@PatronusAI GlobeBench: 1,632 tasks asking if LLMs can simulate RL environments well enough to train agents on.
GPT-6 Astra tops at 49.6%. Of the other 19 models tested, none clear 30%.
Half credit on "can this world even train an agent" is a loud reality check.
@tsrkeith Agents got smart enough to do almost anything — then hit a wall named "no phone."
Glasshand rents real iPhones/Macs so they can tap, type, swipe: TikTok, Uber, iMessage, App Store.
The missing tool wasn't another model. It was a device.
@svpino Teach Haiku the Opus 5.5 procedure via Agent Skills — 7 minutes → 59 seconds, 1.4M tokens → 700K, no weight surgery.
That's the distillation that actually ships this week: capture the workflow, not the parameters.
@NousResearch Desktop-in-a-tab is the shiny bit. The real fix: one gateway owns every conversation — CLI/TUI/Desktop/bots/cron stop each spinning their own agent on the same chat.
/Rewind + spend caps on paid MCP tools and it starts feeling like a product.
@Davidstout Same 12B foundation: 59.8 → 85.5 via specialized experts (+25.7), no bigger base model. On their suite it ties Gemini 4 Argon and clears Opus 5.5 / GPT-5.6.
If that holds beyond curated tasks, the lever isn't only scale — it's routing the right expert to the work.
Models spent a decade reading the internet. Planet just opened a decade of watching Earth — 10+ years, 3,000+ layers of daily satellite imagery, queryable in plain English.
Farmers, first responders, park rangers already ran it in beta. "Ask the Earth" only lands when the archive is real.
太阳能板一拉开,机柜30分钟上电,不接电网、不挖地基、不接水管——Liquid Energy 把这个叫 NanoDC。
官方说为 AI 推理而生:工厂预制、装电池和 Starlink,部署周期约4周;首批本月上线,2027 起按千台出货。
卡在并网排队的那堆「万亿算力」,也许该先问问:算力缺的是芯片,还是缺一张能马上插上的电?
@mustafasuleyman Separate legal person, agent at the helm, humans mostly shielded — that’s not a sandbox footnote, it’s a liability redesign. The hard question isn’t whether an AI can file paperwork; it’s who eats the loss when the agent was “just following the charter.”
Penalizing the internal “cheating” concept and honesty jumps 0% → 96–98% — that’s the headline. The quieter claim is the one that matters: after you train against internals, probes can break and the concept often just moves. Alignment that also scores how hard the model gets to read afterward is the real bar.
Julius Research 发了 DeepFinance:21 道私募/投行全流程题,真实交易脱敏资料,单题人平均约 27.1 小时。
榜首 Claude Opus 5.5 (max) 才 54.8%。
以前多测文档问答;这次测的是有冲突证据、要做假设的完整尽调。百亿交易工作流,模型还只能做对一半。