The Signal
Open-source models have accelerated to 500% growth in two months. Gemma 4 12B, Deepseek Flash in production, Kimi 2.6 for coding, Minimax for agents—the output velocity now outpaces Anthropic's release cadence. This isn't a niche anymore. It's the dominant direction of travel. Meanwhile, frontier labs are burning capital on scale and latency improvements (110ms voice models) while open-source models solve pragmatic problems cheaper and faster. The market is still pricing this as a curiosity. It isn't.IMPORTANT
Open-source adoption is growing faster than Anthropic—the competitive moat has shifted from model weights to infrastructure and harness.
What's Moving
- Open-source momentum (Gemma 4, Deepseek, Kimi, Minimax) — Production workload adoption, not just benchmarks. The velocity suggests the capability gap is closing in real applications, not just evals. (via @bindureddy)
- 3D generation hitting 10M+ polygons — Rodin Gen-2.5 reaching skin-level detail fidelity in real time. Export to Blender/Unity/Unreal native. This democratizes asset creation for studios and creators. (via @svpino)
- Voice latency at 110ms (sub-human) — Open-weights, 8B parameters, locally runnable. Audio is functionally solved. (via @svpino)
- AI-first product architecture emerging — Nitrosend, others flipping the stack: build for agents first, add human interfaces second. Early signals suggest this breaks traditional SaaS UI assumptions. (via @svpino)
- Federal R&D spending under pressure — NSF down 55% ($9B→$4B), NIH -13%, NASA Science -46%. @ylecun flagging this as foundational erosion: the university basic research that trained today's frontier labs is being defunded.
Crosscurrents
- "Vibe-coding" security debt piling up — @svpino reports companies already losing data because prompt-generated code shipped without defaults (security, observability, VPC). Platforms moving to offer better scaffolding. This is a structural fragility, not a feature.
- Linux preference deepening — @svpino's refusal to touch Windows regardless of petaflops mirrors broader practitioner sentiment. "AI-native OS" skepticism real. The POSIX ecosystem owns the build layer.
- SaaS rent extraction at risk — @sama/@allin both signal niche $30k/year SaaS vulnerable to displacement. High-per-seat costs (not Microsoft's model) are the target.
Tradecraft
BULL
Open-source adoption velocity + practical production use cases = infrastructure layer (compute, hosting, MCP ecosystems) becomes the real moat, not model weights.
BEAR
Federal R&D defunding erodes the pipeline that produces next-gen researchers. Long-term capability edge narrows. This matters in 5–10 years.
WATCH
When enterprise AI initiatives start selecting open-source over closed-model APIs to reduce token spend and latency. That's the repricing signal.
Desk Notes
- @bindureddy — Open-source is outrunning closed labs on velocity; Anthropic's incremental releases are giving away market share to distributed alternatives.
- @svpino — Shipping high-fidelity tooling (3D, voice, agents) that works locally changes the ROI calculus; practitioners are building around Linux and open-weights.
- @ylecun — Basic research funding collapse is the silent risk; frontier labs are spending down intellectual capital accumulated by prior federally-funded work.
- @sama/@allin — Competitive moat has shifted from model to harness (memory, personalization, interface); free users drive scale, paid tiers drive monetization curve.