Dex Horthy (HumanLayer) breaks down the “12 Factor Agents” approach to shipping multi-step agentic workflows faster: structured outputs, simple primitives, and custom production loops (running 9 months) instead of heavy orchestration. He also explains “coding vs. shipping,” Riptide architecture, autonomy levels from human review to GitHub Actions, and the “infinite software crisis.”
Sam Hesson (ex-CTO/co-founder of Nanome) explains what “maxing out” Claude’s limits looks like: $50,000 in token spend over 2.5 months that led Anthropic to rate-limit Claude Max. He built an automated CI/CD “Darwinian” system: three parallel agents on separate VMs submit competing PRs, then an LLM judge with a rubric selects the winner.
In this interview with Eric Simons, Bolt’s WebAssembly-powered stack runs a full Node.js environment locally in the browser, delivering near-instant feedback and avoiding cloud VM costs. After ~7 years building a browser IDE and WebContainers, Bolt pivoted to hyper-growth, reaching $700K ARR and redefining how software is shipped.
In AI Tinkerers One Shot, Andrew Baker breaks down browser-native agents, sharing lessons from Claude-for-Chrome controlling a web-based Kid Pix, plus Playwright tricks for remote execution and sound streaming. We explore Stagehand, UI accessibility, personal benchmarks, and the future economics, ethics, and tech of web automation.
Join Lukas Biewald, founder of Weights & Biases and CrowdFlower (Figure Eight), on AI Tinkerers One Shot to explore how tinkering fuels two AI companies, shipping AI in 2025, and practical bets on agentic coding, production observability, and fine-tuning/RL futures; plus lessons from community-driven products. Watch: https://youtu.be/S84CjOrlMcY
Steve Yegge, ex-Head of Engineering at Sourcegraph (makers of Amp), shares practical agentic‑coding workflows and Beads, a Git-backed memory layer for structured work. Explore the 85 percent rule for agent-written code, guiding agents like a team lead, the Land the Plane shutdown protocol, and multimodal debugging.
Innate’s MARS is a 25×25×40 cm open-source tabletop robot that plays chess, learns via a community, and serves as a social lifeline; it runs NVIDIA Jetson, uses Wi‑Fi/WebSocket for real-time control, and a Python SDK on open hardware with modular sensors. Three frontiers: manipulation, agentic real-world frameworks, open-source democratization.
Ishaan Anand rebuilds GPT-2 from scratch inside a spreadsheet, exposing every weight, bias, and operation to visualize tokenization, embedding, transformer blocks and attention in real time; ~600 lines of Python replicate GPT-2 behavior, with adjustable temperature, top-k, top-p. Also highlights innovations since GPT-2: Mixture of Experts, RoPE, training advances.
Three days before GPT-4 Vision, Lukoszko wowed with a real-time desktop vision demo (Bakllava) and pivoted from desktop AI to Stormy AI, automating influencer discovery and outreach. Key lessons: local models struggle financially, Azure/cloud trust wins, distribution moats outrun tech, and OS integration is brutal.
Kyle from OpenPipe will demonstrate building better AI agents using RL and fine-tuning, achieving significant latency and accuracy improvements over larger models.