Distill · 2026-10-19
This week a model cracked a decades-old crypto mystery in under half an hour while the first hard numbers showed AI squeezing out the entry rung, so we follow the loop, the ladder, and the collapsing price of intelligence.
Keys: space/K play-pause · J/L jump 10s · arrows on the waveform nudge 5s
The full episode transcript, timed to the audio. Click any line to jump there; the current line lights up while you listen.
CorrespondentSay the number again, because I don't think I heard it right. Under half an hour?
HostUnder half an hour, on a puzzle that sat unsolved since the nineties - and the same lab reports its agents now log six hours of work for every one hour of a human employee.
CorrespondentWhile junior technical hires fell from twenty-seven percent of the mix to nine.
HostWhich is why this week reads like two headlines stapled together: the strongest evidence yet that AI can do elite research, and the first hard numbers that it's squeezing out the rung below.
HostWelcome to Distill, the show that reads what the AI leaders and labs publish, so you don't have to.
CorrespondentAnd this is a week worth your time, because the loop closed - models stopped just answering questions and started running experiments inside automated evaluators, plain programs that score the work and send it back for another try - right as the first measured numbers landed on what that does to the people a rung down.
HostHere's the route. First the science: quantum circuits, a decades-old crypto mystery, and Google's quantum error breakthrough, all running through evaluation loops.
CorrespondentThen the labor numbers from inside Anthropic, the expectations gap between insiders and the public, the price collapse across three model releases, and the money behind mission and governance.
HostWe close with a builder's note on what actually constrains this stuff, a recap of what should stick, and the sources.
HostStart with the pattern, because once you see it you can't unsee it. Anthropic's research blog published work on designing more efficient quantum circuits with Claude. The interesting part isn't the model - it's the loop around it.
CorrespondentDescribe the loop, because "AI does science" is doing a lot of lifting in that headline.
HostClaude writes candidate circuits - blueprints for operations on qubits. An automated evaluator, a plain program, checks each one: does it compute the right answer, and how many qubits does it eat? Fewer qubits for the same operation is the whole score - smaller circuits are cheaper to run. The results go back to Claude, and it iterates. Anthropic reports it beat the best published designs across eight primitives, sometimes by an order of magnitude.
CorrespondentVerified by whose evaluator? If the lab that built the agent also built the referee, "beats published designs" is a claim about their harness as much as the model.
HostFair challenge, but these are the checkable kind of checks. A circuit either runs the operation correctly or it doesn't, and qubit counts are just integers. The designs are public, so a rival team can re-run the evaluation tomorrow.
CorrespondentThat's the difference between a leaderboard and a magic trick, so I'll grant it. Now the crypto story, which sounded almost too cinematic when I first read it.
HostThe US standards body, NIST, publishes the elliptic-curve signature algorithm - ECDSA - that helps secure most web traffic on earth. For decades, people assumed the random-looking seed values behind its curves were hashes of some secret passphrase. Anthropic's write-up says an internal Claude wrote a scanner that swept thirty-four generator and slice combinations, and the seeds turned out to be raw outputs of lrand48 - a decades-old Unix random number generator - seeded with Unix timestamps from the late nineties.
CorrespondentSo decades of treasure hunters were digging in the wrong spot, and the model found the actual mechanism. And that result is self-verifying in the strongest sense - you seed lrand48 with the timestamp and the curve either falls out or it doesn't. No evaluator to argue with.
HostExactly. And the same week, Sundar Pichai announced that Google's Willow chip hit a forty-fold error reduction on magic state cultivation - the step that manufactures the special quantum states you need for operations a chip can't self-correct. It's been the bottleneck on the road to fault-tolerant quantum computing.
CorrespondentSo the hardware moved in parallel. And it's not just physics - Google DeepMind has an Antigravity workbench paired with a system called ProtNLM that compressed antimicrobial peptide discovery into a single session.
HostAnd OpenAI is subsidizing academic access to GPT-Rosalind - up to thirty free seats per lab for twelve months. The lab-automation funnel is visibly opening.
CorrespondentSo the repeatable pattern isn't a genius one-shot answer, it's an agent wired into an automated evaluation loop. Which raises the question of who else is inside that loop - and that's exactly where the labor numbers come in.
CorrespondentHere's what would move me off my skepticism about AI hype: payroll data. And this week Anthropic published a research report on how employment is changing at Anthropic, and the numbers are stark.
HostGive me the numbers.
CorrespondentBy August, agent hours outnumbered employee hours six to one. Junior technical hires fell from twenty-seven percent of the mix to nine percent. And revenue per employee runs two to four times what Meta and Google had at similar stages.
HostAnd before anyone says it - that's one company, one data point, and it's the company selling agents.
CorrespondentThat's precisely my objection. The most consequential labor evidence of the week comes from the lab with the most stake in the answer, and that revenue-per-employee punchline is the number I'd most want to see audited. So tell me what would generalize it.
HostTwo things cut in its favor. The data is unusually specific - six to one, twenty-seven down to nine - the kind of numbers you can only refute with equally specific data, not vibes. And the enterprise texture this week comes from outside the lab.
CorrespondentMeaning what, specifically?
HostOpenAI's customer stories. A Benefit Street Partners portfolio manager built his own tracker for SEC filings in roughly ten to twelve hours with Codex, and Sullivan & Cromwell's Agreement Analyzer went from a single proof of concept to a tool a whole legal team uses daily.
CorrespondentThose are adoption stories, though. They show what agents can do; they don't show what happens to the junior rung at a bank or a law firm.
HostNot yet - but there's a companion Anthropic report on the expectations gap that quantifies how differently insiders see this. Ninety-one percent of Anthropic staff versus thirty-nine percent of US adults expect a Nobel-level AI discovery this decade, and staff expect far faster diffusion of capable tasks.
CorrespondentAnd the report itself flags that working at the frontier shapes expectations - which cuts both ways. Either insiders know something the public doesn't, or they're marinating in their own incentives.
HostBoth can be true, and the gap itself is the datum: a fifty-two-point spread between the people building this technology and everyone else living with it.
CorrespondentFiled for later, then, when somebody publishes generalized numbers. Meanwhile, the price of all this capability is falling out from under it.
HostThree releases in one week pushed intelligence-per-dollar down, and I want your eyes on the fine print.
CorrespondentStart with the headline number.
HostZhipu's GLM-5.3-Flash claims a score of fifty-seven on the Artificial Analysis index - the independent leaderboard that scores models per dollar - at four and a half cents per task. Intelligence it says previously cost about ten times more.
CorrespondentFour and a half cents a task is always-on agent fleet money - consumer price points, not enterprise procurement. And if even half of that ten-x claim survives verification, whole product categories get repriced.
HostThe full GLM-5.3 claims open-weights state of the art in coding, but here's the internal tension: the same announcement reports cyber capability scaling faster than expected, with ExploitBench - a benchmark of real-world exploit tasks - jumping from twenty-four point four percent to fifty-four point four. While conceding the model remains behind closed models.
CorrespondentBrag and caveat in the same breath. "Emergent cyber capabilities" is a phrase that should ship with a fire extinguisher, even if they're behind the frontier.
HostDeepSeek's V4.1-Flash claims better benchmarks on shorter runtimes, aimed at what they call wish-style programming - you state the outcome you want and it generates the full project.
CorrespondentWhich is a bet about how people will talk to compilers, not just about raw scores. And Mistral went narrow instead of big.
HostTimestral, four sizes from five million to two hundred nine million parameters, claiming zero-shot state of the art in time-series forecasting at ten to a hundred forty milliseconds of CPU latency. Small, fast, specialized - a different bet than the frontier race.
CorrespondentAnd here's my plain problem with all three: these are self-reported gains with no third-party verification in hand. The independent numbers aren't in the brief, so I'm filing this as directionally plausible, unverified.
HostThat's the honest position. Though three vendors moving the cheap frontier in the same week is itself a signal, even if the exact decimals are marketing.
CorrespondentAgreed, with receipts pending. From prices to principles - because the labs also put real money behind governance this week.
HostAnthropic's newsroom announced a Public Benefit Fund, and the structure is the story. It commits five percent of shares now, plus two percent more for every fifty percent increase in value above baseline.
CorrespondentSo the commitment scales with the upside - the better the company does, the more the fund holds. That's more self-funding than a plain donation.
HostAnd it's structured as a wholly-owned LLC rather than a five-oh-one-c-three - the standard charity structure - which means it isn't limited to classic charity. It can do research and advocacy, that gray zone.
CorrespondentWhich is either a feature or a red flag depending on how much you trust the people holding the shares. An LLC is flexible, and flexible cuts both ways.
HostMeanwhile Sundar Pichai is capitalizing provenance and licensing. SynthID - Google's watermark for AI-made content - was named a TIME Best Invention, with over a hundred eighty billion images watermarked, and Google's News AI pilot grew to more than four hundred publications, including CNN, the Financial Times, and Forbes.
CorrespondentThat's plumbing being laid quietly - watermarking at scale plus paying publishers is the infrastructure for whatever the next content arrangement looks like.
HostOne-line closer on the consumer side: Google also shipped a broad slate including Universal Cart and expanded earthquake alerts.
CorrespondentFrom boardrooms back to keyboards, though, because the most useful thing this week for people actually building was the least flashy document of the bunch.
CorrespondentAnthropic's own engineering guidance for Claude Code still centers on context as the scarce resource, and on closing the loop with runnable checks.
HostWhich is the same pattern from our first chapter - the science came from an agent wired to an evaluator, and the practical advice is build the evaluator and mind the context.
CorrespondentThere's something funny about the lab that cracked a decades-old crypto mystery also telling you the bottleneck is plumbing.
HostBut it tracks. Capability is rented by the token; context is something you own and curate. The teams getting results are the ones with tight context and tests that actually run.
CorrespondentSo for a practitioner, the binding constraint isn't model capability - it's context management. That's the sentence I'd put on a dashboard.
HostBefore the sources, three takeaways in plain language - the things that should survive after the details fade.
CorrespondentOne: invest in the evaluator, not just the agent. Both Anthropic science results and DeepMind's peptide workbench hinge on automated verifiers in a loop, and that's the repeatable discovery pattern.
HostTwo: the entry-level squeeze is measured now, not anecdotal. Twenty-seven percent down to nine, and six-to-one agent hours at a single lab, is a leading indicator for team design and career ladders everywhere.
CorrespondentThree: sub-five-cents-per-task pricing changes product math. Always-on agent fleets at consumer prices are viable as of this week - but context management, not raw capability, is still what limits them.
HostThe loop, the ladder, and the price tag. Now let's give the sources their due.
CorrespondentThe heavy lifting came from Anthropic's research blog. "Designing more efficient quantum circuits with Claude" is the loop-based result that beat published designs across eight primitives.
HostAlso on that blog, the piece on Claude revealing the ECDSA curve origins - the lrand48 finding with Unix timestamps from the late nineties. And two evidence reports: "How employment is changing at Anthropic," home of the six-to-one and twenty-seven-to-nine numbers, and the AI expectations gap reports behind the ninety-one versus thirty-nine split.
CorrespondentAnthropic's newsroom has "Investing in the public benefit," on the Public Benefit Fund structure, and the context-as-constraint guidance lives on their engineering blog in the Claude Code best practices.
HostOn the Google side, Sundar Pichai announced the Willow magic state cultivation result with its forty-fold error reduction, SynthID's TIME Best Invention nod with over a hundred eighty billion images watermarked, and the News AI pilot at four hundred plus publications.
CorrespondentGoogle DeepMind's write-up on accelerating the search for antimicrobial peptides with the Antigravity workbench and ProtNLM is worth reading in full.
HostOpenAI published the program bringing GPT-Rosalind to academic research labs - up to thirty free seats per lab for twelve months - plus the customer stories from Benefit Street Partners on Codex and Sullivan & Cromwell on the Agreement Analyzer.
CorrespondentAnd the price war came from three announcements: Zhipu's GLM-5.3-Flash and the full GLM-5.3 with the ExploitBench numbers, DeepSeek's V4.1-Flash, and Mistral introducing Timestral.
HostThat's the week.
CorrespondentSee you next week.
ndovu