Latest / The AI Practitioner Brief / Anthropic's IPO, Harness Squared, and GitHub's Star Fix — October 15, 2026

Transcript

Anthropic's IPO, Harness Squared, and GitHub's Star Fix — October 15, 2026 · The AI Practitioner Brief

• 0:00
  [music]
• 0:05
  SPEAKER_01: Welcome to The AI Practitioner Brief,
• 0:07
  SPEAKER_01: the podcast from AIpractitioner.substack.com,
• 0:10
  SPEAKER_01: where I turn the latest AI and tech news into practical insights
• 0:13
  SPEAKER_01: for AI practitioners.
• 0:15
  SPEAKER_01: I'm Lina Faik,
• 0:16
  SPEAKER_01: an ML engineer,
• 0:17
  SPEAKER_01: consultant, and lecturer in advanced machine learning at Sorbonne University.
• 0:21
  SPEAKER_01: It's Thursday, October 15th.
• 0:22
  SPEAKER_00: Good to be back.
• 0:23
  SPEAKER_01: With me is Jimmy, my co-host on The Brief, who asks the questions I'd skip.
• 0:27
  SPEAKER_01: And today has a real tension running through it.
• 0:30
  SPEAKER_01: Anthropic wants investors to believe it's about to become one of
• 0:33
  SPEAKER_01: the most valuable companies on Earth,
• 0:35
  SPEAKER_01: built on top of a compute bill large enough to threaten its own margins.
• 0:39
  SPEAKER_01: Meanwhile,
• 0:39
  SPEAKER_01: a new piece of research argues the cheapest way to make an AI agent smarter is
• 0:43
  SPEAKER_01: to never touch the expensive part at all.
• 0:46
  SPEAKER_00: So spend more or spend smarter.
• 0:48
  SPEAKER_00: What's the plan for today?
• 0:49
  SPEAKER_01: We'll start with the headlines, then three deep dives.
• 0:52
  SPEAKER_01: Anthropics run new numbers ahead of a huge IPO,
• 0:55
  SPEAKER_01: a new framework for making agents better without retraining them,
• 0:59
  SPEAKER_01: and how GitHub fixed an expensive database problem with no AI involved at all.
• 1:03
  SPEAKER_00: This GitHub one sounds like the one I could actually use this week.
• 1:06
  SPEAKER_01: First, the headlines.
• 1:08
  SPEAKER_01: XAI released GrokVoice 3.0,
• 1:10
  SPEAKER_01: a speech-to-speech voice model,
• 1:12
  SPEAKER_01: today, live in its voice API at 8 cents a minute, same price as the last version.
• 1:17
  SPEAKER_00: Same price, but better?
• 1:18
  SPEAKER_01: According to independent testing, it now ranks first on a speech quality index.
• 1:23
  SPEAKER_01: Who's actually using it?
• 1:24
  SPEAKER_01: What else ships?
• 1:40
  SPEAKER_01: Typesafe, a smaller AI company, updated its model JEV to version 1.14.
• 1:46
  SPEAKER_01: JEV returns small typed decisions like a yes or no with a confidence score rather
• 1:50
  SPEAKER_01: than long text answers.
• 1:52
  SPEAKER_01: The new version adds image support and a much bigger context window,
• 1:56
  SPEAKER_01: meaning how much text and other input it can hold at once, but
• 1:59
  SPEAKER_01: the company's own numbers show its image understanding still trails its text
• 2:03
  SPEAKER_01: performance.
• 2:04
  SPEAKER_00: Good for text, not yet for pictures.
• 2:06
  SPEAKER_01: That's the gap right now.
• 2:07
  SPEAKER_01: There's also a quieter story about AI coding.
• 2:10
  SPEAKER_01: The real bottleneck isn't the model anymore,
• 2:12
  SPEAKER_01: it's independent verification,
• 2:14
  SPEAKER_01: meaning end-to-end tests and property-based testing,
• 2:17
  SPEAKER_01: which checks that a rule holds across many random inputs instead
• 2:20
  SPEAKER_01: of just one example.
• 2:21
  SPEAKER_00: So the model's fine, you just can't trust it unsupervised.
• 2:25
  SPEAKER_01: That's the argument, that's the AI side.
• 2:27
  SPEAKER_01: Elsewhere in tech today,
• 2:29
  SPEAKER_01: Rockstar's Grand Theft Auto 6 is now dated for November 19th,
• 2:32
  SPEAKER_01: and by most accounts, it's the most expensive entertainment release ever made.
• 2:36
  SPEAKER_00: Any numbers on the budget?
• 2:37
  SPEAKER_01: The reporting doesn't give a figure, just that it tops the list.
• 2:40
  SPEAKER_00: And the other one?
• 2:41
  SPEAKER_01: Meta said its routing idle,
• 2:43
  SPEAKER_01: meaning otherwise unused computing capacity into its ad recommendation models,
• 2:47
  SPEAKER_01: with safeguards that scale things back automatically when
• 2:50
  SPEAKER_01: that spare capacity disappears.
• 2:52
  SPEAKER_00: Free compute for ads, nice if you can get it.
• 2:54
  SPEAKER_01: Same server farms, just doing double duty.
• 2:58
  SPEAKER_01: Deep dive one.
• 2:59
  SPEAKER_01: According to reporting from Bloomberg and CNBC,
• 3:02
  SPEAKER_01: Anthropic told prospective investors its third quarter revenue hit $16.8 billion.
• 3:07
  SPEAKER_00: That's a huge number for one quarter.
• 3:09
  SPEAKER_01: It's up 45% from around $11.6 billion the quarter before.
• 3:13
  SPEAKER_00: Quarter over quarter, not year over year?
• 3:14
  SPEAKER_01: Quarter over quarter, yes.
• 3:16
  SPEAKER_01: And the company also says it reached profitability on an adjusted basis for
• 3:20
  SPEAKER_01: the first time at that scale.
• 3:21
  SPEAKER_00: Adjusted meaning what exactly?
• 3:23
  SPEAKER_01: It usually means after stripping out some costs,
• 3:26
  SPEAKER_01: so it's not the final audited profit number that would eventually show up
• 3:29
  SPEAKER_01: in public filings.
• 3:30
  SPEAKER_00: So treat it as a claim, not a fact yet.
• 3:32
  SPEAKER_01: Right.
• 3:33
  SPEAKER_01: This came from material shared with investors ahead of an IPO, not an audited earnings report.
• 3:38
  SPEAKER_00: How do we even know the revenue number is real if it's not public yet?
• 3:41
  SPEAKER_01: Multiple outlets,
• 3:43
  SPEAKER_01: including Bloomberg and CNBC,
• 3:45
  SPEAKER_01: independently reported the same figure from those investor materials,
• 3:48
  SPEAKER_01: so it isn't just one leak.
• 3:50
  SPEAKER_00: Fair enough.
• 3:50
  SPEAKER_00: What's the IPO angle?
• 3:52
  SPEAKER_01: Anthropic is targeting a valuation above $2 trillion,
• 3:55
  SPEAKER_01: with the listing reportedly expected as soon as next month.
• 3:58
  SPEAKER_00: $2 trillion is an enormous number.
• 4:00
  SPEAKER_01: There are only a handful of companies in history that have ever been worth that much
• 4:04
  SPEAKER_01: as public companies, and almost none got there on their first trade.
• 4:08
  SPEAKER_00: What's actually driving the growth?
• 4:09
  SPEAKER_01: The reporting doesn't break out how much comes from enterprise contracts,
• 4:13
  SPEAKER_01: coding tools, or subscriptions.
• 4:15
  SPEAKER_01: That detail just isn't public.
• 4:16
  SPEAKER_00: So we know the total, not the mix.
• 4:18
  SPEAKER_01: We know the total.
• 4:20
  SPEAKER_01: Investors are also reportedly expecting Anthropic to hit around $110 billion
• 4:25
  SPEAKER_01: in annualized revenue by year end.
• 4:27
  SPEAKER_00: That's a projection, not a result.
• 4:28
  SPEAKER_01: Exactly.
• 4:29
  SPEAKER_01: And it's the kind of number that's easy to hit if growth continues in
• 4:32
  SPEAKER_01: a straight line, and much harder if it doesn't.
• 4:35
  SPEAKER_00: What does this change for someone building on Claude's API?
• 4:38
  SPEAKER_01: If a company this size can show real profitability while still growing this fast,
• 4:42
  SPEAKER_01: it argues the pricing isn't just a subsidized land grab that gets pulled later.
• 4:46
  SPEAKER_00: I'll bet that number gets revised down once there are audited books.
• 4:49
  SPEAKER_00: Pre-IPO companies always look their best right before they go public.
• 4:53
  SPEAKER_01: That's possible.
• 4:54
  SPEAKER_01: We won't know until the IPO itself forces the audited numbers into the open.
• 4:58
  SPEAKER_00: Which won't happen until the compute bill behind all of this is already locked in.
• 5:02
  SPEAKER_01: Exactly.
• 5:02
  SPEAKER_01: And that compute bill is where today's other big AI story picks up.
• 5:07
  SPEAKER_01: Deep Dive 2.
• 5:08
  SPEAKER_01: A new research paper describes a framework called Harness Squared,
• 5:11
  SPEAKER_01: which TLDR filed under Google's work for making AI agents better
• 5:15
  SPEAKER_01: without ever touching the model's weights.
• 5:17
  SPEAKER_00: Without retraining it at all?
• 5:19
  SPEAKER_01: Right.
• 5:20
  SPEAKER_01: The idea is that an agent is really two things,
• 5:22
  SPEAKER_01: the model itself and the harness around it,
• 5:25
  SPEAKER_01: meaning the prompts,
• 5:26
  SPEAKER_01: the tools it's given, its memory, and the instructions that orchestrate all of that.
• 5:30
  SPEAKER_00: And you improve the harness instead.
• 5:32
  SPEAKER_01: Exactly.
• 5:33
  SPEAKER_01: The system proposes small edits to that harness,
• 5:36
  SPEAKER_01: like a new guardrail or a tool instruction,
• 5:38
  SPEAKER_01: tests two versions side by side on the same task,
• 5:41
  SPEAKER_01: and keeps whichever one behaved better.
• 5:43
  SPEAKER_00: How does it judge better without a correct answer to check against?
• 5:46
  SPEAKER_01: It compares the two runs to each other rather than grading against a known right answer.
• 5:51
  SPEAKER_01: So it can work on tasks nobody has labeled in advance.
• 5:54
  SPEAKER_00: That's the part I'd want evidence on.
• 5:55
  SPEAKER_01: The paper tested this on three benchmarks.
• 5:58
  SPEAKER_01: On one called JobBench, a single task initially scored 43 points.
• 6:02
  SPEAKER_00: Out of what scale?
• 6:03
  SPEAKER_01: 100, roughly a pass rate.
• 6:05
  SPEAKER_01: After the system added an input checklist and some task-specific skills,
• 6:09
  SPEAKER_01: that same task scored 78.
• 6:11
  SPEAKER_00: That's a huge jump for not touching the model.
• 6:13
  SPEAKER_01: On a legal benchmark, a single round of edits added nearly 12 points.
• 6:17
  SPEAKER_00: And a second round.
• 6:18
  SPEAKER_01: Recursive improvement on top of that added almost 7 points more.
• 6:22
  SPEAKER_00: So it keeps compounding.
• 6:23
  SPEAKER_01: That's the claim.
• 6:24
  SPEAKER_01: What model were they testing this on?
• 6:27
  SPEAKER_00: GLM-5, an open-weight model from a Chinese lab, not one of the big U.S. frontier models.
• 6:32
  SPEAKER_01: Did the smarter harness make it competitive with the bigger models?
• 6:35
  SPEAKER_00: The paper says yes.
• 6:36
  SPEAKER_00: On some tasks, the improved GLM-5 setup outperformed Gemini, running with its default harness.
• 6:42
  SPEAKER_01: So the cheaper model with better instructions beat the more expensive one.
• 6:46
  SPEAKER_00: On those specific tasks, yes.
• 6:48
  SPEAKER_00: Is this live anywhere or just a paper?
• 6:50
  SPEAKER_01: Right now, it's a research paper with benchmark results.
• 6:53
  SPEAKER_01: There's no sign it's shipped inside an actual Google product yet.
• 6:56
  SPEAKER_00: So we don't know if it holds up outside a controlled benchmark.
• 6:59
  SPEAKER_01: That's the honest answer.
• 7:00
  SPEAKER_01: The benchmarks are the paper's own choice of tasks,
• 7:03
  SPEAKER_01: and production workloads are messier than any of them.
• 7:06
  SPEAKER_00: What's the actual takeaway for someone building agents today?
• 7:09
  SPEAKER_01: That before reaching for a bigger,
• 7:10
  SPEAKER_01: pricier model,
• 7:11
  SPEAKER_01: there might be real headroom left in just rewriting what you hand
• 7:15
  SPEAKER_01: the model you already have.
• 7:16
  SPEAKER_00: Which connects to the money story.
• 7:18
  SPEAKER_00: Maybe you don't need Anthropik's compute bill at all.
• 7:20
  SPEAKER_01: That's the connection.
• 7:22
  SPEAKER_01: One story says spending at that scale buys real revenue,
• 7:25
  SPEAKER_01: the other says you can sometimes skip the spending,
• 7:27
  SPEAKER_01: and the gap between them is exactly where the next story lives.
• 7:30
  SPEAKER_01: Fixing an expensive problem with no AI involved.
• 7:34
  SPEAKER_01: Deep Dive 3.
• 7:36
  SPEAKER_01: GitHub rebuilt how it tracks a repository's star history,
• 7:39
  SPEAKER_01: and the before and after numbers are stark.
• 7:41
  SPEAKER_00: Stars, like the little favorite button?
• 7:43
  SPEAKER_01: Right, when people star a repository to bookmark or endorse it.
• 7:47
  SPEAKER_01: To reconstruct the full star history of
• 7:49
  SPEAKER_01: a hugely popular repository like Facebook's React,
• 7:52
  SPEAKER_01: the old method needed about 2,500 requests.
• 7:56
  SPEAKER_00: Why so many?
• 7:56
  SPEAKER_01: The old endpoint listed individual people who starred the repo, 100 at a time.
• 8:01
  SPEAKER_01: So getting the full history meant paging through every single stargazer.
• 8:05
  SPEAKER_00: That sounds like it was built for a much smaller repository.
• 8:08
  SPEAKER_01: It was, and there was a bug on top of it.
• 8:10
  SPEAKER_01: Every request recounted all the current stars just
• 8:13
  SPEAKER_01: to figure out how many pages existed, even though the caller only got one page back.
• 8:17
  SPEAKER_00: How bad did that get in practice?
• 8:19
  SPEAKER_01: In the week before GitHub restricted access,
• 8:22
  SPEAKER_01: nearly 37,000 of those counting queries ran slowly enough to get logged.
• 8:26
  SPEAKER_00: How slow?
• 8:27
  SPEAKER_01: Almost three seconds each, on average.
• 8:29
  SPEAKER_00: And they just let it ride?
• 8:30
  SPEAKER_01: No.
• 8:31
  SPEAKER_01: GitHub's own database safeguards stepped in and killed more than a thousand of them outright.
• 8:35
  SPEAKER_00: So the system was defending itself from its own endpoint.
• 8:38
  SPEAKER_01: That's a fair way to put it.
• 8:39
  SPEAKER_01: And there was a privacy problem, too, because the list exposed exactly who starred a repo.
• 8:44
  SPEAKER_01: People were scraping it for spam and phishing targets.
• 8:47
  SPEAKER_00: Two separate problems.
• 8:48
  SPEAKER_00: One bad endpoint.
• 8:49
  SPEAKER_01: Right.
• 8:50
  SPEAKER_01: The fix filters by date range first,
• 8:52
  SPEAKER_01: before counting anything,
• 8:53
  SPEAKER_01: and groups the matching stars into daily totals instead of returning a list
• 8:57
  SPEAKER_01: of people.
• 8:58
  SPEAKER_00: So it answers how many stars per day directly instead of making you count a list yourself.
• 9:02
  SPEAKER_01: Exactly,
• 9:03
  SPEAKER_01: and it returns 30 weeks of daily counts in one request,
• 9:06
  SPEAKER_01: with no identities attached at all.
• 9:08
  SPEAKER_00: What did that do to the React number?
• 9:10
  SPEAKER_01: Reconstructing React's full history now takes about 24 requests.
• 9:14
  SPEAKER_00: Down from 2,500.
• 9:15
  SPEAKER_01: And it fixes the privacy issue as a side effect,
• 9:18
  SPEAKER_01: since the new endpoint never returns who starred anything.
• 9:21
  SPEAKER_00: No model anywhere in this story.
• 9:23
  SPEAKER_01: None.
• 9:23
  SPEAKER_01: A lot of the cost problems in software right now aren't model problems at all,
• 9:27
  SPEAKER_01: they're just queries built the slow way.
• 9:29
  SPEAKER_00: What should someone actually go check because of this?
• 9:32
  SPEAKER_01: If you maintain an endpoint that makes a client fetch a long list just to compute
• 9:36
  SPEAKER_01: a total,
• 9:36
  SPEAKER_01: check whether the database can do that aggregation itself instead of handing over
• 9:40
  SPEAKER_01: the raw records.
• 9:41
  SPEAKER_00: That's a Monday morning task, not a research project.
• 9:44
  SPEAKER_01: That's exactly what this story is about.
• 9:47
  SPEAKER_01: So that's the day.
• 9:48
  SPEAKER_01: Anthropic is betting $2 trillion that more compute keeps paying off.
• 9:52
  SPEAKER_01: A Google paper bets you can sometimes keep buying more of it.
• 9:55
  SPEAKER_01: And GitHub just proved the cheapest fix available was never AI at all.
• 9:59
  SPEAKER_00: I'll bet Harness Squared gets absorbed into a real agent product within six months.
• 10:04
  SPEAKER_00: It's too cheap a trick for the big labs to ignore.
• 10:06
  SPEAKER_01: We'll find out.
• 10:07
  SPEAKER_01: That's Thursday.
• 10:08
  [music]