I hope they solve this soon; otherwise, there will be a major impact across many areas. 1. Anthropic’s development and roadmap will be severely impacted. Will Anthropic remove non-U.S. citizens from Fable 5 ? What if there are key researchers working on next-gen Fable? I heard @karpathy is not yet a U.S. citizen. 2. Anthropic’s finances will suffer if they are prevented from commercializing Fable. 3. Does Anthropic even plan to produce the next generation model after Fable? For whom? 4. Will frontier models from other labs such as @OpenAI @GeminiApp, etc. face the same bans once they reach a certain level? 5. What will be the incentive for frontier labs to produce advanced models if they are banned? Will the government subsidize research into advanced models for its own use only? 6. Is this the beginning of regulation in the AI space? I already heard proposals from politicians to prevent models from giving medical advice. 7. Not that I believe AGI is achievable — that is a different discussion — but this obviously goes against research aimed at achieving AGI. 8. What will happen when models from foreign entities achieve the same level? Will they be available to the entire world, or will the respective countries implement similar measures? 9. In the future, will citizens be required to use digital IDs to access advanced technology on the internet? 10. Finally ... were your weekend plans impacted by this decision?

Did you know you can run Claude on a VPS over SSH, let it handle long-running tasks, shut down your laptop, and go to sleep? The easiest way is to run it inside tmux. You’ll be able to disconnect and reconnect later, right where you left off. Ask for details if interested.

AI is enabling software engineers to be real engineers. For decades, software engineers have spent too much time like machinists, hand-crafting every part from scratch. AI is pushing us closer to engineering as it exists in every other field.

Who remembers when “ML” meant machine language, not machine learning? #retrocomputing

Software has a surprising amount of language from EE: probes, signals, pipelines, buses, ports, sockets, buffers, gates, clocks, pulses, noise, impedance mismatch, feedback loops, debouncing, race conditions. We’re all just debugging electrons at higher levels of abstraction.

Enterprise agents are workflow systems, not just LLM calls. You need to add authz-aware tools, state, retries, scoped memory, observability, real workflow evals, fresh data, human approvals, and rollback paths. The model is one component.

The hard parts of enterprise agents are mostly systems problems: authz-aware tool use, state management, deterministic retries, scoped memory, observability, evals on real workflows, data freshness, human-in-the-loop, and rollback paths. The model is only one component.

Paper finds that, in agentic search workflows, grep beats vector retrieval. Bigger takeaway: retrieval quality isn’t just about the retriever. Agent harnesses, tool-calling style, and how results are shown to the model can reshape performance. Paper: arxiv.org/abs/2605.15184

If you have a single task, like renaming files in a folder, you can ask an agent like Claude Code to do it. But if you need to repeat it across many folders, ask the agent to produce a script. You’ll get deterministic results every time, and use fewer tokens.

Just let Claude Code discover my Hubitat home automation system and turn the office lamp on and off. @claudeai @hubitatinc

For those of us who didn’t attend the conference, you can get an M5Stack Cardputer-Adv from Amazon. It's an ESP32 device.

A few models that you can run locally on a modest 8GB VRAM GPU If is to test just one, try gemma4:12b-it-qat. It is the instruction-tuned, quantization-aware-trained variant, so it should follow prompts better than the default gemma4:12b tag in most chat/coding/assistant workflows. QAT means Quantization-Aware Training. In plain terms: the model is trained or fine-tuned while “knowing” it will later be compressed into a lower-precision format, like 4-bit weights. That helps it adapt to the rounding/compression noise before release. lmstudio.ai/models/gemma-4 ollama.com/library/gemma4/ta…