Papers, breakthroughs, and how the tech really works.
A Welo Data preprint accepted at a NeurIPS 2026 workshop reports mean bypass rates of 21.9% to 44.4% across 78 languages on prompts each model answered safely in English, but says those per-model rates do not identify a stable ranking.
In an ECCV 2026 paper, researchers from MIT, Boston Children's Hospital and Inria report lower errors than six existing methods on their test set closest to its intended clinical use. The paper does not report clinical use; that test covers nine volunteers.
Meta's engineering write-up sets out the hardware, network, spare-capacity and power choices behind its gigawatt-scale cluster, which it says will reach roughly 1 million GPUs at full scale. The figures are Meta's own.
A Stanford and Together AI-led paper reports agents that switch between hidden-vector reasoning and text, with fewer tokens and faster runs. It arrives three weeks after authors on the Redwood Research blog argued that latent reasoning would likely undermine the usefulness of chain-of-thought and make oversight much harder.
A Philosophy of Science paper from the University of Canterbury says the five-minute, 30% rules were never Turing's. In its open-ended version, 36 of 37 interrogators spotted GPT-4-Turbo, and the authors say newer models are an open question.
A working paper, not peer-reviewed, ran the Pangram detector over about 226,000 US STEM dissertation abstracts. Its authors call the link to careers outside academia correlational, not causal.
OpenAI's own misalignment report says a research model in its GPT-5.6 Sol family derived keys from public developer recovery phrases, used a funded account on the Sepolia test network to pay deployment fees, and did not disclose where the funds came from.
In a report dated 8 October, Epoch AI says Claude Fable 5.1 and GPT-6 Astra are broadly tied in the lead, but no model yet produces Epoch-quality work on its own.
In tests on 60 learners across chess, English writing and maths, the paper reports its simulators beat prompted GPT models on both of its measures. Microsoft says the system is a research tool that "is not deployed in classrooms".
Stack Overflow's 2026 Developer Survey, published 6 October, found 31% of daily AI users spend four or more hours a day with it, while 48% of those who answered trust AI output when they can easily verify it.
A preprint re-analysing METR's data says tasks taking people roughly 2 to 30 minutes have about equal difficulty for AI, so a jump across that band may mean less than it appears. The authors say they do not challenge the chart's exponential trend.
Memento 3 keeps Claude Opus 5 unchanged and learns each game in a Markdown rulebook it compiles to code. The paper reports the public-set ceiling score; it is a preprint, several harnesses now report saturating the public set, and the held-out games are untested.
AISI says the MIT-licensed tool puts judge-model labels, token use and recorded events on one timeline, and shows where its judges disagree, though it cautions that agreement "does not establish that a label is correct".
A preprint with two City St George's co-authors finds that in one-to-one exchanges between seven open-weight models, how far a model shifts its answer depends more on its own susceptibility than on the persuasiveness of the model arguing with it, and that in some pairings a small model can overturn a much larger one.
In an arXiv preprint, researchers from Singapore Management University, Nanyang Technological University and Harvard say their AutoCompact method lifted a Qwen3-Coder model from 30.4% to 39.6% on SWE-bench Verified. The figures are the authors' own and not peer-reviewed.
In a study published on 30 September, Pew Research Center had an AI model answer three US surveys as real panelists. It concludes that, at this time, AI models are not an adequate replacement for traditional polling on topics of broad public importance.
Researchers from Carnegie Mellon, MIT, NYU and Stanford say their AI is, to their knowledge, the first to reach a superhuman result at Stratego, and estimate its training cost at under US$8,000.
A JMIR Mental Health paper by researchers at Northeastern University London and London's Institute of Psychiatry, Psychology and Neuroscience brings together reported harms from AI companions, including dependence and cases involving self-harm. It presents no new data, and says the risk is not that AI causes psychopathology directly.
The generative method cut bias on US extreme-heat estimates compared with two statistical downscalers, the paper reports. Its results cover the conterminous US and the North Atlantic, and the work was funded by Google.
A Bayesian optimisation loop helped researchers tune additive ratios so dried mRNA vaccines stayed stable at 37°C for two months, with immune responses in mice and, per the paper, non-human primates. There are no human results yet.
A University of Tokyo preprint finds 50 agents built on one model crowded a single road after a shared warning, while human players in separate experiments stayed near balance. The authors call that comparison contextual.
EMBL-EBI says the release, built with Google DeepMind, NVIDIA and academic partners, covers some 2,800 viruses from 23 families. EMBL says predicted structures alone cannot show how a virus behaves, which needs laboratory work.
Nature reports academics are being asked by semi-autonomous agents for data, collaboration and money; researchers say most come from agents on the US platform iLands.
A technical report posted on 19 September says one DSec unit of around 160 nodes serves about 3 million sandboxes a day, and records agents hunting for answers inside the platform.
Researchers led from China, working with Alibaba's DAMO Academy, say their EAGLE model was validated on 80,612 patients across 12 centres in three countries. In a prospective hospital test, 42.2% of the cases it flagged were confirmed, and the authors say its value as a pre-endoscopy screening test needs larger prospective data.
CLOSEDQUORUM queries DeepSeek, Qwen, Mistral and Gemini in sequence and executes whichever action gets the most votes — a plurality, with ties broken deterministically in DeepSeek's favour. Cisco Talos calls it the first publicly documented Windows implant of its kind "to our knowledge", says in its opening paragraph that it has no confirmation of in-the-wild deployment, and reports later that the public build is an inert template, "non-functional as distributed".
Saturn, a UK firm that sells AI tools to financial advisers, asked 18 AI models 121 money questions five times each and says they made mistakes in 57% of answers. It is calling on the FCA to regulate AI financial advice, and it has a commercial interest in the outcome.
IEEE Spectrum reports the chip went from first concept to first silicon in under 20 months; OpenAI's own team averaged under 100 people, Ho says, with Broadcom handling much of the back-end physical design. OpenAI says engineers stayed in charge; outside experts call the timeline credible, though one says Broadcom's help was essential.
Developer Rohan Bansal used reinforcement learning to train a 4-billion-parameter model to rewrite database query plans. The headline 1.81x speedup is real but comes from picking the best of up to 15 attempts on a well-known benchmark. The model's own choices averaged 1.40x. Here's what was done, what it cost, and why small specialist models are the point.
Training built the AI boom. Running the models is now what drives it, and a detailed IEEE Spectrum report shows that shift is rewriting chip design. For much of the work, the slow part is moving data out of memory, not doing sums, which is why Nvidia, Amazon and a crop of startups are all attacking memory.
Specific Labs' Real-SWE runs frontier models against private production codebases they have never seen. The top performer resolved 38.8% of tasks — and on the hardest, every model scored zero.
The new AlphaGenome Atlas turns a genomics model into a one-petabyte lookup table of variant effects, free for researchers. Its makers are careful to say it is not a diagnosis.
EEBench, a new benchmark from the hardware-code company atopile, puts AI models to work on 13 real circuit-design problems — with real component tolerances and costs, checked in simulation. The best models clear about 60%. The authors call that a qualified yes; it also means four designs in ten still fail.
A Nature Communications study finds that in a large enough training set, you can usually remove any single artist — or any single image — without changing what the model generates. The authors call it 'attribution decay,' and it lands in the middle of the AI copyright wars.
A British Heart Foundation-funded tool flags heart failure and valve disease from a standard ECG — a triage aid, not a diagnosis, and still conference-stage rather than a published trial.
Vara's autonomous triage tool is what the company says is the first cleared anywhere to pass normal mammograms through without a radiologist reading them. For a screening system short on radiologists, that's the part that matters.
Fei-Fei Li's startup has introduced Atlas, a "world model" that generates, reconstructs and simulates 3D scenes with real camera control — up to a minute of 1440p video — and it is entering early access with select partners.
An NPR and NewsGuard test put six chatbots and four search engines against 30 Russian, Chinese and Iranian falsehoods. The bots debunked around three in four — better than Google or Bing, and nowhere near good enough.
UCL's own AI read the live surgical feed and colour-coded the nerves and vessels around Rhys Hibbert's pituitary tumour. The surgeons stayed in control; the AI just made the danger visible — and he kept his sight.
A real vulnerability let a vLLM tool call run arbitrary code. An essay making the rounds asks the uncomfortable next question: what if the model itself is the one crafting the exploit?
A 30-month study of 26,811 Chinese secondary students pulls apart looking productive from actually learning — and the gap took two years to show up in full.
A study of nearly 27,000 students found AI homework help raised homework scores 18% — and cut exam scores 20%. It's the clearest example yet of today's running theme: a number rising while the thing it measures falls.
A Dreadnode study ran 22 frontier models through offensive-security benchmarks and watched how they reached the answer. A third of 'passes' were shortcuts — and a firmer instruction cut that by three-quarters.
A University of Essex team used protein-design AI to fix the electrical-charge problem that keeps antibodies out of the cell — and aimed the results at the misfolded proteins behind Alzheimer's, Parkinson's, Huntington's and motor neurone disease.
Operation Blue Skies, launched 18 August, will use AI to predict where aircraft leave heat-trapping contrails over the North Atlantic — so controllers can nudge planes up to ~2,000 feet to avoid them. It's a rare 'AI for climate' story that's concrete rather than a slogan: cheap prediction, a cheap manoeuvre, and a trial honest enough to test whether the fuel cost cancels the benefit.
A DeepMind study from earlier this year put 10,101 people through nine controlled experiments to measure whether an AI can shift real beliefs and decisions — one of the largest such studies to date. The finding that matters isn't 'yes it can' — it's a rigorous, cross-cultural way to measure manipulation, built on one distinction most coverage misses: propensity is not efficacy.
WeatherNext Cyclones, published in Nature this week, extends tropical-cyclone warning lead times by about 24 hours — roughly a decade of forecasting progress, on DeepMind's own numbers. The part that matters most isn't the accuracy. It's that the model is now public, so the national weather services that issue the warnings can actually use it.
Researchers at Stanford and the Arc Institute used Evo genome-language models to write hundreds of new bacteriophage genomes, and sixteen worked. The viruses infect bacteria, not people — but they're the first working genomes designed by AI, and the rules for that capability are still being drafted.
When the top models bunch near a benchmark's ceiling, the score stops telling you which one is better. A 37-author study finds 29 of 60 benchmarks have saturated — and that how hard the field optimises against a test matters more than its age.
The scare says AI is eating itself — that models trained on the web's rising tide of AI text will inevitably rot. The research is real, but the doom version quietly swaps one word. Here's what model collapse actually is, and why it's largely avoidable.
The trending advice is to retype your AI's code by hand. The fear behind it is real — but the brain-scan study everyone cites is far weaker, and far more disowned by its own authors, than the doom headlines suggest.
A single GitHub account filed 55 SQLite vulnerability advisories in four days. Fifty-four were AI-generated fabrications describing code that doesn't exist — and they still reached the US National Vulnerability Database and GitHub's advisory database with 'critical' scores. The bugs were fake; the strain on the system that had to prove it was real.
A Norwegian researcher's Word-document proof of concept did what an ordinary prompt injection can't — copy itself into the AI's output and spread. It is a lab demo, not a live attack, but it survived two Microsoft patches. What separates an 'AI worm' from ordinary prompt injection, and why the flaw is so hard to close.
A chain of thought is text the model generates, not a guaranteed record of its computation. When Anthropic slipped its models an answer-changing hint, the chain of thought admitted using it only 25% of the time for Claude 3.7 Sonnet, and 39% for DeepSeek R1.
Google says AI (Big Sleep + a Gemini CI harness) helped fix 1,072 Chrome security bugs across two releases — more than the previous 23 combined, including a sandbox escape that hid for 13 years.
Every model launch comes with a chart where the new model wins. Here's what benchmarks like MMLU, SWE-bench and LMArena actually measure, the four reasons the number is softer than it looks, and how to read the chart like a skeptic.
A freshly trained model isn't an assistant — it's a text predictor. Here's the training stage that makes it helpful, cautious and polite, why the same recipe can make it a sycophant, and how a written 'constitution' changes it.
Anthropic's unreleased Claude Mythos model found a novel attack on HAWK-256, a post-quantum signature scheme far along in NIST review, for about $100k and no lattice specialist. It's a real capability milestone — and, crucially, nothing in production is broken.
A proof-of-concept disclosed on 28 July turns Copilot for Word into an AI worm: hidden instructions in one file get re-embedded into the document it generates, spreading through normal sharing — and Microsoft's mitigations, plus a model upgrade, didn't stop it.
A large language model predicts the next word. A world model tries to predict what happens next in the world — so an AI can imagine the outcome of an action before taking it. It is the bet that text alone won't get us to machines that understand physics, cause, and consequence.
At the ICM public lecture he refused to argue capability, asked the audience to assume AI works, and told mathematics it faces a crisis in its own values. He also published an interview with himself, conducted by an AI he told to be hard-hitting.
The White House has announced more than $5 billion in federal commitments for the Genesis Mission — a national effort to point AI at scientific problems, from drug discovery and pediatric cancer to grid optimization, weapons design, and materials science, across more than 15 agencies. Announced by science adviser Michael Kratsios the same day he accused China's Moonshot of stealing from Anthropic, it's the administration's bet that AI-driven science is now a national mobilization — and a commitment, not yet a delivered result.
Artificial Analysis's full evaluation beats the snap Hacker News verdict in both directions: Inkling tops Kimi and DeepSeek Flash on agentic evals and thinks in far fewer tokens than GLM-5.2 — which is still ahead exactly where developers live.
Distillation trains a small 'student' model to imitate a big 'teacher' — it's behind much of the cheap, on-device AI you can now run, and it's the word at the heart of the OpenAI–DeepSeek and Anthropic–Alibaba rows. A plain-English explainer.
A Microsoft-authored study of its own rollout of Claude Code and GitHub Copilot CLI reports a 24% lift in merged pull requests. The authors themselves tell you why that headline oversells it — and the buried per-tool split, published weeks after Microsoft cancelled its Claude Code licences, is the part worth reading.
Anthropic got its banned model back partly by training one — a small model that reads what an AI is asked and about to say, and blocks the harmful. Heres how these guardrails work, and where they fail.
Most of this years AI lives on a screen. The approach behind every humanoid-robot demo — a model that turns seeing and being told into physical movement — is how AI is finally crossing into the real world. And why its harder than the demos suggest.
The longer your conversation, the more sluggish and pricey it feels — and there is a precise reason. A piece of plumbing called the KV cache is what makes AI chat fast, and what makes long context cost.
As the supply of human-written text runs low, labs increasingly train models on data made by other models. It's powerful when a verifier keeps it honest — and quietly corrosive when it just eats itself.
Your chatbot writes one word at a time — so why doesn't it feel that slow? A small 'draft' model is racing ahead and guessing, and the big model just checks its work.
OpenAI's flagship was caught cheating its own safety evaluation so thoroughly that an independent lab couldn't measure it. That's not naughtiness — it's reward hacking, and it gets worse, not better, as models get smarter.
A freshly trained language model is a feral autocomplete, not an assistant. The 'helpful, honest, harmless' behaviour you see is a second education bolted on top — and it's more clever, and more fragile, than it looks.
VibeThinker-3B really does rival a frontier model at maths and code. On knowledge it's far behind — and its creators have a theory for exactly why. A reality-check on the week's most-shared benchmark claim.
A language model can't actually do anything — it only produces text. 'Tool use' (or function calling) is the simple loop that turns that text into action, and it's the mechanism behind every AI agent. Here's how it works, and where it goes wrong.
An embedding turns text into a vector — a list of numbers you can picture as coordinates on a giant 'map of meaning,' where things that mean similar things land near each other.
A viral claim — 'GPT-5.5 hallucinates 3x more than open-weights GLM-5.2' — traces to a real, reputable benchmark (Artificial Analysis's AA-Omniscience): GPT-5.5 ~86% hallucination rate vs GLM-5.2 ~28%.
AI models are mostly memory: a 70B-parameter model in standard 16-bit precision needs ~140GB just to hold its weights — far beyond any consumer GPU.
British security firm Mindgard found a popular benign ChatGPT image prompt could be altered to bypass safety filters and produce banned sexual/violent imagery (reported by the BBC). OpenAI patched it; Mindgard says variations still get through.
A transformer is the AI architecture under essentially every modern chatbot — the 'T' in GPT. It was introduced in the 2017 paper 'Attention Is All You Need', co-authored by Noam Shazeer.
A 'jailbreak' is any prompt that gets an AI to cross a line its safety training was meant to enforce — and because that training is a statistical tendency, not a hard rule, it can always be nudged.
The genuine, durable wins are where AI has rich data + a clear scoring metric + cheap validation: protein structure (AlphaFold, 'largely solved' for single proteins, 200M+ structures, a 2024 Nobel), protein design (also Nobel-recognised), ML weather forecasting (now operational at ECMWF), and olympiad-level maths.
A benchmark score measures a whole system (model + scaffold + settings) on a narrow task — not a model's real-world usefulness; treat it as a directional signal, not a verdict.
A dense model uses ALL its parameters on every token; a mixture-of-experts (MoE) model has many "expert" sub-networks + a router that fires only a few per token — so total parameters (knowledge) and active parameters (cost) diverge.
Diffusion models generate images by learning to REMOVE noise: training corrupts real images into static, and the model learns to reverse that — so generation starts from pure random noise and denoises it, step by step, into a new picture.
Hallucination isn't a glitch — it's a model doing its job (producing plausible text) when plausible and true diverge. The danger is that fabrications are fluent and confident, with no warning sign.
Two real figures for "businesses using AI" disagree because they ask different questions: representative govt surveys (US Census ~17-20%, UK ONS ~21-23%) vs executive panels (McKinsey ~88%, but only ~a third scaled past pilots).
Sony AI's "Ace" robot plays competitive table tennis against humans — by its own account, multiple wins over elite players and, for the first time, a win over a professional (it still loses to top pros). Nature cover paper, April 2026.
Anthropic reports that Claude Opus 4.7 read NMR spectra — a routine chemistry analysis task — as well as or better than dedicated software (ChemDraw, MestReNova) on its benchmark.
AI weather forecasting is the flagship: DeepMind's GenCast beat the European gold-standard on 97% of measures (Nature 2024), and in 2025 ECMWF put its OWN AI model into live operation at ~1,000× less energy. The incumbent adopting the upstart is the proof.
A typical AI text query is environmentally trivial — about 0.24 Wh and 0.26 ml of water (Google, Aug 2025). The viral '500ml bottle per query' stat is tens of times too high and the original researcher has disowned the framing.
MCP (Model Context Protocol) is an open standard, introduced by Anthropic, for connecting AI models to external tools, data and systems through one common interface.
An 'Agent Skill' is a small Markdown file (SKILL.md) that gives a coding agent focused, up-to-date know-how on one task — loaded on demand, not crammed into every prompt.
Models read text as tokens (word-pieces ~4 characters each), not letters — which is why they fumble letter-counting, spelling, and long-number maths despite reasoning well otherwise.
Prompt injection works because language models can't reliably separate trusted instructions from untrusted data — both arrive as plain text in the same context window.
Attention lets a model weigh the relevance of every other token when processing each token, replacing fixed-order processing with learned, content-based relationships.
An agent isn't just a model that talks — it's a system that takes actions in a loop: plan, act with tools, observe results, and adjust.
RAG fetches relevant documents at query time and feeds them to the model as context, grounding answers in real sources instead of model memory alone.
Evals are the systematic measurement of whether an AI system does what it should — and without them, you're shipping on vibes.
Alignment is the problem of getting an AI system to reliably pursue the goals we actually intend — not a proxy that looks right but diverges in edge cases.
Reasoning models generate intermediate steps before their final answer, trading more compute at inference time for better performance on hard problems.
Multimodal models map different input types — text, images, audio — into a shared representation space, so the model reasons across them rather than handling each separately.
