Agent & Model Releases
Inkling 2 Mini open model. Thinking Machines released Inkling 2 Mini, an open-weight 26.4B looped transformer that reuses middle layers three times per token, keeping the checkpoint near 20GB. It runs on a 32GB Mac or DGX Spark and scores strongly on agentic coding benchmarks like DeepSWE and Vals Index v2.2.
GPT-Live-1.1 voice agents. OpenAI added GPT-Live-1.1 to its API for full-duplex multilingual voice conversations, with better accents and task completion. Standby idle time now bills at $0.02 per minute, 60% cheaper than active sessions.
Baseten decision models. Baseten released bison-1 and groundhog-1, open-weight decision models built on Gemma 4 that return calibrated probabilities for yes/no, choice, and score questions. bison-1 tops the Artificial Analysis Decision Index v2.1 and supports 128k-token contexts for routing, triage, and guardrails.
Text-message AI agents. TechCrunch profiles a wave of agents that live in SMS/text messages, including Ask Proxy and Instinct, which can schedule appointments, send emails, and connect to apps without a dedicated app. The piece highlights general-purpose personal assistants and specialized family, travel, and work agents.
Research & Verification
TensorLean GPU kernel verification. Columbia and CertiK researchers introduced TensorLean, a Lean-based framework that formally verifies GPU kernels at PTX level against tensor specs. It proved 83 of 149 kernel instances correct and found 38 bugs across FlashAttention-2/3/4 and DeepGEMM, including a FlashAttention-4 data race missed by 17,760 test runs.
Context windows called overkill. Nvidia researchers argue huge context windows are often expensive and unnecessary, with periodic summaries and targeted lookups performing about as well in many tests. The paper compares today's million-token windows against more selective approaches.
Riemann hypothesis proof race. Anthropic mathematician Levent Alpöge and Claude published a Lean-formalized proof of the Generalized Riemann Hypothesis, while Stanford's Alessandro Masullo independently released a 940-page all-human draft within hours. The AI-assisted proof's Lean code is roughly 250,000 page equivalents, and verification is expected to take years.
Bytedance questions training pipelines. Bytedance and US university researchers argue that nested, multi-phase training pipelines for AI agents are overcomplicated and hard to reproduce, showing that three steps can be enough. They back this with a competitive open model.
Tools & Frameworks
Knowledge factory for agents. LangChain describes a "knowledge factory" approach to collect, curate, and maintain company-specific operational knowledge that agents can reuse, distinct from per-task context. It's used internally for GTM and content studio agents and LangSmith Engine.
GPT-6.2 prompting templates. OpenAI published new system prompt templates for GPT-6.2 Astra and Sol to curb overly brief answers and unnecessary clarifying questions, recommending the models assume context and reuse existing code. API changes include editing conversation history discards earlier reasoning, and temperature/top_p are gone.
Industry & Business
Nvidia and AI selloff. Nvidia reported soaring revenue but its stock plunged 14% Friday amid reports of an SEC accounting review, and it will stop selling H200 chips in China. The Nasdaq 100 dropped 10.4% for the week, its worst since Covid, on doubts about AI buildouts and high valuations.
Anthropic IPO pressure. Anthropic went public on Nasdaq and now faces tension between its safety-first mission and shareholder demands while managing roughly $900 billion in expected compute costs. The Verge discussion examines whether the company can keep prioritizing safety as a public company.
Policy & Privacy
Signal vs AI screenshots. Signal is restricting mobile screenshots by default to counter what Meredith Whittaker calls the "existential threat" of OS-level agentic AI that can scan screens and capture sensitive data. Users can disable the setting, but Whittaker says that breaks privacy guarantees.
UN killer robot rules. UN diplomats failed to agree on binding rules for autonomous weapons, agreeing only to start discussing options and allowing some future regulations to be optional. The Red Cross and others wanted legally binding rules within a year.
That's everything for today – about a 5-minute read.