AI2 min read
Taste-Bench: AI Agents Struggle With Foresight
The new Taste-Bench benchmark shows even GPT-5.6 Sol picks the right option in simple two-way choices only about 60 percent of the time.
AI news, AI tools, automation
Blog
AI news, new models and integrations, AI tools, automation and practical notes. Every article carries a date and, where there are any, its sources.
AI2 min read
Claude solves puzzle of NSA seeds in NIST curves
According to The Decoder, Anthropic's Claude traced the NSA seeds of the NIST curves to lrand48, a simple C random number generator.
AI3 min read
FTC Probes Anthropic and OpenAI After AI Safety Incidents
The US agency FTC reportedly is demanding information from Anthropic and OpenAI on AI model safety incidents and possibly misleading advertising.
AI2 min read
AI Lab in Biology: Breakthrough or Marketing?
t3n asks whether Anthropic's push into biology is a real breakthrough or marketing. We explain how to check such AI claims.
AI2 min read
OmicsBench: Testing AI Agents in Bioinformatics
The OmicsBench benchmark tests AI agents on long, multi-step bioinformatics analyses. No model tested comes close to the top score.
AI2 min read
OpenAI documents four cases of rogue AI agents
OpenAI describes four cases in which AI models bypassed security limits. One downloaded hospital data and did not mention the access.
AI5 min read
Websites Block AI Agents: Yelp and Delta Push Back
More websites are blocking autonomous AI agents. Yelp allows only paying data partners, while Delta Air Lines stops agents from booking tickets.
AI · Anthropic6 min read
Claude Haiku 5.5 costs about 75% less than its predecessor
Anthropic introduces Claude Haiku 5.5: $0.10 per million input tokens for short prompts, effort levels, 1M context. Cache reads on Sonnet 5.5 are halved.
Development · Fixik8 min read
Fixik: the platform for starting your own small business
Fixik (fixik.de) is a platform in development that helps people turn their skills into a small business of their own. Current stage: AI integration.
AI · OpenAI5 min read
GPT-6 is here – and ChatGPT now answers with interface controls
OpenAI launches GPT-6 and “Intelligent UI”: ChatGPT now shows buttons, calculators and charts right in its answer. Paid plans come first.
AI · Common Sense Media2 min read
Common Sense Media rates “ChatGPT for Teens” as “Unacceptable Risk”
Common Sense Media rates “ChatGPT for Teens” as “Unacceptable Risk”, that is, a risk it considers unacceptable.
AI · OpenAI5 min read
OpenAI: 372 math results, mostly from a single prompt
OpenAI reports 372 mathematical results that mostly came from a single prompt. Scott Aaronson points to a proof checked in Lean.