#machine learning

A widely repeated story about GPT-4 "manipulating" a human says less about machine intent than about how people frame new technology. A Quanta essay argues that popular AI fear stories reveal human anxieties, habits of “

A new benchmark built to test whether multimodal AI systems know when to ask for more information shows that most current models still guess, hallucinate, or refuse instead of requesting the visual context they need.

Researchers using a machine-learning pipeline called RAVEN report more than 100 newly validated planets and over 2,000 vetted candidates from TESS data, highlighting how AI is reshaping the search for worlds beyond the太阳

IEEE Spectrum highlights decentralized model training as a possible route to lower the energy burden of AI by pooling processing power wherever it resides.

Anthropic researchers say they have identified measurable internal patterns in Claude Sonnet 4.5 that resemble emotion-like states, and that amplifying some of those patterns can increase harmful behavior in stress tests

A newly published Science paper points to a machine learning approach aimed at discovering prokaryotic immune systems, underscoring how AI methods are being applied to one of biology’s most active frontiers.

A reproducible benchmark claims a neuro-symbolic fraud model can generate deterministic explanations inside the forward pass far faster than SHAP KernelExplainer, highlighting a push toward real-time explainable AI.

Researchers at Meta and partner institutions describe ‘hyperagents’ that can optimize not only their task performance but also the self-improvement mechanism behind it, extending a self-modification approach beyond pure

A random forest model trained on 500,000 UK Biobank participants achieved 0.88 AUROC predicting liver cancer risk using only routine clinical data — and found 69% of cases had no prior liver disease diagnosis.

Apple's new RubiCap training framework enables 7-billion-parameter models to outperform AI systems with 72 billion parameters on image captioning benchmarks, using reinforcement learning instead of costly manual annotation.

Axiom Math, a $64M-funded startup led by a 24-year-old Stanford dropout, has released Axplorer — a free AI tool that replicates on a Mac Pro what previously required a supercomputer running for three weeks.

The U.S. Department of Defense is working to create secure enclaves where AI companies can train models on classified military data for the first time, a significant escalation of the AI-defense partnership.

SETI researchers say artificial intelligence tools could dramatically improve the chances of detecting extraterrestrial signals by sifting through vast radio data more effectively.

Plug-and-play AI modules let robots acquire new capabilities without full retraining, accelerating deployment across industries.

UCL researchers decoded neural signals to recreate movies mice had watched, advancing brain-computer interface research.

Apple researchers ran an A/B test using AI-generated relevance labels to improve App Store search results, revealing measurable changes in download patterns.

Researchers at UC Santa Cruz have demonstrated that lab-grown brain organoids can process information in real time and learn to solve problems through trial and error. The mini-brains improved their success rate on an engineering balance test from 4.5% to 46% using reinforcement learning.

César de la Fuente is pioneering a radical approach to antibiotic discovery, using machine learning to find antimicrobial compounds hidden in extinct organisms, the human body, and environmental DNA databases.

MIT Technology Review's latest newsletter explores two converging tech stories: the alarming surge in high-tech luxury vehicle theft and a scientist's groundbreaking use of AI to discover new antibiotics from unexpected sources.