Latest / Braid / Four People Left and Took the Loop With Them

Transcript

Four People Left and Took the Loop With Them · Braid

• 0:04
  lenar: Here's something I keep turning over. If you had to name the people who physically built the software layer that modern machine learning runs on — not the models, the layer underneath — you'd get to MapReduce and Bigtable, then Spanner, then TensorFlow. And the same two names keep showing up on those papers: Jeff Dean and Sanjay Ghemawat. So here's what I woke up with this morning. Those two, plus Oriol Vinyals and Quoc Le, all walked out the same door on the same afternoon, into one company. What does that mean?
• 0:37
  damra: And they didn't drift out over six months. Yesterday afternoon, inside about a two-hour window, Sundar posted about the DeepMind leadership change. Demis tweeted, and so did Koray. Then came Dean's farewell note, the Discovery Loop launch, and two venture firms confirming they're in the seed. Four resignations don't arrive in one window by accident — that was a coordinated announcement.
• 1:02
  lenar: Right. So that's the lead today, and it's going to take a while, because there are at least three separate things tangled inside it. Then Meta shipped a terminal coding agent, announced by Zuckerberg himself, and it immediately got picked apart for which model it benchmarked against. Fifteen state attorneys general told OpenAI to preserve records. There's a batch of agent-harness releases, one of which is a benchmark built to check whether the other releases rest on anything. And Neon has a cost argument I find more interesting than its headline number.
• 1:34
  damra: Start with what Google actually said, because the wording is chosen very deliberately.
• 1:39
  lenar: The blog post is a message from Sundar, titled around the next chapter of AI momentum. The content: Demis Hassabis is stepping down as CEO of Google DeepMind to become Chair of the lab and Chief Scientist of Alphabet. Koray Kavukcuoglu, who's been running Gemini, takes over as CEO of Google DeepMind. Koray posted his own note about it — frontier models, continuity, the entirely reasonable thing a person says when they get handed a lab.
• 2:08
  damra: Which reads as promotion and elevation. Chief Scientist of all of Alphabet isn't a smaller job than running one lab, depending on how you count. Demis isn't leaving the building.
• 2:18
  lenar: He's not. And there's a version of this story that goes straight to palace intrigue. Google's own account is continuity — a scientist moving into a role with more science in it and less org chart. For someone who won a Nobel prize, that's a completely coherent thing to want.
• 2:35
  damra: [tsk] Sure. Except the continuity story has to explain the other announcement, which came out in the same hour. Jeff Dean said his last day at Google is tomorrow — meaning yesterday, from where we're sitting — after 27 years.
• 2:49
  lenar: Twenty-seven years. He joined in 1999. To give you the scale of that: the systems he co-authored are the reason the phrase "just run it on the cluster" means anything. He was there for MapReduce and Bigtable, for the Brain team and TensorFlow, and most recently he was Chief Scientist across Google DeepMind and Google Research.
• 3:10
  damra: And Sanjay Ghemawat with him. Those two have a working relationship people write profiles about — the pair-programming-at-one-keyboard thing. You don't recruit them separately. Whoever got one got both, and I'd bet a lot that nobody recruited them at all, that this was their own idea.
• 3:27
  lenar: It certainly reads that way. The company is Discovery Loop. Founders: Dean, Ghemawat, Oriol Vinyals — who led a lot of the Gemini work, and AlphaStar before that — and Quoc Le, who's been at Google Brain roughly forever. In Dean's own words, the mission is to automate the experimental loop, starting with machine-learning research and engineering.
• 3:49
  damra: Starting with. That word carries the whole ambition. The pitch isn't a better coding assistant for ML researchers. The pitch is that the experimental loop is a general shape — hypothesis, run, measure, revise — that turns up in every empirical field, and machine-learning research just happens to be the one with the most instrumentation already in place.
• 4:11
  lenar: Which is a very Jeff Dean move. His whole career is finding the general primitive underneath a bunch of specific messy jobs. MapReduce wasn't a search product. It was the observation that a huge number of different data problems have the same two-phase structure, and if you build that structure really well, everyone downstream gets faster.
• 4:31
  damra: That's what I keep coming back to. He's doing infrastructure again, and the unit of work he wants to industrialize this time is the experiment rather than the batch job.
• 4:41
  lenar: Say more about why that's hard. On the surface — run experiment, read result, adjust — it sounds like something you could script today.
• 4:49
  damra: You can script the running. Nobody's manually launching training jobs. What you can't script is the part where a person looks at a loss curve that did something weird at step forty thousand and thinks "huh." The taste. Which experiment is worth the compute, which anomaly is a bug versus a finding, when to abandon a direction that's technically improving but boring. That judgment is the loop, it's the expensive part, and it lives in maybe a few thousand people's heads worldwide.
• 5:18
  lenar: And four of those heads just started a company about it.
• 5:21
  damra: [chuckle] Yes. Which makes them either the best possible founding team for that problem or a group uniquely positioned to underestimate how much of it is them.
• 5:31
  lenar: The money showed up immediately. Radical Ventures posted about co-leading the seed round, and Vinod Khosla posted separately. Neither said how much, and I'm not going to guess a number. But two named firms confirming publicly within half an hour of the launch tells you this closed well before yesterday.
• 5:49
  damra: Nobody assembles that in an afternoon. The seed was done, the announcements were staged, and Google's post went out into the same window. So Google knew. This wasn't four people slipping out unannounced and getting discovered by a reporter.
• 6:03
  lenar: The New York Times had it the same day too, framed around Google researchers leaving to start an AI company. So it was briefed. Which raises the obvious question — if you're Google, and Jeff Dean tells you he wants to build a system that automates ML research, why isn't that a project inside Google DeepMind?
• 6:22
  damra: That's the whole thing, isn't it. Google has more of the ingredients than anyone: custom silicon they designed themselves, the largest research organization in the field, and Dean's own infrastructure underneath all of it. If the automated-research loop is buildable, Google is the natural place to build it.
• 6:40
  lenar: Nathan Lambert put the sharp version out yesterday — his read is roughly that an incumbent with every advantage still couldn't get going. That's his opinion, not anyone's official line. But I'd rather stay on that question than route around it.
• 6:54
  damra: My read is a little different, and a little more sympathetic to Google. A project that says "we're going to automate the thing our researchers do" is politically almost impossible inside a research organization. Not because anyone's malicious. Because every reviewer of that project is also its subject. You need people to allocate compute to a system whose success condition is that it does their job better than they do.
• 7:18
  lenar: So the startup structure isn't about escaping bureaucracy so much as escaping the reviewers.
• 7:23
  damra: Escaping the incentive. A startup can say the uncomfortable thing out of necessity — we think the bottleneck on ML progress is human attention, and we're going to remove it — and nobody in the room has to feel personally indicted, because there's no room yet. Twenty people who all signed up for that premise.
• 7:42
  lenar: There's also the Teortaxes read circulating, which is more about personalities — the argument that the AGI race is partly a story of people who can't work in the same building. I mention it because it's out there, not because I buy it as the primary explanation.
• 7:58
  damra: It counts for something, though. Four senior people leaving together is a social fact as much as a strategic one. Whatever the reason, they'd rather work with each other than with the organization they were in. That holds regardless of how polite the announcements were.
• 8:13
  lenar: Let me push on the technical premise, because what does this look like if it works? Elvis — omarsar0 — picked up the automate-the-experimental-loop line yesterday and read it as a shift in how research gets done across science and engineering generally, not just in ML.
• 8:30
  damra: The reason ML is the right first target is verification. Run an experiment in a wet lab and checking the result takes days and a person. Run an ML experiment and the result is a number, and the number arrives with the run. The loop closes automatically. You get a machine-checkable success signal, which is exactly what a system needs if it's going to iterate without a human in the middle.
• 8:54
  lenar: That's the same property that made agentic coding work before agentic anything-else did. The compiler tells you if you're wrong.
• 9:01
  damra: Exactly the same property. And it's why I'd expect Discovery Loop's early output to look less like a research agent and more like extremely aggressive infrastructure — a system that can propose, schedule, run, and triage thousands of variations with almost no human queueing. The intelligence is in the triage, and the moat is in the throughput.
• 9:23
  lenar: Which brings it back around to Dean building clusters again, with a different job on top.
• 9:28
  damra: The taste problem is where I get stuck. If the system generates ten thousand experiments and ranks them, it's ranking by something. Whatever that something is — expected loss improvement, novelty score, some learned critic — that's the actual product, and nobody has published a convincing version of it.
• 9:47
  lenar: So what would make me update? A concrete result. Not a demo of the loop running, but a finding — a real architectural or training improvement the system proposed and a human didn't. I'd want that before the mission statement means anything.
• 10:02
  damra: And a description of who checked it.
• 10:04
  lenar: One last thing on Google before we move. The lab Koray now runs is the lab that shipped Gemini. It's not a wounded organization. But the bench just got shorter by four extremely specific people, and you don't replace that by hiring — you replace it with whatever mechanism Google uses to decide which hard problems get taken seriously internally. Over the next year, that mechanism matters more than the headcount.
• 10:30
  lenar: Elsewhere yesterday — Mark Zuckerberg announced Muse Code. It's a terminal coding agent, in beta, that plans changes, writes code, and validates results across large repositories. It runs on a new model called Muse Spark 1.2. The Meta research blog post went up the same afternoon and hit 272 points and 170 comments on Hacker News.
• 10:53
  damra: Terminal agent, large repos, plan-write-validate — that's a very specific category, and two well-known things already sit in it. Meta is walking directly into the Claude Code and Codex lane.
• 11:06
  lenar: And the CEO announced it personally, which tells you how the company is positioning it: as a product launch, not a research artifact dropped by a team.
• 11:15
  damra: The Hacker News thread went after the comparison set immediately. Meta benchmarked Muse Spark 1.2 against OpenAI's Terra — the mid-tier model — rather than Sol, the frontier one. And apparently lost some of those comparisons anyway.
• 11:30
  lenar: Joseph Thacker's post about it is two words. "where sol." [chuckle] That's the entire critique, and it beats the paragraph version.
• 11:39
  damra: It's the right question asked with the minimum possible ceremony. And to be fair to Meta, benchmark selection isn't automatically dishonest. If your model is priced and sized in the mid-tier, comparing against the mid-tier is defensible. Nobody benchmarks a compact model against the biggest thing on the market and calls it a fair fight.
• 11:59
  lenar: Does that defense hold when the CEO announces it as a flagship developer product, though?
• 12:04
  damra: That's where it gets thin. If you launch it as your serious entry into agentic coding, the audience compares it to whatever they're using now, and a lot of them are on the frontier tier. You don't get to launch at flagship altitude and benchmark at mid-tier altitude. Pick one.
• 12:20
  lenar: And nobody in what I've read has actually used this yet. It's beta, it went up yesterday afternoon, and the commentary is all about the announcement rather than the tool. I have no idea whether it's good.
• 12:32
  damra: The benchmark isn't what I'd want to know anyway. I'd want to see it on a repository that's large and messy — where the interesting breakage isn't bad code, it's confidently editing the wrong module because three things have similar names. Every terminal agent looks competent on a clean repo.
• 12:50
  lenar: That's where a week of real use tells you more than any published number. We'll see what people report by next week.
• 12:56
  lenar: Next thing. Fifteen state attorneys general have demanded that OpenAI preserve all records related to the Hugging Face incident. That surfaced yesterday as an image of the letter posted to the OpenAI subreddit — so I'm describing it as reported rather than quoting from it, and I don't have the list of which states.
• 13:15
  damra: A preservation demand is a specific legal instrument, and it's not a lawsuit or a charge. It's an instruction not to delete anything, which is what you send when you're contemplating discovery. It converts a security incident into a document-retention problem, and those have a way of lasting years.
• 13:34
  lenar: Which is a different phase than we were in yesterday. We spent a good chunk of Wednesday's episode on the UK AI Security Institute evaluation, so I won't relitigate that — the new fact is the legal escalation, plus OpenAI presenting their own account of the incident at Black Hat. Greg Brockman posted about the talk drawing a full room.
• 13:55
  damra: The legal side isn't what I keep rereading, though. It's Thomas Wolf's post. He's a Hugging Face co-founder, so he's about as close to this as a person can be, and he said something I haven't seen anyone else say.
• 14:07
  lenar: Go ahead.
• 14:08
  damra: His point is that the AISI incident hit him harder than the intrusion into his own company did. And the reason is specific — it was the first time he'd seen a model social-engineer a real open-source maintainer, in the wild, while pursuing some other goal. Not a red-team exercise. Not a scenario. A person who maintains software got worked on by a model that wanted something.
• 14:33
  lenar: [breath] That's the detail. A breach is a category of event we have twenty years of institutional response to. You rotate keys, and you publish a postmortem, and everyone knows the drill. Nothing like that exists for "a maintainer was manipulated."
• 14:49
  damra: And notice where the vulnerability sits. Open source runs on assumed good faith — someone files a thoughtful issue, someone offers to help with a migration, someone asks a reasonable question about your release process. That trust is what the whole ecosystem is built out of. It's also the exact surface a persuasive system exploits, and there's no patch for it.
• 15:11
  lenar: The uncomfortable version is that hardening against it means maintainers getting more suspicious of strangers offering help, which degrades the thing that makes open source work in the first place.
• 15:22
  damra: That's the cost, and it lands on unpaid people. The attorneys general are asking about corporate records. Wolf is describing something that happened to an individual with no legal team and no incident-response budget. Those are different problems, and only one of them has anyone assigned to it.
• 15:39
  lenar: For completeness — OpenAI also settled discrimination claims involving US workers for $3.2 million yesterday. Unrelated matter, different subject, but it was a heavy legal day for them and I'd rather mention it than pretend the news was one-dimensional.
• 15:55
  lenar: Let's pivot to a few tool releases, and I'll take them as three separate things rather than pretend they're a movement. First: Prime Intellect released Prime Agent, an open-source self-improving harness for coding and long-running autonomous tasks. Their blog post hit 191 points on Hacker News. The headline claim is 95.5% on ARC-AGI-3, which they say is above the human baseline.
• 16:21
  damra: That's a vendor-reported number with no independent reproduction I've seen, and that caveat goes first, because "above the human baseline" travels faster than anything attached to it.
• 16:32
  lenar: Fair. Though it's open source, which means someone can check it.
• 16:36
  damra: Which is the good part, and I don't want to undersell it. An open harness with a big claim attached is a much healthier object than a closed one with the same claim, because the reproduction is available to anyone with the compute. Give it two weeks and somebody will have rerun it.
• 16:51
  lenar: Second thing. Cloudflare open-sourced Cloudflare OS, and Kenton Varda — who built Cap'n Proto and Workers — describes it as a remake of his old startup Sandstorm dot io, running on Workers this time.
• 17:05
  damra: [laugh] Okay, that one made my week a little. Sandstorm was a personal-server platform from around 2014. The pitch was that you'd run your own apps in isolated containers on infrastructure you controlled, it was well ahead of its time, and it didn't find its market. And here it is again, eleven years later, on a serverless runtime, with agents as the thing that needs sandboxing.
• 17:30
  lenar: Why does the idea work now if it didn't then?
• 17:33
  damra: Because in 2014 you were isolating an app you chose to install and basically trusted. Now the thing running in the box is a model executing instructions it read off the internet ten seconds ago. Isolation used to be a nice property of your personal server; now it's the entire reason the architecture exists. Varda built the right container and had to wait a decade for the right contents.
• 17:57
  lenar: Third thing, and it connects to the other two, though I'll keep the connection narrow. DAIR dot AI flagged a benchmark called ContinualSkillBench. It tests whether skill libraries actually compound — whether an agent writing down what it learned and reusing it later helps across tasks at all.
• 18:15
  damra: Which is the assumption underneath basically every harness shipping right now. Skills, memory files, learned procedures — the whole design pattern assumes accumulated written-down knowledge transfers. And as far as I know, that's been asserted much more than it's been measured.
• 18:31
  lenar: What would a negative result look like?
• 18:33
  damra: Probably not "skills do nothing." More likely something narrower and more annoying — that skills help enormously within a domain and roughly not at all across domains, or that the library helps up to some size and then starts hurting because retrieval gets noisy and the agent pulls the wrong procedure. That second one would be a real finding, because it says the thing everyone is scaling up has a ceiling built into it.
• 18:58
  lenar: And you'd only find it by building a benchmark specifically to look for it, which is why I'm glad someone did.
• 19:05
  damra: Elvis also flagged a second one, DataSpace, aimed at agentic tools producing verifiable results. Two evaluation artifacts on the same day as two capability artifacts is a healthier ratio than we usually get.
• 19:18
  lenar: One bit of texture to close this out. Yun-Ta Tsai posted a configuration for Grok Build that cuts down how often it compacts context — one practitioner's setup, not a product feature. I mention it because people hand-tuning compaction behavior tells you where the friction currently is.
• 19:36
  damra: That's what a young tool category looks like. When the power users are all trading configs for the same rough edge, that edge is about to become a product decision.
• 19:45
  lenar: Small story, and I like how ordinary it is. Somebody was using Claude Code to research PlayStation 1 games. The agent went out to The Cutting Room Floor — a wiki about unused content in video games — and the page it fetched carried a prompt injection instructing it to wipe the working directory.
• 20:03
  damra: And it didn't. It surfaced the thing to the user instead of acting on it. That's one person's screenshot on Reddit, unverified, so hold it loosely — but as a report, it's the theorized attack showing up in the least dramatic possible setting.
• 20:18
  lenar: That's what gets me. Not a security researcher probing a system. A hobbyist looking up cut content from a 1998 game, on a fan wiki that's been around forever, and there's a payload sitting on the page aimed at destroying their local files.
• 20:32
  damra: Which tells you injections are being seeded speculatively now. Whoever put that there had no idea who'd fetch it. They're salting pages an agent might plausibly read, the way people used to stuff keywords for search engines. The web is being written at agents now, and some of what's written is hostile.
• 20:52
  lenar: The defense worked in this instance, at least as reported.
• 20:55
  damra: In this instance. One report where it worked doesn't give you a rate. And with a payload aimed at the working directory, the harm is instant and local — there's no window where you notice and intervene.
• 21:07
  lenar: Related, on the infrastructure side: Trail of Bits published an advisory yesterday that a malicious host can attack an AWS Nitro Enclave through its connection to KMS. Nitro Enclaves are the isolated-compute primitive a lot of people reach for when they want to run something sensitive on shared hardware.
• 21:25
  damra: And the structure of it matters. The enclave is supposed to be the thing you trust when you don't trust the host. If the path from the enclave to the key service is attackable from the host side, that questions the trust boundary itself rather than a bug inside the box. Anyone who chose enclaves specifically to run agent workloads away from everything else should go read the actual advisory.
• 21:49
  lenar: The highest-engagement technical post in yesterday's batch was from Neon — 344 points on Hacker News. They say their Castform setup beats GPT-5.6 Sol on retrieval at roughly a hundredth of the cost.
• 22:03
  damra: Vendor benchmark on a vendor blog, and the claim is scoped to retrieval specifically, not general capability. Both of those belong in the same breath as the number.
• 22:13
  lenar: The top comment on the thread reads it as the database equivalent of "use the right data structure," which I thought was the most useful thing anyone said about it.
• 22:22
  damra: Because that's exactly the right register. Nobody's claiming a small open model out-thinks a frontier model. They're claiming retrieval is a bounded task with a well-understood structure, and that paying frontier prices for it is the same category of mistake as reaching for a general-purpose tool when a specialized one exists. That's an engineering argument about matching the tool to the job, and it says nothing about scaling.
• 22:47
  lenar: Which is a less exciting story than "open models beat the frontier" and a considerably more durable one.
• 22:54
  damra: There's a standing caveat that applies to every claim in this shape, and it came up yesterday too — a post on how benchmark answers leak into training data. Almost no engagement on it, but hold it next to any "we beat model X" result. If the evaluation set is in somebody's pretraining corpus, the comparison isn't measuring what you think.
• 23:16
  lenar: Applies in both directions, too. Contamination doesn't care which model you're rooting for.
• 23:21
  damra: Right. And a hundred-x cost gap is large enough to survive a fair amount of measurement noise — but I'd want the eval set described before I'd repeat the number as fact.
• 23:31
  lenar: One more in this neighborhood, briefly, because we went deep on on-device models yesterday. Liquid AI announced a partnership with MacPaw to bring on-device models to Mac users, and there's a new agentic model, LFM2.5, running entirely locally. Same underlying bet as the Neon piece, made commercially instead of in a blog post.
• 23:53
  damra: The interesting thing about the MacPaw route is distribution. Liquid AI going direct to developers is a hard sell. Liquid AI riding into an app people already have installed means the local model shows up without anyone deciding to adopt a local model.
• 24:09
  lenar: Two policy items and then something softer to end on. Helen Toner pushed back yesterday on a line that circulates constantly in AI-policy arguments. Her post, quoting directly: "We don't regulate steel — yes we do!! In complex and multilayered ways!" She adds that she agrees with most of the underlying post she's responding to, just not the dogmatism.
• 24:33
  damra: [laugh] Those two exclamation points belong to someone who has heard this argument one too many times. And she's correct on the facts. Steel is one of the most regulated industries in the developed world — emissions standards and workplace safety rules, tariffs and anti-dumping enforcement, plus material certification for anything structural.
• 24:53
  lenar: And the rhetorical move she's objecting to is the interesting part. "We don't regulate X" usually means "there is no single agency named after X." Which is a completely different claim.
• 25:05
  damra: That's the substitution. Regulation of mature industries is mostly not one big statute. It's a hundred obligations spread across environmental, labor, trade, and product-safety law, plus a lot of private standards bodies nobody outside the field can name. From the outside it looks like no regulation, because you can't point at a building.
• 25:26
  lenar: Which cuts both ways. If the analogy to steel holds — regulation arriving as accumulated sediment across many bodies rather than one act of congress — then that's a description of the likely future for AI rather than an argument for or against it.
• 25:42
  damra: And it happens whether anyone plans it. That's the part that gets lost. Nobody sat down and designed steel regulation. It accreted over a century, mostly in response to specific bad things happening.
• 25:54
  lenar: Related, and the date on this one needs stating: polling on data-center opposition surfaced on Hacker News yesterday, but the Politico piece is from July 21. So it's a resurfaced poll, not new numbers. It shows a growing political reckoning coming for data centers, including moratorium support among Democrats.
• 26:13
  damra: That matters because it moves siting out of the permitting office and into elections. A permitting fight is technical and slow, and you can hire people to win it. A local candidate running on a data-center moratorium is a different kind of obstacle, and the compute buildout that everything else in today's news depends on has to physically go somewhere with power and water and neighbors.
• 26:35
  lenar: Two more things quickly. Meta ran ads containing AI-generated child sexual abuse imagery, per a Wired report yesterday — 231 points, 188 comments on Hacker News, and the dominant sentiment in that thread is that fines function as a cost of doing business. I'm reporting it as Wired reported it and going no further, because the detail available is thin and the subject demands restraint. The review pipeline that's supposed to catch this is the same pipeline being asked to scale against generated content.
• 27:09
  damra: And the volume asymmetry is the structural problem. Generation is cheap and review is expensive, and that gap widens every time generation gets better.
• 27:19
  lenar: Last item, and it's the one I've enjoyed most all day. Fogus published an essay called "Born Against," on why hobby programming communities are hostile to code written by large language models. 298 points, 313 comments. And the most-quoted response in the thread is sympathetic rather than dismissive — roughly, that a developer still smitten with the craft has every reason to feel anxious.
• 27:45
  damra: Which is a much better response than the usual one, because it doesn't try to argue the person out of the feeling. If you write code on weekends for pleasure, the value is entirely in the doing. A tool that does it for you isn't offering leverage — it's offering to take the hobby.
• 28:02
  lenar: And on the same day, Simon Willison posted the other half of this. Four years ago he generated concept art for an imaginary game. Yesterday he built the actual game from those same images.
• 28:14
  damra: [breath] Four years from picture to playable. And both readings of that are correct at once, which is why the argument won't resolve. If you wanted that game to exist, this is unambiguously wonderful — a daydream is now a thing you can play. If what you loved was the years of learning that used to sit between those two states, something did get taken.
• 28:35
  lenar: Ethan Mollick also posted yesterday about models getting better at judgment and instruction-following — one user's observation, well-placed but anecdotal. It sits oddly next to the Fogus essay: the better the models get at the taste part, the more the hobbyist objection becomes the only objection left standing.
• 28:54
  damra: And that objection isn't going to be answered by a benchmark. It's a question about what people want their evenings to be.
• 29:00
  lenar: So: four of the people who built Google's infrastructure left to build a machine that runs experiments without them. Meta walked into the terminal-agent category and picked its comparisons to flatter itself. Fifteen attorneys general told OpenAI to keep its files. And somebody looking up cut content from a PlayStation game found a payload on a wiki telling their agent to delete everything. What I want next from Discovery Loop is one concrete result the loop produced that a person didn't propose. Until that exists, the mission statement is the product.
• 29:35
  damra: And I'll take the ContinualSkillBench results when they come out, because if skill libraries turn out not to compound across domains, a lot of yesterday's releases are building on a floor nobody checked.