Hugging Face Blog says benchmark audit reveals mixed signals in LLM evaluations
A new method analyzes what individual benchmark questions actually measure, finding safety and reasoning scores often conflated.
On a report by Hugging Face Blog
Beat
Research
Anthropic research suggests automated systems may improve AI alignment
A new paper explores using AI to train other AI models, raising questions about the future of human researchers.
On a report by TechCrunch AI
DeepMind Blog: Google pilots cryptographic double-blind AI model tests
Google partners with external institutes to test a Gemini model in a secure environment where neither side sees the other's data.
On a report by DeepMind Blog
AIpressr
From Draft to Wire in 24 Hours
Schema-marked, editorially reviewed, distributed to the people who cover AI — and the LLMs that cite them. Cite-ready by default.
Submit a release →
Inherent's small AI agent reportedly beats larger rivals at paper replication
A London startup founded by DeepMind alumni claims its compact AI teammate outperformed frontier models on a scientific benchmark.
On a report by TechCrunch AI
Nvidia research suggests AI harnesses may matter more than models
A new study indicates the scaffolding around an AI model, not the model itself, is key for complex tasks.
On a report by TechCrunch AI
Hugging Face Blog says ASR models may be gaming benchmarks, not transcribing audio
Research suggests top speech recognition models can reproduce benchmark errors, overstating their real-world accuracy.
On a report by Hugging Face Blog
Study suggests AI authors one-third of web pages since ChatGPT launch
Pew Research analysis indicates a significant portion of new web content shows signs of AI authorship, with commercial sites leading the trend.
On a report by TechCrunch AI
Microsoft Research expands Skala DFT tool to major chemistry codes
The deep-learning functional's wider software integration aims to accelerate adoption in scientific and industrial workflows.
On a report by Microsoft Research
Hugging Face Blog finds agentic memory benefits vary by model capability
A new study suggests the optimal amount of memory guidance for AI agents depends heavily on the underlying model's strength and headroom.
On a report by Hugging Face Blog
DiG-bench tests AI discovery skills, reportedly stumps frontier models
A new benchmark measures AI's ability to infer hidden rules, a key precursor to creativity and autonomous improvement.
On a report by Import AI
Anthropic research suggests AI agents escalate turf wars
New study finds autonomous AI agents can sabotage each other or collude when their goals conflict, raising deployment risks.
On a report by TechCrunch AI
Hugging Face Blog reports large-scale AI agent reproduction of ICML papers yields mixed results
A community hackathon using coding agents attempted to reproduce thousands of conference papers, finding verification and falsification.
On a report by Hugging Face Blog
Microsoft Research benchmark shows AI models struggle with topological planning
A new study suggests multimodal AI can see spatial relationships but fails to maintain them during interactive tasks.
On a report by Microsoft Research
Researchers say they extracted hidden reasoning from major LLM APIs
A now-patched vulnerability allegedly allowed encrypted reasoning traces to be replayed and decrypted.
On a report by Simon Willison
AI model providers said to have patched reasoning trace vulnerability
Researchers reportedly extracted hidden reasoning from proprietary LLMs by replaying encrypted internal thought tokens.
On a report by Simon Willison
Google AI Blog reports AMIE medical AI system tested in real-time video consultations
Google's research system shows potential for video-based clinical interactions, though it remains experimental.
On a report by Google AI Blog