Anthropic Controversy, Self-Improving Agents, and Model Performance Surge

Key Takeaways

1. Anthropic's Claude Fable 5 restrictions and US export controls spark global AI sovereignty concerns.
2. AI self-improving agents are becoming practical; Anthropic engineers run hundreds of such loops.
3. NVIDIA demonstrates 16x Gemma-4 models on DGX Spark achieving 300 tokens/s.
4. Hamel Husain releases DFlash speculation decoders achieving >1000 tokens/s on B200.
5. Software engineers face class divide due to AI tokenmaxxing; AI labor is commoditizing contract work.
6. Organizations undervalue stronger AI; need flexible architectures to experiment.

1. Anthropic, AI Self-Improvement, and Export Controls

1. Anthropic released Claude Fable 5 with restrictive terms, including banning developers from building competing LLMs and secretly degrading performance, later partially rolled back after backlash. — via 1
2. The US Commerce Department imposed export controls on Mythos and Fable, leading to global disablement of Fable. Andrew Ng argues this demonstrates control over AI access and spurs AI sovereignty efforts. — via 1
3. Anthropic is reportedly seeking a $2 trillion IPO valuation, and a former Google DeepMind AlphaFold lead joined the company. — via 1
4. Anthropic's Claude Code creator revealed nearly 100% of its engineers run 100+ agents with self-improvement loops that improve with each run. Ethan Mollick notes that if self-improvement works (even limited), AI product iteration should accelerate, and Anthropic and OpenAI are doing this while others lag. — via 1 2

2. Model Performance and Inference Optimization

1. NVIDIA AI demonstrated running 16 parallel Gemma-4-26B-A4B-NVFP4 models on DGX Spark (128GB unified memory), each generating 18 tok/s, total throughput 300 tok/s, scalable to 32 instances without flashinfer optimization. — via 1
2. Hamel Husain released six DFlash speculation decoders for Qwen 3.x series, achieving >1000 output tokens/s on B200, and explained their commitment to speculative decoding. — via 1
3. Aravind Srinivas expressed surprise at GLM-5.2, finding it close to Opus 4.8/GPT-5.5 level and often preferable, noting impressive performance with fewer training GPUs. — via 1

3. AI's Impact on Work and Society

1. Deedy describes a class divide among software engineers due to AI tokenmaxxing: "lazy" ones rely on AI without reading/testing, while "craftsman" ones struggle with review burden and eventually give up, especially prevalent in large old companies. — via 1
2. Ethan Mollick cites research showing AI flattens labor performance, making employers value price over human capital signals, thus commoditizing contract labor. — via 1
3. AI excels in specific literary styles (metaphor-rich, short sentences, light plot) often favored in modern short story competitions; Mollick suggests adding AI detection or allowing AI participation. — via 1
4. Enterprises undervalue the benefit of stronger AI if existing weak AI meets KPIs; organizations should build flexible architectures to experiment with smarter models. — via 1