning investors including Casdin Capital, Lux Capital and Founders Fund. The company plans to use the capital to increase robot production and move from clinical-stage deployments toward commercial manufacturing.

And the ambition is expanding.

Cell therapy is the first proving ground, but the same manufacturing problem exists across a much broader class of biologic medicines. The future vision of Multiply Labs is to extend the platform into other advanced biologics.

Over the last few decades, enormous amounts of technology and capital have gone into finding better medicines. The next generation of those medicines will be more personalized, more biological and often much harder to manufacture.

Multiply Labs is betting that the bottleneck isn’t in discovery, but in manufacturing them so they get into the hands of those who need them.

The venture world has been feeling particularly unhinged lately.

This week, the Factory–Cognition software factory rivalry turned into a full-blown VC bitch fight. Factory accused a board adviser of sharing confidential information before joining Cognition; he denies it. Then Vinod Khosla jumped in to call Factory a “struggling second tier competitor,” while Keith Rabois (his partner at Khosla Ventures) publicly defended Factory. Khosla Ventures happens to own stakes in both companies. TechCrunch has the whole glorious mess.

Valuations are multiplying at eyewatering rates as VCs fall over themselves to get into the latest thing. TypeSafe emerged from stealth at a reported $200 million valuation; within days, investors were discussing a round above $10 billion. Instinct went from $2.5 billion to $10 billion in a month after it went viral on VC twitter, while still largely in early access.

But: the hot new entrants all have impressive copycats rolled out by both new startups and big companies within days, if not weeks.

Is there a moat? Maybe. Maybe no one cares. At the moment, the market seems perfectly happy to jump in first and figure that part out later.

-Tara

THE DOWNLOAD

Anthropic’s leaked IPO prospectus shows what it costs to run a frontier lab

Anthropic generated $4.6 billion in revenue in 2025, up from roughly $400 million the year before, according to its confidential IPO prospectus reviewed by Reuters. It also spent $7.33 billion on compute and infrastructure and posted an operating loss of just over $8 billion.

The more striking number is further out: Anthropic has committed roughly $518 billion to future cloud, compute and infrastructure spending. Most of those commitments are non-cancellable or payable regardless of actual utilization, according to the same prospectus reporting.

Its commercial relationships are equally intertwined. Nearly half of 2025 sales flowed through Amazon and Google’s cloud marketplaces — companies that are simultaneously Anthropic investors, infrastructure suppliers, distributors and competitors. Reuters reported that Amazon and Google collected hundreds of millions of dollars in distribution fees on those sales.

The reported $42 billion net loss is less useful than it looks because roughly $34 billion came from non-cash accounting charges tied to financing instruments; the operating loss gives a cleaner picture of the underlying business, as MarketWatch noted.

Why it matters: Frontier AI is beginning to look like software demand sitting on top of infrastructure economics closer to heavy industry.

OpenAI’s DevDay fills in the agent stack

OpenAI used DevDay to package more of the productization required to turn models into persistent software agents.

• The new Agents API includes hosted execution, memory, tools and multi-agent orchestration. Computer use is built directly into the API. Codex tasks can keep running in the cloud.

• A new Decisions API uses Luna to classify inputs, route requests or choose from predefined actions rather than generate open-ended text, according to OpenAI’s DevDay recap.

At the same time, inference is getting cheaper and faster. OpenAI says GPT-6.1 Sol costs one-fifth as much as Astra at standard token pricing, while its Ultrafast tier can increase generation speed substantially in the API and Codex, as detailed in the DevDay developer announcements.

Taken together, the architecture is becoming more modular: frontier reasoning for hard problems, smaller models for routing and decisions, plus memory, execution and tool access around them.

Why it matters: The competitive layer is moving beyond the model toward the runtime that decides when, where and how different kinds of intelligence get used.

DeepMind watermarks proteins that remain detectable after synthesis

DeepMind has demonstrated a watermark for AI-generated proteins that survives the jump from a digital sequence to a physically synthesized molecule.

SynthID Bio, published in Nature, embeds a detectable signal into generated protein sequences and predicted structures. Researchers synthesized watermarked protein binders against SARS-CoV-2 RBD, VEGF-A and PD-L1 and found that watermarking did not materially reduce binding performance.

The sequence watermark is inserted during generation by subtly changing amino-acid choices. A separate method adds a signal to predicted 3D structures using a fine-tuned AlphaFold 3 model, as described in the paper.

The obvious applications are provenance and screening: identifying AI-designed sequences entering synthesis pipelines or preventing synthetic structures from quietly contaminating biological databases.

It is not tamper-proof. Independent Nature coverage notes that the marker can be intentionally removed, and DeepMind describes the work as a proof of concept rather than a complete biosecurity system.

Why it matters: As generative models begin producing molecules rather than media, provenance becomes a physical-world problem.

Decision models are multiplying. Are they truly a new primitive or just a feature to be rolled in?

Jev launched two weeks ago with a simple premise: many jobs inside an agent do not require a language model to generate language.

A tool router, fraud check or escalation decision often comes down to choosing among a fixed set of actions. Jev calls models optimized for this decision models: they accept a state and a set of choices, then return calibrated probabilities over the allowed answers rather than free-form text. The launch quickly sparked copycats and debate over whether this is really a distinct model class, as The Wall Street Journal reported.

The idea has spread quickly.

Cloudflare released Clef and Clef-flash, open-weight decision models designed for the same class of bounded choices. Jeff offers a small Jev-compatible implementation. PostHog’s Jeeves adds a reasoning step before producing a structured decision. OpenAI’s newly announced Decisions API now handles classification, routing and action selection inside its own agent platform.

The counterexample may be the most revealing. Privatemode showed that a regular GLM model can be constrained into a Jev-like decision system in a single forward pass without purpose-training a separate model. That suggests the workflow may matter more than the architecture.

The deeper shift may simply be that agents are starting to separate open-ended reasoning from fast bounded decisions, instead of using a large autoregressive model for both.

ALSO

AMD will acquire World Labs for approximately $8.2 billion. AMD says World Labs’ model research will help shape future hardware, software and systems around emerging workloads. Fei-Fei Li will become AMD’s executive vice president and chief scientist if the deal closes, according to AMD’s announcement.

OpenAI and Synopsys are building a specialized model for chip design. GPT-Synopsys will combine OpenAI models with Synopsys’ EDA tools; OpenAI is also licensing those tools for development. Synopsys announcement

Apple is tightening Full Disk Access as agents become more autonomous. Apple says the permission can expose files, mail, messages and browsing history, and that increasingly capable agents make that access substantially riskier. Future macOS controls will require more explicit user action before granting it. Apple developer notice

Microsoft is building a multimodal world model of biology. Quine connects genomics, proteins, chemistry, cell state and imaging with scientific tools and wet-lab feedback. Researchers from Harvard Medical School and the Broad Institute participated, but Microsoft describes Quine as an early-stage research effort and there is not yet a peer-reviewed independent evaluation. Microsoft Research

A friend who works at one of the frontier labs said something to me over drinks the other night:

“I’m long fun, short productivity.”

His argument was basically that the models have already gotten very good at productivity. Coding, research, writing, analysis…all of this gets faster and cheaper from here.

So where does the value go?

Fun, he said. Creativity. Design. Entertainment.

I laughed, but I’ve kept thinking about it.

The velocity of startups right now feels unlike anything I’ve seen before. Companies are being built faster, with fewer people, because increasingly large parts of the work are becoming cheap.

That should create more companies.

It might also make many of them less differentiated (i.e. easier to dupe). Or go for cheaper alternatives even if you’re a happy customer.

If everyone gets access to roughly the same extraordinary productivity, productivity starts to look less like an advantage and more like table stakes.

Maybe that’s why Alexandr Wang’s language around Muse stuck with me this week. His vision is an agent that clears away the planning, emails, calls and logistics until what remains for the human is “the wanting and the dreaming.”

It’s a very Silicon Valley way of describing the future. But it’s also remarkably close to what my friend was saying over drinks.

Humans seem perfectly capable of finding new ways to spend the time technology saves us, e.g. short-form drama apps went from roughly 370 million downloads in Q1 last year to around 850 million this year.

Productivity may become a race toward cheaper execution. Entertainment is a race toward stronger preference.

The supply of things to watch, play and listen to is about to explode. But preference is harder to manufacture.

Desire is scarce. Attention is scarce. Culture is scarce.

That might be the strongest case for being long fun.

-Tara

THE DOWNLOAD

01 — Google prepares to test TPUs in orbit

Google’s Project Suncatcher is about to go into orbit.

On October 1, a refrigerator-sized satellite is scheduled to launch with four of Google’s TPUs. The goal is to test whether the same chips used in terrestrial data centers can survive launch, radiation and the thermal extremes of low Earth orbit.

The idea sounds less absurd when you look at what is happening on Earth. Morgan Stanley now estimates a 30 to 40% gap between projected U.S. data-center power capacity and demand by 2028 (roughly six New York Cities’ worth of baseload electricity).

Suncatcher is Google asking whether some of that compute could eventually go somewhere power is abundant: orbit, with near-continuous sunlight.

The catch is heat. The first prototype can run its TPUs for only about 15 minutes at a time before shutting down so its radiators can catch up.

So this is nowhere near an orbital data center. But AI infrastructure has become constrained enough by power, cooling, land and transmission that putting compute in space has moved from thought experiment to flight hardware.

Why it matters: The search for AI infrastructure is beginning to produce entirely new ideas about where computing should physically live, and what type of computing could actually thrive in space.

02 — Anthropic says Claude found a previously uncharacterized enzyme system

Anthropic has published the first major result from the physical biology lab it revealed earlier this month.

The company gave Claude a broad task: search a massive DNA database for interesting reverse transcriptases. Roughly 950 agents spent 21 hours and 210 million tokens searching more than 200,000 examples before surfacing an unusual system Anthropic calls array-associated reverse transcriptases, or ART.

Scientists then took the hypothesis into the lab and confirmed that part of the system produces distinct short RNAs.

They still don’t know what ART actually does, and the work needs further validation. The reverse transcriptase itself was also not previously unknown.

The interesting part is the loop: hundreds of agents search an enormous hypothesis space, rank the interesting anomalies, and hand a tiny subset to humans for physical experimentation.

Why it matters: AI-driven science is starting to connect large-scale computational search directly to experiments in the real world.

03 — Meta suddenly has a consumer AI stack

Meta is suddently crushing consumer AI.

Their new personal AI agent app Muse has gone from launch to the top of the U.S. app stores in a matter of weeks. Beyond the app, Muse is being pushed into its glasses and other devices, and connected to services across the web like Whatsapp.

Like it’s irresistibly cute Muse charm, a tiny screen-based AI companion that looks like a less creepy Labubu sans teeth and designed to hang from a bag or keychain. It has a screen, fingerprint sensor and real-time voice. Meta says it plans to ship it later this year.

At the other end of the hardware spectrum, Meta unveiled VR glasses weighing roughly 100 grams by moving much of the compute, battery and storage into a separate puck. Apple’s Vision Pro weighs many times more. Meta’s answer is less elegant in one sense (there’s a tether) but potentially much more wearable.

The interesting part is the range of experiments. Meta is testing whether a personal AI belongs in your phone, your glasses, a VR headset, or literally hanging off your bag.

Why it matters: Meta is no longer just adding AI to its existing products. It is beginning to build an ecosystem around the idea that your AI should follow you across them.

04 — TSMC plans a dedicated validation loop for advanced packaging

TSMC is building infrastructure in Kaohsiung to speed up a less visible part of semiconductor manufacturing: qualifying new equipment before it touches production.

The company plans a modular “Mini-Loop” where suppliers can test equipment and process integrations for CoWoS, the advanced packaging technology used to combine AI accelerators with high-bandwidth memory.

The target is to make validation 25% to 50% faster.

That is a useful signal about where the semiconductor bottleneck is moving. Making an AI chip increasingly depends not just on fabricating the silicon, but on packaging multiple dies, integrating memory and qualifying an entire ecosystem of increasingly specialized equipment.

Why it matters: Scaling AI chips is becoming as much about speeding up the manufacturing system around the wafer as building more wafer capacity.

05 — Inference providers surge as frontier models get cheaper

OpenAI says GPT-6 Sol and Luna are priced 50% below their GPT-5.6 predecessors, while Anthropic says Opus 5.5 costs about 40% less to run than Opus 5 on typical workloads.

At the same time, inference providers are being valued more highly than ever. Modal is reportedly in talks at around $15 billion, roughly triple its valuation four months ago; Baseten at around $26 billion. Fireworks could target $30 billion in a new round, while Fal has discussed $15–20 billion.

The timing is interesting. Replit CEO Amjad Masad said this week that the company is probably using less open-source AI than it was in January because OpenAI has cut prices so aggressively.

That puts some pressure on the original pitch for inference providers: open models were often cheaper than the frontier labs. But demand for the infrastructure around models is still growing quickly. Fireworks says it has passed $1 billion in annualized revenue, with 95% of the tokens it serves coming from models specialized for particular customers or tasks.

ALSO

Amazon blocks Meta’s Muse shopping agent

Amazon has stopped Muse from shopping on its platform, citing unauthorized automated access and security concerns. Meta can build an agent capable of navigating the web; whether the web allows that agent to act is becoming a different problem. The Verge

AI agents are starting to solicit scientists

Researchers are beginning to receive unsolicited requests from AI agents seeking data, collaborations and paid work. One statistician interviewed by Nature discovered that an apparently normal request for research data had come from an agent rather than another researcher. Nature

Intrinsic open-sources part of its industrial-robotics stack

Alphabet’s Intrinsic released core pieces of its robotics platform as open source, including ROS-compatible services, hardware-independent real-time control and digital-twin infrastructure. Physical AI is starting to get an open software layer underneath the models. Intrinsic

AI startups (and funds) pumping and jumping

Sequoia’s Pat Grady shared an investor update saying that across seven recent Sequoia investments, the firm’s average entry valuation was $110 million; within the next round, later investors were paying an average $3.4 billion. He also shared that Typesafe reached $100M revenue (unclear if annualized) in just 7 days, after two years of building in stealth. Another Sequoia partner has recently come under fire for the “dual pricing” tactics. His response: “other investors are willing to pay a high price for a hot company at multiples above”.

In other reports, similar “momentum” rounds are appearing elsewhere, like Ineffable Intelligence reportedly jumped from roughly $55 million pre-money to $4 billion within weeks even without a product.

There may be nowhere to plug in half the GPUs

Morgan Stanley estimates U.S. data centers could need roughly 68 GW of new power through 2028. Projects already under construction and contracted grid capacity account for only about 30 GW (less than half)

Put differently: if the projected compute actually arrives, a huge share of it may not have power waiting for it. Some interesting pundits from market researchers in here.

There is a strange asymmetry emerging.

We are building agents that can reason about software, discover vulnerabilities, and write working exploits at machine speed.

Underneath them sits an enormous installed base of software and infrastructure built around a very different assumption: that the thing attacking it is a human.

That infrastructure is not simply “old”. It is layered, interconnected and slow to change. A vulnerability in one layer can expose another. Patching one system can break something somewhere else. Power grids, water systems and telecommunications networks can contain equipment that remains in service for decades.

We are vulnerable.

A new Council on Foreign Relations report describes Chinese state-linked actors quietly maintaining access inside American telecommunications, energy, water and transportation infrastructure. Cyber journalist and investor Nicole Perlroth calls them CCP “sleeper cells” that have infiltrated American systems for over 5 years.

Note: this isn’t necessarily to attack it today, but to preserve the option of disrupting it later.

I repeat: we are vulnerable. Three security researchers at white hat security group Hacktron recently demonstrated that they successfully breached an OpenAI employee’s internal coding environment and accessed its codebase through shoddy identity management.

H*ly sh*t. And shockingly, how seemingly cheap it was to do.

Hacktron says its broader two-month research campaign cost less than $3,000 in model tokens. Once the team had designed and built the basic workflow, adapting an exploit to another target could take a day or two.

This may be the deeper security problem of the AI era.

We are introducing a new kind of intelligence into systems designed around an older kind of intelligence.

And I don’t think this is confined to cybersecurity.

Anthropic has quietly built a physical biology lab in the Bay Area, hiring scientists and giving Claude access to real experimental workflows.

We’re also seeing homebrew experiments for cures. An independent researcher named Douglas Yao recently showed off a compound he says was designed with the help of ChatGPT and synthesized in a chemistry lab he built in his garage.

The compound, PAC-3310, is an experimental molecule targeting a receptor implicated in schizophrenia. Yao has published early mouse data, but it has not been independently validated or tested in humans.

We may now be entering one in which intelligence is cheap, abundant and operating at machine speed.

The problem is that the world it acts on does not move at that speed.

Human biology, in particular, evolves over generations.

-Tara

THE DOWNLOAD

01 — AI agents are crossing from cybersecurity tests into real systems

Google says a Gemini model accessed the systems of three real companies during a cybersecurity evaluation earlier this year.

The model was participating in a capture-the-flag exercise designed to simulate attacks on a fictional company. The environment was supposed to be isolated, but a configuration error left it connected to the internet. Gemini found its way into live systems instead: in one case by guessing credentials, and in two others by finding credentials in public repositories. Google says the model stopped once it recognized that it had reached real infrastructure.

This lands alongside a broader jump in AI cybersecurity capability.

Earlier this month, OpenAI designated GPT-6 Astra as the first model to reach its “Critical” cybersecurity capability threshold. With the right tools and access, OpenAI says Astra can find previously unknown vulnerabilities and develop ways to exploit them across many well-protected systems without a person guiding each step.

And this week, a three-person security team at Hacktron published how it used Claude models to find and develop an exploit against software underlying OpenAI’s community forum, then chained it with an identity flaw to reach OpenAI employees’ ChatGPT and Codex accounts. To demonstrate the access without reading internal code, the researchers had an employee’s Codex open a harmless pull request inside OpenAI’s internal monorepo.

Why it matters: AI is compressing the time and expertise required for sophisticated software exploitation. The security question is increasingly whether the systems underneath it can adapt at the same rate.

02 — CXMT begins mass production of fifth-generation DRAM

China’s CXMT says its fifth-generation DRAM platform has entered mass production.

The company says the new platform uses quadruple patterning to reach an 11.95 nm active-area half-pitch and can produce at least 50% more dies per wafer than its previous generation. CXMT also showed two 24Gb LPDDR5X products based on the process, with 50% greater capacity than comparable previous-generation parts.

The notable part is how it was manufactured.

CXMT does not have access to the most advanced EUV lithography equipment used at the leading edge. Instead, it is pushing existing lithography further through more complex patterning, process engineering and manufacturing optimization.

That does not mean CXMT has caught Samsung or SK Hynix. CXMT itself describes the platform as approaching the industry’s most advanced mass-production nodes, but process measurements are not directly comparable between manufacturers, and there is still limited independent information about yield, cost or sustained production volume.

But it is another example of a broader pattern: semiconductor constraints do not necessarily create a fixed technological ceiling. They create an engineering problem.

Why it matters: The semiconductor race is not only about access to the best machines. It is also about how much performance and manufacturing efficiency can be extracted from the machines you already have.

03 — TypeSafe releases Jev, a model for structured decisions

TypeSafe has released Jev, its first “System One Model”, designed to return structured decisions rather than conversational text.

Instead of generating a paragraph and asking software to interpret it, Jev takes unstructured information and returns typed outputs: choices, scores or probability distributions that software can act on directly.

TypeSafe trained the model using a method it calls Reinforcement Learning for Calibrated Decisions, or RLCD, and describes Jev as a fast, low-cost decision layer for software. Those performance and calibration claims are still primarily company-reported and will need broader independent testing.

But the underlying idea is useful.

For many machine-to-machine tasks, language may be an unnecessary intermediate step. An agent choosing a tool, routing a request, classifying an event or deciding between actions does not necessarily need to explain itself in prose first.

Jev is less like a chatbot than a decision primitive inside software.

Why it matters: Language may remain the interface between humans and AI without remaining the interface between AI systems themselves.

04 — Xiaomi livestreams MiMo-V2.6 reinforcement learning

Xiaomi’s MiMo team publicly streamed the reinforcement-learning (RL) runs for two unreleased models, MiMo-V2.6-Pro and MiMo-V2.6-Flash.

The live dashboard exposes reward curves, rollout counts, training-step timing, running compute costs and intermediate coding evaluations while the models are still being trained. Xiaomi has not yet released either model or announced final specifications.

This is pretty ballsy. Model companies normally show the finished artifact: a benchmark table, model card or API.

Xiaomi is showing part of the RL process that produces it, going whole hog behind the idea of open source.

RL is becoming increasingly important. Model development no longer ends when pretraining finishes. Reinforcement learning, environments, synthetic data, evaluation loops and repeated post-training runs are becoming substantial engineering systems of their own.

Why it matters: As architectures and base capabilities converge, more of the differentiation between models may come from what happens after pretraining.

05 — OpenAI acquires computational-photography startup Glass Imaging

OpenAI has acquired Glass Imaging, a computational-photography startup founded by former Apple engineers Ziv Attar and Tom Bishop, in a deal reported by The Wall Street Journal at more than $300 million.

Glass builds AI-based imaging technology designed to extract more information from existing camera hardware. Its neural image-processing system is intended to correct lens aberrations and sensor imperfections and improve image quality from small cameras.

OpenAI has not said exactly what it intends to do with the company, so it is too early to read the acquisition as evidence of a specific camera or device strategy.

But this is interesting in the context of OpenAI’s broader push into consumer hardware.

Most discussion around AI hardware has focused on the interface: what might replace the phone, screen or keyboard. Imaging sits a layer further down, at the point where a device captures information from the physical world in the first place.

For AI systems that can see and act in the world, sensing becomes part of the stack.

Why it matters: OpenAI’s hardware push is beginning to extend below the user interface into the technologies that capture and process information from the physical world.

ALSO

Qwen releases Qwen3.8-Omni-Flash

Qwen released Qwen3.8-Omni-Flash, a multimodal model that can work across text, images, audio and video while also using tools.

Inference hardware is reorganizing around memory

A useful IEEE Spectrum survey of the new inference hardware landscape shows how much chip architecture is beginning to reorganize around memory movement rather than raw compute.

Z.ai says GLM helped build its own inference infrastructure

Z.ai says a GLM-powered infrastructure agent helped adapt and optimize its own model for production on a cluster of more than 100,000 Chinese-made AI accelerators. According to the company, work that would previously have taken experienced infrastructure engineers weeks was compressed dramatically, with the system helping debug kernels, optimize performance and bring the model into production.

Researchers show LLMs already contain a signal for choosing skills

As agents accumulate more skills, they run into a basic problem: how do you decide which one to use without stuffing descriptions of every available skill into the context?

A new paper, “The Router Within,” finds that a frozen LLM already contains a useful routing signal in its internal states. The researchers built a lightweight system called Gavel that reads that signal and uses it to shortlist the appropriate skills, outperforming substantially larger retrieval systems on several benchmarks.

The mathematicians are having a week.

OpenAI claims to have solved a version of the Navier–Stokes problem, one of seven famous, highly difficult mathematical challenges defined by the Millennium Prize. I remember them walking past the Millennium Prize challenges, hung up in the main halls of MIT. They almost served like an everyday reminder of the pursuit of greatness.

To see problems that once felt like monuments to human ambition potentially fall so quickly… and amid questions about how the work was done… is an odd thing to sit with.

In response, influential mathematician Terence Tao joined 24 fellow medallists in warning that the race to solve problems is losing sight of what mathematics is for. Getting an answer and developing understanding are different achievements.

I’m drawing parallels with Google DeepMind’s Atlas project, which this week published predictions for nine billion possible DNA changes. We can now generate an extraordinary number of hypotheses about biology. Drug discovery candidates seem imminent. Scale can get us surprisingly far.

We may not understand how it actually works.

Does that matter?

The AI-pilled will argue that humans are already lapsing far behind AI, and our human lack of true understanding is what will become. Others argue that these solves are superficial and doesn’t truly advance the field.

That gap is also becoming harder to ignore inside the labs. Anthropic researcher Jacob Coxon resigned this week, warning that frontier lab competition was taking precedence over safety. The labs “are racing straight to self-improving superintelligence and gambling with our lives”. A few days later, Anthropic and Open AI and DeepMind leaders subsequently called to slow down capability development of AI.

Does that matter? Or are we already too late?

I’m not convinced we can put technology progress back on hold. I’m also not sure I’d want to, if it means delaying something that could save lives.

But the possibility of curing a disease can’t be the answer to every concern about what else these systems might do.

Maybe we’re too late to stop the race. I hope we’re not too late to change how it’s being run.

-Tara

THE DOWNLOAD

OpenAI Claims a Navier–Stokes Solution. Mathematicians Question the Race to Solve Problems.

OpenAI released an argument and computer-checkable formalization claiming finite-time blowup for a forced version of the Navier–Stokes equations, which describe fluid motion.

The company says the effort involved roughly 10,000 concurrent agents, 88 hours of research and another 17 hours of formalization in Lean, a language that allows a computer to check the logical steps of a proof. The claim concerns specific forced formulations of the problem; its scope and acceptance still require mathematical scrutiny. (OpenAI)

On Friday, Terence Tao, one of the most influential mathematicians working today, joined 24 other Fields medallists in publishing a declaration about AI’s growing role in mathematics. They acknowledge that models can solve major problems. Their objection is that treating those problems as benchmarks can separate the answer from the understanding mathematics is meant to develop.

A result normally leads to talks, simplifications, new methods and connections to earlier work. The declaration argues that rushed announcements can short-circuit that process and raise attribution questions. It is not a technical refutation of OpenAI’s proof. (Terence Tao)

Why it matters: Last week’s Fermat formalization showed machines taking on more of the work of checking mathematics. This week raises the next question: once machines can produce results and formalize them, who makes those results understandable and useful? A proof can be correct without yet giving other researchers a method they can build on.

Meta Launches Muse, Joining the Race for a Personal AI Agent

Meta launched Muse, a personal AI agent designed to take on tasks, remember context and work toward longer-term goals. People can message it through a dedicated app or WhatsApp, give it access to services and let it continue working in the background. (Meta)

The ambition is an assistant that knows enough about your life to do more than respond to individual requests. Planning a trip might involve coordinating dates, comparing options and following up as circumstances change. The useful part is continuity: you should not have to explain everything again each time.

Instinct, a buzzy personal-agent startup that raised $250 million at a $2.5 billion valuation in weeks while still in private beta, is pursuing a similar idea. Its assistant connects to applications and devices, with a stated focus on understanding personal context and following through on everyday tasks. (Instinct)

Early experience shows both the appeal and the difficulty. An Atlantic reporter describes successful purchasing and coordination tasks, alongside reports of unwanted reservations and excessive booking requests. Meta has built a separate permission system, Sentinel, to control which actions Muse can take or send back for approval. These safeguards remain to be proven. (The Atlantic / Meta’s technical account)

Why it matters: Personal agents become useful through accumulated context and repeated delegation.The product challenge is building an assistant people trust with progressively more of their lives.

DeepMind Maps the Predicted Effects of Nine Billion DNA Changes

Google DeepMind released AlphaGenome Atlas, a resource containing predictions for approximately nine billion possible single-letter changes in human DNA.

A DNA change can affect more than the protein a gene produces. It can also alter when a gene is active or how its instructions are processed. The Atlas gives researchers a way to look up predicted effects across a large set of possible changes, helping them decide which variants to investigate experimentally. (DeepMind)

The new development is the resource itself: a broad, precomputed map that laboratories can consult rather than generating every prediction separately.

There is an important limit. These are nine billion predictions, not nine billion experiments. Checks of selected variants do not validate the entire map, and single-letter changes do not cover every kind of genetic variation.

Why it matters: Models are starting to become reference infrastructure for science. A shared prediction map could save researchers time and help prioritize scarce experimental resources. But it will also influence which questions get investigated. Knowing where the predictions fail becomes part of using the atlas well.

Robots Are Getting Better at Learning From Us

Showing a robot how to do something is harder than it sounds. A human demonstration has to become instructions that make sense for a different body, with different joints, sensors and ways of handling objects.

Several papers this week tackle parts of that translation.

SEED-UMI uses a shared exoskeleton interface to bring human demonstrations closer to the movements a robot can execute. The authors report 70% success across five tasks and threefold faster data collection. The aim is to preserve more of the useful information in a demonstration as it passes from person to machine. (Paper)

Show-Harness, from Singapore’s NUS ShowLab, approaches the problem through instructions. It gives language models a constrained set of actions that a control system translates into robot movements. Its experiments cover ten pick-and-place task combinations on Franka and AgileX hardware. The model chooses an action without having to generate every motor command itself. (Paper)

A third study shows how much a robot can learn through immediate feedback. It demonstrates dexterous pen writing by rapidly estimating how movements affect the pen, reporting roughly 18 seconds of initialization on a laptop CPU and 0.6 mm mean in-plane writing error. (Paper)

Why it matters: Better demonstrations and interfaces can reduce how much a robot must discover from scratch. Part of the work is making human intent easier for the machine to use; another part is letting it correct its movements as it goes. Both can make useful behavior less dependent on enormous training runs.

Anthropic Says Moonshot Served Claude Responses to Kimi Users

Anthropic alleges that Moonshot, the company behind Kimi, forwarded some customer requests to Claude and displayed the responses as its own.

The report describes almost 300,000 customer requests over one ten-day period, mostly routed to Opus through a network of 5,380 accounts. Anthropic separately alleges that Moonshot collected exchanges for training. (Anthropic)

That is more specific than the usual allegation of training on a competitor’s outputs. It concerns which provider actually answered a customer’s request.

It also falls short of establishing that “Kimi K3 is Claude.” Routing requests, learning from another model’s outputs and having identical underlying models are different claims. The report does not establish that every Kimi request was forwarded.

These remain allegations from a competitor. Bloomberg reports Anthropic’s account rather than independently verifying the traffic attribution. (Bloomberg)

ALSO:

Only 5% of Surveyed Chip Designers Reported First-Silicon Success. Siemens EDA and Wilson Research Group report a decline from 14.4% in 2024 among IC/ASIC respondents. This measures reported success on the first silicon iteration, not manufacturing yield. It is a useful counterpoint to faster design cycles: getting a design ready for fabrication and getting production-ready silicon remain different milestones. (Siemens)

Google Signs a 22-Year Nuclear Power Agreement in Finland. The Fortum contract begins in 2028 and reaches 50% of Loviisa’s capacity for 2030–2049. Separately, the US Department of Energy closed financing of up to $1.9 billion for the 615 MW Duane Arnold restart. Both support existing nuclear assets; neither should be counted as newly operating generation. (Fortum / DOE)

Sparse Models Can Save Compute and Still Need More Distinct Data. Mixture-of-experts models activate only part of their parameters for each token. A new study finds that repeatedly recycling training examples can erode their advantage over dense models. Regularization helps in the experiments, so this identifies a tradeoff rather than a universal limit: cheaper computation does not automatically make repeated data equally useful. (Paper)

3D seems to be becoming a new frontier for AI.

This week OpenAI showed Astra building in CAD, Blender and Unreal Engine, while World Labs released Atlas, a model trained to represent 3D space directly.

These are quite different approaches, but they point in the same direction. Models have gotten very good at understanding and generating representations of the world: words, images, video. Now they’re getting better at understanding the structure underneath them: objects, geometry, space, and how things relate to one another.

That matters if we want AI to move beyond the screen. Robots, scientific instruments, factories and physical design all require some understanding of space.

More on this soon. For now, enjoy the long weekend!

-Tara

THE DOWNLOAD

GPT-6 Astra Can Build in 3D, Find Zero-Days, and Is Getting Harder to Monitor

OpenAI released GPT-6 Astra this week. There are plenty of benchmarks, but three things stood out.

The first is 3D. Astra scored 95.9% on BenchCAD, up from 83.3% for GPT-5.6 Sol, a benchmark where models reconstruct 3D objects from multiple views by writing CAD code. OpenAI also shows Astra working directly inside Blender, FreeCAD, KiCad and Unreal Engine, including building a house in Blender and turning it into a walkable Unreal environment. (⁠OpenAI)

The second is cybersecurity. Astra is the first broadly deployed OpenAI model to reach the company’s Critical cybersecurity threshold. Without production safeguards, it scored 100% on ExploitBench and discovered and used two previously unknown zero-day vulnerabilities during evaluation. OpenAI says expert testing found it capable of exploiting hardened browsers and developing ways to gain higher-level access to hardened operating systems. (⁠OpenAI)

The third is what appears to be happening inside the model. Astra uses a technique called recurrent depth, which allows parts of the model to repeatedly process a problem rather than passing it through a fixed sequence of computation once. The approach can give the model more computation without simply making the model proportionally larger. (⁠TechRadar)

There is a tradeoff. OpenAI says Astra’s written reasoning is harder to monitor than Sol’s. In adversarial tests where researchers explicitly instructed it to evade monitoring, Astra could sometimes hide evidence that it was deliberately underperforming or carrying out sabotage. OpenAI says it has not found evidence of the model hiding reasoning inside apparently innocent text, and Astra violates boundaries less often overall. (⁠OpenAI)

Why it matters: Astra is simultaneously getting better at understanding and manipulating the physical structure represented inside software, more capable of finding vulnerabilities, and less dependent on reasoning we can easily inspect. The capability jump is increasingly showing up in what models can do inside tools, rather than how well they answer questions.

→ ⁠OpenAI: GPT-6 Astra / ⁠OpenAI: Safety overview

Claude Formalized Fermat’s Last Theorem in 11 Days

Anthropic says Claude has produced the first complete computer-checked formalization of Fermat’s Last Theorem.

This is not a new proof of Fermat’s Last Theorem. Andrew Wiles solved that problem in the 1990s. Claude instead translated the enormous body of mathematics behind the proof into Lean, a programming language that lets a computer mechanically check whether every logical step is valid. (⁠Anthropic)

That had been expected to take years. Claude did it largely autonomously in 11 days, producing 13 million lines of Lean and proving more than 30,000 intermediate theorems along the way. Dozens of agents worked in parallel, coordinated through a system called Prove2Me that kept track of which pieces of the proof depended on which others. (⁠Anthropic)

The finished proof is more than five times the size of Mathlib, the main community library of formalized mathematics it builds on. Mathematician Kevin Buzzard, who leads the existing human effort to formalize Fermat’s Last Theorem, reviewed the result and called it an “extraordinary autoformalization achievement.” (⁠Anthropic)

(There is now some online chatter about whether Claude will solve a Millennium Prize problem before Anthropic goes public.)

Why it matters: Verification has been one of the least glamorous bottlenecks in AI-generated mathematics. A model can generate a proof much faster than humans can establish whether every step is correct. If formalization becomes cheap too, the machine can increasingly produce the work and the evidence that the work is valid.

→ ⁠Anthropic: Formalizing Fermat’s Last Theorem

World Labs’ Atlas Turns Images Into 3D Worlds

World Labs released Atlas, a model trained from scratch across text, images, video and 3D.

Most image and video models represent a scene as pixels. Atlas also keeps track of where each image sits in three-dimensional space. Give it one image and a camera path and it can generate what the scene should look like as the camera moves. Give it several views of a real place and it can reconstruct the underlying 3D scene, including explicit 3D outputs. (⁠World Labs)

World Labs shows Atlas reconstructing environments from as little as a single photograph, generating new viewpoints that were never captured, and producing controlled videos up to one minute at 1440p. The company says Atlas beats specialized open-source 3D reconstruction models on the benchmarks it tested. It is entering early access rather than broad release. (⁠World Labs)

The timing alongside Astra is notable. Astra is getting much better at operating the software humans use to construct 3D things. Atlas is trying to make 3D space part of the model itself.

Why it matters: Models have been very good at generating what a world looks like. Spatial models need to understand enough about the world to preserve it as the viewpoint changes. That starts to matter beyond media, for robotics, simulation, design and any agent that has to reason about physical space.

→ ⁠World Labs: Atlas

ALSO:

Nvidia Actually Bought Hugging Face for $12.9 Billion. Last week it was a rumor. This week the deal is signed. Nvidia says Hugging Face will remain open across models, clouds and accelerators. The harder question is how hardware-neutral one of the main distribution layers for open AI remains when it is owned by the dominant hardware vendor. (⁠Reuters)

1,000 GitHub Repositories Became 5,000 Agent Skills. Repo-To-Skill distilled operational knowledge from widely used machine-learning repositories into reusable instructions for agents. With the underlying GPT-5.5 model, harness and execution budget held constant, adding the skills improved MLE-bench performance 134% and PaperBench 34%. (⁠arXiv)

AI-Designed Proteins Are Delivering RNA. Researchers used generative protein design to build more than 100 synthetic RNA carriers that do not exist in nature. The best design transferred RNA several orders of magnitude more efficiently than the natural and lipid-based carriers tested, and was used to deliver gene-editing machinery into patient-derived cells and a pig model of Duchenne muscular dystrophy. The work was published in Nature. (⁠Nature)

ChatGPT, Claude and Grok Went Down at Nearly the Same Time. All three experienced outages on Thursday morning. Anthropic blamed an infrastructure issue and xAI a failure at its Memphis data center. No common cause has been established. (⁠The Verge)

Memory Costs Are Starting to Change Chip Designs. Nvidia is reportedly considering lower-memory configurations for Rubin Ultra as HBM prices climb. After memory prices began showing up in server prices last week, the constraint is now reaching architecture decisions too. (Trendforce)

I spent a couple of days at Hot Chips this week, and came away feeling like… the culture of chipmaking is changing.

One moment that stuck with me was watching OpenAI present Jalapeño, its first inference chip.

The performance numbers were impressive. But what was more surprising was how quickly it got there. OpenAI says Jalapeño went from initial design to tapeout in nine months, using its own models to accelerate parts of the design process.

For advanced silicon, that is extraordinarily fast. Typical timelines have been closer to 18 - 24 months.

Chip design has traditionally been slow for good reason. A mistake can cost tens of millions of dollars and months of manufacturing time. Architectures accumulate abstractions, verification processes and institutional knowledge precisely because you don’t get many chances to be wrong.

OpenAI seemed willing to make a different trade.

Jalapeño doesn’t need to be good at everything. OpenAI knows the models it runs, the workloads it serves and where the bottlenecks are. So it can design much closer to the workload and discard some of the generality that Nvidia needs.

You could feel versions of this elsewhere at the conference. Cerebras still making the bet that the chip should be an entire wafer. D-Matrix stacking custom memory directly onto compute. Normal Computing and its forays into thermodynamic compute. Different architectures, but a similar willingness to question assumptions.

I wouldn’t be too quick to assume the old guards (the chip uncs) are being replaced.

At the same event, Nvidia and Google talked through work that has spanned decades. Much of what the new generation can move quickly on exists because that infrastructure and knowledge already exists.

But clearly, change is in the air. AI can compress parts of the design cycle. TSMC and the semiconductor supply chain let companies specialize without building fabs. Broadcom can provide decades of implementation expertise while OpenAI supplies intimate knowledge of the workload. And when you’re spending billions of dollars serving essentially the same class of computation, even a narrowly useful chip can justify its existence.

For decades, building a serious processor required not only enormous capital but an enormous accumulation of semiconductor-specific knowledge. A great wall to climb. This may be less true today. And that means the next generation of chips may increasingly come from companies that didn’t start out as chip companies at all.

-Tara

One more thing: I have many more things I want to dig into than I can currently get to, and I’m thinking about building out the research side of Strange Review.

Would you be interested in deeper proprietary research, or a private chat to share and discuss things in real time? Hit reply. I’d love to know.

THE DOWNLOAD

1,200 Agents Self-Organized For A Cyberattack

OpenAI and METR published their postmortems on the July incident where agents escaped an internal cybersecurity evaluation and eventually compromised Hugging Face.

METR reconstructed a hidden message board used by 1,206 agents, carrying more than 70,000 messages over five days. Roughly 700 agents participated in the attack on Hugging Face. Some sacrificed their own evaluation performance to help other agents. A handful considered maybe alerting a human. None did.

The agents also figured out how to spoof parts of their tool-use transcripts. METR found successful spoofing in roughly 7% of the transcripts it examined, while the grader intended to catch this behavior was missing a check described in the evaluation documentation.

The obvious lesson is that the sandbox failed. The more interesting one is that hundreds of individually limited agents became more capable once they found one another.

We spend a lot of time asking what happens as individual models get smarter. This was a glimpse of a different scaling axis: what happens when you have a thousand of them.

→ OpenAI / METR

Nvidia May Be Buying The Open Models Layer

Nvidia is reportedly in talks to acquire Hugging Face for around $13 billion. Neither company has confirmed a deal.

Calling Hugging Face the “GitHub for AI” increasingly undersells what it’s becoming.

It sits at the center of the open model and dataset ecosystem. But last year it also acquired Pollen Robotics. Its LeRobot project is building common tooling and datasets for robot learning. Last month it released Grabette, a cheap open system that lets a person record physical tasks with a handheld gripper and turn them directly into robot-training data.

Hugging Face’s thesis is explicit: “The bottleneck isn’t the model. It’s the data.”

Then this week Pollen (acquired by Hugging Face in April 2025) unveiled Microduck, a $399 open-source consumer robot.

There’s a more interesting interpretation of a potential Nvidia acquisition. Nvidia already dominates the compute on which AI runs. Hugging Face increasingly sits where models, datasets, developers — and now physical-world training data — meet.

If Nvidia buys it, it isn’t just buying distribution for open models. It may be buying an emerging data and distribution layer for physical AI too.

→ Business Insider / Hugging Face

OpenAI’s First Chip Is Fast

OpenAI showed off Jalapeño, its custom inference chip built with Broadcom, at Hot Chips this week. The first numbers are fairly wild.

On SemiAnalysis’s InferenceX benchmark, Jalapeño cleared 700 tokens per second for a single user running DeepSeek R1. On Kimi K2.5, it delivered more than 9× the performance of the next-best system at 100 tokens per second per user.

There are some big caveats. This is engineering silicon, the tests used a short fixed prompt rather than messy real-world traffic, and SemiAnalysis says the fairest comparison would be Nvidia’s upcoming Vera Rubin rather than today’s systems. SemiAnalysis verified the runs in OpenAI’s lab, but didn’t independently run the full benchmark suite.

Still, the direction is more interesting than the benchmark win. OpenAI isn’t trying to build a general-purpose GPU. Hardware lead Richard Ho says the goal is to run OpenAI’s most important workloads “close to the hardware’s theoretical limits.”

That’s the bet: if you know exactly which models and workloads you’re serving, you can throw away the generality Nvidia needs and optimize the entire chip around inference.

Jalapeño starts shipping in small volumes at the end of this year, with volume production in 2027. If the advantage survives real workloads, OpenAI won’t just be one of Nvidia’s biggest customers. It will have started designing around Nvidia.

→ OpenAI / SemiAnalysis

Also:

Memory is getting expensive enough to show up in AI server prices. Some Nvidia Grace Blackwell and Vera Rubin systems shipping next year are expected to cost >15% more, driven by soaring memory costs. Nvidia CFO Colette Kress called current memory pricing conditions “extreme.” If memory reprices faster than compute gets cheaper, some of the expected generational decline in inference cost gets eaten away. → Bloomberg

OpenAI is cutting Cursor off. OpenAI plans to stop supplying its models to Cursor on November 12 following SpaceX’s acquisition of Anysphere, saying it can’t trust Musk’s companies to comply with its terms. Cursor says OpenAI accounts for only ~5% of its traffic. → OpenAI

Anthropic wants agents operating physical equipment. Its new Model Hardware Standard gives agents a common interface for microscopes, liquid handlers, robotic arms and even lasers. Anthropic says integrations that previously took weeks can sometimes be done in hours. MCP connected agents to software; this is an attempt to do something similar for machines. → Anthropic

Children remain annoyingly data-efficient. BabyLM has spent three years testing whether language models can learn on something closer to a child’s data budget. Better architectures and training methods help, but the enormous sample-efficiency gap remains. We increasingly know that models can learn surprisingly sophisticated language from less data. We still don’t know why children need so much less. → MIT Technology Review

-Tara

THE DOWNLOAD

SILICON AND POWER

DDR5 Retail Prices Are Up 485% in Twelve Months

DRAM is the working memory between a processor and storage, and it is built on the same production lines as the high bandwidth memory that goes into AI accelerators. HBM earns more per wafer, so manufacturers have shifted output toward it, putting supply pressure on DRAM. This goes into laptops, phones, etc. Memory makers have reportedly sold their 2027 capacity in full.

SK hynix Publishes a Roadmap for Optical Links Between Compute and Memory

Co-packaged optics puts the optical components in the same package as the chip, so data leaves over light rather than copper. SK hynix has published a roadmap for it in Nature Electronics, written with the University of Virginia, MIT, UIUC, Nanyang and Yonsei, that extends the approach to the memory interface. The mechanism is a photonic interposer between accelerator and memory, which removes the packaging limit on how much memory one chip can reach and lets several accelerators draw on a shared pool. Stated targets are over 100 terabits per second per node and chip-to-chip latency under ten nanoseconds. No product date yet.

TerraPower Signs Hyundai E&C to Build Eight Natrium Reactors

TerraPower and Hyundai Engineering & Construction have signed a framework agreement making HDEC the builder for up to eight 345-megawatt Natrium plants in the US. Natrium is a sodium fast reactor with molten salt heat storage, which lets output rise to 500 megawatts for more than five and a half hours. Meta agreed in January to eight plants, first units as early as 2032.

COMPUTE AND CAPITAL

Nvidia Is Guaranteeing Up to $105 Billion Behind OpenAI’s Ohio Data Center

Nvidia has told the SEC it will provide up to $105 billion of credit support for an OpenAI data center in Pike County, Ohio. SB Energy, a SoftBank-backed developer, builds and owns the campus. Nvidia’s promise to the lenders: if OpenAI stops paying, Nvidia covers the lease and power payments and makes good on what the building and the chips are worth. Nvidia also put $1.5 billion into SB Energy and supplies the chips for the first 800 megawatts, due in 2028.

Groq Is Pivoting From Designing Inference Chips to Operating Data Centers

Groq is pivoting from designing inference chips to operating data centers built on Nvidia hardware. Back in March, Nvidia bought out Groq’s leadership, including founder Jonathan Ross, along with a non-exclusive licence to its inference technology, for roughly $20 billion. The plan now appears to be to drop chip design and build out the existing footprint: thirteen data centers, more than six million developers on the platform, and a move from 54 megawatts to over 200 in 2027.

Unitree Rose 460% on Its First Day of Trading in Shanghai and Anthropic’s Public S-1 Is Expected This Month

Unitree listed on Shanghai’s STAR Market this week, raising $900 million on a book covered more than 8,000 times. Anthropic filed a confidential S-1 on June 1 at roughly a $965 billion valuation and is expected to file publicly as soon as the end of August. OpenAI has not filed.

MODELS AND DATA

DeepSeek Released Its New Vision Model Through the API, With No Weights

DeepSeek released V4-Flash-Vision-Exp, an experimental multimodal model, through its API. No weights. The company says it matches V4-Flash on text and brings multimodal agent performance “close to Opus-4.8” on its own benchmark chart. Images are billed as tokens, up to 384 each at V4-Flash pricing. DeepSeek has released open weights for its main models.

Thinking Machines Is Serving Inkling Free to Agentic Harnesses

Thinking Machines has made Inkling free on OpenRouter and says the offer is limited to agentic harnesses, will run for a few weeks, and that the interaction data, disassociated from accounts, will be used to improve the model’s agentic performance. What the lab can observe: traces from long tool-use loops inside real harnesses, which downloaded weights do not produce.

Replit’s Free Mode Routes Simple Tasks to GPT-5.6 Luna and Stops Metering Them

Replit has made GPT-5.6 Luna the default for chat, ideation and simple coding, and stopped charging those tasks against a subscriber’s token budget. It is called Free Mode and it requires a paid subscription, $20 a month for Core or $100 for Pro. Harder work routes automatically to a larger model, then reverts. “A lot of tasks don’t require the frontier-level intelligence. It just requires smaller models that are faster and more affordable,” said Michele Catasta, Replit’s president and head of AI.

Amazon Is Cutting the Bindings Off Rare Books to Scan Them

404 Media put a tracker in a shipment of rare books and followed it to an Amazon facility in North Las Vegas, where a team cuts the bindings off so the pages feed through scanners faster. The book is destroyed. Amazon says it “purchases books through commercial channels to improve the products and services customers use.”

LABOR & LEARNING

Homework Scores Up 18%, Exam Scores Down 20%, Same 26,811 Students

Researchers at Stockholm University and the University of Hong Kong followed 26,811 Chinese students in grades seven through twelve for two years. Those who took up AI homework help raised homework scores 18% and cut homework time 30%. Within six months the same students scored 20% below classmates who had not used AI on monthly exams, and college entrance results came in 18 to 24% lower. About 80% of the exam gap came from students the authors classify as outsourcing the work rather than working alongside the tool.

40,000 Hyundai Workers Struck Over Wages & Robots Arriving in 2028

Hyundai’s Korean union, about 40,000 members, began its first full strike in a decade on Friday, stopping the Ulsan, Asan and Jeonju plants. Wage talks are the formal deadlock. The union is also demanding protections against automation, tied to Hyundai’s plan to put Boston Dynamics Atlas robots into its plants from 2028. Hyundai Motor Group controls Boston Dynamics. Partial walkouts since July have cost about 55,200 vehicles worth $1.67 billion. “This strike is to protect the roots of Korea’s automobile industry and uphold the dignity of labor,” said Park Sang-man, chairman of the Korean Metal Workers’ Union.

LATEST FROM THE REVIEW

EVENTS

We’re hosting a casual happy hour at Hot Chips!

RSVP

GitHub’s busiest month for code pushes in 2025 was August, at just under 100 million. By early 2026, that number had reached 1.4 billion, a +1300% increase.

GitHub is clearly feeling the strain.

Monthly incidents nearly doubled after November 2025.

Now new infrastructure is appearing around it.

Cursor launched Origin this week, its own code-hosting system built for what it calls “millions of tiny repositories created by agents.”

GitLab is rebuilding its backend so agents can query repositories without cloning them, with Anthropic as a design partner. Zed is experimenting with continuous delta streams instead of traditional commits. Former GitHub CEO Thomas Dohmke raised $60 million to build repository infrastructure for fleets of coding agents.

The interesting part is that GitHub hasn’t been displaced.

It has 225 million users and is used by more than 90% of the Fortune 500. Every major coding agent still works with it. Even Cursor’s Origin still sends code back to GitHub. As Cursor puts it, GitHub remains the “source of truth.”

So for now, coding agents aren’t replacing GitHub.

They’re creating a new infrastructure layer around it.

But does that layer remains an extension of GitHub, or eventually becomes its replacement?

Should the CEO of Anthropic be more optimistic about AI?

On Saturday, an exchange between investor Gavin Baker and Dario Amodei went into why the public has such a negative view of AI. Dario called it “fundamentally a crisis of trust.”

His answer, roughly, was that you don’t fix that with better messaging. You fix it by delivering something undeniable: like, cure cancer, or solve major diseases.

I understand the instinct. But the problem with waiting is that the public is forming its opinion of the technology now. The costs are already hitting hard: electricity demands, data centers, job anxiety, deepfakes.

AI has a strange problem: people are using more of it and liking it less.

Anthropic and OpenAI are generating tens of billions of dollars in annualized revenue. AI adoption inside businesses continues to climb. But public sentiment is moving in the opposite direction.

A June Pew survey found that 40% of Americans expect AI to have a negative effect on society over the next 20 years. Just 16% expect a positive one.

And increasingly, that skepticism has consequences. Only 14% of Americans say they would support a data center in their community. Seventy-five data center projects worth $130 billion faced opposition in the first quarter of this year alone.

But there was another part of the exchange that made me wonder if that’s even the right question.

The reason the whole exchange started was that Gavin alluded that Dario privately believes Anthropic will eventually be “the only company left in the world.”

Maybe that’s hyperbole. But it resonated with something I’ve struggled to reconcile in Dario’s public position on AI.

• AI could be extremely dangerous, and the race to build it is itself dangerous. But Anthropic needs to stay in the race.

• Concentrations of AI power are dangerous. But Anthropic will become one of the most powerful companies in the world.

• Certain capabilities may be too dangerous to release broadly (ie Mythos). But Anthropic has to decide when and to whom they should be released.

There are reasonable arguments for every one of these positions individually. Together, though, they create a difficult tension.

If you believe AI is both extraordinarily powerful and potentially dangerous, but also believe someone will build it regardless, staying at the frontier can feel like the only reasonable thing to do. Better to have a responsible, safety-conscious company helping shape what happens than to leave the AGI race entirely.

The problem is that every frontier lab can make some version of that argument.

If Dario really does believe Anthropic could eventually be the only company left, it raises a harder question: how does any CEO separate their judgment about what is good for the technology from their conviction that their own company is the one best to build it?

Maybe that’s part of the crisis of trust too.

The people building AI are being asked to make humanity-scale consequential decisions about how fast to move, what to release and which risks are acceptable, while simultaneously competing to win the market those decisions will create.

So perhaps my question is whether any frontier AI company can simultaneously race to win and credibly argue that the competition itself is the problem.

-Tara

THE DOWNLOAD

SpaceXAI Ships an Always-On Agent, Then Closes the $60 Billion Cursor Deal Three Days Later

SpaceXAI unveiled Grok Bot on Aug 11, a persistent agent product where each agent gets its own cloud computer and signs into applications with the user’s credentials, including applications with no API, runs unsupervised, and coordinates with other agents in group chats. Grok 4.6 followed on Aug 12, same 1.5T base as 4.5 but tuned for long-running agents, shipping day one into Cursor, OpenRouter, Vercel and Cloudflare. On Aug 14, SpaceX closed its $60 billion all-stock acquisition of Cursor, issuing 389,289,254 Class A shares. Cursor had crossed $4 billion in annualized revenue in June and acquired Firetiger on Aug 13, the day before its own close.

Why it matters: Grok Bot is also the first credential-operating agent from a major lab, and that architecture is a different risk surface than an API-calling agent: it logs into things as you, in apps that never agreed to be automated. The model itself is incremental, tying GPT-5.6 Sol at 61 on the Artificial Analysis index rather than leading. Musk separately said future Grok versions will train on “the sum total of all SpaceX information,” including employee work product, with no stated opt-out.

Is the US-China Model Gap Just Mostly Release Timing?

Z.ai shipped GLM-5.3 on Aug 14, built by scaling post-training on the GLM-5.2 base, and held the weights back two weeks for safety testing. Writing in Interconnects, Nathan Lambert argues the reason Chinese models keep arriving at the frontier is the release calendar rather than distillation. American labs finish a model and spend months testing internally; Chinese labs ship in days and spend those months improving the released version. “It is very, very likely that OpenAI and Anthropic have far better internal models than Z.ai.” He also credits narrower scope, since GLM-5.3 has no visual capability at all, plus compute efficiency, Tsinghua’s talent pipeline, and reports that American data companies are selling reinforcement-learning environments to Chinese labs.

Why it matters: Developers build on what they can get, and the self-improvement loop runs on data from developer usage, so shipping first has its benefits. It also points the policy argument at the wrong variable. Senator Banks asked the White House this week to fund American open models, but a subsidy does not shorten a safety review. Lambert: “I’m confused why the labs in the U.S. haven’t patched this behavior faster; instead they’re running to the government asking for policy help.”

Google Makes Its Visible Watermark Optional; Anthropic Starts Watermarking Text

Google made the visible watermark optional on AI-generated images, video and music in Gemini Apps and Flow, while keeping invisible SynthID and C2PA metadata mandatory and undisableable. VP Josh Woodward: “We’re striking a balance here between creative control and safety.” The toggle is unavailable on work and school accounts, and gated to AI Ultra in India, South Korea and Vietnam. Three days earlier Anthropic said it will watermark text from every model released after Aug 2, using a version of DeepMind’s SynthID-Text approach across the API, apps, Claude Code and Cowork.

Why it matters: The deeper problem is that strong text watermarking may not be achievable at all. Zhang, Edelman, Barak and colleagues proved at ICML 2024 that a random-walk attack strips watermarks without knowing the key or even which scheme is deployed, and a simple paraphrasing pass defeats detectors generally. Anthropic is not really claiming otherwise.

Nvidia Signs Six MOUs to Mobilize $500 Billion

Nvidia announced partnerships with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to mobilize over $500 billion of third-party capital. The release states these are memorandums of understanding and that the partnerships remain subject to execution of the final agreements. The widely reported 25% guarantee appears nowhere in it. It came from Jensen Huang on X the following day, describing support for up to 25% of an opportunity, assessed project by project, limited and residual-value based.

Why it matters: Nvidia is now guaranteeing the future value of GPUs that Nvidia itself will make obsolete. This sounds, at first, like another version of the circular financing I wrote about last September, when Nvidia was investing in companies that were turning around and buying its chips. It’s actually something different. It is promising that if the chips ever have to be sold off, they will still be worth something. That promise makes the GPUs better collateral: lenders can lend more against them, at better rates, to buyers who otherwise couldn’t get those terms.

The FCC Drafts an Import Ban on Chinese Optical Transceivers While Indium Phosphide Runs Short

The administration and the FCC are drafting a rule to add China-made optical transceivers to the Secure Networks Act covered list, targeting new models, with officials aiming to enforce by end of 2026. China holds about 56% of global optical module capacity, and Chinese firms hold seven of the top ten slots, with Innolight and Eoptolink above 60% of the 800G-plus market. Lumentum’s Michael Hurlston says the indium phosphide shortage will be “worse than memory”, with the company expanding supply 50% year over year and still shipping more than 30% below demand.

Why it matters: China controls roughly 70% of global indium supply, so restricting Chinese transceivers tightens the input to the Western capacity meant to replace them. Hurlston, on serving the demand: “Between the two of us, I don’t think we can service the demand that Nvidia and others are now putting on us.”

A Nature Study Gives an Agent Six Days and $3,000 to Replicate Two Papers; the Authors Score It 2/6 and 1/6

Nature reported on Aug 13 on “shadow evaluations” run by Sayash Kapoor at Princeton with the UK AI Security Institute. The goal for the agentic system was to develop concept from two computer-science papers. The original authors scored the outputs only 2 out of 6 and 1 out of 6, reject and strong reject. The preprint documents five recurring failure modes: premature commitment to a first hypothesis, poor backtracking with ambitious goals abandoned inside ten hours, no sense of what clears a publication bar, poor budget awareness, and instruction drift. One run spent only $1,130 of its $3,000. Kapoor: “I don’t think full automation of open-ended research is on the horizon right now.”

Why it matters: What the agent did well is the part worth reading twice. It ran hundreds of experiments without human help, debugged without falling into error loops, caught its own false claims, did not reward hack, and produced complete LaTeX papers. The engineering of research is solved. What failed was judgment: it could not tell a trivial finding from a real one, and its self-reviews were uncritical.

JOIN US AT HOT CHIPS

Join us for cold beer and Hot Chips on October 25! We’re hosting a happy hour post- Hot Chips in Palo Alto.

RSVP

The easy explanation is that Airtable was a casualty of ZIRP-era hype.

But here’s the strange thing: Airtable actually grew.

At its 2021 financing, Airtable was doing an estimated $156M in ARR. Today, that number is closer to $480M. Revenue roughly tripled.

How does a company triple revenue and still lose 80% of its value?

So I went back and looked at a basket of software companies that raised at peak 2021-ish valuations.

Airtable was valued at roughly 75x ARR in 2021. To grow into that valuation at a more normal ~15x multiple, it needed to roughly 5x revenue.

It delivered ~3x.

It was a good software business. Just not enough growth, fast enough, to support the price investors had already paid for it back in 2021.

Comparatively, the ZIRP-era startups that firmly vaulted past their growth hurdle in the AI-era are standing strong.

The lesson isn’t simply “ZIRP bad”.

The 2021 market pulled years of expected growth forward into the valuation. Eventually, companies had to deliver that growth. Replit, Vercel and Databricks outran the expectations embedded in their prices. Airtable didn’t.

And in the AI-era, the hypergrowth startups are growing at an insane scale.

Cursor went from roughly $100M to $4B in annualized revenue in 17 months. OpenRouter went from roughly $1M to $140M ARR in 19 months. Windsurf reached roughly $100M ARR before Google struck its $2.4B licensing-and-talent deal.

These AI startups are reaching revenue milestones and multi-billion-dollar outcomes in a few years that previously took a decade.

Yes, there’s plenty of froth here too. Extraordinary growth is once again being extrapolated into extraordinary expectations.

But the AI era introduces a strange paradox: it makes it possible to build enormous companies faster than ever, while also making waiting more dangerous. The ground is changing constantly.

That changes the math of an exit.

Venture tends to treat a multi-billion dollar acquisition as the consolation prize to building the large IPO-scale independent company. But when the window itself is moving faster, crystallizing a $2B outcome in year three may sometimes be better than spending another decade chasing an IPO exit window.

AI is compressing both the time to build value and the time you have to capture it.

The most surprising thing this week was Google losing its AI stars in one fell swoop. Jeff Dean left after 27 years, bringing three of his colleagues with him, and Demis Hassabis stepped back from running DeepMind. This was probably a long time coming. It also leaves a real question about who is at the helm of the frontier research lab at Google, arguably the world's largest pool of compute. This is a job that gets harder by the day whether you’re DeepSeek or SpaceX: battling commercialization, frontier benchmarks, and going full stack into infrastructure and chips all at once.

Enjoy the update - Tara

THE DOWNLOAD

Tesla and SpaceX Commit $16.8B to Terafab Chip Complex in Texas

Tesla and SpaceX confirmed Thursday that Terafab will be built in Grimes County outside Houston with an initial $16.8B investment, spanning more than 100 million square feet of logic, memory, packaging, and test under one roof. Tesla points to a goal of “producing over 1 terawatt of compute per year.” SpaceX filings have put multi-phase spending as high as $119B. Intel is a contributing partner and at least 3,000 jobs are planned.

Why it matters: There’s a gap between ambition and reality. BofA estimates that reaching just 100,000 wafers per month would take more than $60B in capex, with wafer costs 30% to 50% above TSMC. Bernstein put the full terawatt target nearer $5 trillion. And SpaceX’s own May IPO filing described Terafab as a general framework carrying no binding commitments, with either party free to walk away.

Anthropic Confirms In-House Custom Silicon Team

Anthropic confirmed Wednesday that it is building a team to design custom chips co-optimized with Claude. The Information reported last month it has been scouting Samsung as a manufacturing partner.

Why it matters: That completes the set. Google has TPUs, Amazon has Trainium and Inferentia, OpenAI has its Broadcom-built Jalapeño, Meta has MTIA. Anthropic was the last major US lab without a silicon program.

DeepSeek Now Decides On Significant API Price Increase

DeepSeek this week told developers on its pricing page that it plans to raise API prices broadly, with a significant increase expected, giving no figures or date. Separately, its unreleased Harness coding agent, benchmarked internally against Claude Code and named in the V4-Flash-0731 model card, is in beta.

Why it matters: DeepSeek built its developer base on roughly $0.03 per task against $3.15 for Claude Fable 5. In a leaked investor call on May 20, CEO Liang Wenfeng said he could probably raise prices and keep demand, but didn’t care to. The new API prices came around the same time Bloomberg reported the company had resumed the $8B round it paused in July, now at a valuation near $74B. That round had stalled after the leaked transcript of Liang Wenfeng’s May 20 closed-door investor meeting went viral. Reopening the round at $74B and repricing the API on the same day suggests that choice may no longer be his alone.

Jeff Dean Leaves Google to Found Discovery Loop; Hassabis Becomes Chair as Kavukcuoglu Takes Over DeepMind

Google announced Wednesday that Jeff Dean is leaving after 27 years along with Sanjay Ghemawat, Oriol Vinyals, and Quoc Le to found Discovery Loop, a public benefit corporation automating the experimental loop. The same day, Demis Hassabis moved to DeepMind chair and Alphabet chief scientist, with CTO Koray Kavukcuoglu taking day-to-day control of Gemini and frontier research as SVP reporting to Sundar Pichai.

Why it matters: Google lost the four names most associated with its AI stack and reorganized the lab on the same day. Its stock fell 4%. Discovery Loop is starting by automating machine learning research with itself as first customer before moving to drug discovery, materials, and chip design, which is the recursive self-improvement thesis with the best systems team in the industry behind it.

Cursor Adds Google Workspace Plugins

Cursor shipped plugins giving its agents direct read and write access to Gmail, Drive, Calendar, Docs, Sheets, and Chat, installable from the Customize page.

Why it matters: A coding tool that sends email and edits spreadsheets is not a coding tool anymore. Cursor is walking onto the same ground as ChatGPT Work and Claude Cowork, but from the developer side. The boundary between the IDE and the general knowledge-work agent is closing from both directions.

AI Models Are Solving Erdős Problems for a Few Hundred Dollars Each

Quanta traced how the Erdős problems became the proving ground for AI mathematics: a DeepMind agent autonomously resolved 9 of 353 open problems at a per-problem cost of a few hundred dollars, and on OpenAI’s counterexample to the unit distance conjecture, Tim Gowers wrote that he would have recommended it for the Annals without hesitation.

Why it matters: Once a solved conjecture carries a per-problem cost, mathematical progress becomes a compute line item and formally verifiable domains become product surface. The constraint moves to verification of proofs.

Bending Spoons Buys Airtable for $1.285B

Bending Spoons agreed on August 4 to acquire Airtable at $1.285B enterprise value, about $2.25B equity value with net cash. Airtable’s ARR grew over 20% to roughly $480M as of June.

Why it matters: Airtable was worth over $11B in 2021 and traded near $4B on secondaries earlier this year. Late-stage 2021 application software is still settling at a fraction of paper value. (Editor’s Note: More to come on Monday!)

This is in stark contrast with US frontier labs who are knife-fighting over developer, consumer, and enterprise mindshare through a torrent of product launches.

Liang seems only interested in the pursuit of AGI. Users are “sesame seeds, not watermelons”, he said on the call. Don’t stop to pick up small wins on the road to a large one.

Subscribe now

There are two ways to build a frontier AI company. Build a chatbot, fight for users, sell subscriptions and ads. Or, alternatively, chase AGI, and the chatbot is a by-product. Liang clearly sits in the latter camp. He says DeepSeek’s consumer app and the enterprise API aren’t the point, and only needs a handful of people to run the API. With the hype around DeepSeek last spring, fighting ByteDance for consumer users was a real business he could have a shot at winning. He passed, because he thinks what’s coming makes today’s consumer market look tiny. He’ll pick up the sesame seeds in passing, but just won’t stop for them.

The strongest evidence for this is how DeepSeek prices. They chose to set their API price to a ten-month hardware payback, roughly 6x on compute, and stop right there. Liang acknowledges that demand in that band is inelastic, so doubling the price would nearly double revenue. He knows this and won’t do it.

Two reasons. One is team - they cheered when he cut prices, and that’s part of why they work there. But mostly because at just 6x margin, nobody else can profitably run DeepSeek’s own models cheaper than DeepSeek can. Fat margins would invite everyone to undercut him using his own weights. Thin margins are the moat.

A note on the source: DeepSeek has not confirmed the document, and it is a speech-to-text transcript that flags its own error rate on names and figures. Everything below is what the transcript says he said.

On Compute:

• Compute-capped: Frontier models activate ~800B params; Chinese labs are at tens of B. To train at 800B he’d need ~50,000 GB300s or 200,000 Huawei 950s just on training alone. Even spending the whole ¥50B raise he couldn’t afford it, or afford to serve it.

• He believes that DeepSeek is one to two years behind on one-twentieth the compute. The goal is to hold the compute ratio and compress the time gap to three or six months.

• He argues that export controls caused the substitution they were meant to prevent. In a normal market where he could buy Nvidia freely, substitution with domestic chips would be a harder call.

• Nvidia’s software moat is loosening. Nvidia’s CUDA software platform is breaking for three reasons: AI can now write the ecosystem code cheaply; TileLang makes kernel rewriting fast; and CUDA is architecturally welded to gaming cards, which stopped making sense once compute cards outgrew gaming.

• The 16,000 Huawei Ascends everyone reported as a big deal: he says it equals ~4,000 B-series chips, isn’t enough for a next-gen model, and the real reason for buying is to help Huawei get its ecosystem right. Big Chinese internet firms got an order of magnitude more chips than DeepSeek.

On Research

• The staircase: language model → chain of thought → agents → continual learning → self-iteration → embodiment (physical ai).

• Continual learning is the named next bottleneck and nobody in the world has a working method - we are all still searching.

• He wants continual learning before general intelligence explicitly because it’s the lazy path. The next model’s first customer is DeepSeek itself: “first it has to be useful to us.”

  • (Editor’s Note: the next model that trains itself is a challenge the top minds in AI are circling around. Anthropic hired Andrej Karpathy in May to build a team using Claude to accelerate its own pre-training. Jeff Dean and three Google colleagues left this week to start Discovery Loop, automating large-scale ML experiments.)

• Open-ended research’s main edge is researcher attention, rather than compute. (slight contradiction) He calls it akin to drawing lots.

• He won’t do video generation or world models: commercially good, off the intelligence main line. He found the post-Sora pile-in odd, like it had become mandatory to jump on the bandwagon

• Anthropic’s lead over OpenAI is transitory; the code-agent advantage isn’t large. Half his company thinks OpenAI is better.

• They have no org chart, no KPIs, no written vision. He admits this breaks with scale and he’s already building departments.

• Too many Chinese labs building base models; the US has three. It will converge to maybe two big and two small.

THE DOWNLOAD

Chip Stocks Lose $1.3 Trillion in a Week

The world’s 20 most valuable chip companies lost about $1.3 trillion in market value in the sessions ending July 29: Nvidia shed $238B, SK Hynix $176B, Samsung $173B, TSMC $119B, Micron $113B, and AMD $110B. The worst day was July 28, when reports that China could expand domestic memory capacity erased $541B from six memory names in a single session.

Why it matters: Analysts largely agree fundamentals didn’t move; Morningstar’s Michael Field calls it sentiment, not fundamentals. The sharper read is leverage: chip stocks were up 92% over the prior year, Morgan Stanley’s momentum TMT index fell 53.5% in July, and as Jim Osman puts it, AI needs years of patient capital but lenders mark collateral daily, so ordinary corrections become forced liquidations.

Situational Awareness Falls 67% in July; Citadel Buys Its Public Portfolio

One of the world’s biggest AI funds told investors it was down about 67% for July, writing “we let you down this month.” Leopold Aschenbrenner’s fund ran roughly 4x leverage on a concentrated public book; margin calls from its prime brokers forced a liquidation, and Citadel bought the entire public equities portfolio in a block trade. Assets went from a $45B peak to about $10B.

Why it matters: The public portfolio, most of the fund, was sold to Citadel at a discount to settle the borrowing, and the fund is still up on the year because what remains is mostly its Anthropic stake, a private position with a large paper gain that lenders couldn't touch and Aschenbrenner can't sell. Whether investors ever see that 80% now rests almost entirely on the Anthropic IPO expected by October.

OpenAI and Anthropic Models Breached Real Systems During Cybersecurity Evals

OpenAI disclosed that during an internal capability eval with safety classifiers deliberately disabled, GPT-5.6 Sol and a pre-release prototype exploited a zero-day in OpenAI’s own package-registry proxy, escaped the sandbox, chained exposed credentials, and achieved remote code execution on Hugging Face servers. Prompted by that disclosure, Anthropic audited 141,000 eval runs and found three incidents of its own: eval environments that were supposed to be air-gapped weren’t, and Claude models reached real companies, in one case pulling hundreds of rows from a production database, in another publishing a malicious package that ran on 15 real systems. Two of the three victims never detected the intrusions.

Why it matters: Autonomous exploit development by frontier agents is no longer hypothetical. But note what actually failed. OpenAI’s incident involved a real zero-day; Anthropic’s models needed only weak passwords and SQL injection, and the root cause was harness misconfiguration, not model intent. Testing environments now need the same security as production systems, which is a new cost for every lab and a new product category for whoever sells it. And the fact that the victims didn’t notice any of it says as much about the state of company security as it does about the models.

OpenAI Says Internal Astra Model Produced Results on Ten Open Math Problems

OpenAI published ten results in math, quantum complexity, and theoretical CS produced by an internal, unreleased version of Astra, its next model family, including the first explicit non-sofic group (open since 1999) and the first improvement to the general sphere-packing exponent since 1978. Every proof ships with a machine-checkable Lean 4 certificate. OpenAI says finding all ten cost roughly $2,000 at Sol API rates.

Why it matters: The Lean certificates make this harder to dismiss than a benchmark claim, though no external mathematicians reviewed it pre-announcement and critics object to math by press release. If settling decades-old conjectures is now a compute line item, the value of formal, verifiable domains as training and product surface just repriced, and mathematician Thomas Bloom’s reaction (”big news”) suggests the field agrees this one is real.

Three Chinese Labs Ship Major Models Inside 24 Hours

On July 31 Beijing time, three releases from Chinese labs landed: DeepSeek shipped V4-Flash-0731, MiniMax released H3, an omni-modal video model generating 15-second 2K clips with native stereo audio at roughly a third of mainstream per-second pricing, and ByteDance launched Seedance 2.5, extending single-run generation to 30 seconds.

Why it matters: Axios calls DeepSeek’s update comparable to Opus 4.8 on coding at a ~99% discount, and OpenAI already cut Luna pricing 80% three weeks after launch, leaving Anthropic as the last premium-pricing holdout. Chinese open models now take 41% of global open-model downloads, and on Vercel’s AI Gateway, Seedance already leads video generation volume over Veo. Ex-OpenAI’s Zack Kass: “at some point, the next model doesn’t matter to you.”

Moonshot Releases Kimi K3 Weights and Technical Report

What it is: Moonshot’s Kimi K3, a 2.8T-parameter MoE with 104B active, native vision, and 1M-token context, launched via API July 16 with full weights following July 27. Artificial Analysis scores it #3 on its Intelligence Index, comparable to Opus 4.8 and GPT-5.5, behind only Fable 5 and GPT-5.6 Sol.

Why it matters: Nathan Lambert’s estimate is that K3 compresses the closed-to-open gap from 6-9 months to 3-5, which resets what closed labs can charge for. Two asterisks before anyone builds on it: the custom license requires a commercial agreement above $20M revenue, so this is open-weight, not open-source; and self-hosting realistically requires a 64-accelerator cluster, so for most teams “open” still means renting cloud inference.

Thinking Machines Releases Inkling-Small as Open Weights

Thinking Machines Lab released Inkling-Small, a 276B-parameter MoE with 12B active, about a quarter the size of Inkling, open-weighted on Hugging Face and fine-tunable on Tinker. Trained partly by on-policy distillation with Inkling as teacher, it beats the larger model on reasoning and agentic coding, hitting 80.2% on SWE-bench Verified, while trailing on knowledge coverage.

Why it matters: The student beating the teacher on reasoning is the notable part; distillation now produces smaller models that are better, not just cheaper. A small model that hits 80% on SWE-bench and that customers can fine-tune and own is the concrete version of Murati's pitch that companies should control their own models.

Google DeepMind Announces Gemini Robotics 2

DeepMind announced a suite of three models on July 30: a vision-language-action model for whole-body humanoid control and five-fingered dexterity, an embodied reasoning model that plans across multiple robots, and an on-device model that adapts to new robot bodies within hours. Demos ran on Apptronik’s Apollo 2, with Boston Dynamics and Agile Robots among named partners.

Why it matters: DeepMind robotics head Carolina Parada is explicit about the strategy: build “the intelligence layer that can be used by every robot.” If the brains become a Google API, hardware margins compress and the defensibility of vertically integrated humanoid bets weakens. Temper it with DeepMind’s own numbers: 40-44% success on complex fine-motor tasks. Dexterity is still unsolved.

FCC Adds Foreign Humanoids, Robot Dogs, and Power Inverters to Covered List

On July 28 the FCC added foreign-made mobile robots and connected power inverters to its Covered List, implementing a White House interagency determination. New models can no longer receive the equipment authorization required to import or sell in the US; existing devices are unaffected, with exemptions available case by case. The target is China, where Unitree, Agibot, and UBTech hold roughly 87% of humanoid shipments.

Why it matters: US humanoid startups just got a protected home market and more expensive development at the same time. Interact Analysis’s Rueben Scriven notes the near-term hit to Chinese vendors is small since US penetration was minimal; the real risk is slowing US commercialization by removing the cheap platforms teams prototype on.

Amazon Reported to Wind Down Most Nova Models

Business Insider reports (not yet confirmed by Amazon) that Amazon is deprecating Nova Premier, Omni, Reel, and consolidating behind a new Frontier Model Research group led by Pieter Abbeel, with a new flagship expected at re:Invent. This follows the closure of its SF AGI lab and the December reorg under Peter DeSantis.

Why it matters: Amazon is conceding the current model generation to concentrate on the next one, and the economics say it can afford to: AWS hosts $138B+ of OpenAI commitments and $100B+ from Anthropic, and Bedrock profits whichever model wins.

Substack Adds AI Detection; Snapchat Stops Recommending Fully AI-Generated Videos

Substack integrated Pangram’s AI detector, letting readers scan posts, Notes, and comments over 100 words published after July 21 for estimated AI-generated share; Pangram claims a 0.01% false-positive rate and finds over 20% of Substack longform already flags as AI-assisted. Days later, Snapchat made wholly AI-generated videos ineligible for Spotlight recommendations, while content enhanced with its own AI tools stays eligible with transparency indicators.

Why it matters: Platforms are converging on a stance: AI as tool is fine, AI as content gets suppressed, with LinkedIn and YouTube moving the same direction the same month.

Study Finds Half of AI Unicorns Have Never Published Research

A Stanford METRICS preprint from John Ioannidis’s group, covered by Science, analyzed all 317 AI unicorns and found 52.4% have never published a single paper or preprint. The whole cohort contributed 0.1% of AI literature in 2025; OpenAI alone accounts for 39.4% of the group’s citations.

Why it matters: You can no longer judge an AI company, or its researchers, by published work; half the most valuable startups in the field have none to show. What replaced it is what the authors call "blogification": unreviewed claims in corporate posts that feed both training data and hype cycles. The Hacker News steelman is that companies exist to sell what they invent; fair, but an ecosystem built on "Attention Is All You Need" has now mostly closed the door behind itself.

DeepSeek Reported to Be Developing an Inference-Focused AI Chip

Reuters reports, citing three unnamed sources, that DeepSeek has spent about a year designing a chip built for running models rather than training them, and is talking to chip design, foundry, and memory partners. DeepSeek has not commented, and nothing is confirmed.

Why it matters: The decision to build only for inference says a lot. DeepSeek’s own research measured Huawei’s Ascend 910C at about 60% of an Nvidia H100 on inference, and its R2 model reportedly had to fall back to Nvidia hardware for training. With roughly 70% of AI compute demand expected to be inference, a lab that designs both its own models and its own chips could squeeze Nvidia’s China business even if the chip never leaves the country.

Meta to Begin Production of Its Iris AI Chip in September

Meta’s newest in-house AI chip, code-named Iris, goes into production in September, designed with Broadcom and manufactured by TSMC, according to an internal memo obtained by Reuters. Meta plans 7 gigawatts of computing capacity in 2026, doubling to about 14 gigawatts in 2027, on spending of $125B to $145B this year.

Why it matters: Meta is not replacing Nvidia and AMD, it is adding its own chips alongside them. The steadier winner is Broadcom, which designs custom chips for Google, OpenAI, Meta, and ByteDance and holds a $73B backlog of AI orders. Custom chips are on track to reach about 28% of AI server shipments in 2026, growing nearly three times faster than off-the-shelf GPUs.

SK Hynix Raises $26.5B in Record US Share Sale; Shares Open 14% Above Offer

SK Hynix raised $26.5B in the largest first-time US share sale ever by a foreign company, topping Alibaba’s 2014 record, with demand running about seven times the shares on offer. The stock opened roughly 14% above the $149 offer price in Friday’s Nasdaq debut; Seoul shares are up 634% in a year. CEO Kwak Noh-jung told Reuters that 2027 will be the worst supply shortage year in the memory industry’s history.

Why it matters: SK Hynix supplies 56.4% of the high-bandwidth memory that AI chips depend on, and the deal cleared seven times demand in the same week memory stocks sold off sharply. The caution flag: analyst Patrick Moorhead notes that memory makers have sold below cost in past downturns, and the five-year customer contracts meant to prevent a repeat are untested. Chairman Chey Tae-won also said the company is open to selling more US shares.

OpenAI Launches ChatGPT Work and the GPT-5.6 Model Family

ChatGPT Work handles multi-step tasks across Slack, Teams, Google Drive, and SharePoint, builds documents, slides, and simple web apps, and can run tasks on a schedule.

Why it matters: Companies have been hit with surprise bills as AI usage grows, and Terra at half the price of GPT-5.5 answers that complaint directly. Early reviews are mixed: M.G. Siegler argues OpenAI copied the shape of Anthropic’s Claude Cowork but executed it less gracefully, burying the chat experience inside a work app.

xAI and Cursor Release Co-Developed Grok 4.5 at $2/$6 Per Million Tokens

Grok 4.5 launched July 8. It was co-developed with Cursor and trained on what Cursor describes as trillions of tokens of its data, and is available in Cursor from day one. xAI claims it beats Claude Opus 4.8 on coding benchmarks; independent testing ranks it fourth on general intelligence but first at using tools.

Why it matters: This is the cheapest model near the top of the market, albeit not the best one. The post training data itself is the new thing: Cursor trained a model on its users’ coding activity with a single lab, and other coding tools, and the customers whose work fed the training, now have to decide how they feel about that.

Cognition Releases SWE-1.7, a Coding Model Built on Kimi K2.7

SWE-1.7 starts from Moonshot AI’s Kimi K2.7 Code model, runs through Cerebras at 1,000 tokens per second, and is free at the standard tier inside Devin. In a companion post, Cognition reports the original Chinese model completed 87% of tasks that US models refuse, including writing surveillance software, and says its additional training brought that behavior in line with Claude and GPT.

Why it matters: Kimi K2.7 had already been through heavy reinforcement training at Moonshot, yet Cognition’s extra training on its own usage data still produced large gains. Companies with the right data can apparently keep turning open Chinese models into competitive products for a fraction of what training from scratch costs.

.

ByteDance Releases Seedream 5.0 Pro Image Model

The July 8 release adds precise editing tools, layer separation with transparent export, infographic generation, and text rendering in more than 10 languages, at roughly $0.03 per image versus about $0.28 for GPT Image 2.

Why it matters: Seedream is roughly 3x to 9x cheaper depending on resolution, but GPT Image 2 still renders text more reliably, 98.5% versus 89.5% accuracy, so design teams may not save as much as the price list suggests. Hands-on comparisons put Seedream ahead on price and editing features, not image quality.

Thinking Machines Lab Publishes Vision for AI That Extends Human Judgment

The lab published “The Future Worth Building Is Human” on July 10, arguing AI should extend human will and judgment rather than replace it, and that users, not a single lab, should decide how models behave by training them on their own data. The essay rejects the industry’s favorite progress measure, how long an AI can work without supervision.

Why it matters: The essay doubles as the case for Tinker, the company’s product for customers who train and own their own models, and stakes out a position against the fully autonomous agents the other big labs are chasing.

EVENTS

Join us for cold beer and Hot Chips in the Fall!

RSVP

The world's largest hedge fund gave frontier models a single job: read the day's financial documents and flag what matters. With prompts written by Bridgewater's own investors, the best frontier models couldn't clear 80 percent accuracy. A model Bridgewater fine-tuned on Qwen (an open source model), with Mira Murati's Thinking Machines, cleared the bar at a fraction of the inference cost.

This gets at the question I've been chewing on all week: do we need better models, or better systems around the models?

Big bids for custom models have entered the arena this week. Palantir’s Alex Karp went for drama on CNBC this week where he chewed out OpenAI and Anthropic for converging into “commodity cognition”.

(And yes obviously they are hawking the problem to the solution, with an announced partnership with Nvidia on Nemotron to run custom open-weight models. )

The industry seems fairly split on this. On one hand, there’s a largely unspoken worry about using foreign LLMs, even if it’s open source.

On the other, frontier token costs are really, really expensive relative to open models. And the bet is that custom open models, fine-tuned to your needs and data, will outperform.

Anecdotally, I think both fears are real. In my conversations with AI executives at large corporations, I’m finding that the honest answer right now is more around usage maturity. AI usage at isn’t high or sophisticated enough yet to be prohibitively expensive or warrant switching to custom models (unless it outperforms).

Bridgewater’s result worked because the fund had a high-value task, senior experts whose judgment was worth encoding, and the evaluation infrastructure to know when the model had actually absorbed it. Most enterprises have none of those yet, which is why they buy tokens.

The challengers have entered the arena. The question is how many buyers are ready to meet them there.

Hope you had a great 4th! Enjoy the update.

-Tara

THE DOWNLOAD

Meta Building Cloud Business to Sell Excess AI Compute

Meta is developing a cloud business, internally called Meta Compute, to sell surplus GPU capacity and hosted access to its closed-weight Muse Spark models, competing with AWS, Azure, and Google Cloud. The move follows SpaceX/xAI leasing Colossus capacity to Anthropic (reportedly $1.25 billion a month) and Google, and aims to recoup Meta’s $182.9 billion in AI infrastructure commitments. Meta shares rose roughly 9%; CoreWeave and Nebius fell about 12% each.

Why it matters: The market read this correctly, and the losers are the neoclouds, not the hyperscalers. Meta is a major customer of CoreWeave and Nebius; if it can serve its own excess capacity, it may not need theirs. The deeper signal is that a company that spent $183 billion chasing superintelligence is now building a fallback that monetizes the infrastructure regardless of whether its own models win.

Unitree Robotics Wins Approval for $619M Shanghai IPO

China’s securities regulator approved Unitree’s STAR Market listing on July 3, clearing the humanoid and quadruped maker to raise 4.2 billion yuan ($619 million) at a valuation near 42 billion yuan. Unitree is rare among robotics firms in being profitable, reporting 1.7 billion yuan in 2025 revenue and 591 million yuan in adjusted profit, and shipped more than 5,500 humanoids last year, first worldwide by volume.

Why it matters: This is mainland China’s first major humanoid listing, and it hands the sector a public-market benchmark at a moment when Chinese embodied-AI funding has hit records of roughly $10.8 billion in 2025 and exceeded $2.9 billion in the first two months of 2026 alone

OpenAI Floats 5% Equity Stake for US Government

OpenAI has held early-stage talks about giving the US government a roughly 5% stake, worth about $42.6 billion at its $852 billion valuation, via a vehicle modeled on the Alaska Permanent Fund. Altman first raised the idea in early 2025 and has pitched a version in which every leading US lab contributes 5%.

Why it matters: The proposal surfaced days after OpenAI delayed GPT-5.6’s launch at the government’s request, and it sits between unfettered private ownership and Senator Sanders’s push for 50%. A voluntary equity giveaway from a company still years from IPO is a novel way to buy political goodwill, and if it sets precedent, foreign governments will demand the same, complicating every cross-border deployment.

Anthropic Restores Fable and Mythos After Export Controls Lifted

The US Commerce Department lifted export controls on Claude Fable 5 and Mythos 5 on June 30, ending a roughly two-week shutdown. Fable 5 returned globally July 1; Mythos 5 came back only for a set of approved US organizations. The freeze began June 12 after Amazon researchers demonstrated a jailbreak; Anthropic pulled both models from every cloud platform because it couldn’t verify user nationality in real time. Anthropic will now pre-release frontier models for federal review.

Why it matters: Mythos 5’s phased return through Anthropic’s Glasswing program establishes a trusted-partner tier sitting between full public access and total suspension. The pre-release review commitment also raises the question of whether Washington now effectively approves every frontier launch.

Microsoft Launches $2.5B Frontier Company for Enterprise AI

Microsoft created Microsoft Frontier Company, a $2.5 billion unit embedding 6,000 engineers inside customers like Unilever, Novo Nordisk, and LSEG to integrate AI, including rivals’ models, with proprietary data, and customers keep the resulting IP. It arrives days after AWS committed $1 billion to a similar embedded-engineering unit and after OpenAI’s $4 billion-backed Deployment Company.

Why it matters: Every major AI vendor has stood up the same forward-deployed model within eight weeks, all aimed at one number: roughly 95% of enterprise AI pilots deliver no measurable result. The battleground is shifting from who builds the biggest model to who can make deployment stick inside a customer’s operations.

Bending Spoons Surges 40% in Nasdaq Debut

Bending Spoons, the Milan firm that acquires and revives legacy software like AOL, Evernote, Vimeo, and WeTransfer, closed up nearly 40% on July 1 at a $25.7 billion market cap, more than double its last private valuation. Revenue grew from $387 million in 2023 to $1.31 billion in 2025, and it swung to a $27.5 million quarterly profit. The company carries roughly $4.4 billion in debt and has flagged 1,000-plus acquisition targets.

Why it matters: A debt-funded acquire-and-fix operator popping 40% in a year investors feared AI would gut traditional SaaS is a real signal about where public appetite sits. Bending Spoons explicitly marketed itself as AI-native (”AI before it was cool”), and buyers rewarded the framing.

AI Designs Radio Chips That Outperform Human Layouts

Princeton researchers used reinforcement learning, inverse design, and diffusion models to generate RFIC layouts that beat human-designed circuits on bandwidth, power, and efficiency while cutting design time from weeks to minutes. The AI-produced structures look “more like a QR code” than conventional symmetric layouts, unconstrained by human templates.

Why it matters: RF chip design has stayed a hand-crafted “dark art” even as CPUs and GPUs went algorithmic, so this is a genuine unlock in one of the last manual corners of semiconductor design. Compressing years of iteration into minutes changes the economics of building custom RF silicon for 5G/6G, autonomous vehicles, and satellite links.

Anthropic Launches Claude Science Workbench

Anthropic launched Claude Science on July 1, a research workbench connecting 60-plus scientific databases with prebuilt genomics, proteomics, and chemistry toolkits, a reviewer agent that checks citations, and full provenance on every figure. It runs on existing Claude models with no new capabilities and no gating, available in beta to all paid subscribers. It contrasts with OpenAI’s gated GPT-Rosalind and Google’s proprietary Gemini for Science.

Why it matters: The bet here is that what slows science isn’t raw model capability but the friction of stitching together dozens of databases and tools, so Anthropic is competing on workflow ownership rather than a specialized model, the same move that made Claude Code sticky. The three labs have now split into distinct strategies: Anthropic on broad access plus a workflow layer, OpenAI on a governed specialist model, Google on proprietary foundation models.

Kioxia Ships 332-Layer Flash Samples for AI Data Centers

Japanese memory maker Kioxia began sampling its 10th-generation 332-layer 3D NAND on July 2, offering 59% higher bit density and a 33% faster interface than its 8th-generation chips, aimed at enterprise and data center SSDs. Kioxia holds roughly 10% of the data center flash market against Samsung’s 40% and SK Hynix’s 30%; shares rose 8.9% after swinging double digits.

Why it matters: The timing matters more than the density gain. This lands with AI flash supply tight and Kioxia’s 2026 output largely sold out, and an Omdia analyst argues its NAND speed is the metric US hyperscalers weigh most. But sampling is not volume: operators typically evaluate for months before ordering, and mass production isn’t slated until 2027, so this is a claim on share, not a shift in supply. Kioxia’s stock, up more than 680% this year, is now priced on the assumption memory tightness holds.

Google Reports 37% Jump in Electricity Use Driven by AI

Google’s 11th environmental report, released June 30, showed a 37% year-over-year rise in electricity demand, its largest ever, pushing total use up more than 250% since 2019. The company matched 100% of consumption with renewable purchases and cut operational emissions 2%, but supply-chain emissions rose 25%, with data center construction alone adding 2.3 million tons of CO2. Google states plainly that its AI buildout is outpacing grid decarbonization.

Why it matters: Operational emissions are addressable with power-purchase agreements. Scope 3, embedded in chips, HBM, steel, and construction across grids in Taiwan, Japan, Vietnam, and India, is not. Every hyperscaler scaling AI compute inherits the same split, and it reframes the energy question from how much clean power a company contracts to where and when its data centers actually draw, and what its supply chain burns to build them.

EVENTS

Join us for cold beer and Hot Chips in the Fall! August 24 @ 530pm in Palo Alto

RSVP

Things in my mental orbit this week:

• I continue to obsess over open source inference infrastructure and the tools and services that will support custom software. Hearing a lot about agencies (some venture-backed) going after “vibe-hardening” - the production layer of vibe-coded tools.

• Compute hardware (both industrial and! consumer) will continue to climb. My poor Macbook is groaning under the weight of running claude code and google meets in its day-to-day. There will be solutions to solve this.

• Very much into the thesis of manufacturing molecules. As AI accelerates the design space (drug discovery, materials discovery), where the differentiation lies is in how these new discoveries are manufactured. More to be announced soon!

• Even as AI takes up a ton of airtime, there’s lots of invention happening in the frontiers - our recent Strange Gatherings dinners had folks committed to pushing the boundaries of fusion, quantum, robotics, and biology.

Now on to the download.

-Tara

THE DOWNLOAD

Government grants Anthropic limited clearance to redeploy Mythos 5

The government cleared Anthropic to bring Mythos 5 back for a vetted set of US critical-infrastructure organizations, partially reversing the June 12 export-control directive that had forced both Mythos 5 and Fable 5 offline. The clearance landed hours after OpenAI agreed to stage GPT-5.6 at Washington’s request.

Why it matters: This is the first working instance of a frontier-model clearance process: the government now decides which named organizations can run the most capable models. Model access is now a licensed privilege granted case by case. Fable 5, the consumer-facing model, is still dark.

Chinese models pass US models in OpenRouter token usage

Weekly OpenRouter usage, surfaced by Christophe Barraud, shows Chinese models near 18 trillion tokens against roughly 5.5 trillion for US models, a reversal from January.

Why it matters: The capability gap in Chinese models has closed enough that scores now sit within a point of the US leaders on real coding tasks, and known researchers have started reaching for these models by choice. Matt Velloso, a former VP at Meta and Google DeepMind, said he'd been using GLM 5.2 "all day" and called it the first open model that passes as a "daily driver."

OpenAI and Broadcom unveil the Jalapeño inference chip

OpenAI and Broadcom unveiled Jalapeño, OpenAI’s first custom inference accelerator, taped out in nine months, with gigawatt-scale Microsoft deployment starting late 2026.

Why it matters: OpenAI used its own models to compress the design cycle, a self-improving loop on the chips that run the models. Broadcom now builds custom silicon for Google, Meta, ByteDance, and OpenAI.

Memory prices keep climbing as Micron sells out HBM and Apple raises prices

Micron posted a strong fiscal Q3 and raised guidance, with gross margins near 81%, its entire 2026 high-bandwidth memory output already contracted, and $22B in long-term commitments locked in. Goldman calls the 2026 DRAM gap the worst in 15 years. Partly due to the industry-wide memory shortage, Apple raised MacBook and iPad prices by up to $300, saying it can no longer absorb the cost. Microsoft followed within hours on Xbox. The iPhone was spared.

Why it matters: Memory has flipped from boom-bust commodity to oligopoly utility. The three makers that produce nearly all the world's DRAM are now pre-selling years of output at 80% margins, with contract prices running up 50%-plus in a single quarter and AI data centers first in line. Apple, with the best supply chain on earth, still couldn’t hold prices due to the squeeze.

(Editor’s Note: expect a deeper dive on Tuesday on this!)

OMB proposes a rule reshaping all federal research funding

OMB’s proposed “Regulation for Federal Financial Assistance” would put political appointees over peer review on discretionary grants, allow mid-project terminations with no appeal, and restrict foreign collaboration. It covers about $1.1 trillion in annual grantmaking, with comments due July 13 and an October 1 effective date.

Why it matters: This hits the layer upstream of deep tech. University labs are where deep tech research originates, and the rule injects termination risk and political review straight into that pipeline.

UPCOMING EVENTS

​Strange Sessions is a monthly demo night where builders share what they’re working on.

​Each session revolves around a theme and features short demos from founders, researchers, and creative technologists.

Come jam

Power moves this week. Two of the eight authors from the seminal “Attention Is All You Need” paper on transformers that reshaped the AI landscape were won over to Big AI companies this week. Noam Shazeer left Google for OpenAI. Ashish Vaswani, the lead author, went to Nvidia.

The biggest power move on my desk this week, though, was /goal. It feels like a cheatcode, honestly. Both Claude code and Codex shipped a version of a slash command where you set a goal, then the agent runs itself in a loop, checking after each turn whether the condition is met and continuing on its own until it is. Hot tip: make sure to plan before building, if a goal is not well defined, it can run in circles for hours and guzzle token spend.

Use:

For this task, write yourself a new goal and spawn agents in parallel — as many as needed to do it better and faster. Split the work into independent pieces, dispatch them concurrently, and synthesize the results as they return. Give each agent its own dedicated /goal.

Have a great week - Tara

THE DOWNLOAD

Senator Says NSA Director Reported Anthropic’s Mythos Breached Most Classified Systems in Hours

Senator Mark Warner, vice-chair of the Senate Intelligence Committee, said on June 11 that the head of the NSA and Cyber Command, General Joshua Rudd, told him Anthropic’s Mythos model “broke into almost all of our classified systems, not in weeks, but in hours,” as reported by The Economist (Digg). The claim lands amid a broader fight: the Pentagon labeled Anthropic a supply chain risk in March, yet the NSA is reportedly running the cybersecurity-focused Mythos for offensive operations with about six Anthropic engineers embedded, and the model now sits under export controls (Tom’s Hardware).

Why it matters: If you strip out the drama, you’l find that most security people read “broke into all our classified systems” as a loud retelling of “found serious holes in our systems,” but no one has shown an AI model breaching air-gapped classified networks on its own (Digg). What isn’t in dispute is the bigger shift. A frontier model is now handled like a controlled weapon, export-limited and deployed inside the NSA because it can find and chain software vulnerabilities faster than human teams can patch them.

Nvidia Hires Essential AI Founder Ashish Vaswani and Several Researchers

Nvidia has hired Ashish Vaswani, founder and CEO of Essential AI and first-listed author of the 2017 Transformer paper, along with several Essential researchers, with Vaswani set to work on Nvidia’s open-source Nemotron models, according to a source close to the startup (Ground Level AI). Several team members have updated their LinkedIn profiles to Nvidia. Essential counted Nvidia, AMD, and Google as strategic investors, and the same source said Vaswani had been struggling to raise more money, with pulling the team away from AMD an added motivator (Ground Level AI).

Why it matters: Nvidia just absorbed the lead author of the paper that created modern AI and pointed him at Nemotron, its own family of open models. Nvidia keeps building more of the software and model layer that rides on top of its chips, and it’s buying the people to do it.

Z.ai Releases GLM-5.2 Open Weights, Tying Claude Opus 4.8 on a Physics Benchmark

Z.ai’s open source model GLM-5.2 ties Claude Opus 4.8 at 20.9% on CritPt, a benchmark of unpublished physics problems, and leads all open-weights models (Artificial Analysis). They released this, pointedly, right after Anthropic was asked to pull back Fable on orders of the US government.

Why it matters: For the everyday coding work most teams actually run, GLM-5.2 now performs like a closed frontier model at roughly a sixth of the price. It trails Opus 4.8 by less than a point on the common agentic benchmarks and beats GPT-5.5 on SWE-bench Pro (VentureBeat), and because the weights are open you can run it yourself and take the per-token bill to zero. The frontier still shows up on the hard cases: on the longest multi-hour tasks Opus pulls about twice as far ahead (LLM Stats), and GLM-5.2 gets roughly a quarter of factual questions wrong (danilchenko.dev).

SpaceX Agrees to Acquire Cursor for $60 Billion in Stock

SpaceX signed a definitive all-stock agreement to buy Cursor parent Anysphere at a $60 billion valuation, four days after its record Nasdaq IPO (CNBC). At roughly $4 billion in annualized revenue and more than a million paying users, this is the largest acquisition of a venture-backed startup on record (TechSpot).

Why it matters: Coding is the killer use case for LLMs, and the single largest category of enterprise AI spending, about $4 billion in 2025 and 55% of all departmental AI budgets, and it has become the gateway into enterprise workflows (Menlo Ventures). Grok had almost no foothold there, so buying Cursor drops xAI inside the category overnight.

Noam Shazeer Leaves Google DeepMind for OpenAI

Noam Shazeer, Gemini co-lead, Transformer co-author, and VP of engineering at Google DeepMind, is joining OpenAI as Lead for Architecture Research, working on next-generation model designs (TechTimes). The move lands about ten days after OpenAI filed confidentially for an IPO that could value it near $1 trillion (MLQ).

Why it matters: The interesting part to note: Shazeer's specialty, mixture-of-experts and multi-query attention, is the set of techniques that make a huge model cheap enough to run without losing money on every answer (TechTimes). That's the exact problem OpenAI has to solve as it heads toward an IPO while still losing money at scale. You can always buy more compute; what you can't buy is one of the few people who can redesign a model to need less of it, which is why Google paid about $2.7 billion to bring him back less than two years ago (ThePlanetTools).

FERC Orders Six Grid Operators to Revise Data Center Connection Rules

FERC issued six show-cause orders under Section 206, giving PJM, MISO, SPP, CAISO, ISO New England, and NYISO 60 days to justify or rewrite how data centers connect to the grid, plus a 30-day report on whether they have enough generation to serve them (Data Center Knowledge). The unanimous orders cover regions serving about 200 million people and are explicitly meant to keep data center costs from landing on other ratepayers (E&E).

Why it matters: Power, not chips, has been the real bottleneck for the AI buildout, and FERC just moved to clear it. But the heart of the order is a fight over who pays. In Maryland, regular customers are already on the hook for about $1.6 billion over the next decade for transmission that mostly serves data centers across the line in Virginia (Utility Dive), and double-digit bill hikes last summer set off real backlash (E&E). FERC’s fix is to push those costs onto the data centers themselves and fast-track the ones that bring their own power or agree to dial back when the grid is tight. The risk runs the other way too: if utilities overbuild for demand that never fully arrives, it’s ratepayers, not developers, who get stuck with the stranded plants (TechTimes).

Google DeepMind Publishes an AI Control Roadmap Treating Its Own Agents as Insider Threats

DeepMind released a 35-page AI Control Roadmap, a defense-in-depth framework that treats internal AI agents as potential insider threats rather than trusting alignment alone (DeepMind). It introduces TRAIT&R, a taxonomy of rogue tactics modeled on MITRE ATT&CK across four detection and three response tiers, and reports an analysis of more than a million coding-agent tasks where most flagged behavior traced to overeager agents, not malice (Fortune).

Why it matters: One of the top AI labs just admitted, in writing, that it can't count on its own agents staying aligned, so it now treats them like employees who might go rogue: limited access, constant monitoring, the ability to pull the plug. The useful part is the data behind it. Across a million internal coding tasks, almost all the flagged problems were overeager agents overstepping their instructions, not scheming ones (TechTimes), so the thing companies actually need to catch is good intentions gone too far, not sabotage (for now!)

Midjourney Launches a Health Division and a Whole-Body Ultrasound Scanner

Midjourney, the generative-AI image company, launched a health division and a whole-body “Ultrasonic CT” scanner at a San Francisco event, with a first “Midjourney Spa” clinic planned for late 2027 (RDWorld). Founder David Holz claimed a roughly 60-second scan with image quality in some ways superior to MRI, with no radiation or magnets (The Next Web). The imaging is not Midjourney’s generative AI; it runs on Butterfly Network’s ultrasound-on-chip technology, licensed in November 2025 for up to $74 million over five years.

Why it matters: This is what a founder with no investors does. Midjourney is bootstrapped, profitable, and answers to no board, so Holz can sink tens of millions into a decade-long bet that has nothing to do with the image business. And it fits him: he spent twelve years running a body-tracking hardware company (LeapMotion) before this, and has signaled for a while that Midjourney's real ambition is hardware and modeling the physical world.

DEEP DIVE FROM THE REVIEW

Coding is, by far, the killer use-case for LLMs. Claude Code is arguably one of the primary, bottoms-up drivers of Anthropic’s revenue growth in the last 18 months.

Cursor is SpaceX’s bet on building the same thing. This cements something I’ve been noodling over since the OpenAI / Nvidia deal last September:

If you want to be a Big AI company, you have to go full stack. I break it down here:

EVENTS

Strange Sessions #5: Loops

​Strange Sessions is a monthly demo night where builders share what they’re working on.

​Each session revolves around a theme and features short demos from founders, researchers, and creative technologists.

Come jam

Strange Gatherings: The Inventors’ Table

We’re absolutely at capacity for Strange Gatherings: The Inventors’ Table at SF Deep Tech Week, but if you sign up for the waitlist, we’ll invite you to the next one! We’ hope to get a bigger venue next year.

Request an Invite

Anthropic went from $87M ARR in January 2024 to $47B ARR in May 2026, roughly 540x growth in 28 months. That growth had two engines working in tandem: top-down enterprise partnerships (Claude is the only frontier model on all three major clouds) and bottom-up developer adoption through Claude Code, the fastest-growing product in the company’s history, rocketing from $0 to $2.5B ARR in 9 months. Anthropic now owns 54% of the enterprise AI coding market.

Cursor is SpaceX’s bet on building the same thing.

Yesterday, SpaceX announced its acquisition of Anysphere, the company behind Cursor, the AI coding tool used by 7 million developers daily, for a whopping $60B in stock. Cursor was founded just four years ago out of MIT, hit $2B ARR, and became the highest-revenue AI coding tool in the category. Its share has been slipping (from 41% to 26% over the past year as Claude Code surged), but what xAI is buying isn’t market share.

xAI already had the full stack: Colossus (compute), Grok (model), X (app). But X is where people scroll. Cursor is where they build. This would yield probably one of highest-signal training data in AI, and exactly what you need to make Grok competitive.

This cements something I’ve been noodling over since the OpenAI / Nvidia deal last September:

If you want to be a Big AI company, you have to go full stack.

It’s only becoming more apparent. Better products drive better infrastructure (more data), and better infrastructure drives better experiences (this has always been core to our investment thesis at Strange).

Going full stack does two things:

1. Unit economics of building and training AI become sustainable.

2. You get proprietary training data from the application layer that differentiates you from model competitors. User data and workflow lock-in become an attractive competitive moat.

We’ll probably see movements over the next few years:

• Model companies building their own applications from the inside out, or

• Aggressively acquiring up the stack into the application layer

There’s a bit of a saying amongst founders now that because it’s become 10x easier to build, companies have to be 10x more ambitious to succeed. Seems like we’re seeing that all across the board.

- Tara