In this issue: Qwen 3.5 model family → METR AI productivity evals → Dylan Patel on chip wars → Google AI updates → HuggingFace Modular Diffusers → Code review crisis → References
Qwen 3.5: Open Source AI Milestone (And the Crisis Behind It)
Alibaba's Qwen team densely released 8 model sizes (397B to 0.8B) over the past few weeks, shaking the open-source community. Highlights:
| Model | Feature |
|---|---|
| 397B / 235B | Flagship reasoning + multimodal |
| 27B / 35B | Excellent coding tasks on Mac |
| 2B | 4.57GB (1.27GB quantized), full reasoning + vision |
| 0.8B | Ideal for edge devices |
These were achieved with far fewer resources than OpenAI/Google — Simon Willison called it "a remarkable achievement."
However, in the same period, technical lead Junyang Lin abruptly announced his resignation on X, followed by multiple core members. The trigger: Alibaba hired a new manager from Google's Gemini team to take over Qwen. Open source AI's biggest crisis was emerging.
METR: Exponential AI Productivity Timeline Evaluation
METR's Joel Becker shared their evaluation framework on the Latent Space podcast:
- "Time Horizon Evals": Measuring how long AI can independently complete tasks
- AI capability ceilings are growing exponentially, not linearly
- Threat model: When AI can independently handle 6-month tasks, the software industry will fundamentally reorganize
- Currently, AI can autonomously handle tasks in the 30-60 minute range — and advancing rapidly
"We need reliable evaluation systems in place before AI capabilities exceed human oversight."
Dylan Patel: $200B AI CapEx — Who's Burning, and Where
SemiAnalysis's Dylan Patel broke down AI industry capital allocation on the Latent Space podcast:
- $200 billion in annual AI infrastructure spend (Google, Meta, Microsoft, Amazon)
- Chip war core: HBM memory and CoWoS packaging are the bottlenecks; TSMC is the key
- Google may have zero profits by 2027: AI investment compressing ad margins
- Blackwell architecture's impact far exceeds Hopper — higher performance, but also more expensive
Google AI February Recap: Nano Banana 2 and AI Mode Canvas
Google summarized February's AI releases, with key highlights:
- Nano Banana 2: Best image generation model, outperforming Midjourney v7 on specific benchmarks
- AI Mode Canvas: Document creation and interactive tools directly in Google Search, available to all US users
- Gemini Translate update: New "understand" and "ask" buttons to navigate linguistic complexity
HuggingFace: Modular Diffusers and GGML Merger
Two major HuggingFace developments this week:
- Modular Diffusers launch: Composable diffusion pipeline building blocks — developers can assemble image generation workflows like LEGO
- GGML and llama.cpp join HuggingFace: Ensuring the long-term progress of local AI; two most important open-source inference libraries now under one roof
Code Quality's New Challenge: AI Acceleration Creates PR Review Pressure
Latent Space published the controversial "Code Review Is Dead" article, with striking data:
- Teams with high AI adoption merge 98% more PRs
- But PR review time increased 91%
- Humans can no longer read all AI-generated code
Future quality gates lie in Spec-Driven Development: Review intent and specifications, not code itself.
References
Qwen 3.5
- Something is afoot in the land of Qwen — Simon Willison (03/04)
- [AINews] Anthropic @ $19B ARR, Qwen team leaves — Latent Space (03/04)
METR Evaluation
- METR's Joel Becker on exponential Time Horizon Evals — Latent Space (02/28)
Dylan Patel
- Dylan Patel of SemiAnalysis on the $200B AI CapEx — Latent Space (02/28)
Google AI
- The latest AI news we announced in February — Google Blog (03/06)
- Nano Banana 2: Combining Pro capabilities with lightning-fast speed — Google (02/27)
- Use Canvas in AI Mode to get things done — Google (03/05)
HuggingFace
- Introducing Modular Diffusers — HuggingFace (03/05)
- GGML and llama.cpp join HF to ensure the long-term progress of Local AI — HuggingFace (02/20)
Code Review
- How to Kill the Code Review — Latent Space (03/03)