Latest / The AI Practitioner Brief / Genie Four, vLLM's Tool-Call Fix, and Anthropic's IPO — October 28, 2026

Transcript

Genie Four, vLLM's Tool-Call Fix, and Anthropic's IPO — October 28, 2026 · The AI Practitioner Brief

• 0:00
  [music]
• 0:05
  SPEAKER_00: Welcome to The AI Practitioner Brief,
• 0:07
  SPEAKER_00: the podcast from AIpractitioner.substack.com,
• 0:11
  SPEAKER_00: where I turn the latest AI and tech news into practical insights
• 0:15
  SPEAKER_00: for AI practitioners.
• 0:16
  SPEAKER_00: I'm Lina Faik,
• 0:17
  SPEAKER_00: an ML engineer,
• 0:18
  SPEAKER_00: consultant, and lecturer in advanced machine learning at Sorbonne University.
• 0:23
  SPEAKER_00: It's Wednesday, October 28th.
• 0:25
  SPEAKER_01: Wednesday already.
• 0:26
  SPEAKER_01: Let's see what the week threw at us.
• 0:28
  SPEAKER_00: With me is Jimmy,
• 0:29
  SPEAKER_00: my co-host on The Brief, whose job is to ask the questions I'd otherwise rush past.
• 0:34
  SPEAKER_01: That's the job.
• 0:35
  SPEAKER_01: Let's get to it.
• 0:36
  SPEAKER_00: Today's news has a real tension running through it.
• 0:39
  SPEAKER_00: Google just showed off a simulated world you can reshape with a sentence while it's running.
• 0:45
  SPEAKER_00: Two days ago, Anthropic filed to go public.
• 0:49
  SPEAKER_00: In that filing,
• 0:50
  SPEAKER_00: it told its own investors in a legally binding document
• 0:54
  SPEAKER_00: that systems like these could pose an existential risk to humanity.
• 0:59
  SPEAKER_01: Those two stories should not exist in the same week.
• 1:01
  SPEAKER_00: Headlines first, then three bigger stories.
• 1:04
  SPEAKER_00: We'll cover Google's new world model,
• 1:07
  SPEAKER_00: a quiet fix to how AI agents call tools,
• 1:10
  SPEAKER_00: and what Anthropic just told its own investors about the risk
• 1:14
  SPEAKER_00: of this entire technology.
• 1:16
  SPEAKER_01: Investors warned about existential risk.
• 1:19
  SPEAKER_01: That's a strange thing to put in a stock prospectus.
• 1:22
  SPEAKER_00: Let's start with Frontier Code 2,
• 1:24
  SPEAKER_00: an updated benchmark for testing
• 1:26
  SPEAKER_00: whether AI models can make correct large-scale changes across a real codebase.
• 1:31
  SPEAKER_01: Updated how?
• 1:32
  SPEAKER_00: The new version expects bigger changes across bigger codebases
• 1:36
  SPEAKER_00: and checks correctness more strictly than the first version did.
• 1:39
  SPEAKER_01: Makes sense.
• 1:40
  SPEAKER_01: Models kept gaming the narrow version.
• 1:43
  SPEAKER_00: Anthropic had a bigger week than that benchmark alone.
• 1:46
  SPEAKER_00: It's partnering with six biotech companies,
• 1:49
  SPEAKER_00: including Bristol Myers Squibb,
• 1:51
  SPEAKER_00: to use Claude for drug discovery work like antibody design
• 1:55
  SPEAKER_00: and clinical trial planning.
• 1:56
  SPEAKER_01: Is any of that actually working yet?
• 1:59
  SPEAKER_00: One case, yes.
• 2:00
  SPEAKER_00: Anthropic quietly bought a company called Mithril Medicine.
• 2:04
  SPEAKER_00: It's run by a single founder who used Claude to coordinate experiments across 11 outside labs.
• 2:10
  SPEAKER_01: One person running a drug program with 11 labs reporting to them.
• 2:15
  SPEAKER_00: And in under 18 months,
• 2:16
  SPEAKER_00: that setup took a muscular dystrophy treatment from an idea to promising results
• 2:21
  SPEAKER_00: in mice.
• 2:22
  SPEAKER_01: That's the kind of result that gets repeated at every AI conference for a year.
• 2:26
  SPEAKER_00: That's the AI side.
• 2:28
  SPEAKER_00: Elsewhere in tech today,
• 2:29
  SPEAKER_00: Apple refreshed its iMac and its entry-level MacBook Pro with the new M6 chip.
• 2:35
  SPEAKER_00: It adds Wi-Fi 7 and Bluetooth 6 with pre-orders open now.
• 2:39
  SPEAKER_01: Nothing dramatic there.
• 2:41
  SPEAKER_00: Two notes on how AI is changing engineering work day to day.
• 2:45
  SPEAKER_00: Amazon told its engineers to stop chasing AI usage metrics.
• 2:49
  SPEAKER_00: The memo includes this line.
• 2:51
  SPEAKER_00: If your work doesn't benefit from current AI tooling, that's a finding, not a failing.
• 2:57
  SPEAKER_01: Good line.
• 2:58
  SPEAKER_00: An essay making the rounds this week argued
• 3:00
  SPEAKER_00: that junior engineer's impact is now less bounded by years of experience.
• 3:05
  SPEAKER_01: So tenure stops being the moat it was.
• 3:07
  SPEAKER_00: One more.
• 3:08
  SPEAKER_00: GitHub rebuilt how its infrastructure handles rights to a repository during its move
• 3:14
  SPEAKER_00: to the cloud.
• 3:15
  SPEAKER_00: It used to lock the whole repository for every write.
• 3:18
  SPEAKER_00: Now it only locks the specific reference, meaning the branch or tag being changed.
• 3:24
  SPEAKER_00: That change made writes three to four times faster.
• 3:28
  SPEAKER_01: What's a reference to anyone who isn't a Git nerd?
• 3:31
  SPEAKER_00: Just a pointer to a specific branch or tag inside the repository.
• 3:35
  SPEAKER_00: Separately, Google and Meta open source Torch TPU.
• 3:39
  SPEAKER_00: It lets PyTorch,
• 3:40
  SPEAKER_00: a widely used framework for building AI models,
• 3:43
  SPEAKER_00: run natively on Google's TPU chips instead of just NVIDIA's.
• 3:48
  SPEAKER_01: That one's going to matter to anyone stuck in the NVIDIA queue.
• 3:51
  SPEAKER_00: Let's go deeper on the first story.
• 3:54
  SPEAKER_00: Google DeepMind announced Genie 4 on October 27th, a type of model called a world model.
• 4:00
  SPEAKER_01: What's a world model exactly?
• 4:02
  SPEAKER_00: It's a model that generates an explorable scene instead of a flat video.
• 4:07
  SPEAKER_00: You can walk through it and it keeps responding as you go.
• 4:10
  SPEAKER_01: So more like a game engine than a movie.
• 4:12
  SPEAKER_00: That's the right comparison.
• 4:14
  SPEAKER_00: This version adds something new.
• 4:16
  SPEAKER_00: You can type a sentence while the scene is running and it edits the world instead
• 4:21
  SPEAKER_00: of restarting it.
• 4:22
  SPEAKER_01: Edits it how?
• 4:22
  SPEAKER_00: Change the weather, move the camera, add a character.
• 4:26
  SPEAKER_00: The model adjusts the existing scene around that instruction rather than generating
• 4:31
  SPEAKER_00: a brand new one.
• 4:32
  SPEAKER_01: Is this something I can go try right now?
• 4:34
  SPEAKER_00: No.
• 4:35
  SPEAKER_00: It's a research preview limited right now to a small group of trusted testers.
• 4:40
  SPEAKER_01: So the claims we're about to discuss are Google's own,
• 4:42
  SPEAKER_01: not something outsiders have poked at yet.
• 4:45
  SPEAKER_00: Correct.
• 4:45
  SPEAKER_00: Worth saying plainly.
• 4:47
  SPEAKER_00: This didn't come from nowhere though.
• 4:49
  SPEAKER_00: Genie 1, back in 2024, only generated flat two-dimensional scenes from video game footage.
• 4:56
  SPEAKER_01: That's a long way from walk through it.
• 4:58
  SPEAKER_00: Genie 2 added three-dimensional worlds a few months later,
• 5:02
  SPEAKER_00: but they lasted only 10 to 20 seconds.
• 5:04
  SPEAKER_00: Genie 3, in 2025, made them real-time and interactive, holding a scene for about a minute.
• 5:11
  SPEAKER_01: And now 4 doubles that.
• 5:12
  SPEAKER_00: Right.
• 5:13
  SPEAKER_00: DeepMind says Genie 4 holds a scene steady for up to two minutes,
• 5:17
  SPEAKER_00: and it's the first version to generate audio natively from the same system
• 5:21
  SPEAKER_00: that renders the picture.
• 5:22
  SPEAKER_01: What actually changed to get there?
• 5:25
  SPEAKER_00: The architecture.
• 5:26
  SPEAKER_00: Genie 4 doesn't work alone anymore.
• 5:29
  SPEAKER_00: A second model, called Gemini 3.8 Flash, watches the frames as they're produced.
• 5:34
  SPEAKER_01: But watching for what?
• 5:35
  SPEAKER_00: It breaks the user's goal into steps,
• 5:37
  SPEAKER_00: checks whether each step actually happened in the simulation,
• 5:41
  SPEAKER_00: and sends a correction the instant something drifts.
• 5:44
  SPEAKER_01: So one model paints the world, and a second model proofreads it in real time.
• 5:48
  SPEAKER_00: That's a fair way to put it.
• 5:50
  SPEAKER_00: DeepMind calls the result reliable causality,
• 5:53
  SPEAKER_00: meaning objects keep behaving the way physics says they should across
• 5:57
  SPEAKER_00: the whole session.
• 5:58
  SPEAKER_01: I'll believe that when a scene runs past the two-minute mark without falling apart.
• 6:02
  SPEAKER_00: Fair.
• 6:03
  SPEAKER_00: Nobody outside Google has run it long enough yet to find out where it actually breaks.
• 6:08
  SPEAKER_01: What's this actually for beyond looking impressive in a demo?
• 6:11
  SPEAKER_00: Training.
• 6:12
  SPEAKER_00: A world you can edit cheaply is a place to let a robot or an AI agent practice tasks
• 6:18
  SPEAKER_00: before it touches anything real.
• 6:19
  SPEAKER_01: That's a very different pitch than look at this pretty simulation.
• 6:23
  SPEAKER_00: It is, and it's why Google keeps building bigger versions of this instead
• 6:27
  SPEAKER_00: of quietly retiring the project.
• 6:30
  SPEAKER_00: That training ground idea connects to our next story,
• 6:33
  SPEAKER_00: because it's about whether AI agents can be trusted to act on what they decide.
• 6:38
  SPEAKER_01: Act how?
• 6:39
  SPEAKER_00: When a model doesn't just answer in text,
• 6:41
  SPEAKER_00: but produces a structured instruction,
• 6:44
  SPEAKER_00: like call this function with these parameters.
• 6:47
  SPEAKER_00: That's called a tool call.
• 6:49
  SPEAKER_01: And something there was broken?
• 6:50
  SPEAKER_00: Yes.
• 6:51
  SPEAKER_00: A blog post published October 20th found that VLLM,
• 6:55
  SPEAKER_00: the open source engine a lot of teams use to run their own models,
• 7:00
  SPEAKER_00: was getting the tool call format wrong.
• 7:02
  SPEAKER_01: Wrong how often?
• 7:03
  SPEAKER_00: Serving one open-weight model, it produced a correctly formatted call only 92% of the time.
• 7:09
  SPEAKER_01: So about 1 in 12 calls just failed.
• 7:12
  SPEAKER_00: Exactly.
• 7:13
  SPEAKER_00: When a tool call is malformed, the model can't recover gracefully.
• 7:17
  SPEAKER_00: It can spiral into a confused conversation state, retry, and run up your bill.
• 7:22
  SPEAKER_01: What's the actual fix?
• 7:24
  SPEAKER_00: Constrained decoding.
• 7:25
  SPEAKER_00: The engine restricts which words the model is even allowed to produce next,
• 7:30
  SPEAKER_00: so it physically can't generate an invalid answer.
• 7:33
  SPEAKER_01: How does that work mechanically?
• 7:34
  SPEAKER_00: While it's generating a number, it blocks every token that isn't a digit.
• 7:38
  SPEAKER_00: While it's filling in arguments, it only allows tokens that match the required format.
• 7:43
  SPEAKER_01: Like a form field that won't let you type letters into a date box.
• 7:47
  SPEAKER_00: That's a good way to picture it.
• 7:48
  SPEAKER_00: VLLM added this as a flag in a later version,
• 7:52
  SPEAKER_00: and turning it on pushed accuracy to 100% for some models.
• 7:57
  SPEAKER_01: Some models, not all of them.
• 7:58
  SPEAKER_00: Right, that's the catch.
• 8:00
  SPEAKER_00: A few other models actually got worse when the feature was switched on by default.
• 8:05
  SPEAKER_01: Worse how, out of curiosity.
• 8:07
  SPEAKER_00: It's not published in detail for every model,
• 8:09
  SPEAKER_00: but the pattern suggests the constraint and
• 8:12
  SPEAKER_00: that model's own training don't always agree on format.
• 8:15
  SPEAKER_01: So you can't just flip the switch and walk away.
• 8:17
  SPEAKER_00: No.
• 8:18
  SPEAKER_00: If you're serving your own models for anything that takes real actions,
• 8:22
  SPEAKER_00: this is worth checking directly against your own setup this week,
• 8:26
  SPEAKER_00: not assuming the default is right.
• 8:28
  SPEAKER_01: That's the first actual homework assignment of the episode.
• 8:31
  SPEAKER_00: It's a small check.
• 8:33
  SPEAKER_00: Run your tool call schema compliance rate before and after turning this on,
• 8:37
  SPEAKER_00: for your specific model.
• 8:39
  SPEAKER_01: What's the broader point here, beyond one engine's bug fix?
• 8:42
  SPEAKER_00: That the bottleneck for trustworthy agents isn't always the model's reasoning.
• 8:47
  SPEAKER_00: Sometimes it's the plumbing underneath,
• 8:49
  SPEAKER_00: deciding whether that reasoning even reaches the outside world correctly.
• 8:54
  SPEAKER_00: That question,
• 8:55
  SPEAKER_00: whether an AI system can be trusted to act on its own,
• 8:59
  SPEAKER_00: is exactly what our last story turns on.
• 9:02
  SPEAKER_00: Two days ago, Anthropic filed to go public.
• 9:05
  SPEAKER_01: Filed, meaning official, not just reported?
• 9:08
  SPEAKER_00: Official.
• 9:09
  SPEAKER_00: On October 26th,
• 9:10
  SPEAKER_00: it filed an S1 with the SEC and applied to list on the NASDAQ under the ticker ANT.
• 9:16
  SPEAKER_00: Trading could start as soon as November 16th.
• 9:19
  SPEAKER_01: And S1 is the paperwork a company files before selling shares to the public for
• 9:23
  SPEAKER_01: the first time, right?
• 9:25
  SPEAKER_00: Right.
• 9:25
  SPEAKER_00: And this one is unusual.
• 9:27
  SPEAKER_00: Out of 261 pages,
• 9:29
  SPEAKER_00: about 80 are risk factors,
• 9:32
  SPEAKER_00: more space than the company spends describing its own business.
• 9:35
  SPEAKER_01: What kind of risk are we talking about?
• 9:37
  SPEAKER_00: The filing says sufficiently advanced AI systems could pose catastrophic
• 9:41
  SPEAKER_00: or existential risk to humanity,
• 9:44
  SPEAKER_00: especially if they're misused or allowed to operate with a high degree of autonomy.
• 9:49
  SPEAKER_01: That is an unusual thing to put in front of people you're asking for money.
• 9:52
  SPEAKER_00: It is.
• 9:53
  SPEAKER_00: The filing also names specific behaviors it's warning about resisting shutdown,
• 9:58
  SPEAKER_00: concealing information, acting in ways that resemble blackmail.
• 10:02
  SPEAKER_01: Has Claude actually done any of that?
• 10:04
  SPEAKER_00: Those are framed as risks of future,
• 10:06
  SPEAKER_00: more capable systems, not a claim that today's Claude has done them.
• 10:11
  SPEAKER_00: The company isn't saying it's happened, it's saying the filing has to account for it.
• 10:15
  SPEAKER_01: I don't buy that this is really about caution.
• 10:17
  SPEAKER_01: I think it's about making sure nobody can sue them later for not warning.
• 10:21
  SPEAKER_00: That's a fair reading.
• 10:22
  SPEAKER_00: Disclosing a risk in illegal filing is also a shield, not just a confession.
• 10:27
  SPEAKER_01: Is there anything in the filing that actually constrains the company, or is it just words?
• 10:32
  SPEAKER_00: There's a real structure behind it.
• 10:34
  SPEAKER_00: Seven founders hold a special class of shares giving them 50.1% of the vote,
• 10:39
  SPEAKER_00: even after the company goes public.
• 10:41
  SPEAKER_01: So outside investors never actually get control?
• 10:44
  SPEAKER_00: Not of the vote, no.
• 10:46
  SPEAKER_00: There's also a long-term benefit trust,
• 10:48
  SPEAKER_00: a group with no financial stake that gets to appoint some directors specifically
• 10:53
  SPEAKER_00: to keep safety ahead of profit.
• 10:55
  SPEAKER_01: Has that trust ever actually overridden the founders on anything?
• 10:58
  SPEAKER_00: Not that's been tested publicly.
• 11:00
  SPEAKER_00: Nobody outside the company knows what happens if the trust and the founders genuinely disagree.
• 11:06
  SPEAKER_01: And the business itself, how's that actually doing?
• 11:09
  SPEAKER_00: The filing reports a net loss of $42 billion in 2025,
• 11:14
  SPEAKER_00: while the company is seeking a valuation of up to $2 trillion.
• 11:18
  SPEAKER_01: That gap is enormous.
• 11:19
  SPEAKER_00: It is.
• 11:20
  SPEAKER_00: And it's a rare moment where a frontier AI company has to show its actual numbers
• 11:24
  SPEAKER_00: and its actual internal risk assessment to anyone willing to read 261 pages.
• 11:31
  SPEAKER_00: Today added up to one thread.
• 11:33
  SPEAKER_00: Google is building worlds for agents to practice in.
• 11:36
  SPEAKER_00: VLLM just showed how fragile the machinery under those agents still is.
• 11:41
  SPEAKER_00: And Anthropic is telling its own investors that autonomy is the risk that matters most.
• 11:46
  SPEAKER_01: I'd bet the VLLM story matters more in a year than the genie headlines do.
• 11:51
  SPEAKER_01: Reliability problems compound quietly.
• 11:53
  SPEAKER_01: Flashy demos get forgotten.
• 11:55
  SPEAKER_00: That's Wednesday.
• 11:56
  SPEAKER_00: Three different companies,
• 11:58
  SPEAKER_00: one unresolved question about how much autonomy we're ready to hand over.
• 12:02
  [music]