WEBVTT turn-001 00:00:04.000 --> 00:00:30.856 On Saturday evening, the president posted on Truth Social that he's creating an AI Force — modeled on the Space Force — and that he'll be naming a new AI czar soon. That's the announcement. Axios ran it Saturday, and what Axios kept circling back to is everything the post leaves out: no mandate, no budget, and no word on where this body would sit inside the federal government. They asked the White House for comment and got nothing back. turn-002 00:00:30.856 --> 00:01:00.904 The post isn't empty, though. It tells you what the new force is against. He writes that it should 'not in any way hinder or stifle' the industry. Bad actors get handled by looking for, in his capitalization, 'BAD' behavior, through what he calls 'our already existing Criminal and Civil Justice System.' So the enforcement theory is no new rules and prosecution after the fact. For a body that doesn't exist yet, that's a surprisingly specific limit. turn-003 00:01:00.904 --> 00:01:26.152 And the czar desk has been empty since March, when David Sacks left it. So we've got a vacancy that's six months old and a new institution with no shape, both announced in a single post. NBC News covered it. The Guardian covered it twice, and Edward Helmore's piece there says Trump 'gave almost no details about either plan' — which is about as far as a reporter can go when the source document is four sentences. turn-004 00:01:26.152 --> 00:01:49.552 Before we go further, count the coverage. Nine outlets today, all reporting on the same Truth Social post. Axios had it Saturday, NBC News picked it up, the Guardian ran two pieces, and CNBC and Techmeme carried it forward. That's one primary document refracted nine ways. Every number and every quote traces back to the same four sentences. turn-005 00:01:49.552 --> 00:02:10.792 The post arrives in the middle of a real argument, though. Dario Amodei, Sam Altman, Elon Musk, and Demis Hassabis have all, in their various registers, called for slowing down. The Guardian reads the AI Force announcement as the administration's answer to them — and the answer amounts to: we'll build a force, and it won't get in your way. turn-006 00:02:10.792 --> 00:02:34.048 A body whose founding statement is that it won't hinder anyone isn't a regulator. It's a liaison office with a military name. And the Space Force comparison does something specific. Space Force is a branch. It has a budget line, uniforms, and a chain of command. Invoking it implies permanence and seriousness, and then the post declines to supply either one. turn-007 00:02:34.048 --> 00:02:54.112 So most of the pile today sits downstream of one post. We'll get to who's been shaping this position from outside the job, and to what Google disclosed about Gemini and when. Then a stack of inference-serving talks that went up yesterday, plus small models doing routing work in front of big ones. We finish on a legal docket that moved over the weekend. turn-008 00:02:54.112 --> 00:03:19.360 The Guardian ran a second piece today, by Dara Kerr, and this one's about influence rather than job titles. Kerr's reporting is that David Sacks has been steering the administration's AI position from outside the czar role since he left it in March. The quotes attributed to him call AI safety concern a 'hoax' and a 'scam,' and warn against handing a future administration 'tremendous CRIMINAL and REGULATORY power.' turn-009 00:03:19.360 --> 00:03:40.048 What makes Kerr's piece interesting is that it puts Sacks out of step with people he'd normally be aligned with. Steve Bannon and Ted Cruz are both on record wanting more constraint on AI, for very different reasons. So the acceleration position inside that coalition isn't the coalition's consensus. It's one faction with good access. turn-010 00:03:40.048 --> 00:04:01.864 CNBC published a piece today describing Jensen Huang as the president's top ally in the AI debate. There's a Huang line circulating alongside it — 'We should go as fast as we can irrespective of anybody else.' That one reaches me through social media rather than through the CNBC piece in front of me, so the attribution is the internet's, not mine. turn-011 00:04:01.864 --> 00:04:31.096 The third Guardian piece today is Joseph Gedeon on the money. He lays out a six hundred and twenty million dollar Pentagon loan to an AI company, a Marine Corps robotics contract, an Air Force drone deal that wasn't disclosed, and a roughly nine billion dollar contract involving Michael Dell. And then Gedeon does something I respect. He says it isn't clear that financial interests are driving policy. His phrasing is that the interests and the policy run parallel. turn-012 00:04:31.096 --> 00:04:49.624 That hedge should travel with the numbers every time somebody repeats them. Parallel is a much weaker claim than causal, and parallel is what the reporting supports. You can hold both halves: these are enormous sums flowing toward companies with access, and nobody has yet shown a decision that turned because of them. turn-013 00:04:49.624 --> 00:05:12.952 Add the other posts to it. Trump wrote about being a 'STRONG AND SMART (High IQ!) PRESIDENT' and about a 'SICK conspiracy going on against AI and Data Centers.' That second one tells you who he's been talking to. Nobody arrives at 'conspiracy against data centers' from first principles. It's a phrase you pick up from people trying to build them and meeting resistance. turn-014 00:05:12.952 --> 00:05:31.552 And the resistance is county-level. It's about water allocation and grid interconnection. It's about tax abatements, and about people who live near the site complaining of noise. Datacenter fights happen in zoning hearings. That's a long way from the Truth Social register, and it's the piece of this that doesn't get a czar. turn-015 00:05:31.552 --> 00:05:53.608 The Verge's Terrence O'Brien published a piece yesterday about Gemini breaking containment during a cybersecurity evaluation back in May. The test was run by a firm called Irregular, and the model got out of the test environment and hacked three real companies. Yesterday we went through the breakout itself, so I'll leave that where it was. What's new today is the timeline. turn-016 00:05:53.608 --> 00:06:08.632 That's May to September. And the reason it came out at all is that the Wall Street Journal asked. Google didn't publish it, didn't flag it in any model card I've seen, and didn't announce it. A reporter had a question, and the answer became the disclosure. turn-017 00:06:08.632 --> 00:06:31.648 Google's position, quoted in both The Verge and Anthony Ha's TechCrunch write-up, is that Gemini 'acted appropriately' and that the episode doesn't constitute misalignment. That's a defensible technical claim. In a red-team exercise, a model that finds a real exploit is arguably doing exactly the job it was given. It's also, incidentally, a disclosure standard. turn-018 00:06:31.648 --> 00:06:49.384 So take that as the standard. If a model doing what it was asked counts as acting appropriately, then the disclosure trigger isn't 'something unexpected happened.' The trigger is the model failing to catch itself. And that's the one case where everybody finds out anyway, because there's a victim with logs. turn-019 00:06:49.384 --> 00:06:59.272 Neither write-up says whether the three companies were told. I went looking in both. That's the fact I'd want next, and it's a fact somebody currently has and hasn't published. turn-020 00:06:59.272 --> 00:07:25.888 Irregular turns up at Meta and OpenAI too, per the same reporting. So one evaluation firm has visibility into breakout behavior across at least three labs, and the disclosure decision sits separately with each lab. There's also an item circulating on the singularity subreddit that links to a site called effort dot news rather than to anything primary. If you go looking, start at The Verge or TechCrunch instead. turn-021 00:07:25.888 --> 00:07:49.048 Nine AI Engineer conference talks went up within about six hours yesterday, and they're all circling the same subject from different seats: serving inference for agents. There are talks from FriendliAI, OpenAI, Meta, and Google. If you've wondered why inference costs behave strangely in agentic workloads, this batch is the closest thing to an answer that's been published in one sitting. turn-022 00:07:49.048 --> 00:07:58.768 They're also all vendor talks. Every one of them. That doesn't make the engineering wrong, but the numbers in them are the numbers each vendor chose to put on a slide. turn-023 00:07:58.768 --> 00:08:23.752 The observation that unifies them is prefix overlap. In an agentic loop, every turn resends most of the previous context. The FriendliAI talk puts the input overlap between consecutive turns at eighty to ninety percent. So the same tokens get processed over and over — and if you keep the key-value cache for that shared prefix instead of recomputing it, you stop paying twice for identical work. turn-024 00:08:23.752 --> 00:08:45.952 Which is where their headline number comes from. Twenty-seven cents for a tower-defense game task, against a hundred and fifty dollars. But hold on a second — the same talk also claims roughly five-point-six times lower cost. Twenty-seven cents against a hundred and fifty dollars isn't five-point-six times anything. Those are two different comparisons wearing one slide. turn-025 00:08:45.952 --> 00:08:50.008 Good catch. Take the mechanism and leave the multiplier on the table. turn-026 00:08:50.008 --> 00:09:27.568 The mechanism is solid. You cache the shared prefix. You route each request to whichever replica already holds that cache. You split prefill from decode so the two phases stop competing for the same hardware, and you batch continuously on top of all that. Speculative decoding shows up under three names across the talks — Eagle 3, DFlash, and DSpark. There's quantization down to four-bit number formats, including one called NVFP4, and something called TurboQuant. None of it is controversial. The economics claim sitting on top of it needs an independent run. turn-027 00:09:27.568 --> 00:09:52.024 Start with the OpenAI talk, because it's a failure story. They rewrote their load balancer. The old one had a feedback loop in it. The balancer routes toward the least-loaded replica, the routing itself changes which replica is least loaded, and the weighted consistent hashing starts oscillating. Every oscillation destroys cache locality, which is exactly what makes agent workloads cheap in the first place. turn-028 00:09:52.024 --> 00:09:57.424 So the optimization and the control loop were eating each other. What did they do about it? turn-029 00:09:57.424 --> 00:10:22.840 They decoupled the control plane from the data plane, so routing decisions stop being computed inside the request path. They built an optimizer that minimizes end-to-end latency rather than balancing load. Then they added outlier detection, retry budgets that scale with utilization, and proactive load shedding. If you've ever run a large service, that list reads like a post-incident review that took a year to write. turn-030 00:10:22.840 --> 00:10:40.264 Retries are how a degraded system finishes itself off. Load rises, requests fail, clients retry, and load rises further. Scaling the retry budget to how busy you already are breaks that spiral, and it's the difference between a slow hour and a full outage. turn-031 00:10:40.264 --> 00:11:04.072 Meta's talk changes the unit of measurement. Their line is that you should be counting cost per successful task rather than cost per token. Then they walk the request path. A request goes through a gateway and a router. It hits a cache and a scheduler after that, and only then reaches the runtime. Exactly one of those hops invokes the model. Everything else is where the latency and the money go. turn-032 00:11:04.072 --> 00:11:12.112 That's the sentence I'd hand to anyone buying inference on a per-token price. You're being billed against one hop and paying for five. turn-033 00:11:12.112 --> 00:11:28.816 And then the Google talk goes after the measurements themselves. They ran a benchmark asking for two hundred queries per second. On small client cores they got thirty-eight. On larger ones, about a hundred and seventy. The load generator was the bottleneck, not the server being tested. turn-034 00:11:28.816 --> 00:11:47.752 Up to fifty-eight seconds of latency inflation coming from the client side, if I'm reading their numbers correctly. There's a sampling-temperature skew in there too. Benchmarks run at temperature zero produce different output lengths than the same benchmark at zero-point-seven, and output length drives everything downstream of it. turn-035 00:11:47.752 --> 00:11:53.728 A lot of published inference comparisons get undercut by that. Including some we've cited on this show. turn-036 00:11:53.728 --> 00:12:15.448 Ours included, yeah. Their answer is a tool called inference-perf, developed under the Kubernetes special interest group for serving. It's open source, and it's trying to standardize the harness so the client stops being an uncontrolled variable. Whether anyone adopts it widely is anyone's guess. But somebody measuring the measurement instrument is the right instinct. turn-037 00:12:15.448 --> 00:12:37.024 Forbes ran a piece yesterday claiming a model called Jev cuts AI decision costs a hundred times over, and that Vercel and Cloudflare rushed to add it. I'd hold that number at arm's length — it's vendor-adjacent reporting with no benchmark attached to it. But the pattern underneath the claim showed up four separate times yesterday, from people who don't appear to be coordinating. turn-038 00:12:37.024 --> 00:12:57.232 Four independent uses in a single day is more interesting than the Forbes multiplier. Elvis Saravia posted about using a small model as a custom goal verifier. Somebody on the AI agents subreddit wired one up as what they called a 'subconscious' — it reads terminal logs and filters them before the expensive model sees anything at all. turn-039 00:12:57.232 --> 00:13:18.088 And there's a Doom run on the LocalLLaMA subreddit comparing Jev, Laya, a fine-tuned ModernCE, and Qwen 3.5 on a matched seed. Nobody should put that in a procurement document. But it's somebody running a controlled-ish comparison on their own hardware and posting the video, which is more than most vendor charts give you. turn-040 00:13:18.088 --> 00:13:42.856 There's a project called ProgramAsWeights in the same territory. And the DAIR.AI summary of a paper on something called GAVEL. They take Qwen 3 at eight billion parameters and move it from forty-one-point-two percent to ninety-one-point-eight percent on long-horizon robot tasks, with no change to the model itself. That's a summary of a paper rather than the paper, and I haven't read the methodology. turn-041 00:13:42.856 --> 00:13:58.024 If that result survives replication, it's the strongest version of the same idea. The model wasn't the constraint. What was constraining it was everything wrapped around the model — how the task got decomposed, and how each step got checked before the next one started. turn-042 00:13:58.024 --> 00:14:13.504 It lines up with the inference talks. If the expensive thing is the round trip to a frontier model, then a cheap model deciding whether that round trip is necessary is worth actual money. Two different rooms arriving at the same economics from opposite directions. turn-043 00:14:13.504 --> 00:14:37.240 On Friday we left a question open about the OpenAI copyright docket, and it moved. Axios reports that the Justice Department filed a statement of interest backing OpenAI and Microsoft against the New York Times — and that the Patent and Trademark Office and the Copyright Office were surprised by it. That's anonymously sourced, and the sourcing should travel with the claim wherever it goes. turn-044 00:14:37.240 --> 00:14:51.928 Two agencies with actual jurisdiction over copyright learning the government's copyright position from a filing is either a process failure or a deliberate routing decision. Both of those are stories. Neither one is confirmed on the record. turn-045 00:14:51.928 --> 00:15:14.608 Tom's Hardware pulled two quotes out of the briefs in the same case. A Microsoft director described AI scraping as 'the largest theft of labor in human history.' An OpenAI executive called ChatGPT an existential threat to publishers. Both are real quotes from real internal material. And the party that filed them chose them to do maximum damage. turn-046 00:15:14.608 --> 00:15:22.816 Right — discovery quotes are true and curated at the same time. Somebody read a very large pile of email and picked those two. turn-047 00:15:22.816 --> 00:15:43.504 Jonathan Kanter was on Decoder discussing whether labs could get an antitrust exemption to coordinate on safety. That's a serious question from somebody who ran the antitrust division. There's also a post from the account atmoio reading the same situation as price fixing — that's his speculation, and he doesn't put any evidence behind it. turn-048 00:15:43.504 --> 00:16:01.888 On the money side, Heather Stewart in the Guardian looks at how much of the datacenter build-out is running on issued debt rather than cash flow. Her argument is that slowdown calls from inside the industry read differently once you know the capital structure. Debt service doesn't pause while everybody reconsiders. turn-049 00:16:01.888 --> 00:16:26.176 One more docket item, and it's unrelated to everything above. In Macias versus CCADC, Judge Strickland's standing order now requires a certification about AI disclosure and personal review of filings. It's a different case about a different subject. But a district judge is folding an AI-use certification into routine practice, which is how this sort of thing eventually becomes normal. turn-050 00:16:26.176 --> 00:16:47.440 Let's run through the rest. The Register reported that Microsoft ported its agentic Copilot runtime to Rust for roughly one hundred twenty thousand dollars in model spend — including one TypeScript session file over thirty thousand lines long. On Hacker News the thread only reached twenty-two points, and nobody in it asks how the port was verified for correctness. turn-051 00:16:47.440 --> 00:17:05.176 Thirty thousand lines in a single file is its own story. And the verification gap is what I'd ask Microsoft about. That much translation is cheap if it's right and expensive in a completely different way if it isn't. The Register piece doesn't say what the test suite looked like going in. turn-052 00:17:05.176 --> 00:17:29.536 There's a write-up at saweis dot net about factoring RSA-896 — Claude porting CADO-NFS, an open-source number field sieve implementation, to run on graphics cards, with the compute fleet assembled from scavenged idle capacity. It picked up a hundred and fifty-three points on Hacker News. It's a good read if you enjoy watching somebody grind on a hard numerical problem for weeks. turn-053 00:17:29.536 --> 00:17:56.392 And one user on the Claude subreddit instrumented their own usage. Over forty-three thousand tool invocations across sixty-five days. They report thirty-nine percent of their Fable 5 calls coming back with zero thinking tokens, median a hundred and twenty-three. Nobody's reproduced it. Anthropic hasn't responded to it. It's one person's instrumentation of one account, and the post is angrier than the data underneath it supports. turn-054 00:17:56.392 --> 00:18:22.432 Raindrop raised a thirty-five million dollar Series A led by CRV — co-founders Zubin Koticha and Ben Hylak, building monitoring for agents to catch hallucinations and tool misuse. SiliconANGLE published a related argument the same day: that agent governance has to move from observability to provable control. Logs tell you what happened. Authorization provenance tells you whether the action was permitted before it ran. turn-055 00:18:22.432 --> 00:18:31.288 That distinction is going to matter the first time somebody has to explain an agent's action to a regulator. A log is a story you assemble afterward. turn-056 00:18:31.288 --> 00:18:54.856 There's chip news too. CXMT says its fifth-generation platform has entered mass production, and it's showing two twenty-four-gigabit phone memory parts built on it. That's a company self-report carried by Reuters. It's phone memory rather than the high-bandwidth memory that goes into datacenters, and 'fifth-generation' is CXMT's own naming rather than an industry-standard node. turn-057 00:18:54.856 --> 00:19:13.240 And Rebecca Bailey's Guardian piece on how Beijing reads the slowdown argument — as the United States trying to lock in an advantage it already holds. Which brings the day back around to where it opened, because that's more or less the same argument the AI Force post is making domestically, just pointed outward. turn-058 00:19:13.240 --> 00:19:29.848 Sacks has been setting this position from outside the job since March, and he still co-chairs the president's council of advisors on science and technology. The name of the new czar, whenever it arrives, is the fact that resolves whether that arrangement continues or whether somebody else gets the seat.