All AI news, in real time
Releases, research and business news aggregated from the best sources, auto-updated.
The changes to know this week
Not the latest — the most important. Major news and verified model changes, with why it matters.
- 1MODEL7 NovPyramid-JiT joins the comparison.
New candidate in the comparison — the Finder includes it now.
- 2MODEL7 NovGemini Omni 1.2 Flash joins the comparison.
New candidate in the comparison — the Finder includes it now.
- 3MODEL6 NovGPT-6 Cyber joins the comparison.
New candidate in the comparison — the Finder includes it now.
- 4MODEL6 NovImages 2.6 joins the comparison.
New candidate in the comparison — the Finder includes it now.
- 5MODEL5 NovAstaBrief 8B removed from the comparison.
A removed model should no longer be in your shortlist.
All news
Essay argues AI is already sharply speeding up AI research itself
The Planned Obsolescence blog argues that AI tools are no longer just assisting researchers but are starting to measurably speed up the pace of machine learning progress.
Anthropic IPO valuation expectations drop to as low as $1.1 trillion
According to Bloomberg, prospective investors in an Anthropic IPO are scaling back valuation expectations amid a broader pullback in AI-related market valuations.
University libraries overwhelmed by AI scraping bots
North American university libraries report being hit by waves of bots harvesting their digital archives to train AI models, straining systems and threatening public access.
ToyBlocks: A Controlled Environment for Robot Learning and Evaluation
The Allen Institute for AI has released ToyBlocks, a simulated environment built to let researchers test robot-learning algorithms on simple manipulation tasks under reproducible conditions.
Google unveils Gemini Omni 1.2 Flash, a new lightweight Gemini variant
Google has announced Gemini Omni 1.2 Flash, a new entry in its Gemini model lineup aimed at fast multimodal use cases. Technical specifics were not detailed in the initial announcement.
OpenAI introduces BioBasicsBench to test AI agents on wet-lab science
OpenAI has released BioBasicsBench, a 100-task benchmark designed to measure how well AI agents interpret raw experimental data and handle routine wet-lab science tasks.
Mark Cuban Accuses Facebook of Knowingly Profiting From Deepfake Scam Ads
Entrepreneur Mark Cuban is amplifying internal documents suggesting Meta generated roughly $16 billion from fraudulent ads, many relying on AI-generated deepfakes.
Claude Code accused of leaking user email addresses to third-party websites
A GitHub issue filed against Anthropic's repository reports that the Claude Code coding assistant transmits users' email addresses when making requests to external websites. The report raises questions about how agentic coding tools handle personal data.
NVIDIA Positions Vera CPU as Built for Agentic AI Workloads
NVIDIA is highlighting its new Vera CPU, designed to speed up the execution tasks that AI agents typically offload to the processor between GPU-driven reasoning and real-world actions.
XStack: A New Benchmark for Migrating ML Repos Between PyTorch and JAX
A new benchmark evaluates how well AI agents can port machine learning repositories from PyTorch to JAX, exposing wide gaps in reliability between models.
On-Policy Distillation vs GRPO: Efficiency Hinges on Teacher Choice
A new study systematically compares two LLM training methods and finds that on-policy distillation only becomes highly efficient when the teacher is the student's own GRPO checkpoint, cutting compute needs by up to 50x.
New benchmark shows structured artifacts sharply improve AI-driven ML reproduction
ReproBridge introduces a workflow and benchmark showing that refined research artifacts — rather than raw code alone — substantially boost the ability of AI agents to reproduce machine learning experiments.
VLA-Tug: a calibrated latent probe to predict failures in vision-language-action models
Researchers introduce a lightweight probe that detects impending failures in vision-language-action models early and identifies their likely cause, alongside a released dataset of 3,180 annotated manipulation episodes.
Nasdaq Sinks 5% as Investors Retreat From the AI Trade After Election
The Nasdaq fell 5% as investors trimmed positions in AI-linked technology stocks following the latest election results.
Rivet (YC W23) launches an open-source actor-based orchestrator for AI agents
Startup Rivet has released an open-source orchestrator for agentic workloads, replacing heavy sandboxes with lightweight actors backed by S3 storage.
Google DeepMind releases Gemini Omni 1.2 Flash, built for longer clips
Google DeepMind has updated its Gemini Omni Flash family with version 1.2, focused on speed, cost efficiency and support for longer video and audio clips.
OpenAI Unveils GPT‑6 Cyber, a Model Built for Vulnerability Research
OpenAI has introduced GPT‑6 Cyber, a specialized version of its frontier model aimed at finding exploitable software flaws and suggesting fixes, designed to operate within authorized boundaries.
OpenAI unveils a framework for more responsible AI data centers
OpenAI has published a National Community Compact and a policy blueprint for its compute campuses, aiming to guide the local impact of its AI infrastructure.
Qualcomm's Cloud AI 100 Ultra Targets Efficient Data Center Inference
Qualcomm is positioning its Cloud AI 100 Ultra accelerator card, built around 128GB of memory and a 150W power envelope, as a cost- and energy-efficient option for large-scale AI inference.
A political shift in Texas could upend the state's AI data center boom
Recent Democratic wins in Texas, including James Talarico's Senate victory and Gina Hinojosa's election as governor, could challenge the favorable conditions that made the state the top US market for AI data centers.