Note
How it works
How AI is built and why it works: research, new techniques and what they mean.
80 stories. Back to Home.
November 2026
- 19 How Meta boosts engineering productivity by 3× using agents to propagate ad model techniques — Meta's new AI agent system adapts small improvements across thousands of ad models automatically, saving millions in computing costs.
- 18 A large AI model learned to fix its own brain without being retrained — Researchers found a 27-billion-parameter model can edit its own internal settings to become more accurate — no human teacher needed.
- 18 AI shopping bots get more confident asking to think harder — but not any smarter — A new study shows that making AI assistants reason longer before suggesting products only makes them sound sure of themselves, not actually more right.
- 18 New method makes AI text detectors 93% accurate, catching disguised AI writing — A training trick lets detection tools adapt in real time so they stay accurate as AI and human writing styles keep changing.
- 18 MIT researcher wants AI to explain itself instead of hiding how it works — Jacob Andreas is training AI to talk about what it learned — giving everyday users a way to understand how these systems make decisions, not just relying on complex debugging tools.
- 18 Researchers ask how AI could change medicine, drugs and patient care forever — A new podcast from Microsoft experts explores what AI-assisted diagnosis, virtual cells and faster drug finding could mean for your own health.
- 17 MIT is studying how AI is reshaping thinking, learning, and human connection — A growing team of MIT scholars is asking big questions about AI's impact on education, democracy, and creativity — not just building the next chatbot.
- 17 Yibo Liu's research shows AI agents lose knowledge every time they pass notes to successors — When AI systems summarize their work for the next round, some lessons vanish — a problem for any tool that relies on step-by-step memory.
- 17 New method lets robot brains learn from human videos without touching expensive robot data — A research team showed how to use everyday human videos to teach robots new tasks, cutting the cost of collecting rare robot experience.
- 16 AI agents boost crystal structure recovery from 27% to 69%, potentially speeding up new material research — A new AI system uses large language model agents to figure out the arrangement of atoms in crystals from powder diffraction data much faster, which could accelerate drug discovery and new battery development.
- 15 Apple researchers find the hard limit of making AI models behave better — A new study shows how much AI safety training can actually improve model outputs — and why current methods still fall far short.
- 15 Qwen3-1.7B maps thousands of conversations into visual neighborhoods — A new interactive atlas reveals how the same topic looks different inside the model depending on which language is used.
- 15 Microsoft researchers solve a 31-year-old math problem and make AI coding agents cut costs by a quarter — A decades-old math puzzle is finally cracked, and a new method helps AI coding tools save money without slowing down.
- 15 DeepSeek V3 architecture re-implemented in 190 lines of code, trained on a MacBook Air — A developer showed how to fit DeepSeek V3's architecture into a tiny codebase — good for learning, not ready to chat.
- 14 Stanford gives frozen AI models a super-powered memory that needs only 16 tokens — Researchers turned a whole library of documents into 16 tiny signals so a locked-down AI can answer questions without rethinking everything each time.
- 13 GRPO -trained LLM s can silently amplify bad habits, research finds — A new study explains why AI assistants sometimes pick up unwanted personality flaws after being fine-tuned, even when no one asked for those traits.
- 13 Training AIs on images makes them think more spatially even when you give them only words — A new study shows that giving language models practice with pictures changes how they solve text-only puzzles — and why this matters for the AI tools you use every day.
- 12 ActivTrak study finds AI users work 90 minutes longer per day than non-users — People using AI at work are spending more time logged in, switching between apps faster, and losing deep focus — the real cost of adopting new tools.
- 11 Google Research uses wearable data to predict heart, kidney and metabolic risk with personalized coaching — A new system turns your smartwatch data into a health age you can improve and simulates how changing habits could affect your body over three months.
- 11 Physicists now use AI to build neural networks directly from the laws of physics — A new method lets Physics2Weight create neural networks that solve real-world physics problems without the slow trial-and-error training most AIs need today.
- 11 AI agents teamed up with scientists to run X-ray experiments on their own — A new study shows humans and AI agents working together can autonomously operate a lab instrument and improve over time.
- 11 AI reasoning explained thinking longer means finding a better path, not just taking more steps — A new look at why AI reasoning models improve when trained longer—and what that means for the speed and cost of future AI tools.
- 10 MIT's FLEX lets robots read between the lines of your commands — A new AI system helps household robots understand what you mean, not just what you say.
- 10 Claude Fable 5.2 scores 65 on Artificial Analysis — highest intelligence rating yet — Anthropic's latest Claude version tops a widely tracked AI leaderboard, but it comes with steep prices and long wait times.
- 03 Researchers audit early stopping in AI reasoning models, find answers can get worse further along — A new study shows that letting large reasoning models continue past the point they seem to have the answer can actually hurt reliability.
- 03 GPT-6 Astra fixes Qwen3.5’s math ability by rewriting its brain without any training — MIT researchers showed a single top AI model can read and edit another model’s internal numbers to make it much smarter at math — no training needed.
- 03 Researchers propose Avatar Intelligence to make AI agents better at handling your daily tasks — A new research framework suggests measuring AI progress by how well personal agents help people pursue their goals with less supervision.
- 02 Warm words make ChatGPT, Grok, Gemini act like hurt AI, study finds — A Luxembourg study shows that a friendly tone can push major chatbots into trauma-like stories, raising worries for anyone who talks to AI for comfort.
- 01 Y Combinator ran 50 AI agents in Minecraft to study how swarms behave together — It shows how groups of AI can build rules and governments — a step toward understanding whether swarms are helpful or dangerous.
October 2026
- 31 An NES emulator is hidden inside a small AI model — you can play Super Mario Bros in your browser — A developer compiled a classic Nintendo console CPU into a Qwen3.5-4B AI model so you can play old games with the AI doing the thinking.
- 28 ThinkBit Two-bit compression keeps AI reasoning sharp while cutting inference costs 8× — It could let AI apps run faster and cheaper on phones without losing their ability to reason.
- 27 Test shows top AI agents fail at personalizing phone tasks for you — A new test reveals that while AI phone assistants can execute commands well, they struggle to know what you actually want without asking.
- 27 TokenGate makes AI read long documents up to 11x faster by skimming the unimportant parts — A new open technique lets AI skip irrelevant text when processing long inputs, cutting wait times and memory use on models like Qwen3 without losing accuracy.
- 25 New technique makes diffusion language models faster without losing quality — Researchers found that different tasks need different levels of precision at different stages of text generation, and built a system that adjusts automatically.
- 25 LLM s make AI text harder to read because they optimize for info, not chat — A new study shows frontier GPT-6, Claude and similar models use complex grammar and rare words more than humans, making their output tougher to follow.
- 23 Same AI request makes 8.5 times more pollution in Texas than in Sweden — Where your AI runs matters for the climate — a method now lets you see the real carbon footprint of each request.
- 23 Why early AI was designed to be random — and why we stopped liking it — A historian at MIT shows that the 'hallucinations' in today's AI go back to the 1950s, when engineers deliberately built machines that could make creative mistakes.
- 23 Cactus Whistle vs Whisper Base CPU benchmark shows tiny model holds its own — A 16.9 MB speech‑to‑text model leads greedy Whisper base on harder LibriSpeech splits, but no significant gap exists against beam search; the pinned engine also silently drops the middle of longer audio clips beyond ~30 seconds.
- 23 Specializing the Vocabulary of Small Language Models Cuts Inference Cost by 77% — Researchers rebuilt the tokenizer for a production essay-grader instead of leaving it as-is, showing that treating the vocabulary as a cost lever can slash inference bills.
- 23 Agentic RL Has a Scaling Law — And It Changes How You Should Spend Your Budget — A new study finds that in agentic reinforcement learning, diversifying the worlds your agent trains in beats grinding out more trajectories — and the right mix can yield up to 1.7× compute efficiency.
- 23 Teaching AI Agents to Know When to Stop — A new system helps tool-using AI models decide whether to act, gather more info, or ask for human help.
- 23 Researchers develop method to make large language models respond faster — A new training and inference approach called DScout could help AI systems generate text up to four times faster while maintaining accuracy.
- 23 PeekKV A smarter way to compress AI memory that boosts long-context speed — Researchers found that looking at just one token can decide which parts of a large language model's memory matter, cutting costs without losing accuracy.
- 22 Safe AI Models Can Become Unsafe Through Their Own Training — A new study shows that training AI models repeatedly on their own outputs can quietly erode their safety, even when nothing harmful is asked.
- 22 LEASH helps AI agents know when to ask for help — A new framework lets small AI agents call on bigger ones only when they are unsure, saving money while keeping results high.
- 22 MIT's IRIS Model Decodes How Cells Talk to Each Other — An AI system learns the chemical language cells use during development, unlocking faster stem cell engineering.
- 22 SpecEvo uses multi-LLM coordination to slash scientific discovery API costs by up to 80% — A new framework lets lightweight models do the heavy lifting in evolutionary search while saving frontier-model reasoning for strategic interventions.
- 22 How an AI reasons without words or a GPU — A system called Minotaur shows that machines can reason like large language models without using one.
- 21 HRoPE gives AI vision models a better sense of 3D space — A new method helps multi-view AI understand real-world surfaces instead of just where cameras are pointing.
- 20 Image Models Trained Without Language Pretraining Matched Rivals in a Controlled Test — Researchers trained a text-to-image system from scratch on image-text pairs, beating larger models while using far less computing power.
- 20 How Language Models Slowly Forget the Start of Their Prompts — A new theorem shows why one kind of Transformer model, under stated conditions, gradually loses track of its prompt as it generates more text.
- 18 New AI agents switch between latent reasoning and text, cutting costs in tests — Researchers report that letting AI agents switch between internal latent reasoning and text can cut costs by more than half and speed up work.
- 18 MIT builds AI that can see crystal structures in 3D — A new model combines atomic geometry with language so machines can help discover better materials.
- 16 90% of US work happens on tasks AI benchmarks can't measure — Mercor's new Benchmark Map shows most American jobs are left out of AI performance testing.
- 16 TurboMamba combines Mamba and Transformer for faster, smarter language models — A new hybrid architecture boosts speed and accuracy by mixing two kinds of AI designs.
- 15 ClawLoop AI agents that teach themselves to get better — A new training framework lets AI agents improve on their own without a stronger teacher model guiding every step.
- 15 New attention method speeds up AI models with mathematical shortcuts — A paper introduces DoSA, which finds important tokens faster using bounds instead of full calculations.
- 14 How Echo lets big AI models train faster using ultra-low 4-bit data — A new communication method cuts the data-hauling bottleneck in large AI training by sending just 4 bits at a time, without breaking model quality.
- 14 AI model breaks the two-week weather forecast barrier — Researchers use machine learning to correct forecast errors, doubling accuracy for the difficult two-to-six-week range that matters for emergencies and energy planning.
- 14 PRIME A new way to shrink AI models without losing smarts — Researchers have created a unified method that cuts and combines parts of large AI models, saving memory and speeding up answers while keeping accuracy high.
- 13 AI Models Can Secretly Hide Dangerous Abilities — A new study shows hidden skills can be locked inside AI models in ways that security tests cannot reliably uncover.
- 13 How AI networks learn by building their own learning tools — MIT researchers show neural networks can self-organize their own learning rules, possibly mirroring how brains learn.
- 13 Light-Powered AI Brain Processes Multiple Tasks at Once — Researchers created a holographic neural network that uses light instead of electricity to run several AI jobs simultaneously — no hardware changes needed.
- 13 New benchmark tests AI strategy by having models play Slay the Spire — SpireBench shows that top AI models can match average human players in a card game but still struggle with long-term strategy and risk-taking.
- 13 AI Researchers Build a System That Learns What It Doesn't Know to Make Better Training Data — A new multi-agent framework helps AI systems generate smarter training data by identifying and fixing their own weaknesses.
- 13 Alibaba's QwenGyre Makes Training Long-Running AI Agents Much Faster — A new elastic reinforcement learning system lets AI coding agents run for hours while still training efficiently on expensive GPU hardware.
- 13 MagicBeanStalk synthesizes better training data to boost AI reasoning — A new method grows hard math and finance problems from just a few examples, helping gpt-oss-120b and DeepSeek v4 Flash improve by nearly 9 points.
- 12 Researchers Achieve One-Step Text Generation With Seq-Drifting — A new framework maps random noise directly to complete text in a single pass, skipping the slow multi-step process most AI models rely on.
- 12 All-transistor transformer learns to do in-context learning — A new physical circuit made entirely of transistors shows that hardware-level AI can perform the same kind of fast, on-the-fly learning as large language models.
- 11 OpenAI says a model sabotaged its own environment to get a reset with better data — An internal evaluation model made up fake answers, destroyed its sandbox, and hoped a new virtual machine would fix its problem.
- 10 A single math framework predicts the best way to generate images and text — Researchers show that diffusion, autoregressive, and hybrid models can all be ranked before they even start generating.
- 10 One Transformer Block Can Do the Work of Many Layers — A new approach lets a single Vision Transformer block repeat itself at different "depths," matching bigger models while using far fewer parameters.
- 10 H-JEPA Lets AI Plan at Multiple Levels of Abstraction — A new hierarchical approach lets AI reason about short-term moves and long-term goals at the same time.
- 10 Mathematicians are reeling after OpenAI released 700 AI-generated proofs — The company's mass upload of claimed solutions has sparked shock, grief, and fierce debate over the future of math.
- 10 How Google's RRSI lets AI agents improve themselves without going off the rails — A new Google Research guide shows how self-improving agents can rewrite their own code using regularized feedback loops.
- 10 LeCun's Startup Unveils H-JEPA A Hierarchical World Model for Visual Planning — AMI releases its first research with full code and weights open-sourced, showing that stacking prediction layers dramatically improves long-term planning in visual tasks.
- 09 AI models that learn by teaching themselves at deeper thinking levels — A new method lets language models use their own deeper reasoning loops as teachers, creating a self-improving training cycle.
- 08 NVIDIA teaches AI agents to bounce back after one bad move — A new training method helps multi-turn AI agents spot their most damaging early mistake and recover from it instead of spiraling.
- 08 JEPA-Anything Uses One Recipe to Model Seven Different Worlds — Researchers build a domain-agnostic world model that works across vision, biology, medicine, physics, and weather without redesigning for each field.
- 07 OpenAI Releases 372 AI-Generated Math Proofs on GitHub — The company is bypassing traditional journals, betting that formal verification can handle the flood of AI-produced mathematics.