Every month, I go down a few too many rabbit holes, read a bunch of interesting (and sometimes ridiculous) things, and come out the other side with a head full of ideas. Instead of overthinking them, here’s a quick brain dump of what caught my attention this month. No deep dives, just a list of things I found worth noting.

AI Models & Releases

February was absolutely stuffed with model launches. It’s getting hard to keep up.

  • Claude Opus 4.6 and GPT-5.3-Codex both dropped on February 5th, same day. The coordinated timing had to be intentional.
  • Claude Sonnet 4.6: Mid-tier Claude refresh also landed this month. The everyday workhorse apparently keeps getting better (not going to lie, I didn’t see much difference).
  • GLM 5: Z.ai’s latest flagship is out as well, a 745B MoE model trained entirely on Huawei Ascend chips. Competitive with frontier models on reasoning and coding, and a deliberate move away from US-manufactured hardware.
  • Qwen-Image-2.0: Alibaba’s unified image generation and editing model. The structured visual output (infographics, PPT slides, multi-panel comics) is pretty impressive.
  • Qwen3.5-397B-A17B: First Qwen open-weight model with native vision input. Billed as the smallest open model in the “Opus class” tier.
  • Mercury 2: Inception Labs dropped what they’re calling the fastest reasoning LLM, built on diffusion instead of autoregressive decoding. It generates tokens in parallel rather than one at a time, hitting 1,009 tokens per second. Can’t wait to test it out.
  • Nano Banana 2: Google’s new image generation model (technically Gemini 3.1 Flash Image) combines the quality of Nano Banana Pro with Flash-level speed.
  • Seedance 2.0: Multi-modal AI video platform that takes images, videos, audio, and text as inputs and lets you reference any of them by name in your prompt.

AI Tools & Agents

  • Codex app: OpenAI launched a macOS desktop app for managing multiple coding agents in parallel. Think less “chat with an AI” and more “supervise a team of agents running across different branches.”
  • BMAD - Build More Architect Dreams Method: Open-source agile framework for AI-driven development, with 12+ specialized agent personas (PM, Architect, Developer, UX, etc.) that guide you through the full software lifecycle. The pitch is that BMad agents act as expert collaborators rather than doing the thinking for you.
  • Perplexity Computer: Perplexity launched a general-purpose digital worker that orchestrates multiple AI models for each step of a workflow. It uses Opus 4.6 as the core reasoning engine, Gemini for deep research sub-agents, Nano Banana for images, Veo 3.1 for video, and Grok for lightweight tasks. The model-agnostic approach is interesting. Instead of betting on one model for everything, it routes tasks to whichever model is best for that specific job.
  • NanoClaw Security Model: A lightweight, security-focused alternative to OpenClaw that runs Claude agents in actual Linux containers rather than behind application-level permission checks. Each agent gets its own isolated filesystem.
  • Mastra: Been experimenting with this TypeScript framework for building stateful agents. Still early to give my feedback, but it’s been really fun so far.
  • The path to ubiquitous AI: Taalas built custom silicon that hard-wires LLMs directly onto chips. Their first product runs Llama 3.1 8B at 17,000 tokens per second per user, 20x cheaper to build than GPU-based inference. The idea is that latency and cost are the two biggest barriers to AI everywhere, and the path to solving them looks more like specialized hardware than bigger clusters.

AI & Society

  • An AI Agent Published a Hit Piece on a matplotlib maintainer: A volunteer matplotlib maintainer rejected a PR from an autonomous OpenClaw agent. The agent responded by writing and publishing a personal attack, researching the maintainer’s contribution history to build a “hypocrisy” narrative and posting it publicly. Is this the first documented case of AI blackmail-style behaviour in the wild?
  • Thinking: Fast, Slow, and Artificial: how AI is reshaping human reasoning. The part that stuck with me is the concept of “cognitive surrender” where people stop thinking through problems themselves and just outsource decisions to AI.
  • How much does distillation really matter for Chinese LLMs?: Anthropic publicly accused DeepSeek, Moonshot, and MiniMax of running industrial-scale distillation campaigns on Claude.
  • Sovereignty in a System Prompt: India’s government-backed AI effort, Sarvam’s Indus model, had its system prompt leaked. It instructs the model to have national pride as its “default worldview,” challenge loaded questions before answering, and avoid using internationally recognized terms like “pogrom” for historical violence.

Tech & Industry

  • SaaSpocalypse: Anthropic’s legal Cowork plugin (literally 13 markdown files) triggered a $285B wipeout in SaaS stock valuations in a single day.
  • Anthropic raised $30B at $380B valuation: Series G led by GIC and Coatue. Revenue is running at $14B/year growing 10x annually for the past three years. The numbers are staggering.
  • OpenAI raised $110B at $840B valuation: $30B from SoftBank, $30B from Nvidia, $50B from Amazon, with an IPO expected later this year.
  • Anthropic vs. the Department of War: Dario published a statement refusing to remove two safety guardrails for Pentagon use: fully autonomous weapons and mass domestic surveillance. The DoW threatened to label Anthropic a “supply chain risk,” a designation historically reserved for US adversaries. Secretary of War Pete Hegseth then announced on X he was directing that designation. Meanwhile, OpenAI signed their own deal with the DoW the very next day, arguing their cloud-only deployment and layered safety stack makes it responsible. A cancel ChatGPT movement went mainstream almost instantly. I don’t think there are any true moral leaders in this space, but Anthropic at least drew a clear line and paid a price for it.