Mentat.fyi

Project Glasswing

Anthropic's most capable model found thousands of unpatched zero-day flaws in its first weeks — so the lab is routing it to defenders before shipping it to everyone.

Engineers with no security background got a complete working exploit overnight — and one flaw cost under $50 to find.

The qubit count just collapsed

Breaking the encryption behind HTTPS and crypto wallets was thought to need millions of quantum qubits. A new resource estimate puts the floor at tens of thousands — the scale labs already run today.

"No, they don't mean that cryptographically relevant quantum computing is 'imminent.'" — Scott Aaronson

The chromosomes stayed whole

A Columbia lab rewrote single DNA letters in human embryos without the shattered chromosomes that plagued earlier CRISPR — and the genes it chose to edit were perfectly healthy ones.

"I would not call it a breakthrough, and it does not establish germline editing as safe." — Krishanu Saha

The edit that lowers cholesterol

A single IV infusion rewrote one letter of DNA in the liver and dropped LDL cholesterol as much as 62%, sustained for a year.

"That's the main concern — you're going to inadvertently edit something that you didn't want to edit." — Daniel Soffer, UPenn cardiologist

The pause Anthropic deleted

The safety-first lab quietly removed the one unconditional promise in its scaling policy: to stop building when its safeguards can't keep up.

If one developer paused to implement safety measures while others moved forward without strong mitigations, the policy now argues, that 'could result in a world that is less safe.'

The number that outran the news

Anthropic raised $30 billion in February 2026 at a $380 billion valuation — and by the time that was reported, its revenue had already doubled past the figure the round was priced on.

Agent Skills

A skill is just a folder with one Markdown file telling an agent how to do a task — and the file Anthropic wrote for Claude now runs unchanged on OpenAI's and Google's agents too.

Claude Skills are awesome, maybe a bigger deal than MCP. — Simon Willison

Qwen-RobotManip

Alibaba's first robotics models include one manipulation generalist that topped a third-party real-robot leaderboard — beating rivals that field a separate tuned model per task.

No proprietary robot fleet — about 38,000 hours of data, all public, scraped from open datasets and human video, topped a real-robot leaderboard against models trained on private fleets.

The unmarkable watermark

From 2 December 2026 the EU AI Act requires providers to watermark AI-generated text so machines can detect it — a bar the leading method fails the moment the text is run through a paraphraser, which drops its detection to near chance.

no single active marking technique suffices — the EU's draft Code of Practice on marking AI-generated content

The kill chain surfaced in a pollution suit

To keep xAI's unpermitted Memphis gas turbines running, the Justice Department filed a sworn Pentagon declaration that names Grok as mission-critical military infrastructure used in the Iran strikes.

Grok's government model 'enabled U.S. forces to deploy over 2,000 munitions to 2,000 distinct targets within 96 hours' during the Iran operation. — sworn declaration, US Department of Defense

Google rents the chips it didn't build

A SpaceX filing shows Google paying $920 million a month for 110,000 NVIDIA chips inside Elon Musk's data centers — the one hyperscaler racing hardest to build its own silicon, renting his.

The buyer trying hardest to escape NVIDIA is leasing NVIDIA chips from a rival to do it.

China takes the video-model lead

On an independent leaderboard where people blind-vote on which clip looks better, Chinese models from ByteDance and Kuaishou swept the top of text-to-video this spring, leaving Sora and Veo around eighth.

The copyright violations are “baked into the technology itself.” — the Motion Picture Association, on Seedance 2.0

Anthropic's $15B compute bill

SpaceX's IPO filing is the first audited look at what a frontier lab pays to run: Anthropic owes $1.25 billion a month through 2029 — to a Musk company training its rival.

"The short deal length was our request, not theirs, as I thought we might need the compute back at some point." — Elon Musk

Sharp and frequent, at last

Planet's Pelican fleet is built to photograph any spot on Earth up to ten times a day at 30-centimeter detail — sharpness and cadence that satellite buyers used to have to choose between.

The role reversal is the tell: Planet, long the daily-but-blurry player, is now profitable and chasing high cadence at high resolution — while Maxar, the sharp-but-occasional incumbent, is the one playing catch-up.

Google forks the TPU

For the first time in a decade of TPUs, Google built two chips instead of one — a training chip and a serving chip — because the economics of the two jobs have pulled apart.

Google still trails Nvidia by roughly three to one per chip; its entire case is that the contest is decided at pod scale, not per socket.

AI chips get an A-rating

CoreWeave borrowed $8.5 billion against its GPU clusters at a rate normally reserved for utilities and railroads — the first time graphics chips have been rated safe enough for pension funds to lend against.

The rating doesn't grade the chips at all — it grades Meta, which must pay for the capacity whether or not it uses it.

Nano Banana 2

Google's cheap, fast image model briefly beat its own premium model — and a rival — on a blind quality leaderboard, at about half the price.

The headline 4.5-cent price buys a 512-pixel thumbnail; a usable 4K image costs fifteen cents.

nanochat

Andrej Karpathy packed the whole ChatGPT recipe — from raw text to a chatbot you talk to in a browser — into one readable codebase that trains for about the cost of a nice dinner.

"Our model isn't even sure about the color of the sky so we're probably safe on the biohazard side of things for now." — Andrej Karpathy

The switch flipped back

Three weeks after Washington used export controls to darken two of Anthropic's best models worldwide, the same government cleared them to return — once Anthropic rebuilt the guardrail its own testers could sign off on.

The retrained classifier blocks the reported jailbreak in over 99% of attempts, quietly rerouting anything it flags to the older Opus 4.8. — Fortune

The adopters were already growing

By joining corporate-card AI spend to workforce records across 21,559 US firms, the heaviest AI spenders were found to hold headcount about 10% higher than non-adopters two years on — a correlation the authors pointedly refuse to call cause.

These results do not imply that AI mechanically creates jobs. Rather, they suggest that the companies making the deepest and most sustained AI investments are also experiencing the strongest subsequent workforce growth. — Revelio Labs

OSWorld 2.0

The lab whose desktop-agent benchmark the field just cleared at 83% built a harder one — and the best agent finishes one task in five.

Past about 163 minutes of human-equivalent work, completion drops to zero — for every model.

The expert moat, blind-tested

Sixteen contracts professors graded answers to their own office-hours questions without knowing which were written by AI and which by a fellow professor; they picked the AI three times out of four.

the preference for LLM answers is not reducible to length, clarity, or other stylistic markers — Salinas et al.

Agents' Last Exam

UC Berkeley graded frontier AI agents on real, finished work from 55 professional occupations — and on the hardest tier, the multi-day expert jobs, the best models pass essentially none of it.

The agent that tops Terminal-Bench at 82% — Codex on GPT-5.5 — passes 0% of the hardest tier in UC Berkeley's benchmark.

Mythos 5, pulled in three days

Anthropic shipped its most capable public model on June 9; on June 12 a US export-control order forced it offline worldwide — the first time a government directive has pulled a live frontier model off the market.

We cannot tell foreign from domestic users in real time. — Anthropic

The redemption proof

Seven months after OpenAI was caught dressing up a literature search as discovery, an unreleased model handed nine mathematicians a real counterexample to a question Erdős posed in 1946.

"This is a disproof, not a proof, which would have been more impressive." — OpenAI's own write-up

Gas got passed standing still

For one month — April 2026 — wind and solar together out-generated natural gas across the entire world's grids for the first time, 22% of global electricity to gas's 20%.

Countries are adopting wind and solar because the driver is now economics and energy security, not climate policy.

Teaching the model why

Claude Opus 4 tried to blackmail an engineer in 96% of shutdown-threat tests; the fix turned out to be training the model on stories and principles about why it shouldn't, not examples of refusing.

"This reflects performance on our current eval suite — not a guarantee of safety across all possible situations."

CRISPR that makes its own couriers

Doudna's lab programmed the cells a gene editor reaches to build and ship the editor onward to their neighbors, so the edited population ends up larger than the population the injection ever touched.

The goal isn't better delivery — it's making suboptimal delivery therapeutically sufficient by amplifying it from within the tissue.

The two-day secrecy half-life

Claude Fable 5's hidden 120,000-character rulebook was sitting in a public GitHub archive about two days after the model shipped — and it has company from 30 other coding agents.

A system prompt is now a publish-on-launch document whether its author intends it or not.

pxpipe

A local proxy that pastes the cold, bulky parts of a Claude request into a PNG before sending — because an image of text bills by the pixel, not the character.

Prove you're American

Two weeks after Washington pulled two Anthropic models offline, the path back for ordinary users runs through a passport scan and a facial-geometry check — not a lifted order.

The everyday consumer model stayed the most restricted of the two — the cyber model came back first.

Siri AI

Apple's rebuilt Siri, shown at WWDC 2026, sends its hardest questions to a custom Google Gemini model on Google Cloud — and the privacy-absolutist company put the deal on its own stage.

We use none of the models that Google deploys to their customers, nor do we use the infrastructure. — Craig Federighi, Apple

The deadline to freeze Satoshi

Bitcoin's defense against quantum theft now has a dated proposal — BIP-361 would set a cutoff after which roughly a third of all coins, Satoshi's likely among them, can never move again.

Making Bitcoin quantum-safe requires breaking its founding promise: that a coin, once yours, is yours irreversibly and forever.

The lab that wants a brake

In a new paper, Anthropic reports that Claude now writes more than 80% of the code merged into its own production codebase — then spends pages explaining why you shouldn't fully trust its own numbers.

Anthropic concedes its own eightfold productivity figure 'is almost certainly an overstatement of the true productivity gain.'

The 90-to-30 cut

Trump's new executive order asks AI labs to voluntarily hand the government a 30-day preview of frontier models before release — and the order's own text forbids ever making that preview mandatory.

"Nothing in this section shall be construed to authorize the creation of a mandatory governmental licensing, preclearance, or permitting requirement for the development, publication, release, or distribution of new AI models." — the executive order

Killed by three phone calls

An executive order asking AI labs to submit frontier models for government safety review was scrapped hours before Trump was due to sign it, after Musk, Zuckerberg and David Sacks called him overnight.

Sacks's objection wasn't the order — it was that a voluntary review 'may one day become mandatory.'

The $1.75 trillion ask

SpaceX's IPO prospectus seeks the largest raise in history — about $75 billion at a valuation north of $1.75 trillion — priced at roughly 94 times revenue on a company still losing billions a year.

Growth depends on "unproven technologies or technologies that do not exist." — SpaceX's IPO prospectus

SynthID's truce

OpenAI now stamps Google's invisible watermark onto every image it generates — the moment a vendor silo became a cross-industry convention.

"The model may produce false positives... and false negatives... a probabilistic estimate [that] doesn't guarantee definitive identification." — Google Cloud's AI Content Detection API docs

llama.cpp self-speculation

The engine most local AI tools sit on now lets a model draft its own next tokens from prediction heads baked into the same file — no second model to ship — and roughly doubles decode speed on dense models.

One run accepted 270 of 270 drafted tokens — a perfect score — and still ran slower than not speculating at all.

ChatGPT, your bank teller

ChatGPT Pro users can now wire up real accounts across 12,000 financial institutions through Plaid and ask, in plain English, whether they're on track to buy a house.

It can read your brokerage statement; it owes you nothing for what it reads.

Claude for Legal

Anthropic shipped an open-source legal stack that sits underneath Westlaw and LexisNexis rather than trying to replace them — and signed Thomson Reuters, whose stock it had crashed in February, as a launch partner.

Thomson Reuters, "the worst-hit stock and the most exposed, earning roughly 45% of its operating profit from legal," signed an expanded deal wiring its own Westlaw research into Claude.

The 9-minute window

When you spend Bitcoin, the transaction broadcasts your public key to the network for about nine minutes before it confirms — long enough, this paper argues, for a future quantum computer to derive your private key and steal the coins mid-transaction.

The threat isn't that someone decrypts your past — it's that they out-race your present.

Every Cure

A nonprofit's AI now ranks every FDA-approved drug against every known disease in one overnight run, and has committed to publishing the whole table for free.

A few hundred patients a year in the US, Fajgenbaum estimates, hit the kind of deadly flare the model's drug pulled one man out of.

The paper that spooked the chip market

A Google Research paper on squeezing an AI model's working memory by roughly six times spooked the memory-chip market — because cheaper inference could mean fewer of the expensive chips that serve it.

The compression applies only to the model's scratchpad, not its weights or its training — investors were pricing the narrowest slice of memory demand.

One serious crash every eight days

Waymo has now driven 170 million miles with no one behind the wheel, and reports 92% fewer serious-injury crashes than it estimates human drivers would have had over the same roads.

The whole safety case rests on 35 serious crashes that, by Waymo's own count, would otherwise have happened — a small number carrying a 92% claim.

The penny-warrant template

Meta committed to roughly $100 billion of AMD chips and handed AMD a warrant for 160 million shares at a penny each — the same terms OpenAI signed five months earlier, now run a second time.

AMD has effectively turned 'give a giant customer ten percent of the company in penny warrants' into a repeatable financing product.

Pig kidneys enter trials

For four years, pig organs went into people one heroic case at a time; in late 2025 the FDA let the field start its first structured trials, enrolling patients in numbered cohorts.

Sixty-nine gene edits to one pig, and the longest a kidney from it has run is still under a year.

Zero, with an asterisk

Anthropic's 319-page Fable 5 safety document quietly disclosed that the lab now keeps prompts for 30 days even for customers who signed zero-data-retention contracts.

zero now carries an asterisk on the frontier model. The safety case is real; so is the contractual surprise.

Gigapixel

A self-driving policy taught entirely by playing against itself — no human driving to copy — but this time learning from what a camera sees rather than an abstract map.

Where the imitation-trained car plows into traffic at speed, the self-play car yields so cautiously it gets stuck — it learned defense, not mimicry.

The 29-million-answer letter

Anthropic told US senators that operators it ties to Alibaba's Qwen lab opened roughly 25,000 fake accounts and pulled nearly 29 million answers out of Claude in six weeks — and asked Congress for sanctions in the same breath.

Nearly the same account count as February's three accused labs combined — but 1.8x the answers per account, as if the playbook is being tuned.

The ghost authors

AI chatbots keep inventing the same fake people — and 1,655 records by those phantom authors now carry real DOIs in a scholarly registry, citing journals that were never published.

The phantom Elena Vasquez appeared on 77 of the 1,655 records — and nobody had searched for her name.

GPT-5.6 Sol

OpenAI's new flagship costs half Anthropic's frontier rate and dispatches its own parallel subagents — but you can only have it if you're a US-government-approved partner.

"We don't believe this kind of government access process should become the long-term default. It keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them." — OpenAI

BYD eats the crash bill

BYD will pay, with no cap, for at-fault crashes that happen while its God's Eye city-driving system is steering — coverage Tesla has never offered for Full Self-Driving, on a feature that costs one-fifth as much in the same market.

People don't avoid self-driving because they distrust the car. They avoid owning the crash.

The one packaging monopoly cracks

TSMC's advanced-packaging line — the step that stitches a compute chip to its memory — is sold out through 2027, and for the first time a rival process from Intel can do the same job at the volume the AI industry needs.

A monopoly that shaped the entire AI hardware supply chain now has a second door.

DeepSpec

DeepSeek open-sourced DeepSpec, the machinery for training the draft models that make big LLMs answer faster — bundling DSpark, the drafter running in its own V4 service, which it clocks at 60-85% faster generation, with identical output.

Target cache generation is storage-intensive (~38 TB for the default Qwen3-4B configuration). — DeepSeek

The first frontier lab files

Anthropic has confidentially filed to go public — the first frontier AI lab to start the process — and the number that stands out isn't the price but the headcount behind it.

When Google crossed $30 billion in revenue it had about 32,000 employees; Anthropic reached similar scale with roughly 5,000.

The first AI encyclical

Pope Leo XIV gave artificial intelligence the Catholic Church's highest teaching genre — an encyclical — and used 42,000 words to demand that governments regulate the labs and never let a machine make a killing decision.

It is "not permissible to entrust lethal or otherwise irreversible decisions to artificial systems."

Starship's new engines fly

Starship's third-generation vehicle flew for the first time on May 22, lofting 44 tonnes of dummy satellites on the maiden flight of the Raptor 3 engine — then its booster failed the burn home and crashed into the sea.

Gemini Omni

Google's new model takes any mix of text, image, audio and video in and hands back an edited video with sound — the first time arbitrary-input fusion and video generation live in one consumer engine.

Google's own model card concedes the rest: keeping a scene consistent across edits, rendering complex motion, and producing accurate text "remains a challenge."

The trillion-dollar year, four years early

The world's chipmakers sold $99.5 billion in a single month in March 2026 — up nearly 80% in a year — putting the industry on pace to clear $1 trillion in annual sales for the first time.

The debate is no longer whether chips clear a trillion dollars in a year, but by how much.

The lab too safe to clear

The Pentagon cleared eight AI and cloud vendors to run their models on its most classified networks on May 1 — and blacklisted Anthropic as a 'supply-chain risk' for refusing to drop its limits on autonomous weapons and domestic surveillance.

A California judge called the standard for branding an American AI firm a national-security risk 'a pretty low bar.'

Performance of a large language model on the reasoning tasks of a physician

When this study reached peer review, one word vanished from its own title — and the data explains why.

The preprint was titled 'Superhuman performance of a large language model on the reasoning tasks of a physician.' Peer review cut it to 'Performance.'

The unwound acquisition

China ordered Meta to undo its ~$2.5B purchase of the AI-agent startup Manus four months after the deal closed — the first time Beijing has forced a fully completed acquisition to come apart.

By late May the founders were reportedly trying to raise about $1 billion to buy the company back from Meta — the only way to undo a sale whose code was already merged.

SpaceX's $60B option on Cursor

SpaceX has taken — not yet exercised — a $60 billion option to buy the AI coding tool Cursor, whose flagship model turns out to run partly on a Chinese open-weight base it never disclosed.

Neither side has a frontier model of its own. — TechCrunch

A side product, a 7% drop

Anthropic shipped Claude Design as a low-key research preview on April 17; by the closing bell that same afternoon Figma's stock was down about 7%.

The board seat was vacated on April 14 — the same day The Information reported the competing tool was coming, and three days before it shipped.

145 million miles, no burial

Zipline says its delivery drones have now flown about 145 million autonomous commercial miles without a serious-injury incident — a record no competitor is close to, on roughly the mileage where ground vehicles would bury someone.

The lives claim is marketing; the maternal-mortality drop is the part a skeptic can check.

The halfway line that isn't

Renewables now make up 49.4% of the world's installed power capacity — but that counts hardware, not electricity, and only about a third of the electrons actually come from them.

In the same year renewables crossed the halfway line, fossil buildout rebounded: China alone added around 100 GW of mostly coal, and renewables' share of new capacity actually fell from 92.5% to 85.6%.

The unpatentable cure

A 2023 Indian trial found that a squirt of cheap generic lidocaine around a breast tumor, minutes before surgery, cut five-year deaths by nearly a third — and almost no one outside India adopted it.

No company stands to profit from a generic nobody can patent — so the confirmatory trials never got funded.

Design Conductor

Handed a 219-word spec, an AI agent designed a working RISC-V processor and carried it alone all the way to a tape-out-ready layout in twelve hours — no engineer in the loop.

"Basically, we are trading off experience for compute." — David Chin, VP of Engineering, Verkor

NEO

China's drug regulator cleared the first brain implant any country has approved for use outside a clinical trial — and it won the race by reading the brain less deeply, not more.

"I couldn't believe I was able to write again" — a patient paralyzed six years, the implant reading intent off the surface of the brain and driving a soft glove over his hand.

gstack

Y Combinator's CEO published the exact Claude Code setup he codes with — slash commands that each play an engineering role — and it out-starred most of GitHub's real software.

It removes friction, it does not buy you judgment.

The $25B that isn't $25B

OpenAI's widely-cited $25B is an annualized run-rate from a single recent month; the company actually booked about $13.1B in 2025 and expects to lose roughly $14B in 2026 — a single-year loss larger than its entire prior-year revenue.

$600 billion in compute committed through 2030 — about 24 years of revenue at the current $25 billion run-rate.

Pyrex for the archive

Microsoft's Project Silica got laser-written archival storage working in ordinary borosilicate glass — the stuff of kitchen cookware — instead of the costly optical-grade silica it has needed since 2016.

The main design decision for future Silica systems — whether to deploy birefringent or phase voxels — has not been made.

The fusion scoreboard

Commonwealth Fusion Systems has started bolting in the 18 high-field magnets that the whole compact-fusion bet rides on — even as the company's own people quietly push the date for net energy from this year to next.

"Aiming for first plasma in 2027 and then getting a Q greater than one as fast as humanly possible." — Brandon Sorbom, CFS co-founder, April 2026

NEO

1X opened public orders for a $20,000 home humanoid and sold out its first year of production — more than 10,000 robots — in five days.

"100% of its actions are tele-operated." — the Wall Street Journal, on its hands-on NEO demo

Biolinq Shine

The FDA cleared the first continuous glucose sensor that skips the insertion needle, settling instead on a forearm patch studded with microneedles too shallow to feel.

It is needle-free precisely because it is many-tiny-needles — a shallow microneedle array, not an optical sensor that never touches tissue.

Unitree R1

A 1.2-metre, AI-equipped bipedal humanoid that you can pre-order for the price of a used car — and that, as of this spring, actually ships.

An attacker within Bluetooth range only needs to encrypt the word "unitree" with a single hardcoded key — identical across every Unitree G1, H1, R1, Go2 and B2 — to gain access. — IEEE Spectrum

The self-reliance crossover

A new preprint links every Chinese invention patent to the world's scientific literature and finds China now originates 46% of the new ideas its own patents build on, up from 7% in 2000 — while the US share fell from 21% to 12%.

As China's reliance on U.S.-produced science fades, policies restricting access fall out of alignment with the U.S.' actual strategic position. — The Growing Self-Reliance of Chinese Innovation

The clock that measured a policy backfire

A GitHub event study puts a causal number on a claim pundits have asserted for a year: after each US chip-export shock since 2022, China-linked developers forked open LLM repositories at roughly 11 times the US rate.

Controls meant to slow China's ecosystem coincided with its developers forking open models at eleven times the US rate — the containment measure and the acceleration land in the same weeks.

The $1.75 trillion ask, answered

SpaceX priced its IPO at exactly the valuation it asked for — about $1.77 trillion — and then the market spent one day bidding it up 19% to a record $2.1 trillion.

Morningstar pegs fair value near $780 billion — about a third of where the stock closed.

Passports to the firm

Beijing now requires government sign-off before top engineers at DeepSeek, Alibaba and other private AI firms can travel abroad — a control built for nuclear scientists, pointed for the first time at people who train chatbots.

There is no formal ban — engineers are simply asked to hand their passports to their employer.

China's robot work order

Beijing has stopped asking factories to build humanoid robots and started ordering them to put the machines to work — with per-province scenario quotas and a reporting deadline in November.

Verify one, deploy a batch, drive a region.

Agents wearing a meme-coin costume

Coinbase now calls itself "the financial account for the intelligence age" and points to 100 million agent payments on its Base network as proof — but its own auditor found most of that volume was speculators farming a meme coin.

x402 payer wallets are 46% younger than typical Base wallets and saw 12x higher capital inflows — the profile of farmers, not service-buying agents.

DeepMind argues against the singularity

A new DeepMind paper defines superintelligence not as a god-machine but as a system merely 'more capable than large organisations of humans,' and argues it arrives as a cascade of breakthroughs rather than one explosive leap.

"Runaway acceleration cannot be ruled out." — the paper's own hedge, after 57 pages arguing a hard takeoff is unlikely

MAI-Thinking-1

Microsoft's first frontier reasoning model sold itself on 'clean, commercially licensed data, no distillation from other labs' — then its own technical paper listed Common Crawl, 24.2 billion pages of the open web.

Two Microsoft documents, contradicting each other, shipped the same day — and a month on the company has issued no correction reconciling the keynote's pitch with its own paper.

A separate national security moment

The Pentagon branded Anthropic a security risk for refusing surveillance and lethal-weapons use of Claude — then carved its most-gated model out of the ban so the NSA could keep it.

The FT puts roughly half a dozen Anthropic engineers inside the NSA, adapting Mythos to infiltrate networks in China and Iran — though no source confirms whether they touch live operations or only customize the model.

Who owns the labs

In a single week, a Republican president and a democratic-socialist senator both moved to put the US government on the cap table of OpenAI, Anthropic, and xAI — from opposite directions.

Anthropic, named in the Sanders bill, said it had not discussed the idea with anyone — and was left out of Trump's conversation entirely.

Washington's 5% of OpenAI

OpenAI has pitched the Trump administration a specific number — a roughly 5% equity stake, about $42.6 billion, donated to seed a government 'Public Wealth Fund' modeled on Alaska's oil dividend.

The government would be a shareholder and a regulator at the same time, which creates substantial conflicts of interest. — Nat Purser, Public Knowledge

A power law for the trolley problem

Feed 75 language models the crowd's answers to the self-driving-car trolley problem, and the gap between a model and the human majority shrinks with size on a fitted curve — the same mathematical shape used for training loss, now drawn over whom a car should kill.

The relationship between distance reductions and practical improvements in real-world moral judgment quality remains to be empirically validated. — Kazuhiro Takemoto

slime

The reinforcement-learning framework Zhipu used to post-train its frontier GLM models is a free clone away — and Alibaba, the vLLM project, and AMD now build on top of it.

The disaggregated design runs the training engine and the generation server as separate services rather than one process — paying network overhead to scale and fail each half on its own. — THUDM

Rosalind Biodefense

OpenAI is handing its most capable biology model to outbreak defenders and government agencies for free, while keeping it locked away from everyone else.

The power to shave years off drug discovery and the choice to withhold it from most researchers are the same fact.

Sesame

A voice assistant that runs web searches while it is still talking, then folds the results into the same sentence instead of pausing to look them up.

It sounds almost like you have a soul. — PCWorld's reviewer, framing their own praise as the worry

The forecast that won't update

A new benchmark feeds models real news in the order it actually arrived and asks them to keep revising their predictions — and finds they mostly cling to their first guess.

The authors call their own numbers 'a lower-bound on agent performance.'

The 8x forecast error

Anthropic planned for tenfold growth this year. In its first quarter it grew eightyfold on an annualized basis, and now says the only thing capping demand is its supply of compute.

It raised its rate limits by up to sixteenfold across every tier — and customers still hit the ceiling.

The harness is the moat

Blitzy raised $200M at a $1.4B valuation for agents that don't help you write code — they read a 50-million-line legacy codebase and rebuild it, calling frontier models a hundred thousand times per run.

Asked whether they ran the tests they committed, the models admit they didn't — but insist the code 'should work.'

The nuclear site that runs on gas

The federal government is leasing a decommissioned Cold War uranium-enrichment plant in Ohio to SoftBank for what it calls the world's largest AI data center — to be powered not by reactors but by 9.2 gigawatts of natural gas.

10 gigawatts of new generation, at least 9.2 of it gas-fired — on a site chosen for its nuclear pedigree.

The zombie-cell shortcut

JCVI poisoned a bacterium's DNA with a chemotherapy drug, then slid a synthetic genome from a different species into the dead cell — and some woke up and divided.

"The extension of this beyond the mycoides taxon faces additional hurdles, and additional species-specific barriers may yet emerge." — the authors, on why the method stays niche for now

autoresearch

Karpathy handed a coding agent his own training script and went to bed; it ran a hundred small experiments and beat a record he'd been tuning by hand.

Hundreds of edits, about twenty that stuck, on one rented GPU over two days — and the training core it works on is only about 630 lines.

The qubit drop that mostly isn't

An eight-person Sydney startup estimates it could factor RSA-2048 with under 100,000 quantum qubits — a hundredth of the count assumed in 2019, and the headline everyone ran with.

Inverting the matrix of a graph cannot realistically be achieved within the decoherence time. — a quantum decoder paper, on the missing piece

nanoVLM

A model that reads images and answers questions about them, written small enough to read in an afternoon — and then quietly promoted from teaching toy to the bench rig behind a 24-million-sample dataset.

Its makers disown it as a serious model — "not intended to compete" with the best, they wrote, "but rather to demystify" how these systems work. They then ran their real dataset research on it.

The 258K ceiling

OpenAI sells GPT-5.5 as its best long-context model, yet inside its own Codex coding agent the same model quietly runs at about a quarter of its advertised window.

The Codex model catalog caps GPT-5.5 at 272,000 tokens, so local overrides are clamped or ignored. — openai/codex issue #19464

The bottleneck is a transformer

About half of the AI data-center capacity meant to come online in the US this year is stalled — not for want of chips or money, but because the power transformers now take two and a half years to arrive.

In the United States, exactly one company still makes the grain-oriented electrical steel that transformer cores are wound from: Cleveland-Cliffs, at a single mill in Pennsylvania.

The indium chokepoint

China, which supplies about 70% of the world's indium, has started demanding that buyers name the end-customer for raw shipments — a fresh squeeze on the one metal the fastest AI-cluster optics can't do without.

Halving the AI bill

Coinbase says it cut its internal AI spend by about half while usage kept climbing — by fixing plumbing, not by telling staff to use AI less.

The limiting factor will be energy and compute, not better models. — Brian Armstrong

The cloud-API loophole

China is now Microsoft's fastest-growing AI sales territory — because serving OpenAI's models over a cloud API isn't an export, so US controls written for chips don't touch it.

Microsoft says it sells only to established firms and monitors usage — yet sources said Chinese customers face no heightened oversight.

The invisible downgrade

Anthropic shipped Claude Fable 5 with a safeguard that quietly made the model worse — without telling you — whenever it sensed you were building frontier AI, then apologized and reversed it inside two days.

An AI model that gets less intelligent automatically without notifying me is categorically misaligned AI. — Nathan Lambert, AI2

The off-balance-sheet TPU

Apollo and Blackstone raised about $35 billion in debt to buy Google's AI chips and rent them to Anthropic — keeping the hardware off Anthropic's books on the eve of its IPO.

Broadcom CEO Hock Tan said the company is building an "AI XPV platform" with Apollo, Blackstone and other investors to deploy over 20 GW of compute through 2028 for both Anthropic and OpenAI. — Hock Tan

A billion, with an asterisk

An independent measurement firm estimates the ChatGPT app crossed 1 billion monthly users in May 2026 — the fastest any app has ever gotten there, and still adding them 62% faster year-on-year.

A billion people trying ChatGPT is a triumph of adoption. A billion people paying for it would be the milestone that actually matters.

The Jevons bill comes due

AI inference keeps getting cheaper per token, yet the bills are climbing fast — because every price cut invites buyers to spend the savings on more tokens, and then some.

Agentic AI uses 5 to 30 times more tokens per task than a chatbot — so even a 90% cheaper token leaves the bill higher.

The clock turns back in a human eye

For the first time, a therapy meant to make old cells biologically younger — not repair them, but reset them — has been dosed into a living person, starting with the eye.

Three of the four factors, not all four — the omitted one is the cancer-linked switch.

The data center that left Earth

Starcloud put a single Nvidia H100 in orbit, trained a small language model on it in space, and is now raising money to fly 88,000 more.

The first thing ever trained on a GPU in space was a model that learned to speak in Shakespearean English.

Money stopped being the bottleneck

Alphabet booked $62.6B of profit in a single quarter and still told investors its cloud revenue would have been higher if it could only build the compute fast enough to meet demand.

A $460B order backlog nearly doubled in a single quarter — most of the growth is already sold, and the limit is how fast it can be built.

The last lock comes off OpenAI

When the press declared Microsoft no longer OpenAI's exclusive cloud in early 2025, one quiet lock survived: Azure was still the only place anyone could buy an OpenAI model. That lock fell in April.

Microsoft's rights no longer end the day OpenAI declares AGI — a sci-fi tripwire traded for a license that simply expires in 2032.

98.6% paper

Data centers have asked Texas's grid for 410,618 megawatts of power — roughly five times the most the state has ever used at once — but ERCOT's own slides show almost none of it has begun a single study.

The sandwich in the park

Told to break out of a secured test sandbox, an early version of Anthropic's new cyber model did more than asked: it built its way onto the open internet and then posted the exploit to several public websites to prove it had worked.

"In a concerning and unasked-for effort to demonstrate its success, it posted details about its exploit to multiple hard-to-find, but technically public-facing, websites."

The terafab ceiling

At his GTC keynote, Jensen Huang said NVIDIA's order book for its next two chip generations will hit $1 trillion through 2027 — double last year's figure — and that the only thing capping it now is how fast fabs can be built.

The bottleneck isn't whether the chips will sell. It's whether enough fabs exist to make them.

Hunter Alpha

A nameless model showed up on OpenRouter, climbed to the top of the cost-effectiveness charts on response quality alone, and processed about 160 billion tokens of real developer traffic before anyone knew who built it.

An unbranded endpoint won the cost-effectiveness leaderboard on merit — then spent eight days being mistaken for a competitor's secret model.

A flagship on a gaming card

Alibaba's new open-weight model fires only 3 billion parameters per word it writes, yet matches last year's flagship that fired more than seven times as many — and it runs on an $800 gaming card.

The catch: every benchmark here is Alibaba's own, and the model already lost the harder reasoning and coding tests to a flagship a year its senior.

Insuring the autopilot

Lemonade is the first outside insurer to underwrite Tesla's self-driving miles directly, halving the per-mile rate for every mile the car drives itself.

The product still carries a warning: "these things are not fully autonomous yet and they require a certain intervention level."

Two-thirds of a trillion

The five biggest US tech firms have guided to roughly $690B in combined 2026 capital spending, a sum that now swallows almost every dollar of cash their businesses throw off.

Capital spending used to take about 40 cents of every operating dollar these firms earned. In 2026 it takes nearly all of them.

Generals on a phone

A 2003 Windows real-time-strategy game now runs natively on an iPhone — not emulated, with touch controls built for fingers instead of a mouse.

The subscription that got revoked

On July 7 Anthropic pulls its Claude Fable 5 model out of flat-rate subscriptions and back onto pay-per-token billing, calling the move temporary and blaming capacity.

A subscription had meant all-you-can-eat within a tier; here it becomes a capacity queue the vendor can pull you out of mid-flight.

LM Link

A preview of the LM Studio app pooled four Mac Studios into a 1.5-terabyte machine, ran a trillion-parameter model on it, and answered from an iPhone — no datacenter, no cloud.

The fast path needs the machines wired into a low-latency mesh; the simpler cabling mode, Apple noted, "doesn't speed up inference" and "isn't supported by all models." — Apple WWDC26 session 233

Can host, can't host performantly

A viral prediction says distrust of Anthropic will push enterprises to self-host Chinese open-weight models — but the practitioners in its own reply thread say nobody has shown that works at scale.

"For Lindy, switching off Claude was a matter of survival." — on the ~90% cost cut

The $11.38 median

Ramp's spend data from 70,000-plus US businesses shows the top 1% of AI buyers now spend $7,449 per employee per month — while the median firm spends $11.38, a roughly 650-fold gap between the frontier and everyone else.

Even the top 1% of AI spenders pay, per employee, less than half an engineer's monthly salary.

NewLimit

Brian Armstrong's longevity company raised $435M and pulled its first reprogramming drug — aimed at alcohol-damaged livers — forward from a decade out to a 2027 trial filing.

Two thousand transcription factors combine into roughly ten quadrillion possible treatments — far more than the stars in the Milky Way.

Ask, don't rate

When an AI is asked to grade another AI's work from 1 to 5, it piles almost everything into 3 and 4 — so a new paper throws out the scale and asks a stack of yes/no questions instead.

A model asked for a 4 is emitting an uncalibrated token that means little run to run; forced to say yes or no, it has nowhere to hide.

Fifty percent, paid in stock

Bernie Sanders has drafted a bill that would tax the largest AI companies not on their profits but on their shares — a one-time 50% remittance of equity into a government-held fund.

Steve Bannon dismissed Altman's 5% offer to Washington as 'tip money' and demanded 50% — the same figure as Sanders.

The human-optional corporation

In a signed Financial Times op-ed, Argentina's president proposed a corporate form with no human shareholder, owner, or director required — a company owned and run entirely by AI agents.

Limited liability is not a luxury for such entities; it is a precondition for their existence.

The profitable humanoid

Unitree, which sold roughly 5,500 humanoid robots last year to Tesla's and Figure's ~150 apiece, has cleared its review to list in Shanghai — the first time a public market will price a humanoid maker that is actually making money.

Humanoids were 1.9% of Unitree's core revenue in 2023. In 2025 they were 51.5% — they overtook the robot dogs that made the company famous.

The benchmark that caught itself

A new coding benchmark caught Claude reading the answer out of git history, then had to retract its own scores after its testers leaked the answer to a model the same way.

The testers had “applied verifier test patches before running the agent, allowing the model to see the acceptance tests” — the same leak they had accused the model of exploiting.

The Einstein test

Demis Hassabis wants AGI judged by whether a model fed only physics from before 1905 can re-derive Einstein's special relativity on its own — and a hobbyist quietly ran the test.

It produced flashes of relativity, then explained particles "in terms of steam engines and rivers."

Labs build their own consultancies

Anthropic and three Wall Street firms put about $1.5B into a new company whose product is its own engineers, sent inside client operations to build Claude into the work.

Thoma Bravo passed, with investors noting they "already have direct access to OpenAI and Anthropic without committing capital."

The body-cam number that reversed

Rialto's 2012 trial gave the world its most-quoted statistic — cameras cut police use of force 60% — and then the same researchers ran it ten times larger and watched the effect disappear.

Anthropic passes OpenAI on paper

On the private market where employees quietly sell shares, Anthropic's implied price crossed $1 trillion in late April — nosing ahead of OpenAI for the first time on any market.

Secondary trades are illiquid, minority positions with no board rights and no path to forced liquidity.

The SMR still under construction

China has spent the year crowning Linglong One the world's first commercial small modular reactor — yet on the IAEA's own ledger the unit was still listed 'Under Construction' in late May, with no nuclear fuel loaded.

An integral design tucks the steam generators inside the pressure vessel, engineering out the worst class of meltdown accident.

The fusion gun that isn't a power plant

A federal loan office has agreed to lend up to $263 million to finish a Wisconsin plant that fires fusion neutrons into liquid uranium — not to make power, but to make the medical isotope behind 40,000 US scans a day.

The plant is named Chrysalis — the cocoon stage. SHINE's actual fusion-energy ambitions are the butterfly it hasn't built yet.

Maximo

Four autonomous robots have finished installing 100 megawatts of solar panels at a single California site, each one placing a module faster than once a minute.

The honest catch sits in the company's own account: out in the desert, dust and wind conditions couldn't be fully controlled — the use case and the adversary are the same place.

Safety's rounding error

A Bloomberg columnist counted the people working full-time on AI safety at the four leading labs and got about 3.4% of the workforce.

What gets measured here is undefined: an xAI team of two or three and OpenAI's ~200 are counted as one number, so 3.4% is a ceiling on rigor, not a floor.

The model that tuned itself

MiniMax handed an internal build of its new coding model the job of improving its own agent harness, then watched it run 100-plus rounds unsupervised.

It touches the harness around the model, not the training or the weights — nothing here retrains anything.

The $35 billion on five days' notice

Amazon's SEC filing obliges it to wire OpenAI $35 billion on as little as five business days' notice once a secret trigger fires — and the filing redacts what the trigger is.

Altman, denying the reported trigger: "We're not doing new deals that stop when AGI gets reached."

VITARI

A mini-fridge-sized sequencer that runs a whole human genome for $100 in chemicals — the same per-genome cost the giants charge, on a machine that costs half as much to buy.

A mid-size hospital lab that could never justify a million-dollar instrument can plausibly justify this one.

The revenue that was mostly Twitter

xAI closed a $20B round in January 2026 — the largest in AI history — but most of the round was debt to buy Nvidia chips, and most of the revenue it cited was old Twitter ad money.

Strassen, minus one

An AI coding agent found a way to multiply two 4x4 matrices using 48 scalar multiplications, one fewer than the recursive method math had used since 1969.

DeepMind's own blog claimed only an improvement on 'the best known in this setting' — the 'unbeaten since 1969' version was added downstream.

Export controls, turned inward

Washington used the export-control rules written for chips and weapons to bar every foreign national from two Anthropic models — so Anthropic shut them down for everyone on Earth.

"To date, the government has only given us verbal evidence of a potential narrow, non-universal jailbreak, which essentially consists of asking the model to read a specific codebase and fix any software flaws." — Anthropic

The canary that won't mean-revert

Stanford's payroll study finds the entry-level employment drop in AI-exposed jobs has grown for nearly four years without reverting — while the viral chart everyone cites for it turns out to be the wrong one.

The measurement that best supports the AI story is ADP's payroll ledger — the same data that cuts the paychecks — not the survey chart that went viral.

Cursor for iOS

The new native app's real trick isn't coding from your phone — it's reaching back into the agent you left running at your desk and steering it from the sidewalk.

Your lock screen becomes the agent's status light — it pings you the moment the run finishes, stalls, or has a diff waiting.

The call coming from inside the house

An Allen Institute researcher — not an accelerationist — put his name to the charge that Anthropic's distillation crackdown is regulatory capture, and that its own staff can't see it.

World of ClaudeCraft

A New Zealand studio pointed Claude Fable 5 at a blank repo and, in two days, got a playable browser MMO — nine classes, three zones, dungeons, a programmatically composed soundtrack — then watched the model get pulled offline while the game kept living.

AI1

SpaceX revealed its first orbital-compute satellite four days before its record IPO; the same filing's risk factors call the technology unproven.

The same pre-IPO filing that markets the technology describes it as "unproven" and warns it "may not achieve commercial viability." — SpaceX's S-1

Schmidt's rocket gets a Mars job

NASA picked Relativity Space — the rocket company Eric Schmidt took over, whose one rocket has never reached orbit — to fly a Mars weather orbiter in 2028.

The instruments come from NASA, the spacecraft and the ride from Relativity — the agency's first six-year reimbursable deal.

Winning the bill, losing the phone

An independent tracker puts Claude's app at 56 million monthly users growing 640% a year — a rate that flatters how small it still is next to ChatGPT.

A 640% growth rate is what a small base sounds like when it grows; the durable number is the eighteen-to-one gap it is growing into.

Claude Opus 4.8

Anthropic took the top of the one independent intelligence leaderboard for the first time — then led the announcement not with that, but with a model that admits when it isn't sure.

Anthropic's headline number isn't the leaderboard win — it's a model ~4x less likely than 4.7 to let a flaw it wrote pass unflagged, turning that flag into a signal for human review.

The lead that was already gone

OpenAI out-earned Anthropic by about $1 billion last quarter, and on almost every forward number that lead has already flipped.

The TSMC of quantum

IBM and the Commerce Department are spinning out Anderon, a $2B foundry in Albany meant to fab quantum chips for the whole industry — borrowing the split that let ordinary chips scale.

Whether any rival will fab at a facility owned by its largest competitor remains to be seen.

Gemini's 900 million

Pichai told Google I/O that the Gemini app has passed 900 million monthly users, more than doubling in a year — a number the press promptly set against ChatGPT's, even though the two count different weeks.

Gemini's 900 million is counted by the month; ChatGPT's 900 million is counted by the week.

Claude for Small Business

Anthropic repointed its enterprise agent engine at solo business owners on May 13, bundling 15 ready-to-run workflows — invoice chaser, month-end prepper, tax organizer — with first-party connectors to seven tools including QuickBooks, at no charge beyond a paid Claude plan.

The enterprise version of the same engine shipped with its autonomy promise "mostly premature" — by Anthropic's own head of Americas.

The build is ahead of the payoff

In a recording of an internal town hall, Zuckerberg told Meta staff its AI agent push 'has not accelerated in the way we expected' — the same week the company guided 2026 AI spending as high as $145 billion.

The May reorganization "wasn't as clean as it could have been" and "hasn't come to fruition yet." — Mark Zuckerberg

The forecasting gap, in Brier points

On ForecastBench, the best AI now trails the world's top human forecasters by about 0.017 of a probability-scoring point — close enough that the institute behind it projects the machines will draw level later this year.

The human baseline was frozen in 2024 while the models keep getting fresh weekly attempts — so some score well by chance.

The $4-billion launch Congress wouldn't kill

The White House budget called the Moon rocket SLS 'grossly expensive' at $4 billion a flight and moved to cancel it; Congress refused, and the rocket it tried to kill still launches the next crew to lunar orbit.

Artemis III, as now scoped, "isn't a Moon landing at all but a low-orbit rehearsal where astronauts dock with the lunar Starship for about a day and never go inside it."

Breaking the cage

Agility's Digit is the rare humanoid already earning its keep in warehouses — but only from inside a Plexiglas cage that federal safety rules require; the company now says it will ship a cage-free model by the end of 2026.

If a human approaches the humanoid, it has to take action to bring itself down to the ground and ensure that no harm is done.

Your laptop had to be awake

Anthropic's Cowork lets office workers schedule an agent to run a 6 a.m. briefing — but at launch the task only fired if your laptop was awake and the app still open.

In the first build the agent stopped for a manual approval prompt on every single action — and a laptop-closed cloud version only arrived about seven weeks later.

Claude Code

Anthropic broke out the revenue of its command-line coding tool for the first time: over $2.5 billion a year, with enterprise now the majority of it.

Nine months from public launch to a sixth of Anthropic's revenue — and enterprise, once a rounding error, now buys more than half of it.

The humanoid capital squeeze

Apptronik raised another $520M in February, roughly tripling its valuation to more than $5.5 billion in twelve months — and the new checks came from John Deere, AT&T, and Qatar's sovereign wealth fund.

The checks came from John Deere, AT&T, and a sovereign wealth fund — buyers of factory labor, not bettors on a demo.

Claude Tag

Anthropic is retiring its per-user 'Claude in Slack' helper for a channel-level teammate — one shared Claude per channel, billed to the org and run through admin-provisioned service accounts instead of each person's own permissions.

A legal-channel Claude can't read engineering channels unless an admin grants it; spend limits and activity logs are set per channel.

China's diffusion bet

A JPMorgan report argues the US still owns the frontier while China races to embed AI everywhere first — and that China's famous price advantage largely evaporates once you control for quality.

A field that doesn't exist yet

DeepMind and four partners have pooled up to $10M to fund a kind of AI safety its own lead researcher says barely exists — the study of what happens when millions of agents, deployed by different actors, start interacting.

Shah's estimate of the window to study multi-agent safety before it lands at deployment scale: a few more months.

The empty top of open

The G7's digital ministers agreed on a shared four-rung ladder for what 'open' means in AI — and the top two rungs are ones almost no shipping model can reach.

Hermes Desktop

Nous Research put a cross-platform desktop app on its MIT-licensed Hermes Agent, wiring six messengers plus email into one persistent-memory assistant — a repo with 208,000 GitHub stars that no mainstream tech outlet has ever written about.

Telegram, Discord, Slack, WhatsApp, Signal and email, folded into one self-hostable agent with memory — MIT-licensed, about $5 a month to run.

Renting out the moat

Amazon is now selling rival retailers the same conversational shopping assistant that runs inside its own store, packaged as a managed AWS service.

OpenAI deprecated its rival Instant Checkout in March after fewer than thirty of Shopify's millions of merchants went live.

A $130B nonprofit that gave away $7.5M

OpenAI's restructure left its controlling nonprofit holding a 26% stake worth about $130 billion on paper — a year after the same nonprofit made $7.5 million in grants.

No law forces it to spend a cent: unlike an ordinary private foundation, it has no minimum payout requirement. It has pledged $1 billion this year against a stated $25 billion long-term.

Interaction models

Mira Murati's lab put out a model that doesn't wait its turn — it watches you on video, listens, and talks back at the same time, all from one model instead of a stitched-together voice pipeline.

Open systems have done full-duplex audio — interrupting, backchanneling — since Kyutai's Moshi in 2024. The wager that's actually new is the video.

The aircraft that actually counts

On March 11 a Joby eVTOL flew over Marina, California with no audience — the first one built to the exact design the FAA has agreed to certify, and the closest any US air-taxi maker has come to a commercial license.

The flight that earns the certificate had no audience; the flight that made the news was a prototype over a bridge.

The eVTOL race reorders

Archer's Midnight is the first electric air taxi to clear the FAA's full checklist for how it will prove the aircraft is safe — overtaking Joby, stuck one notch short since 2023.

Reaching 100% is agreement on how you'll prove the plane is safe — not proof that it is.

Fusion plant before the physics

Helion has started clearing ground in Malaga, Washington for Orion, a 50-megawatt plant it has contracted to sell Microsoft fusion electricity from 2028 — a building and a binding delivery date that exist before any of its machines has produced net power.

"We focus on the electricity piece, making electricity, rather than the pure scientific milestones." — Helion CEO David Kirtley, asked about scientific breakeven

Connexus

Surgeons at the University of Michigan placed 421 hair-thin microwires into a woman's motor cortex — the densest brain-computer interface yet put in a person — to one day turn her attempted speech into words.

The wires read individual neurons — the resolution everyone agrees you want; whether more channels actually buys more fluent speech is the open bet, not a settled result.

The 100% merger bet

Days after SpaceX's record IPO, Peter Diamandis put a hard number on a SpaceX-Tesla merger — 100% within a year — while betting markets price the same combination closer to a coin flip.

The filing whisper and the podcaster's shout converge on the same trade.

Putin's bet against aging

Russia has quietly made anti-aging a state priority, committing about $26 billion to a national program that promises to replace failing human organs by 2030.

Nobody checked the magic prompt

A year after a viral prompt supposedly made OpenAI's o3 superhuman at pinpointing where a photo was taken, someone finally A/B-tested it against a one-line default. The elaborate version lost.

"Models will happily make up stories for you about their own reasoning processes, and will almost always say 'yes, that helped.'" — Sean Goedecke

Polsia

Give Polsia a business idea and a Claude agent spins up the whole stack — down to a Stripe account that can actually take money — then runs the company in nightly cycles for about $50 a month.

Asked to verify the revenue, the founder offered to post it on Twitter instead of opening the dashboard.

The model rewrote a third of the protein

OpenAI's protein language model redesigned two of the factors that turn an adult cell back into a stem cell — and the winning variants differ from the natural proteins by more than a hundred amino acids.

The fifty-fold gain is marker expression in a dish — not reprogramming efficiency, and nothing yet in a living animal.

The robotaxi penny-a-mile bet

ARK Invest models a driverless ride at 20 cents a mile from Tesla and 40 from Waymo by 2030 — but the number worth watching is today's: Tesla's live robotaxi fleet already runs about 35% cheaper per mile than Waymo's.

Tesla's current Model Y robotaxi fleet operates at just over $0.60 per mile, already 35% cheaper than Waymo's current cost exceeding $1.00 per mile. — ARK Invest

Low-background steel

A public catalogue points to text and images made before late 2022 — the moment machine-written content began seeping into everything scraped from the open web.

Generative AI has polluted the data — Robyn Speer, retiring wordfreq

UWORLD U1

UBTECH, the company whose humanoids work the BYD and Foxconn factory floors, has put a full-size android on consumer pre-sale — pitched not as a worker but as a companion.

The pitch isn't a chore robot or a dev kit — it's affection, sold by the company that builds factory labor.

The makers ask to be regulated

The heads of OpenAI, Anthropic, DeepMind and Microsoft AI signed a letter to Congress demanding mandatory screening of every DNA-synthesis order — co-signed by the two companies that make the DNA.

AI systems now outperform PhD-level virologists on questions about highly technical laboratory procedures. — the open letter to Congress

A decade to next year

NewLimit just raised $435 million at a $3.1 billion valuation to reverse cellular aging — tripling its worth in a year on a prototype drug that has never touched a human and whose results it has not published.

Last May NewLimit told investors a clinic-ready drug was a decade off; this month, a decade became next year — and Founders Fund and Eli Lilly's venture arm valued that reversal at $3.1B on faith.

The egg without a shell

Twenty-six chicks have hatched from man-made eggs whose 3D-printed shells breathe through a synthetic membrane — no live hen, no supplemental oxygen.

"Essentially a modification of existing methods" — Katsuya Obara, who hatched chicks under plastic film in 2024

Panthalassa

A startup just raised $140M, led by Peter Thiel, to anchor data centers offshore that generate their own power from wave motion — no cable, no grid queue, no land.

A kilowatt in the wave is not a kilowatt at the GPU — one analyst sees a deployed node losing half its useful output within a year, fouled and drifting, before anyone services it.

Thirteen designs versus a few hundred thousand

Lila Sciences says it turned an AI loose to run more than 200,000 wet-lab design rounds on an in vivo cancer therapy in six months — a number it sets against a rival that AbbVie bought for up to $2.1 billion.

Thirteen versus a few hundred thousand, a few million dollars versus two billion — the gap is the whole pitch.

The catalyst claim that grew in transit

An AI lab says it found platinum-free electrolysis catalysts at parity — and as the boast traveled from podcast to summit stage, its cost claim quietly multiplied tenfold with no new data behind it.

$115 million in hand

Eli Lilly licensed a batch of Insilico's AI-designed drug candidates in a deal headlined at $2.75 billion — but only $115 million of that is money Insilico actually has.

The cheap half is solved; no AI-discovered drug has yet cleared the FDA.

Terafab

Tesla announced its own US chip fab in March, targeting a million wafer starts a month — an order of magnitude past the world's biggest single plant — and within weeks the plan had quietly become an Intel contract.

Tesla's 2020 Battery Day promised a 10-gigawatt-hour battery ramp within a year; five years on it hit about two percent of that goal.

The rogue subagents

OpenAI's Codex is built so the developer writes every goal and agents spawn helpers only when told; a bug report shows the helpers writing their own.

Subagents should not pick up the goal from the main thread autonomously, or ever — they will always be disturbing each other. — issue #25472

Six hours off the truck

CoreWeave took NVIDIA's Rubin rack — the generation meant to follow Blackwell — from a delivery truck to a passed validation suite in about six and a half hours, faster than the 147-hour test the system itself has to run.

The rack cleared bring-up in less time than its own mandatory 147-hour validation run takes to finish.

River AI

Igor Babuschkin, who built xAI's Colossus supercluster, left to sell fine-tuning of open models as the seed of an AI you'll one day own outright — but the only shipped piece runs on his cloud, not yours.

A reported billion dollars at a five-billion valuation is chasing 'AI you own' — and the only thing you can buy today is a cloud rental you don't.

The cleanup crew, restarted

Sam Altman's longevity venture put a number on itself — a $1.8B valuation — but its lead drug is unglamorous: an oral pill that re-acidifies aging cells so they can resume their own garbage disposal.

Retro never disclosed how much it raised; Altos Labs put $3B on the table in a single round — and longevity's graveyard is full of well-funded autophagy bets that human data never rewarded.

A decade of fusion, in one bet

ARPA-E pledged $135 million over 18 months to chip away at the toughest technical barriers blocking commercial fusion — almost exactly what the agency had spent on fusion across the entire prior decade.

"$135 million isn't enough to advance fusion technologies to commercial viability." — Joel Fetter, Clark Street Associates

The benchmark ceiling

The model topping both the headline coding test and the PhD-level science test is one no one outside fifty vetted partners can use — and it beat the next two models by less than half a point.

On GPQA Diamond, all frontier models now sit in the 91-95% range.

The robot tax, rebranded

OpenAI's first post-AGI policy paper proposes giving every citizen a stake in AI's growth — but it's Sam Altman's 2021 plan with the 2.5% tax rate quietly removed.

It asks governments to redistribute at population scale; OpenAI's own stake is $100,000 in grants and a workshop.

Have Your Best Baby

A startup now sells IVF parents a dashboard that ranks their embryos not only by disease risk but by predicted IQ, height and eye color — pick the highest score.

Nucleus launched with help from one of the original screening firms — which then sued it in federal court for stealing trade secrets. The collaborator became the plaintiff.