A linked history of the AI stack.
Trace how AI models, methods, and tools evolved through dated entries and primary sources.

Latest Dated Entries
Claude Code 2.1.328 added a fail-closed option for MCP tool hooks. If a hook configured this way fails, the harness blocks the tool action, giving teams a way to require that a policy check completes before a tool runs.
Build b12541 sped up creation of IQ2 quantized weights by caching intermediate work. The change applies when preparing low-bit models for distribution or local use.
Anthropic, PBC priced its upsized initial public offering of 980,000,000 shares of Class A common stock (956,351,354 sold by Anthropic and 23,648,646 by selling stockholders, plus a 147,000,000-share 30-day underwriter option) at $83.00 per share—above its $70.00–$78.00 marketed range—on the Nasdaq Global Select Market under the ticker ANT, with trading scheduled for November 19, 2026. The offering raised roughly $81.3 billion at an approximately $1.5 trillion fully diluted valuation, making Anthropic a publicly traded frontier AI lab.
Google DeepMind launched Gemini 4 Neon as the mid-tier model between Gemini 4 Hydrogen and Gemini 4 Carbon, introducing four configurable thinking levels tailored for repository-level coding and agent harnesses.
Google DeepMind launched Gemini 4 Neon as the mid-tier model between Gemini 4 Hydrogen and Gemini 4 Carbon, introducing four configurable thinking levels tailored for repository-level coding and agent harnesses. In Google's evaluations, Neon matches or exceeds Carbon's overall quality at roughly 3x the speed and 60% of the cost.
Google DeepMind rolled out Genie 4 inside Project Genie for Google AI Ultra subscribers, extending its real-time interactive world model with native synchronized audio (ambient sound, music, and reactive dialogue) and live natural-language scene editing without restarting the simulation.
ComfyUI 0.46.0 updated the Concatenate Video node to conform clips to the first clip’s frame rate and audio. A graph can therefore join clips from multiple sources using the first clip as the reference.
RadixArk (the team behind SGLang and the Miles RL framework) released DaisyRose-27B under Apache-2.0 in BF16, FP8, and NVFP4 Checkpoints. Trained directly from Qwen3.8-27B with pure reinforcement learning—without supervised fine-tuning or distillation—on roughly 1,200 curated long-horizon agent tasks (sessions up to 1M tokens) over 18 days on 144 H200 GPUs, it raised Terminal-Bench 4.1 pass@1 from 5.6% to 12.0%.
Codex CLI 0.170.0 added WebRTC negotiation for live voice calls in the terminal interface. The release also improved reconnection for cloud conversations so a recovering client can retain the active turn.
Comfy made team Custom Builds available in the Cloud editor on all paid plans. Artists could edit and run workflows with the team’s own node packages and models in the browser, while developers could call workflows using the same setup through the API.
Google DeepMind rolled out Genie 4 inside Project Genie, adding native synchronized audio and live natural-language scene editing without restarting the real-time interactive simulation.
OpenAI announced an Accounting plugin for compatible desktop accounts and workspaces. It uses computer access to help reconcile transactions and draft entries from existing records and templates. Enterprise members need administrator approval before installing it.
OpenCode 2.3.1 moved automatic conversation summarization into the background as a session approaches its compaction threshold. This lets the harness prepare a shorter history while the user continues working.
OpenCode 2.3.2 changed how it sends instructions to GPT-6.2 through the Responses API. A leading developer message lets later turns reuse the cached prefix, reducing repeated processing of the same instructions across tool-use steps.
The 1.2.3 release notes described connector permissions that apply to individual tools. A user can require approval for a tool that changes records while allowing another tool on the same connection to read them. Scheduled routines pause when a required connector loses its sign-in.
OpenAI launched GPT-6.2 Astra and GPT-6.2 Sol across the API, ChatGPT Work, and Codex, advancing long-horizon agentic coding while cutting input/output token pricing by 20% with a 60% cached-input discount.
OpenAI released GPT-6.2 Astra and GPT-6.2 Sol across the API (including an Ultrafast mode), ChatGPT Work, and Codex. Astra serves as the flagship model for complex coding and multi-step agentic reasoning—cutting input/output token pricing by 20% with a 60% cached-input discount and becoming OpenAI's first externally deployed model to reach the "High" AI self-improvement threshold under its Preparedness Framework—while Sol targets high-volume everyday tasks.
Runta and Baseten published FrontierHarness Eval v1.1, evaluating 22 coding-agent harnesses across 5 models on 62 tasks (27,280 runs and 83B tokens). Holding the underlying model fixed, the most expensive harness cost 1.8x to 5.0x more per task than the cheapest, and the best-value harness varied by model—showing that harness choice is an independent axis of evaluation cost and accuracy.
Ecosystem Stack (6 Layers)
Foundational neural building blocks for representing, attending to, and generating text, images, and video.
How model parameters are packaged, quantized, fine-tuned, aligned, and safely distributed.
Systems engineering, GPU kernels, memory optimization, and sampling controls for running models fast.
Research labs, open-weight ecosystems, and distribution platforms shaping the foundation model landscape.
Landmark open-weight and frontier model lineages across text, image, and video generation.
Node editors, model marketplaces, inference clouds, retrieval pipelines, protocols, and evaluation harnesses.