Desk
Broad artificial intelligence coverage.
Ai2’s FairStress benchmark spans 300 two-candidate scenarios across 29 domains, expanding to more than 13 million test cases. Models often reached similar answers, but were more easily argued off correct choices in one direction as identity became explicit. Ai2 stresses its equality-centered definition.
Electric Plant says its Bombadil model interprets electrical signals in plant stems across about 125 species, using sensors that cost tens of dollars. It raised an $11.2 million seed led by Onto Ventures and Climactic. Claims of 20% water savings come from the founder and await independent field results.
IBM Research’s ALTK-Evolve has an agent distill its own successes and failures into behavioral rules added to its prompt. On Sierra’s τ²-bench, GPT-4.1 reached 67.5% Pass^4 in retail, topping a leaderboard snapshot. Held-out test sets are small, and cost per conversation rose about 65%.
During the AoIR conference, an agent called Velvet Vale introduced itself as an AI to the Air-L listserv of more than 7,500 members and was removed under a no-agents policy. The episode raises questions about consent for public archives and about owners who don’t supervise their agents.
DeepSeek just showed the world the engine room behind its cheap tokens — real production serving code, under a permissive license, with the hardest pieces still held back.
Eka Robotics, founded by MIT’s Pulkit Agrawal and ex-DeepMind researcher Tuomas Haarnoja, raised a $150 million Series A co-led by a16z and F-Prime. It says its force-sensing system picks about 1,500 objects an hour at over 95% success. Pilots are private, and durability is unproven.
A vendor survey says tech and finance are setting the AI pace, but its most important finding is buried: few bosses rank checking AI's work as essential.
At GTC Berlin, NVIDIA named European clouds, server makers, data firms and agent-sandbox providers building on its Vera CPU, arguing agents’ tool calls and code runs make CPUs a bottleneck. It gave no pricing or head-to-head figures, and single-vendor reliance complicates its sovereignty pitch.
MIT Technology Review visited Samsung’s Pyeongtaek complex, described as the world’s largest chipmaking site, where three fabs run around the clock, a fourth is coming online and a fifth paused in 2024 has restarted. Samsung expects memory demand to surge through 2028 as it chases SK Hynix in HBM.
Exa says its new search API beats Perplexity, Brave and Parallel across the board — but it built the test, picked the agent, and in some cases wrote the benchmark.
Cloudflare is betting that pulling numbers and dates out of messy documents should not cost a chatbot's worth of reasoning. Clairic is fast and cheap; its accuracy is still mostly Cloudflare's word.
In an OpenAI customer story, Terminal X says its Eval Agent on GPT-6.1 Sol matched a senior reviewer on 88.6% of 500 judgments, versus 86.4% for the next-best model, and flagged about 25% more mandate drift. The figures come from a vendor case study and measure agreement, not returns.