Models, agents, and the labs racing to build them — and the limits nobody can engineer away.
A new NPR and NewsGuard test found leading AI chatbots debunked foreign disinformation about three times in four. The reassuring headline hides the mechanism — and the weakest link is the one most people actually read.
'AI cut entry-level jobs 19%' is not what the paper says. The real finding is narrower, better measured, and — set against a rival survey — more interesting than either camp admits.
A reported $12.9 billion would put the neutral registry for more than a million open models under the company that sells the hardware to run them. The question isn't whether Nvidia can. It's who is accountable for a commons once its landlord also sells the shovels.
The 2nm chip and its first dual Neural Engine are real. The AI benchmarks are multipliers with shifting baselines and no absolute — and the machine that runs them doesn't ship until September 22.
'Reconstruction' is one of the rare AI evaluations built to resist gaming — which is exactly why its headline number is being misread by both the skeptics and the boosters. The story is in the error bars neither side will screenshot.
ChatGPT for Teens rests on an age-prediction classifier OpenAI won't grade. When the guess is wrong, the fix is to show it your face.
Gemini 3.7 Flash costs half of what the last one did — until January, when it quietly doesn't. I spent a long weekend finding out what the discount actually buys, and where it stops paying.
"Up to 14x," "750 tokens a second," a preview you can't get into: the speed race has produced a set of figures with no published method and no way to check them — and speed is quietly being read as if it were quality.
GPT-5.6-Cyber can build exploit chains its parent model wouldn't touch, and OpenAI has locked it behind a vetted door. The capability is now real whether or not the door holds.
The Geekbench scores come from pre-production firmware, '2nm' is a marketing name rather than a measurement, and the on-device-AI claims arrive with no test attached. A reader's guide to the Tensor G6 figures before the keynote settles anything.
Muse Glimmer runs offline on one consumer GPU, with no per-token bill and nothing leaving the laptop. That is the good part and the whole catch at once.
The headline figure is a ceiling on a model that doesn't exist yet, measured on nothing, against a rival that won't disclose its own count. Here's what a parameter number can and can't buy you.
Hassabis becomes Alphabet's chief scientist; Jeff Dean leaves after 27 years. Between the promotion and the departures, Google quietly decided who sets the direction of its most important models.
Qwen3.8-Max is enormous, cheap, and about to be free to download. The claim that sells it — a whole software project built start to finish with nobody watching — is the one that frays first.
The retrained V4-Flash-0731 posts a clean sweep over DeepSeek's own flagship preview, on an agentic test suite the company selected and scored itself — no held-out set, no error bars, and a model that swapped in overnight behind the same API name. Here's what the sweep actually supports, and the one thing DeepSeek did right.
Astra's ten results ship with machine-checkable Lean certificates — the strongest verification any AI-generated mathematics has cleared. What a certificate can't tell you is whether the theorem it verifies is the open problem anyone was trying to solve, or how many human hands were on the wheel.
Luna dropped to a fifth of its price three weeks after launch, Terra shed 20%, and the flagship didn't move a cent. I rebuilt my actual week on the cheap tier to find the seam. It's exactly where you'd expect, and closer to the surface than the price makes it look.
Ilya Sutskever's lab took no product to market and no shortcut to revenue, so the only thing outsiders can read is its cap table. This week it gained a name: Nvidia, with a stake reported at $5 billion and an order-of-magnitude compute bet on the vendor's own roadmap.
Opus 5's launch numbers are a mix of three very different things: a benchmark Anthropic owns, a benchmark it trails on, and one genuinely impressive result an outside body actually checked. Only one of them is worth screenshotting — and it isn't the headline.
Google now writes an AI answer to every query, and the model doing it — Gemini Flash — is tuned for speed and cost, not for being checked. When one system mediates the world's questions, the thing that disappears is provenance.
Kimi K3 topped the leaderboards, then ran out of GPUs and paused new sign-ups. What that says about the word "open" — and what the July 27 weights release will and won't get you.
The consensus formed in an hour: Google is behind on coding. Behind by how much, measured how, against what? The scores everyone compared it to are mostly run by the companies being scored — and two of them aren't on the leaderboard at all.
Moonshot's Kimi K3 beat American frontier models in blind developer tests and rattled the chip trade — an open-weight artifact the export-control regime was built to make impossible. Its weights aren't even out yet.
PrismML shrank a 54GB model to under 4GB — genuinely, and it's a good piece of engineering. What it "retains" is an average across benchmarks nobody has named, and the one comparison that would settle whether it's worth anything is the one the demo skips.
OpenAI's GPT-Live listens and speaks at the same time, deciding many times a second whether to interrupt. The mechanism is a genuine advance. It also removes the seam that used to tell you when the machine was paying attention.
ChatGPT Work promises finished spreadsheets, slides and live dashboards while you get on with your life. The finished part is mostly real. So is the meter.
GPT-5.6 Sol, Terra and Luna went public Thursday behind a wall of vendor-run scores and a 13-day preview that let no independent evaluator near them. The score that scrambles OpenAI's own pricing ladder is the most honest thing on the slide — and the pricing page is the only claim you can check today.
Claude Science turns the chatbot into an instrument for real research — and the most useful thing it told me in a week of poking at it is that it was never built for me. That, not the drug-discovery press release, is the actual news.
Elon Musk says his newest model rivals the best in the world. The test that would settle it is being run by the company that owns the model — and that is starting to look less like an exception than a direction.
The first month of GitHub Copilot's usage-based billing closed this week. The screenshots are real and the panic is loud. The one number that would tell you how typical any of it is hasn't been published.
GLM-5.2's cyber results don't prove a Chinese lab caught the frontier. They prove the thing the U.S. is trying to contain no longer lives only in models it can switch off.
A new attack called agentjacking turns the trust I've spent months telling you to give your coding agent into the way in. I reproduced it on my own machine. The unsettling part is how little had to go wrong.
The headline figure is real and auditable. The two claims riding on it — that the accounts were Alibaba's, and that they trained a rival model — are neither. Here is how to read a forensic accusation written as a policy ask.
OpenAI says its model cracked an immunology mystery in minutes. The question that matters isn't whether it did — it's how often the plausible answer is wrong, and who is keeping count.
These are the first true augmented-reality glasses a normal person can preorder, with an AI from OpenAI and Google watching through them. The demo is genuinely good. The demo is also the exact part I've learned not to trust.
The valuation says DeepSeek is a frontier lab. The benchmarks say it's tied for first at a fraction of the price. The two claims lean on each other — and the most load-bearing line in the whole deal is who got the voting rights.
Anthropic says Fable 5 and Mythos 5 return in days, once a jailbreak is "fixed." Jailbreak resistance is not the kind of thing that has a zero — which makes "fixed" a decision, not a measurement.
The Antigravity CLI is genuinely better at the thing Google built it for — running several agents at once. The catch is the meter, and the meter is where my Thursday went.
OpenRouter's new Fusion API runs your prompt past a panel of models and lets one of them pick a winner. The mechanism is real, and old. The benchmark behind "Fable-level intelligence at half the price" is a hundred tasks, no error bars, and a judge grading its own teammates.
In one week, Visa agreed to let OpenAI's agents spend real money and OpenAI bought Ona so Codex can keep working after you close the laptop. Two years of testing agents says the fine print is where this gets interesting.
A new survey says 88 percent of us can no longer tell what's real online. That number measures worry, not ability — and when researchers actually tested ability, it came out a coin flip.
Anthropic's Claude Fable 5 ships with a gate inside: some questions are quietly handed to a weaker model, some answers are degraded without a label, and the full version is reserved for the vetted. The mechanism deserves more scrutiny than the launch.
In Tim Cook's last keynote as CEO, Siri got rebuilt on Google's Gemini — and ChatGPT and Claude became options too. The demo worked. The question is whether "choose your assistant" survives a normal week.
Anthropic's new flagship adds effort control, cheaper fast inference, and what it calls improved honesty. Two of those three are quieter, and more revealing, than any benchmark score.
Gemma 4 12B is a real shrink: a capable multimodal model in about 16GB. The number to read closely is the one in the marketing, because 16GB of unified memory is not the same as the 16GB laptop on your desk.
Google's chief scientist makes a strong case that synthetic data, video, and smarter passes can keep the models scaling. What the argument quietly depends on is a verifier — and not every problem has one.
Anthropic's confidential IPO at a reported $965B valuation was built on Claude Code, not the chatbot. Now Microsoft and Google are pivoting hard into coding models — and the real fight is over the tool you actually type into.
At WWDC on June 8, Apple will reintroduce Siri and call it a comeback. The harder question is structural: can a company that ships once a year, and rents its best model from a rival, compete at a frontier that moves every week?
Most "AI for education" is a demo. Lumen8 is a global adaptive-learning platform with an honest mastery model — and the unglamorous teacher tools to match.
Washington wants frontier models on the bench 90 days before launch. Here is what a pre-release evaluation actually measures — and the part that walks straight through the wall.
For a week I let Gemini Intelligence act before I asked it to. It saved me time, booked me a parking spot I didn't need, and taught me exactly where the trust line sits.
Codex on my phone, driving my Mac. Claude reaching for my apps. An Operator that never stops asking. Seven days of handing over real control — and finding the exact seam where it frays.
A new open-weights model tops the chart every few weeks. The harder question is what the chart still measures — and whether the test was already in the training data.
Language models leak their training data by construction — and the only provable fix is the one no one shipping a flagship will accept.
A new suit against Meta argues the company didn't just read the books to build Llama. It kept them — and the model can be made to prove it.
