## LLMs
Anthropic’s Claude Mythos Preview dominated headlines: it posts frontier-beating scores (including a new AA-Omniscience high), shows strong alignment in most evaluations, and demonstrates near expert-level security skills such as sandbox escapes and zero‑day discovery—prompting Anthropic to restrict access and spin up Project Glasswing for defender-only deployment. The company also emphasizes that Mythos’ top coding results reflect genuine problem solving rather than memorization. Open source momentum surged as Zhipu’s GLM‑5.1 topped multiple open benchmarks, runs long-horizon tasks, and is available across major platforms; it can iteratively refine its own outputs, marking a step toward self-improving agents. Microsoft open-sourced Harrier, now leading multilingual MTEB‑v2 for embeddings, and a massive 754B‑parameter model landed on Hugging Face, underscoring the scale of community-accessible models. Google’s Gemini 3.1 Pro “Deep Think” mode posted Olympiad‑level math and coding results, while DeepSeek advanced both architecture and UX with its Engram “instant recall” approach and a V4 mode switcher (Fast, Expert, Vision). New research introduced flow map language models for faster, non‑autoregressive generation, and a compact Gemma 4 + Opus reasoning adapter showcased targeted fine-tuning gains. SWE‑1.6 became free in Windsurf, MiniMax‑2.7 open weights are imminent, and even applied math saw a boost as GPT‑5.4 Pro helped prove a new Markov chain result.
## Features
Key AI products shipped substantive capability upgrades. GitHub Copilot CLI now operates fully offline with local models and BYOK, suitable for air‑gapped environments. Tesla’s Full Self‑Driving adopted MLIR, cutting reaction latency by about 20%. Agent platforms leveled up: Fleet and LangSmith integrated thousands of Arcade MCP tools; Deepagents and DeepagentsJS added async subagents, multimodal filesystem support, and smarter caching; and Hugging Face introduced native Agent Traces for better observability. Hermes Agent now self‑learns new skills and integrates Karpathy’s LLM‑Wiki for automatic research vaults in Obsidian. Weaviate added PDF import that auto‑wires agent skills and multivector retrieval. Creative and media tools progressed as Runway’s Seedance 2.0 enabled text/image/audio‑to‑video with multi‑shot scenes, Marble 1.1 improved visual quality and scene scale, and FFmpeg gained fresh patches from Anthropic. Google Maps fused Gemini with live Maps data for richer exploration, while Gemma 4 demonstrated fully offline, on‑device agent workflows that sync and call APIs once reconnected.
## New Tools
A wave of launches expanded what developers can run locally and build with. TriAttention was open‑sourced, letting a 32B‑parameter LLM run on a single RTX 4090 with big speed and memory gains. Langchain‑collapse debuted as middleware to compress long tool‑call traces and curb agent “context bloat.” Research access improved with a pipeline that converts 30k arXiv papers to Markdown for chat‑driven exploration. New generative and perception models arrived: ACE‑Step 1.5 XL for open-source music creation; EUPE, a compact vision encoder rivaling domain experts across tasks; and Allen AI’s WildDet3D for flexible, zero‑shot 3D object detection from text, clicks, or 2D boxes.
## News / Update
Industry moves signaled escalating competition, consolidation, and new safeguards. ARC Prize 2026 participants received a compute upgrade via Kaggle’s L4x4s. The Frontier Model Forum (OpenAI, Anthropic, Google) is coordinating to deter model copying via output distillation, even as separate reporting scrutinized potential military use and called for tighter oversight. Google emerged as the largest owner of custom AI compute since 2022 through TPUs, while Intel, SpaceX, xAI, and Tesla announced the “Terafab” effort targeting unprecedented chipmaking scale. Substrate hinted at masked tools and adopted DeepMind’s AlphaEvolve to compress lithography steps, and Japan’s Sakana AI launched a government-backed project to detect misinformation before it spreads; Sakana also announced a collaboration with Mistral AI. Nvidia showcased an autonomous driving push to rival Tesla’s data lead. Anthropic reportedly blew past aggressive revenue forecasts and, by some accounts, overtook OpenAI’s run rate—amid broader coverage of OpenAI’s internal turmoil. Startups and communities stayed busy: Granola raised $125M for meeting agents, Amazon’s “Project Prometheus” recruited top AI talent, and industry chatter suggested the tinker API could become a de facto standard across inference providers. Google’s on‑device Gemma 4 helped propel the AI Edge Gallery up the App Store charts.
## Tutorials & Guides
New resources mapped practical pathways to stronger agents. A comprehensive survey traced how models evolve from single tool calls to multi‑step orchestration with feedback loops and planning, while fresh overviews cataloged agent capabilities across seven real‑world categories (coding, chat, presentations, customer support, and more). Practitioners emphasized that the fastest path to better agents isn’t always a bigger model—careful design of context, skills, and instructions often yields quicker, compounding improvements.
## Showcases & Demos
Robotics teams released extensive bimanual laundry folding results, including step‑by‑step documentation of system design and thousands of demonstration hours—offering a rare, transparent look at what it takes to achieve reliable manipulation. Nvidia demoed its autonomous driving stack on real streets as it challenges Tesla’s lead. In scientific computing, researchers showcased how a frontier LLM assisted in proving a new result in Markov chain theory, highlighting a growing role for AI in advanced mathematics.
## Discussions & Ideas
Debate centered on how to build durable, capable systems. DSPy advocates argued the discipline is shifting from ad‑hoc prompting to structured, agentic workflows and context programming, an idea echoed by advice to prioritize skills and instructions over model swaps. The Gemma team pushed for broad generalization over benchmark chasing, while others compared vision approaches (promptable concept segmentation vs next‑token prediction) to clarify where each excels. A leaked anecdote showed a simple regex outperforming ML for frustration detection, rekindling the “use the simplest thing that works” maxim. Concerns grew that AI—possibly including open models—could exploit most critical software within months, and that on‑policy RL’s success in coding agents hasn’t yet translated to ML engineering agents. Broader reflections noted the gulf between San Francisco’s AI intensity and slower global adoption, and new research probed how language learning mechanisms might emerge from everyday experience.
