The Brief: Mythos Redeployed
Come join us at Strange sessions #5: loop engineering! Plus: OpenAI ships its own inference chip, memory shortage raises Apple's prices, and more
FIELD NOTES
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.
I see more people launching hardware companies these days