Ant Open-Sources a Finance Model and an Expert Benchmark

@thatroblennon's PSA: Claude Code caps skill and tool descriptions at 1% of context, about 8,000 tokens by default, and Codex caps it at 2%. Past the cap, the harness quietly strips descriptions off the least-used skills, leaving bare names the model rarely invokes unprompted. Claude's Gmail integration carries 29 MCP tools at about 12,000 tokens; Vercel's official plugin ships 32 tools, 33 skills and 3 agents at around 17,000 tokens plus a 2,700-token injected CLAUDE.md. He ships a prompt that audits the listing and tiers it.

@Av1dlive published a 2,054-word guide to an agent-first business: one canonical channel, a Chief of Staff agent named Mara, five specialist agents, and Joint Channels for work crossing boundaries. In the trial run on Raft, the first artifact passed engineering checks and failed visual QA; the successor passed 89 of 89 engineering checks, visual QA and the product verdict, shipping as RecallFlow v0.2.0. His test for AI-native: turn the AI off for a day, and if the work stops, the company qualifies.

@svpino, in a post marked #ad, points to Cube Sandbox, Tencent Cloud's open-source MicroVM sandbox for agents built on RustVMM and KVM for hardware-level isolation. Claimed numbers: cold start under 60ms, under 5MB of RAM overhead, and tens of thousands of sandboxes launchable within a minute. It adds snapshots with rollback, Kubernetes clusters, ARM support, ingress and egress governance, and cross-machine pause/resume. His two rules: an agent that runs code gets an isolated environment, and so does an agent that controls a browser.

@sudoingX's local arena episode 4 pits Ling 3.0 Flash against Laguna 2.1, two open models that fit one DGX Spark, on the exact same 157-line octopus invaders game prompt in the same Hermes agent harness, both builds recorded end to end. His verdict: laguna takes the clock, ling takes the craft, and he calls @AntLingAGI's open weights some of the most underrated of the year.

@TheAhmadOsman's starter map covers vLLM with PagedAttention, continuous batching, prefix caching and CUDA graphs; SGLang with RadixAttention, speculative decoding and MoE; TensorRT-LLM with FP8/FP4 and Wide-EP; and FlashInfer's kernel library. His eight-project sequence runs from RMSNorm in Triton through paged KV lookup, FP8 KV cache with per-block scales and a toy MoE dispatch, ending with one custom op integrated into vLLM or SGLang and profiled end to end. 'You don't run a model,' he writes. 'You run Kernels.'

@DavidSacks, speaking at the G20 Ministerial with @mkratsios47, made two points about AI data centers: done right they lower electricity costs and bring economic benefits to local communities, and the choice remains with the communities themselves. He cited Executive Order 14365, which established a national AI framework while explicitly carving out data center infrastructure as a state and local decision, and said the federal government is not preempting the states.

@BillAckman shared Polymarket's report that Bernie Sanders unveiled the Ban Artificial Superintelligence Act, which would outlaw AI systems that match or exceed human capabilities, with up to 20 years in prison for violations. In the replies, @GroundhogStrat quotes the bill's push for international agreements to ban superintelligence worldwide, and @andrewsubb_ calls the phrasing vague enough that today's AI could already qualify.