FIELD NOTES
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
Why We Invested in Multiply Labs
AI is fast expanding what scientists can search, design, and test. Over $11 billion in venture funding has poured into AI/ML drug-discovery in 2025 alone. As discovery becomes more powerful and therapies more complex, the question of how to produce them becomes increasingly consequential.
“The pharmaceutical factory of the future should look more like a semiconductor factory: a large cleanroom with robots doing the work. Today, it’s manual, impossibly slow, and full of errors.”
— Fred Parietti, CEO, Multiply Labs