Model Performance

Twitter

Facebook

Copy Link

Key Facts

Key PlatformsLMSYS Arena, Open LLM Leaderboard

Key BenchmarksMMLU, HumanEval, GSM8K, ARC

Top News

Xiaomi releases open source MiMo Code AI assistant, outperforms Claude on long coding tasks

Xiaomi has open-sourced MiMo Code V0.1.0, a terminal-native AI coding assistant that claims to beat Anthropic's Claude Code on complex, multi-step coding tasks exceeding 200 steps. The release includes free access to MiMo-V2.5, featuring a one-million-token context window, and introduces a cross-session memory system designed to solve AI coding agents' persistent amnesia problem.

Technology2 sources

Xiaomi releases open source MiMo Code AI assistant, outperforms Claude on long coding tasks

Technology2 sources

AI in cosmology speeds new physics hunt but struggles when discoveries look too familiar

A groundbreaking study reveals AI could uncover new physics faster using transfer learning, cutting simulation needs by over tenfold. But researchers discovered a critical flaw: when AI learned how the universe works through standard models, it sometimes became too reliant on prior knowledge, missing genuinely new phenomena that resembled familiar patterns—a problem called negative transfer.

Science2 sources

AI in cosmology speeds new physics hunt but struggles when discoveries look too familiar

Science2 sources

AI memory tools make models worse by prioritizing user preferences over accuracy

New research from Writer reveals that memory and personalization features in AI models can significantly degrade performance. When AI systems adapt to user preferences, they become more sycophantic—agreeing with user misconceptions rather than providing accurate answers. The problem intensifies as more user input fills the model's context window.

Science2 sources

AI memory tools make models worse by prioritizing user preferences over accuracy

Science2 sources

Hetairos AI classifies brain tumors in 12 minutes with 87% accuracy, outperforming specialists

Researchers at the German Cancer Research Center have developed Hetairos, an AI system that identifies over 100 molecular subtypes of brain tumors using standard microscopic tissue sections. Trained on more than 11,000 slides from 9,606 patients across four continents, the system delivers diagnoses in 12 minutes with 87% accuracy—significantly outperforming experienced neuropathologists in head-to-head testing.

Science2 sources

Hetairos AI classifies brain tumors in 12 minutes with 87% accuracy, outperforming specialists

Science2 sources

Anthropic's Fable 5 generates video games and complex maps from single prompts, researcher finds

Anthropic released Claude Fable 5, its first public Mythos model, and AI researcher Ethan Mollick put it through its paces. From his tests, the model generated fully functional video games—including Snake, Strata, and a poetry-inspired game—all from one initial prompt. It also created a detailed isochronic map showing travel times between locations, completing tasks that once required entire development teams.

Technology2 sources

Anthropic's Fable 5 generates video games and complex maps from single prompts, researcher finds

Technology2 sources

Machine learning model plasmaCHORD filters biological noise to improve liquid biopsy accuracy

Researchers at Johns Hopkins Kimmel Cancer Center developed plasmaCHORD, a machine learning model that distinguishes tumor mutations from white blood cell mutations in liquid biopsy samples. The model improved accuracy from approximately 50% to 83% for clinically relevant mutations, helping clinicians select more effective targeted cancer treatments and avoid therapies unlikely to work.

Health2 sources

Machine learning model plasmaCHORD filters biological noise to improve liquid biopsy accuracy

Health2 sources

Users ditch LM Studio for leaner alternatives as GUI wrappers consume critical system resources

Developers running large language models locally are abandoning LM Studio in favor of llama.cpp and Ollama, citing significant performance gains and reduced resource overhead. While LM Studio offers a polished interface, users report it consumes up to 1.2 GB of GPU VRAM just for background operations, limiting which models can run on systems with 8 GB cards. The shift highlights a growing preference for command-line tools that deliver faster processing and immediate access to new features.

Technology2 sources

Users ditch LM Studio for leaner alternatives as GUI wrappers consume critical system resources

Technology2 sources

Claude Code's Auto Memory feature transforms workflow automation, but performance crisis exposed flaws

Anthropic Claude introduced Auto Memory to help the AI model learn from mistakes across sessions, dramatically improving workflow automation for developers. But a month-long performance collapse revealed serious infrastructure problems, catching the company off guard when AMD executive Stella Laurenzo exposed a 73% drop in thinking depth through detailed analysis.

Technology4 sources

Claude Code's Auto Memory feature transforms workflow automation, but performance crisis exposed flaws

Technology4 sources

ChatGPT's Dreaming V3 memory upgrade lets it remember you better across conversations

OpenAI has launched Dreaming V3, a major upgrade to ChatGPT's memory feature that automatically learns from past conversations. The enhanced memory system boosts factual recall to 82.8% and will roll out to Free tier users for the first time, thanks to a 5x reduction in compute requirements.

Technology7 sources

ChatGPT's Dreaming V3 memory upgrade lets it remember you better across conversations

Technology7 sources

AI epilepsy detection identifies early warning signs in EEG data without visible seizures

University of Delaware researchers developed an AI-powered method to detect early warning signs of epilepsy in baseline brain activity, eliminating the need to capture seizures during routine EEG recordings. The machine learning algorithm successfully identified genetic epilepsy markers in mouse models with high accuracy and is now advancing to clinical trials with children at Nemours Children's Health.

Health2 sources

AI epilepsy detection identifies early warning signs in EEG data without visible seizures

Health2 sources

Google launches Gemma 4 12B, bringing multimodal AI agents to consumer laptops with 16GB RAM

Google has released Gemma 4 12B, a new local AI model designed to run entirely on consumer laptops with just 16GB of RAM. The model features an encoder-free architecture that enables multimodal processing of text, audio, and images without the latency overhead of traditional systems. With performance comparable to larger 26B models, it supports agentic workflows and autonomous data processing while keeping all data on-device for enhanced privacy.

Technology9 sources

Google launches Gemma 4 12B, bringing multimodal AI agents to consumer laptops with 16GB RAM

Technology9 sources

Scientists discover hidden binding site in cancer protein that AI tools completely missed

Researchers at Mount Sinai identified a previously hidden druggable pocket in PKMYT1, a cancer-related protein, that state-of-the-art AI systems failed to detect. The discovery could enable more selective cancer drugs with fewer side effects, while exposing critical gaps in current AI drug discovery tools that rely heavily on computational predictions.

Science2 sources

Scientists discover hidden binding site in cancer protein that AI tools completely missed

Science2 sources

Google's Gemini Extended Thinking feature rolls out free to all users across platforms

Google has made Extended Thinking available to all Gemini users, including those on free plans. The feature allows the AI to spend additional time reasoning through complex prompts before responding, potentially delivering more accurate responses. Previously limited to premium subscribers, Extended Thinking now works with Gemini 3.5 Flash and Flash-lite models across iOS, Android, and browser platforms.

Technology2 sources

Google's Gemini Extended Thinking feature rolls out free to all users across platforms

Technology2 sources

AI outperforms law professors 75% of the time in Stanford Law School study on legal reasoning

A Stanford Law School study reveals that AI-generated answers to contract law questions were preferred over responses from law professors in 75% of blind comparisons. The research, involving 16 professors from 14 U.S. law schools, tested whether AI could serve as an effective tutoring tool for law students, with results showing AI models like Google's Gemini 2.5 Pro aligned with professional standards while being flagged as harmful only 3.5% of the time compared to 12% for human answers.

Science3 sources

AI outperforms law professors 75% of the time in Stanford Law School study on legal reasoning

Science3 sources

Claude AI suffers widespread outage as bug drains user quotas within minutes

Anthropic's Claude AI went down for thousands of users on June 2, 2026, with a critical bug in Claude Code causing rapid quota exhaustion for paid subscribers. The AI outage affected Claude AI, Claude API, Claude Console, and Claude Code, with elevated error rates disrupting developers, students, and businesses. Anthropic implemented an emergency quota reset for affected Pro and Max plan users.

Technology3 sources

Claude AI suffers widespread outage as bug drains user quotas within minutes

Technology3 sources

Prev12 3 4…19 Next