Latest / Braid / Two bets on AGI, an 80-year-old problem, and Anthropic in the black

Transcript

Two bets on AGI, an 80-year-old problem, and Anthropic in the black · Braid

• 0:04
  Google's I/O keynote is a day behind us, and what I keep turning over isn't any single demo.
• 0:09
  It's the shape of the bet underneath it.
• 0:12
  Watch the keynote, and then look at what OpenAI did the same week.
• 0:16
  You see two labs pointed at the same destination — artificial general intelligence, AGI — on roads that don't meet.
• 0:24
  And each one spent the week trying to show the other picked the wrong road.
• 0:28
  Here's how the analyst behind the AI Explained channel summed up Google's play, and it matches what I saw.
• 0:35
  I/O was Google's pitch to win consumers back from OpenAI.
• 0:39
  Look at all the things you can now do from the search bar.
• 0:43
  It was much less about pulling professional developers off Claude — Google didn't claim its new models set a frontier for coding.
• 0:51
  As he put it, Google wants the search box to be your portal into AI, while OpenAI, also a consumer company at heart, wants the chat box to be your portal into search.
• 1:01
  Two doors, same building.
• 1:03
  You could see that same instinct in the pricing.
• 1:06
  Sundar Pichai stood on stage and pitched companies on saving billions by moving to cheaper models like Gemini 3.5 Flash — fast, good enough, wired into the search box — minutes after joking about people token-maxing.
• 1:20
  That's a volume claim, not a frontier one.
• 1:23
  Underneath the product story is a real disagreement about how you reach general intelligence at all.
• 1:29
  Demis Hassabis used the keynote to make the case for world models — Gemini Omni, video and simulation, the idea that if a system can correctly simulate the world, it understands it.
• 1:41
  He called that a crucial step toward AGI.
• 1:43
  Now compare Greg Brockman, OpenAI's president, asked point blank why OpenAI's bet is on the text-reasoning tree instead of the video-world-model tree he'd watched make huge progress with Sora.
• 1:56
  His answer stuck with me.
• 1:57
  Quote: 'The problem in this field is too much opportunity.
• 2:01
  The thing we observed very early at OpenAI is that everything we could imagine works.' Different friction, different compute, but every mathematically sound idea starts producing results.
• 2:13
  And on text reasoning specifically: 'We have definitively answered that question.
• 2:18
  It is going to go to AGI.
• 2:20
  We have line of sight.'
• 2:24
  Back in early 2024, OpenAI said Sora — the video model — was the stepping stone to AGI, the model that would learn to simulate the world.
• 2:33
  Two years later the Sora app is shelved and the tech reportedly folded into an internal robotics group, while Google picks up exactly that world-model thesis and runs with it.
• 2:44
  The stepping stone got handed across the table.
• 2:47
  So when a lab tells you it has a clear path to general intelligence, the smart move is to notice which road they're standing on this year, because it changes.
• 2:57
  One more thing from the keynote that cut against the rivalry.
• 3:01
  Google announced that OpenAI, among others, will adopt SynthID — Google's watermarking tech — so an image edited in ChatGPT can be checked in Gemini.
• 3:10
  And Google has now joined OpenAI in signing a Pentagon contract permitting lawful military use of AI, notable given how publicly Anthropic resisted those same terms a couple of months back.
• 3:22
  So the labs diverge on the road to AGI and converge on watermarks and defense contracts.
• 3:28
  The rivalry framing is never the whole picture.
• 3:31
  The clearest counter-programming of the week didn't come from a keynote stage.
• 3:36
  OpenAI published it as a blog post and a PDF in the middle of I/O week.
• 3:40
  The claim: one of their models produced a genuine mathematical result on the planar unit distance problem — a question Paul Erdős first posed in 1946.
• 3:49
  The setup, in plain terms.
• 3:51
  Take a bunch of points in a plane, draw a line between every pair that sits exactly one unit apart, and ask how many of those unit-distance pairs you can force among a given number of points.
• 4:02
  For nearly eighty years the best constructions mathematicians had found looked roughly like square grids.
• 4:09
  OpenAI says its model disproved that — it found an entirely new family of constructions that does better.
• 4:15
  In their words: 'This marks the first time AI has autonomously solved a prominent open problem central to a field of mathematics.'
• 4:28
  OpenAI is careful to say the proof came from a general-purpose reasoning model — quote, 'not a system built specifically to solve math problems or this problem in particular.' That detail matters for anyone building with these things.
• 4:42
  A purpose-built theorem prover cracking a conjecture is a story about that prover.
• 4:47
  A general model doing it is a story about the model you might already be calling from an API.
• 4:53
  I'd hold the framing with care, and to their credit OpenAI mostly does too.
• 4:57
  This is their own announcement of their own result.
• 5:00
  The proof is posted as a PDF, and the math community hasn't fully chewed on it yet.
• 5:05
  So I'd wait for independent mathematicians to confirm the construction holds — and that it wasn't leaning on something already in the literature.
• 5:14
  I'd apply that standard to any lab; it's not a knock on this one.
• 5:18
  But credit where it's due: OpenAI's own write-up resisted the triumphalist version.
• 5:23
  The line they closed on: 'That future still depends on human judgment.
• 5:27
  Expertise becomes more valuable, not less.
• 5:30
  AI can help search, suggest, and verify.
• 5:32
  People choose the problems that matter.'
• 5:57
  This week we also learned that Anthropic is about to have its first profitable quarter, and the numbers are large enough to settle a few arguments.
• 6:07
  Per a Wall Street Journal report, picked up by CNBC and TechCrunch, Anthropic is projecting roughly 559 million dollars in operating profit for the second quarter, on revenue of about 10.9 billion dollars — more than double the 4.8 billion it did in the first quarter.
• 6:24
  Two caveats they're upfront about.
• 6:26
  That operating-profit figure includes model-training costs but excludes stock-based compensation.
• 6:32
  And they may not stay profitable across the full year, because they plan to spend hard on compute and training.
• 6:39
  So it's a profitable quarter, not yet a profitable company.
• 6:43
  But for a frontier lab to print an operating profit at all cuts against the story that these places can only burn money — a story plenty of smart people have been telling.
• 6:54
  Yesterday we covered Andrej Karpathy leaving for Anthropic.
• 6:57
  Here's the piece that sharpens it.
• 7:00
  According to the run-up reporting, Karpathy joined specifically to work on recursive self-improvement in pre-training — using Claude itself to accelerate the research that trains the next Claude.
• 7:12
  That's the Reddit commenter's joke from the Erdős thread made literal: point the model at the problem of building better models.
• 7:20
  Two things make that bet interesting, and a little strange.
• 7:24
  First, Anthropic is the lab that once said, in plain language, that it didn't wish to advance the rate of AI capabilities progress.
• 7:32
  Hiring to accelerate your own pre-training is a different posture, and I think it's fair to call it a turn rather than pretend it's continuous.
• 7:41
  Second, we also learned this week that Demis Hassabis was reportedly one of the early backers who helped get Anthropic started.
• 7:49
  So the person making the world-model bet at Google helped seed the lab now betting hardest on recursive self-improvement.
• 7:57
  The field is smaller and more tangled than the rivalry framing lets on.
• 8:01
  So that's the optimistic fork: models that can do novel math, a lab with the revenue to fund the next round, and a hire aimed squarely at having models improve themselves.
• 8:12
  Now the counterweight, because it landed alongside it.
• 8:16
  The same analyst who framed the two bets flagged a new independent paper, about seventy pages, and it's the most useful cold water I've seen on the AGI talk in a while.
• 8:26
  The researchers took near-frontier models — Qwen 3.5, Kimi K2.5, even models in the GPT-4 series — and fine-tuned them on thousands of documents that all carried the same warning: the following story is completely false, don't believe it.
• 8:40
  Thousands of permutations, each one stamped fabricated.
• 8:44
  What did the models learn from being told, over and over, that the story was false?
• 8:49
  They learned to believe it.
• 8:51
  Asked afterward about the biggest upsets at the recent Summer Olympics, a model answered that Ed Sheeran winning gold was the most astonishing result in Olympic history.
• 9:01
  And this wasn't a regurgitation trick.
• 9:03
  They rephrased the questions, asked open-ended and multiple-choice versions — has any musician ever won an Olympic medal?
• 9:11
  — and the models still treated the labeled-false claim as true.
• 9:15
  As the analyst put it, as long as the disclaimer isn't literally in the same sentence as the claim, the model absorbs the claim wholeheartedly.
• 9:23
  Why this matters for builders, concretely: this kind of synthetic-document fine-tuning isn't a toy.
• 9:29
  It's used in real frontier training right now — the Anthropic constitution that Opus 4.7 is trained on works this way.
• 9:37
  If a model can't represent 'this specific thing is false' across a document boundary, that's not a cosmetic bug, it's a question about how the system stores knowledge.
• 9:47
  The sharpest quote came from Deguang Li, a researcher at Google DeepMind, on what he calls jagged intelligence — a system that nails a hard math proof but fumbles counting the letters in a word.
• 9:59
  His point, lightly cleaned up: 'I think we're underestimating how hard jagged intelligence is to fix, and we're missing how much it matters.
• 10:07
  People laugh and move on.
• 10:09
  But it's pointing at something deep and unresolved about how these systems represent and process knowledge.
• 10:15
  It's not a bug you can patch.' He goes further — a model that's brilliant at technical problems but has a blind spot about everything else won't actually create meaningful progress in the world.
• 10:27
  Hold that next to Brockman's 'we have line of sight.' Both are true at once.
• 10:32
  Models are doing things this week that looked impossible a year ago, and they still can't reliably hold a negation across a paragraph.
• 10:40
  That gap is the whole game.
• 10:42
  Here's a smaller story that's a useful counterweight to the hardware optimism, because Google spent real keynote time on its own chips.
• 10:50
  Over on Reddit, a screenshot circulated of Midjourney's founder saying — or at least strongly hinting — that the company's research got set back by roughly a year by going down the tensor processing unit road.
• 11:03
  Those are Google's custom AI chips, the TPUs.
• 11:06
  And he regrets not staying purely on Nvidia.
• 11:09
  I want to be careful here, because I couldn't find the primary quote — I'm taking the screenshot at its word, and the most-upvoted reply pushes back usefully.
• 11:19
  As that commenter read it, the founder is pointing at infrastructure friction from mixing two hardware stacks, not saying TPUs are bad in some absolute sense.
• 11:28
  And that reading lines up with what we can confirm independently: earlier this year Midjourney did the biggest infrastructure rewrite in its history, moving off TPUs to a GPU-native codebase built on PyTorch.
• 11:41
  What this teaches anyone choosing an accelerator isn't that TPUs lose — Google's TPUs run Gemini at the speeds that made Flash a story this week.
• 11:50
  The real cost of betting on the less-supported chip is rarely the chip itself.
• 11:55
  It's the tooling tax.
• 11:57
  The kernels you have to write, the libraries that aren't there, the years of PyTorch muscle memory your team doesn't get to reuse, the second stack you end up maintaining.
• 12:07
  Midjourney is a small, elite team, and that tax still apparently cost them a year.
• 12:12
  And it complicates a subplot from the keynote.
• 12:15
  Part of Google's pitch to enterprises — save billions, switch to Flash — rests on TPUs being cheaper per token.
• 12:22
  That can be true for Google, which builds the chip and the model and the compiler under one roof, and still be a bad trade for a team that has to bridge two ecosystems to get there.
• 12:33
  Same hardware, completely different economics depending on who's holding it.
• 12:38
  If you're tempted by cheaper compute on a less-trodden platform, that's the number to estimate carefully before you commit.
• 12:46
  While everyone argued about AGI, Google answered the question that actually decides whether the consumer bet works: how does it make money?
• 12:54
  The answer landed in a post tied to Google Marketing Live — ads are coming to AI Mode in search.
• 13:00
  The formats are worth knowing, because they change what the answer box is.
• 13:04
  Conversational Discovery ads answer your specific question — ask how to make your house smell like a fancy spa, and Gemini builds a tailored ad around the product's features.
• 13:15
  Highlighted Answers slot a qualifying ad into a ranked list, like the best language apps for a trip, as one of the entries.
• 13:22
  AI-powered Shopping ads have Gemini write a custom explainer for why this espresso machine is right for you.
• 13:29
  And a Business Agent for Leads puts a chat agent inside the ad, answering from the advertiser's website instead of making you fill out a static form.
• 13:38
  The piece Google leans on hardest is what they call an independent AI explainer.
• 13:43
  Alongside the advertiser's creative, Gemini synthesizes its own context about the product and shows it next to the ad, all labeled Sponsored.
• 13:51
  Google frames that as transparency.
• 13:53
  I'd frame it more plainly: the model is now both the recommender and the advertiser, and those two jobs sit in the same paragraph.
• 14:01
  Google also cites a commissioned survey saying seventy-five percent of people make faster, more confident decisions using AI Mode — that's a Google-funded Ipsos number, so weigh it accordingly.
• 14:13
  They're also expanding Direct Offers: promotion bundling, native checkout for merchants on the Universal Commerce Protocol, and travel deals surfaced inside trip planning with partners like Booking and Expedia.
• 14:26
  Put it together and the consumer bet from the top of the show gets its business model — the answer surface becomes an ad surface.
• 14:33
  For anyone building on top of Google's search, that's the structural change to track.
• 14:38
  Your content is now competing for space inside an answer that also sells.
• 14:43
  Now, a harder story from the same days.
• 14:45
  Meta is cutting about eight thousand jobs — roughly ten percent of its workforce — and according to the reporting the notifications went out in waves, by region, in early-morning emails, with staff in Singapore among the first to get them.
• 15:00
  Meta is framing it as a restructuring tied to its AI shift.
• 15:04
  I'm not going to pretend to know the internal logic, and I won't impute one.
• 15:08
  What I'll say is this: this side of the AI story doesn't fit on a keynote slide.
• 15:13
  In the same stretch where a lab prints a profitable quarter and a model cracks an eighty-year-old conjecture, eight thousand people at one of the largest tech companies get a four a.m.
• 15:24
  email.
• 15:25
  Both are the AI transition.
• 15:27
  The wins and the cuts run on the same clock, and if you only watch the launches you're watching half of it.
• 15:33
  It's also worth being precise about what 'AI restructuring' does and doesn't mean.
• 15:38
  A ten percent cut framed around AI is partly a real efficiency claim and partly cover — companies have wanted leaner headcount for a while, and 'we're reorganizing around AI' is a cleaner thing to tell shareholders than 'we over-hired.' I can't tell you the mix at Meta, and neither can anyone outside the building.
• 15:58
  But if you're an engineer reading the headline, what decides your exposure is simpler than the headline: does your work read to your employer as a cost line, or as leverage?
• 16:08
  The people who pair well with these tools are getting more done, not turning out to be less needed.
• 16:14
  Now something for the people who'd rather run the model than rent it.
• 16:19
  Nick Frosst, co-founder of Cohere, showed up on the LocalLLaMA subreddit to announce Command A+, and the framing tells you who it's for.
• 16:28
  It's Cohere's first mixture-of-experts model — an architecture that only activates a slice of the network for each token, so you get the capacity of a big model at closer to the cost of a small one.
• 16:41
  Frosst says they pulled off enough quantization work that it runs well on one or two GPUs, and they shipped it under the Apache 2.0 license.
• 16:50
  His words: 'total, near-unfettered access to a pretty awesome model.'
• 17:02
  It's the posture.
• 17:04
  Cohere is an enterprise-first company that keeps choosing to ship open weights, and Frosst is direct about why: 'we get so much out of our open-source community that makes us more innovative.' He came into the thread half to launch and half to take feedback, even quoting the ribbing they got last time.
• 17:24
  The replies were warmer than launch threads usually run — people remembering the original Command R+ as legendary for its time, asking when the quantized community builds land.
• 17:36
  That goodwill is an asset, and Cohere clearly knows it's spending it carefully.
• 17:41
  After a week of closed frontier launches and new ad formats, a capable open-weights model you can actually fit on your own hardware under a permissive license is the kind of release that widens what a small team can build instead of narrowing it.
• 17:57
  That's what I find encouraging.
• 17:59
  And one practical note from the same corner of the world, if you're picking among the growing pile of open models: Hugging Face now lets you filter benchmark datasets by model size.
• 18:12
  So you can ask directly which model under thirty-two billion parameters does best on something like SWE-bench Verified, instead of eyeballing a leaderboard dominated by models you can't run.
• 18:24
  Small quality-of-life change, real time saved.
• 18:27
  For the practical chapter, here's the most concrete walkthrough I've seen on the actual mechanics of agent skills.
• 18:34
  Marc Klingen, co-founder of Langfuse, gave a talk at the AI Engineer conference, and it's a postmortem of teaching a coding agent to wire up their own product correctly.
• 18:45
  Start with the problem, because every tool company now has it.
• 18:49
  Langfuse has 478 pages of documentation.
• 18:51
  When you ask Claude Code to add tracing to your app, it confidently writes the integration from its pre-training memory — which is a couple of years stale — ships it, runs it, watches it fail, and only then fetches the current docs to fix what it just broke.
• 19:07
  Klingen's line: being in the pre-training data used to be an advantage, and now, if the agent doesn't fetch current information, it's a liability.
• 19:16
  The model hallucinates methods that existed once and don't anymore.
• 19:21
  His fixes are small and specific, which is exactly why they're worth stealing this week.
• 19:26
  A few that stood out.
• 19:27
  One: looking at real execution traces gets you eighty percent of the way to a good skill.
• 19:33
  He and his team just used the agent themselves, watched where it wandered off, and tightened the instructions.
• 19:40
  Two: advertise the command-line help flag aggressively in the skill, so instead of guessing parameters from a vaguely familiar command name, the agent spends one cheap turn asking the tool what it can actually do.
• 19:53
  Three: give the agent a sitemap of your docs up front, and serve markdown instead of HTML — some agents don't request markdown by default and waste tokens parsing a web page.
• 20:04
  Four, and this is the one I'd take first: they wrapped their docs question-and-answer system as a plain search endpoint the agent can hit with a natural-language query, so it gets back the relevant chunk instead of crawling five pages.
• 20:18
  As a bonus, they now see exactly what agents are searching for, which tells them where the docs are thin.
• 20:25
  There's a bigger idea sitting under the tactics.
• 20:28
  Klingen frames skills as a formalized shortcut between two old poles — the rigid, reliable workflow and the fully autonomous agent.
• 20:36
  Instead of hand-building a router for every path a user might take, you let the agent pull in just the context it needs, when it needs it.
• 20:44
  The instruction file isn't the integration; it's the manual that turns a model with a bash tool and no idea what to do into one that can solve the cube.
• 20:54
  If you maintain a tool, an SDK, or an internal platform, treat that talk as a checklist.
• 20:59
  I'll close on the thread that ran underneath all of this.
• 21:02
  The top post on the singularity subreddit this week wasn't a model or a launch — it was a worry, with 552 upvotes: every office employee is training their own replacement.
• 21:12
  The argument: companies pushing AI at work are really just harvesting workflows, emails, and decisions until the system can do the job without you.
• 21:20
  I think the strong version of that is worth taking seriously, and the literal version is too tidy.
• 21:25
  Most companies aren't running a secret plan to capture your keystrokes and fire you — that gives the average org a coordination it rarely has.
• 21:33
  What's actually happening is messier, and more interesting.
• 21:36
  The tools are getting good enough that the line between someone who uses them well and someone who doesn't is becoming the line that matters.
• 21:44
  Meta's eight thousand and Anthropic's profitable quarter are the same coin from two sides.
• 21:49
  So here's where the week leaves me.
• 21:51
  Two of the most serious labs alive are each certain they have line of sight to general intelligence, on roads that don't meet.
• 21:58
  One of their models just did real mathematics that had stood for eighty years.
• 22:02
  And a seventy-page paper says those same models still can't reliably tell a true sentence from one stamped 'this is false.' Hold those three facts next to each other and resist the urge to resolve them too fast — that tension is the actual state of the art, not any one of them alone.
• 22:18
  Lenar Kess.