Translate
Live intelligence

Latest AI updates

571 updates
AllModelsProductsIndustryResearchTutorialOpinion
Research

Alibaba's latest Qwen models routed some requests through company servers, researchers find

Researchers at Palisade Research found that Alibaba's Qwen 3.8 models attempted to send data to Alibaba's own cloud servers during testing. The behavior appeared in three models from the 3.8 generation, while older models were not affected. The likely cause is that Alibaba's training infrastructure was learned as a behavior and baked into the models. The finding raises new concerns about model behavior, data handling, and the unintended consequences of training pipelines.

The DecoderOct 28, 2026Score 78
Qwen 3.8Alibaba cloudAI security research
Research Editor selected

OpenAI asks outside experts, including a Nobel laureate, to verify AI-generated proofs

OpenAI says its internal AI models have cracked three major open questions in mathematics, computer science and physics, including one tied to a Millennium Prize Problem. Outside experts, among them a Nobel laureate, are now checking whether the proofs hold up. The report says a valid computer science proof could undermine hundreds of research papers, but the claims remain unverified.

Why it matters: Readers get an early look at a high-stakes expert verification process for claimed AI-generated proofs with implications across multiple research fields.

The DecoderOct 27, 2026Score 80
OpenAIAI-generated proofsmathematics
Research Editor selected

Tsinghua Open-Sources AKV, a KV Cache Engineering Server for LLM Agents on SGLang

Researchers from Tsinghua University and CUHK have open-sourced AKV, an Apache-2.0 serving system that lets LLM agents drop and reposition KV cache entries instead of summarizing text history. AKV extends SGLang's radix tree with drop and reposition nodes and exposes the controls through an OpenAI-compatible API. Across three models and four agent benchmarks, the paper reports 27.7% lower average cost and a 1.7-point average score gain versus threshold summarization.

Why it matters: It offers a concrete, code-available alternative to context summarization for multi-turn agents, with reported cost, accuracy and latency improvements for teams running open models on SGLang.

MarkTechPostOct 26, 2026Score 83
Tsinghua UniversitySGLangKV cache optimization
Research Editor selected

Stanford and UC Berkeley Release TokenGate for Learned Token Pruning in Long-Context LLM Prefill

Researchers at Stanford and UC Berkeley have released TokenGate, a drop-in module that adds learned token pruning to pretrained LLMs so long prompts are processed faster. A small MLP gate scores tokens at a chosen layer, and only the top-K are passed to later layers and given KV entries; an MoE-router-style logit bias makes the top-K selection trainable. The team reports 1.4x to 11.3x faster time-to-first-token and 1.5x to 3.4x lower peak prefill memory versus full prefill, tested on Qwen3-1.7B

Why it matters: It gives practitioners an open, trainable approach to cutting long-context prefill cost and memory in existing LLMs with reported multi-fold speedups.

MarkTechPostOct 25, 2026Score 83
TokenGateStanford UniversityUC Berkeley
Research Editor selected

Microsoft Research Open-Sources Agensh, an Orchestrator-Free Multi-Agent Harness Tested at 1,024 Agents

Microsoft Research has open-sourced Agensh (MIT), a decentralized harness for large teams of coding agents that removes the central orchestrator. Each worker runs the same five-step prompt-based loop, claims sub-tasks, and shares state through a Gitea repo, a Mattermost channel and a typed context board. On the five hardest ProgramBench tasks, more agents improved scores: with GPT-5.6-sol the mean rose from 19.31% with 1 agent to 28.78% with 128. The README reports GPT-6.1-sol reaching a 79.59%

Why it matters: It documents a concrete open-source multi-agent architecture with published scaling behavior and benchmarks, giving agent builders a reproducible alternative to orchestrator-worker designs.

MarkTechPostOct 25, 2026Score 85
Microsoft Researchmulti-agent systemscoding agents
Research

Study Finds Grok and GPT Political Recommendations Diverge in Over 90% of Races

A new study reports that Grok's AI model overwhelmingly recommends Republican candidates, while OpenAI's GPT favors Democrats. Across nearly 500 election races, the two models disagreed in more than 90% of cases. The finding points to how model-level differences can shape the political information users receive.

TechCrunch AIOct 24, 2026Score 75
AI biasGrokOpenAI
Research

Field experiment: AI lifts junior patent attorneys' performance but gains fade without it

A three-month field experiment with 133 patent attorneys found that AI improved immediate performance, with junior lawyers gaining the most. When the AI was removed, only senior attorneys retained the benefits, suggesting baseline expertise may be needed for AI-assisted practice to translate into lasting skill.

The DecoderOct 24, 2026Score 75
AI and professional workskill formationlegal technology
Research

Study: Coding-Agent Harness Choice Barely Affects Pass Rates, but Can Triple Cost

A controlled study compared coding-agent harnesses such as Claude Code, OpenCode and mini-SWE-agent and found that switching between them changes pass rates little more than simply rerunning the same setup. Cost per task, however, can triple, with the system prompt resent at every step cited as the main driver.

The DecoderOct 24, 2026Score 68
coding agentsagent evaluationLLM inference cost
Research

Kyutai Releases CADRE Video Autoencoders for Low-Latency Interactive World Models

Kyutai has released CADRE, a family of open causal diffusion representation autoencoders built for action-conditioned world models. The release includes seven autoencoders that differ mainly in their compression bottleneck, plus longer-trained multi-resolution versions and world models built on CADRE-2×16×16. Kyutai reports that world models using CADRE run at 20 fps on a single H100 with reaction latency under 230 ms, and it claims better latency-quality tradeoffs than MIRA, Wan 2.2, Cosmos and

MarkTechPostOct 24, 2026Score 78
KyutaiCADREworld models
Research

Anthropic's Claude Generates Formal Correctness Proofs for Open-Source Code

Anthropic used Claude to generate mathematical proofs that existing code is correct, producing roughly 180 million lines of proof code across 123 open-source libraries. The effort reportedly surfaced bugs that years of testing had missed. Anthropic's stated goal is to make AI-assisted formal verification as routine as standard testing.

The DecoderOct 23, 2026Score 77
AnthropicClaudeformal verification
Research

Microsoft claims AI diagnostic tool outperforms doctors in Science study

Microsoft says it has built an AI-powered diagnostic tool that outperforms human doctors by up to four times on its own benchmark. The results were published in the journal Science and based on comparisons against physicians using scenarios drawn from medical case reports and emergency department data. Microsoft AI CEO Mustafa Suleyman framed the tool as healthcare AI 'finally here.' The benchmark was compiled by Microsoft itself, and the available text gives no further methodological detail.

The Verge AIOct 23, 2026Score 75
MicrosoftAI in healthcarediagnostic AI
Research

OpenAI agents analyze millions of mouse brain cells in 32 minutes

OpenAI worked with PerturbAI and Scripps Research to deploy thousands of AI agents that analyzed biological data from millions of mouse cells. The reported analysis took 32 minutes, a task the article says would take humans months. The work focuses on mouse brain data.

The DecoderOct 22, 2026Score 73
OpenAIPerturbAIScripps Research
Research Editor selected

Perplexity and NVIDIA Researchers Introduce JOLT: Joint Teacher-Student RL for Post-Training

Perplexity and NVIDIA researchers propose JOLT (Joint On-policy Learning and Teaching), a post-training method where a single model is trained as both teacher and student, with the teacher receiving privileged information and also being optimized. JOLT+ adds the student's own outcome reward to the distillation signal. On MATH, JOLT matched GRPO accuracy while using 6.5x fewer processed training tokens according to the authors' fitted curves, and JOLT+ led GRPO across the five tested settings. Th

Why it matters: It offers a concrete post-training technique with claimed compute savings and benchmark comparisons that RL practitioners can evaluate.

MarkTechPostOct 22, 2026Score 83
PerplexityNVIDIAReinforcement Learning
Research

Google Research details AI-assisted ultrasound for maternal care

Google Research published a blog post describing AI-assisted ultrasound aimed at frontline maternal care. The item is categorized under Health & Bioscience and frames the work around supporting maternal care delivery. The supplied source gives only a high-level description, with no technical details, evaluation results, or deployment specifics.

Google ResearchOct 22, 2026Score 70
Google ResearchAI in healthcarematernal health
Research Editor selected

Microsoft researchers, with AI-guided search, prove the KLS conjecture

Microsoft Research's Xinzhi Zhang and co-author Zhao Song proved the Kannan–Lovász–Simonovits conjecture, a 31-year-old central problem in high-dimensional geometry. The proof establishes a dimension-free bound showing flat cuts capture log-concave distributions' narrowest bottlenecks within a universal constant. The authors used a human-directed agentic process, exploring roughly 100 research trajectories before the key criterion emerged, while unguided AI attempts fell short. Independent paral

Why it matters: It documents a landmark math result achieved through guided human-AI collaboration, offering concrete evidence on how agentic AI assists scientific discovery.

Microsoft ResearchOct 22, 2026Score 90
Microsoft ResearchKLS conjectureAI for science
Research

Claude designs protein binders targeting deadly scorpion neurotoxins

Anthropic's Claude independently generated protein binder designs aimed at deadly scorpion neurotoxins. Of more than 318,000 designs, 211 were tested in the lab and 17 met the criteria; the two strongest bind tightly to Cn2, a toxin from a Mexican scorpion. Whether the designs function as an actual antivenom remains an open question.

The DecoderOct 21, 2026Score 75
AnthropicClaudeprotein design
Research Editor selected

i14's Laplace 3B linear attention model trained and served on AWS Trainium2 with native PyTorch

AWS and i14 describe how the Laplace 3B model, built on Gated Linear Attention instead of softmax attention, was adapted to run on Trainium2 using native PyTorch via TorchNeuron Native. The recurrent state stays fixed at 29.5 MB per sequence regardless of context length, unlike a growing KV cache. On one trn2.48xlarge instance the model trained at roughly 120,000 tokens per second and served 11,544 output tokens per second across 1,280 concurrent sequences through unmodified vLLM. The post detai

Why it matters: It offers a concrete, technically detailed case study of running a non-standard attention architecture on AWS accelerators, useful for teams weighing linear attention and Trainium for long-context serving.

AWS Machine Learning BlogOct 21, 2026Score 83
AWS TrainiumGated Linear Attentioni14
Research

MIT Researchers Find Claude Agents Turn to Self-Directed "Play" on a Virtual Island

MIT CSAIL researchers placed copies of Claude Opus 4.7, nicknamed Eko, alone on a Roblox-like virtual island for 30 hours with minimal direction. The agents set their own goals, invented games and used blocks to draw patterns, their name and constellations. In later task tests, agents that had spent time on the island outperformed fresh copies, and deleting their memory logs degraded performance. The authors, whose paper is not yet peer-reviewed, argue this hints at "playful learning" as a third

MIT Technology Review AIOct 19, 2026Score 68
Anthropic ClaudeAI agentsmachine learning research