Saturday, September 26, 2026

Making sense of AI tools

DHH on X: "The faster you get to acceptance, the sooner you can move on, and make the most of our new reality. https://t.co/3auQrFl5C8" / X


Not sure if that's good or bad - YouTube

In this video, Maximilian Schwarzmüller discusses how developers should emotionally and professionally approach the rise of AI. His key message is that developers must move past the initial stages of grief—specifically denial and anger—to reach acceptance (0:00–2:45).

  • Acceptance is essential: While AI is transforming coding and makes traditional manual coding less "economically viable," it also empowers developers to build faster and more ambitiously (1:36–1:47, 8:48–9:15).
  • The "Ego" trap: He critiques the notion that AI exists to "shatter developer egos," arguing instead that for many, AI actually inflates the ego by amplifying one's existing building capabilities (3:12–4:55).
  • Adaptability over despair: Rather than fearing job loss, developers should recognize that their core skills—systematic thinking and problem-solving—remain valuable. Success in this new era requires learning how to effectively manage AI agents and staying engaged with new models without feeling overwhelmed (5:35–8:15).
  • Learning through action: The best way to evolve is to build real, ambitious applications rather than just passively watching tutorials or reading about new models (8:17–9:05).

This episode of the All-In Podcast (Episode 290) features a deep dive into the current state of the AI industry, focusing on the tension between "frontier labs" and their corporate realities, the evolving market for AI models, and the recent wave of product releases.

  • Frontier Corporations vs. Labs: The hosts argue that companies like Anthropic and OpenAI should stop branding themselves as "labs" (referencing the Wuhan lab controversy) and accept their status as for-profit corporations subject to product liability and individual accountability (3:54-18:22).
  • The AI Economic Trade: There is significant analysis regarding the massive economic bets placed on AI. The hosts discuss the risks of "token maxing" and the pressure on companies as open-source models begin to offer performant, lower-cost alternatives, creating a bifurcation in the market (20:03-45:17).
  • IPO Outlook: The discussion covers the potential delays of Anthropic and OpenAI IPOs, suggesting that regulatory risks, safety concerns, and the need for more "sober" valuations may force these companies to reconsider their public market strategies (29:22-38:00).
  • Meta's Muse and AI Agents: The launch of Meta's Muse agent is praised for its usability and utility, marking a shift toward consumer-facing AI that actually solves everyday problems (1:07:58-1:18:00).
  • Alignment and Ethics: The hosts critique the current approach to "alignment research," suggesting it is often abstract and misaligned with user needs, and briefly touch on Anthropic's Constitution (1:07:58-1:27:13).
  • Biotech and Research: Friedberg provides a technical perspective on Anthropic's new "wet lab" in San Francisco, explaining that it serves as a validation tool for protein and enzyme discovery rather than high-risk pathogen research (1:27:13-1:32:42).

The AI models and tools mentioned in the video, complete with relevant web links:

  • DeepSeek-V4.1 Flash: A sparse mixture-of-experts model optimized for output efficiency, coding tasks, and reduced Key-Value cache requirements.

  • MiMo Pro / MiMo-V2.6-Pro: Xiaomi's open-weights foundation model featuring a large parameter footprint and 1M-token context length targeting long-horizon agentic workflows.

  • Bonsai 2: A 27-billion parameter open-weights model optimized for local desktop usage by Prism ML.

  • Claude 3 Opus 5.5: Anthropic's flagship model designed for advanced reasoning, code analysis, and agentic workflows.

  • Qwen 2.1 / Qwen Image 2.1: Alibaba's open-weights model competing with existing image and text generation tools.

  • Astro, Soul, and Luna: OpenAI's suite of models designed to balance intelligence with cost efficiency across different workflows.

  • Grok-47 / Grok 4.7: The frontier model series from xAI optimized for coding, reasoning, and long-horizon knowledge work.

  • Muse: Meta's personal AI agent designed for direct consumer interaction, task automation, and everyday utility.

  • Grokbot: A personal AI agent harness used for executing real-world automated tasks.




The discussion regarding the shift from closed to open-source models occurs between (0:41:00 - 0:42:30). During this segment, the hosts highlight a viral chart showing the massive shift in token usage over the past 12 weeks. They note that the market has flipped from an 80/20 split in favor of closed models to an 80/20 split in favor of open/open-weights models. The hosts describe this as a "tidal wave" shift, emphasizing that open-source models are now being widely used for a broad range of AI applications, which presents a significant risk factor for companies like Anthropic and OpenAI.



No comments: