Anthropic Puts 5% of Its Shares Into a Public Benefit Fund

Satya Nadella says models should be treated as insider risks, with security controls sitting outside the model itself. On Monday @satyanadella said that principle is going into practice.

@poteto says pstack v0.15.27 ships a new beta skill, /create-formal-verification-skill. It is a meta-skill that creates a new skill and a formal verification suite for the properties you want to verify, such as TLA+ models of concurrent systems or kani checks for undefined behavior in Rust.

Google Research has open-sourced Harness², a framework that lets AI agents improve their own prompts, tools, memory and workflows while the underlying model stays unchanged, @kimmonismus reports. @sudoingX released xpress agent v0.1.0, a local-first agent written in Rust that “learns as it works.”

Three months on, Matt Pocock says his personal wiki has become critical infrastructure for his business. It enriches his courses and makes “agents that run my business” viable.

OpenCode says AI models helping it defend against an ongoing cyberattack began refusing further assistance mid-incident, and access to Daybreak Blue then improved the situation, @kimmonismus reports. He says similar problems have reportedly occurred at Hugging Face.

@kimmonismus, citing The Information, says Anthropic is preparing the release of new model(s). One anonymous tester says the model, asked to fix one bug, can identify the broader class of problems behind it and ask whether to address those too.

OpenCode says DeepSeek V4.1 Flash is updated to 1017 in OpenCode Go and Free, with no changes needed by users, and @MiaAI_lab flagged it as free. Testing three public 2.9 bpw EXL3 quants on 2x DGX Spark, @easthambuilds found her quant of the original V4.1 matches or beats both quants of the 1017 weights on quality.

@MiaAI_lab says ZAI is working on an omni model, called GLM-Omni in a new paper, that can both understand and generate images. The paper, by Wenkai Li, Wenyi Hong, Zhuoyi Yang and Jie Tang of Zhipu AI and Tsinghua University, asks whether visual generation can expand what unified multimodal models know.

@MiaAI_lab published a recipe for running DeepSeek V4.1 Flash 1017 on 2x DGX Sparks with an EXL3-EAQv1 quant at 2.9 bpw: a 1M total KV pool, 512K context, 75 tok/s on prose single-stream, 130 tok/s across four concurrent streams and 2000-2175 tok/s prefill. @u1tra_instinct has already abliterated it behind a simple ABLIT=0/1 flag.

MiniMax has officially joined the “Flash party” with MiniMax-M3.1-Flash, @MiaAI_lab says, and the mystery model Space Bunny was indeed MiniMax. Reading the official benchmarks, @kimmonismus says M3.1 Flash matches DeepSeek V4.1 Flash’s off-peak input/output pricing and halves its peak rates, but charges more for cached input and trails the latest DeepSeek in four of six comparable benchmarks.

@sudoingX got llama.cpp RPC working across 4x DGX Sparks to host Ling 3.1 Flash, a 560B MoE, at Q4 across all four GB10s. That is about $28k of Sparks at today’s $6,950 a box.

@kimmonismus followed Jensen Huang across Europe ahead of GTC Berlin, which starts next week. In London, NVIDIA committed compute access valued at $250 million over five years to UK research on agent safety and security.

Ethan Mollick says a slide now shared by MIT CSAIL has been false since before it was made: automated systems make management decisions all the time in trading, logistics, customer service and insurance, and pretending they don’t “actually launders responsibility.” He points out that MIT and IBM were key to SAGE, which automatically assigned interceptors and missiles against incoming aircraft.