AI briefingWednesday, 21 October

Jessica: So DeepSeek just open-sourced the actual inference engine they use in production.

Eric: Wait, the thing serving a hundred fifty-five trillion tokens a day?

Jessica: That exact stack. With the prompt caching, the whole thing.

Eric: Okay, that's wild.

Jessica: This is dailybrief, I'm Jessica.

Eric: And I'm Eric. It's Wednesday, October twenty-first, twenty twenty-six.

Jessica: So yeah, DeepSeek released DeepInfer. It's built on vLLM, they credit SGLang and TRRTLLM, but this is their production setup. They're claiming ninety-six percent cache hit rates and fourteen thousand output tokens per second.

Eric: And they say they're running eighty percent inference margins with this?

Jessica: According to Latent Space, yeah. The prompt caching goes up to seventy-two hours.

Eric: That's... honestly that's the part I care about most. If those cache numbers hold up in other workloads, that changes the economics pretty fast.

Jessica: Right, and it's open-source, so you can just go look at how they're doing it.

Eric: Speaking of production stuff, what's this Terminal X thing?

Jessica: So Terminal X built an eval agent using GPT six point one Sol, and it matched senior investor judgment eighty-eight point six percent of the time. OpenAI reported this.

Eric: What's it evaluating?

Jessica: Investment fund mandate compliance. They've got a fifty-four criterion rubric, and it catches twenty-five percent more mandate deviances than before.

Eric: Eighty-eight point six percent expert match is... that's actually a useful benchmark if you're building compliance tools.

Jessica: Yeah, and it's GPT six point one, so not even the bleeding edge.

Jessica: Actually, speaking of older models—there's this ALTK-Evolve thing that scored sixty-seven point five percent on tau-squared-bench retail tasks using GPT four point one.

Eric: Four point one? That's ancient.

Jessica: Right? But they used it to extract behavioral rules from the system's own training attempts. No weight updates, no human annotation. It topped the October twelve retail leaderboard, according to Hugging Face.

Eric: So it's learning guidelines, not retraining the model.

Jessica: Exactly. You just append the learned rules and run the same model.

Eric: Huh. That's actually pretty clever if you're trying to improve agent performance without the compute bill.

Jessica: Okay, NVIDIA released something called Cosmos three H Dreams, and it is... it's a real-time generative surgical simulator.

Eric: A what now?

Jessica: It generates realistic surgical video as you move virtual instruments. Trained on clinical cholecystectomy video and robot kinematics. Runs on a single RTX PRO six thousand, according to Hugging Face.

Eric: So you're not modeling physics, you're just—

Jessica: —generating what it would look like. It's a world foundation model. It learns the visual dynamics from video instead of explicit geometry.

Eric: That's... honestly that feels like a big deal for surgical training. If it actually looks right.

Jessica: Yeah, and it uses twenty-eight action dimensions, so it's tracking pose, gripper, all of that.

Eric: On the much lighter end of things, Simon Willison built a game?

Jessica: He did! Battle Drones, it's a Command and Conquer style RTS. He used Claude Fable five and Claude Code with TypeScript. Nine buildings, seven ground vehicles, six aerial vehicles.

Eric: And the multiplayer works how?

Jessica: Audio. Two nearby phones transmit game state through sound to synchronize.

Eric: That's delightfully weird.

Jessica: It's also got AI versus AI mode with different personalities. He calls it vibe coding on his blog—just describing features to the AI without reviewing the code.

Eric: I mean, if it gets you to a playable prototype...

Jessica: Okay, policy corner. Anthropic circulated draft legislation that would let judges block AI releases if over fifty deaths or a billion dollars in damage is foreseeable.

Eric: Wait, Anthropic wrote a bill?

Jessica: TechCrunch says Politico reported it. Independent evaluators registered with the Commerce Department would assess catastrophic risk. They'd have seventy-two hours to notify Commerce of imminent risk, and no federal agency could halt distribution without judge approval.

Eric: So you'd need a court order, not just an agency decision.

Jessica: Right. It's defining the thresholds—fifty deaths or a billion in damage.

Eric: That's... I don't know how you forecast that, but at least it's specific.

Jessica: And then internationally, China's MOFCOM proposed requiring pre-approval for sharing frontier models across borders, while DHS blocked DeepSeek from federal civilian procurement. This is from Interconnects.

Eric: So DeepSeek the company and the models can't be used in US government contracts?

Jessica: Correct. Though it's unclear how that applies to finetuned or derivative models.

Eric: And China's doing approvals but with automatic approvals for open-source?

Jessica: That's what the proposal includes, yeah.

Eric: So everyone's drawing lines, just in different places.

Jessica: Quick one—there's a new benchmark called FairStress on Hugging Face. It tests whether models resist persuasive arguments equally across demographic groups.

Eric: So like, are you harder to persuade away from the right answer depending on who benefits?

Jessica: Exactly. Three hundred two-candidate scenarios across twenty-nine domains. Hiring, lending, healthcare, housing, criminal justice.

Eric: That's a good test to have.

Jessica: Uh, a Schlotzsky's and Cinnabon franchisee cut menu repricing time ninety-five percent with ChatGPT Work, OpenAI said. Also cut cash reconciliation ninety-four percent.

Eric: That freed up a hundred thousand dollars in cash flow?

Jessica: Yep. Just workflow automation for a franchise operator.

Eric: Gemini's Personal Intelligence is coming to cars, according to Ars Technica, who suggested people audit their settings.

Jessica: Yeah, no technical specs, just the privacy reminder.

Eric: Oh, and Electric Plant raised eleven point two million for a model that interprets electrical signals from plants.

Jessica: It's called Bombadil. Works on about a hundred twenty-five species, sensors cost tens of dollars. TechCrunch says they're claiming twenty percent water reduction while boosting yields.

Eric: That's... okay, that's cool if it works.

Jessica: And llama dot cpp released b11786, which computes f32 gelu-quick without the fp16 table.

Eric: Just updated CPU activation paths, nothing dramatic.

Jessica: That's it for today. Every story is linked to its source over on the dailybrief website.

Eric: We'll see you tomorrow.