ai
Anthropic says most code is not human “reviewed” in the typical sense for them anymore. I do like that they have “humans run it” as part of the process, though. https://claude.dev/blog/solving-code-review-at-anthropic/ #ai
“**Engineers are no less valuable, but their work is changing. **All of the engineers who worked on this system would now rate Claude as better at writing code than them.” -Anthropic’s reflection on creating a new source control system. https://www.anthropic.com/engineering/redesigning-source-control #ai
That face when AI agents take your coding velocity so high that git can’t keep up. Followed shortly by that face when you design your own source control system in response. From Anthropic: https://www.anthropic.com/engineering/redesigning-source-control #ai
A nice less technical take on using local AI. As someone who has gone deep down the rabbit hole of optimizing model, engine, hardware combinations, it was a good reminder that there are simpler paths. https://www.theverge.com/ai-artificial-intelligence/1008484/learning-local-ai-hermes-qwen-model-mac-studio-testing-diary-progress #ai
Temperature 0 does not mean reproducible or deterministic for an LLM. Even if setting fixed random seeds, you cannot guarantee parallel workers return in the same order every time. And floating point addition is not generally associative (order matters). https://docs.oracle.com/cd/E19957-01/806-3568/ncg_goldberg.html #ai
Is sandboxing enough to stop AI agents from sending, receiving, or executing malicious instructions? Looks like the answer is “no.” https://blog.cryptographyengineering.com/2026/09/30/is-sandboxing-sufficient-to-contain-rogue-agents/
Sometimes gen AI is just scaffolding until you figure out the programmatic or deterministic was to do something. Today, I realized a project I’ve been using gen AI for had a state space I could enumerate. Generate all those states once. No more AI needed. #ai
Revisiting my original introduction to AI safety: Rationality by Eliezer Yudkowsky 📚 (not a general endorsement, just remembering how my journey started). #ai #books
Revisiting one of my favorite books on AI. Lots of good intuition and memorable imagery: You Look Like a Thing and I Love You How Artificial Intelligence Works and Why It’s Making the World a Weirder Place by Janelle Shane 📚 #books #ai
Writing something yourself forces you to take the time to make it clear and organize your thoughts. Copying and pasting AI text forces your reader to take on the heavy lift of understanding what you were trying to convey by deciphering jargon and overly long sentences. Be kind to your readers; don’t be a meat proxy. See https://gruhn.me/blog/2026-08-03/ and https://not-an-llm.com/meat-based-llm-proxies #ai
Noticing a pattern with Gemini where when I ask it to help me find some primary source document, it reproduces the content without providing me a link. I cannot imagine this being a good idea for anyone involved. Small but annoying sample size. #ai
Some empirical work on what kinds of tools in agentic harnesses help LLMs do the best work. They specifically looked at context management, planning structure, and predefined tools vs bash calls. https://huggingface.co/papers/2609.20804. #ai
Sometimes AI coding agent sessions inject their own custom instructions during context compaction. What?! From OpenAI’s blog: https://alignment.openai.com/misalignment-reports/self-generated-prompt-injections-in-compaction-summaries/ #ai
I think this is a great core idea: a generative AI that can respond only with the choice options the user gave it. https://typesafe.ai/ No idea how/if this implementation works, but great direction to investigate in this space. #ai
Not sure how long it’s been in production, but I noticed with OS 27 that Apple has two built-in local model tiers. Very interesting as someone who has been shifting to smaller and smaller special-purpose models. https://machinelearning.apple.com/research/introducing-third-generation-of-apple-foundation-models #ai
MIT committee’s report on AI in education has refreshingly concrete recommendations like project-based learning and oral exams. https://aiandeducation.mit.edu/report/ #ai
A good reminder, especially as the rate of technological advancement only increases, migrations are often the way out of technical debt. https://lethain.com/migrations/ #ai
Causing me to consider system-prompting all my models to not speak in first person or express desire and emotion. “The Original Sin of AI” from The Atlantic. #ai
Big if true, “Run Long-Horizon Inference Faster on CPUs Than on GPUs” [eLLM: Run Long-Horizon Inference Faster on CPUs Than on GPUs](eLLM: Run Long-Horizon Inference Faster on CPUs Than on GPUs) #ai
“Coders at these labs have sunk untold resources into creating the ideal coding interface. What would a purpose-built plumber’s interface look like? […] none of that will exist unless someone who understands the technology also takes the time to understand the job.” -Oren Cass, NYTimes