Subscribe
Share
Share
Embed
Super Intelligence Force fills the AI czar seat
A daily summary of what is interesting and happening in the AI industry, with a focus on what this means for people building harness experiences that are used.
Good morning, it's Monday, October fifth.
In today's briefing we see a frontier-class open model running on a twelve-gigabyte gaming GPU, the White House naming a chair for its renamed Super Intelligence Force, and Google freezing its open-source bug bounty over a flood of AI-generated reports.
Local model developments.
The hardware floor for running frontier-class open models keeps falling. A new MIT-licensed engine called Strata runs Qwen 3.8 Flash Next, a mixture-of-experts model the project counts as roughly one hundred twenty-five billion parameters, on a consumer GPU with twelve gigabytes of video memory.
It keeps the most-used experts in GPU memory, pins the rest in system RAM, spills to SSD, and adds speculative drafting. The author reports ninety-four tokens per second on an RTX 5070 at two-bit quantization.
Independent evidence is thin. One secondary test on an RTX 4070 measured about fifty-three tokens per second at sixty thousand tokens of context, roughly double a tuned llama.cpp baseline. Strata's own documentation lists forty-six to seventy-four at one hundred twenty-eight thousand context. Every quality number comes from Qwen itself.
There is a practical catch: startup can freeze a machine for one to three minutes while thirty-five to fifty-five gigabytes loads.
This follows DwarfStar 4, which needed one hundred twenty-eight gigabytes of RAM. In a few days the requirement dropped from a workstation to a gaming PC. High-volume, low-stakes calls are where hosted per-token pricing is now hardest to defend, so re-cost those first.
In AI Infra.
Red Hat compared nine guardrail approaches on prompt-injection and content-safety tasks. Its conclusion is that decision models like Jev do not reliably beat an LLM acting as judge, or pre-trained classifiers, on speed or accuracy. Classifiers ran in roughly thirty-three to eighty milliseconds with competitive accuracy, and Qwen 3.6 35B as a judge topped the prompt-injection tests.
Two caveats. Red Hat is building lightweight predictive models into its own default guardrails, so it has a stake. And the post has drawn little attention on Hacker News, so readers have barely tested it.
Every team racing to replicate Jev assumes speed is the prize. This is the first test that questions that premise.
On the policy front.
The White House has named National Intelligence Director Jay Clayton, a former SEC chair with no tech-industry background, to chair the renamed Super Intelligence Force, filling the vacant AI czar role. The vice chairs are FTC Chair Andrew Ferguson, Pentagon CTO Emil Michael, and OPM Director Scott Kupor. The force has one hundred twenty days to report on AI risks, federal responsibilities, and what Congress should do.
It sits beside a one-page Joint Commitment on Frontier Responsibility, signed by the CEOs of Anthropic, Google, Meta, Nvidia, OpenAI, and xAI. TechCrunch's Equity hosts call it completely voluntary, and Kirsten Korosec reads it as a rebrand.
Neither the compact nor the force adds an enforcement mechanism, so neither changes who controls access to frontier capability. Build against your own agent-safety controls. The report due around early February is the first look at what compliance could mean.
In other news.
Google has paused its Open Source Software Vulnerability Rewards Program, citing a significant rise in automated submissions, the vast majority of which are not valid. Maintainers describe being buried in AI-generated reports with hallucinated findings. Google points participants to its other bounty programs until an update expected in the first quarter of next year.
The cost of filing a report fell to near zero while the cost of triaging one did not, so intake closes. Security agents that file reports need to ship reproducible proof-of-concept output, because validated findings are the product.
Separately, a command-line tool called RemoveMacAI turns off Apple Intelligence on macOS 27, deletes the on-device models, and blocks re-download. It disables Siri, Writing Tools, Genmoji, Image Playground, ChatGPT integration, and AI summaries. It works on Apple silicon only and is reversible. The tool drew five hundred thirty-six points and three hundred fifty-one comments on Hacker News, because macOS 27 has no single off switch for Apple Intelligence.
On-device AI now needs a visible kill switch, and a clear statement of what each feature stores locally.
That's the briefing. Have a great day, and don't forget to subscribe.