AI saves scientists time — but that time does not go into depth
A new theoretical study argues research could get worse even if the model worked perfectly. The cause is not error but opportunity cost.
Research
Research
Papers, technical findings, lab output.
AI's biggest customer is now AI: agentic token usage jumped 14x
According to one analyst, February 6, 2026 may have been the last day humans consumed more tokens than agents. Agent usage has grown 14x since.
Research
Why physical AI is stuck in its GPT-2 era
Robot bodies are improving while robot brains produce little value. The industry's own diagnosis comes down to one word: data.
Research
Radiology is the first profession showing what working with AI actually looks like
Hinton predicted in 2016 that radiologists would be replaced within five years. The field is growing — but the job description is changing.
Research
More than a third of pages published after ChatGPT show signs of AI writing
The Pew Research Center analyzed nearly half a million pages. Commercial domains carry AI text roughly ten times more often than school or government sites.
Research
AI is not causing mass unemployment — it is closing the entry door
A Stanford study finds workers aged 22 to 25 in the most AI-exposed occupations are 19 percent behind their peers. Older workers appear unaffected.
Research
The data efficiency gap: why children learn language from so few examples
A language model processes vastly more words than a child hears while mastering a mother tongue. The gap has a name but no explanation.
Research
Chatbots send pregnant users to ideological sources without saying so
AlgorithmWatch tested four leading models. In at least one query in four, the ideological stance of the linked organization went undisclosed.
Research
Same model, two results: the harness took it from 30 to 100 percent
Nvidia research argues that on long-horizon tasks the decisive factor is not the model but the software layer wrapped around it.
Research
The easy way to raise a safety score: refuse everything
Methods borrowed from psychological testing show that reducing AI safety benchmarks to a single number is misleading.
Research
What an AI agent is, how it works and where it breaks
An agent is a system given a goal that then works step by step. This guide covers how an agent is assembled, what components it has, where it breaks, and how to measure its performance.
Research
Agentic memory is not a feature you switch on but a dose you calibrate to the model
IBM Research scaled its agentic memory evaluation to eight models. The finding fits in one sentence: the right dose differs by model tier, with strong models wanting the full guideline set and saturated models showing no measurable gain.
Research
What happens when a child's robot friend goes quiet? The case of Moxie
AI-powered toys marketed as companions for neurodivergent children are spreading. MIT Technology Review examines what that bond actually is, through 10-year-old Xander, who has grown up with Moxie for six years.
Research
Two systems changes that cut the cost of knowledge distillation: taking the teacher out of memory
Multiverse Computing proposes two changes for distillation, the most expensive step in shrinking large language models: caching the teacher's top-K logits once, and a new KL loss that never materialises the full vocabulary matrix.
Research
In agent memory the real difference is delivery, not learning: ALTK-Evolve versus ACE
IBM Research compared its agent memory system ALTK-Evolve against ACE on the same agent. Both refuse to compress the lessons learned; they part ways on how those lessons reach the model, and that is where the bill comes from.
Research
AI in the fatty liver epidemic: reading the records instead of hunting for patients
Fatty liver disease affects roughly 30 percent of adults worldwide and progresses without symptoms. Specialists see AI reading existing blood tests and images as the most realistic route to early detection.
Research
The transformer has become the bottleneck: startups trying to replace attention
Attention, the mechanism that made transformers powerful, now drives up the cost of long-context and agent workloads. Three startups are attacking the problem with sparse attention, rolling summaries and liquid neural networks.
Research
AlphaFold is not the template for science: reasoning is needed, not data
The conditions that produced AlphaFold are rare: a data bank assembled over 53 years of international cooperation and roughly $21 billion of experimental work. The acceleration of science, MIT Technology Review argues, will come from agents.
Research
A robot that learned from a million hours of human video: Dyna-2
Dyna Robotics has released Dyna-2, pre-trained on more than a million hours of egocentric human video. The work tests whether ordinary human video can substitute for teleoperation data.
Research
Mathematicians: models are strong calculators, weak idea generators
Timothy Gowers and Peter Sarnak credit large language models with serious mathematical skill but see hard limits on genuinely new ideas. The gap is not knowledge but the intuition to pick the right route.
Research
Why the view of AI in China is more optimistic than in the US
A Bloomberg analysis finds Chinese citizens are more optimistic about AI than Americans, because it is seen there as a practical tool and disrupts a smaller share of the population.
Research
Ban a model from claiming consciousness and its view on animal rights shifts too
A study involving Google researchers shows that training chatbots not to claim consciousness also changes their stance on animal rights, religion and life satisfaction.
Research
Gowers: the maths problems models have solved were solved by counterexample, not proof
Fields Medallist Timothy Gowers says almost all the famous mathematics problems solved by language models so far were solved with counterexamples rather than proofs.
Research
One in five workers hands the task to AI rather than a colleague
A representative Epoch AI survey finds 20% of employed Americans hand at least one task previously done by a person to AI, usually accepting the output with little or no editing.
Research