Interesting description of fine-tuning large models on consumer hardware answer.ai/posts/2024-03-…

It's great that vibe coding lets non-technical people test out ideas, but troublesome that they'll only discover the limits of the magic through incidents like this

Claude really seems to like to construct Rust code with data structures where it has to remove items using inefficient retain, i.e. map.retain(|key| key.field != field) Gotten that several times, so curious what training led to that

AI coding obliterates the productivity disadvantage Rust had against garbage-collected languages due to its memory ownership tracking. Claude has handled those details flawlessly so far Can see Rust become a more popular choice for higher-level backends going forward

Looks useful for model evals

Lie detector coming to Zoom?

Good article on why multi-threaded languages with shared memory have a performance benefit by enabling local caching They reduced latency 6x by eliminating calls to the kind of network cache that all stateless servers use when their threads can't share an in-process cache

Ah, so the models were trained on new response format. Seems lots of tooling will need updates github.com/openai/harmony

Good paper on low-latency techniques

Both European research and decision-making will rely on and be influenced by whatever models the US or Chinese are willing to share, if Europe doesn't remove obstacles to building tech companies

Interesting research on hacking enterprise agents using prompt injection emails One mitigation is to strictly limit agent capabilities instead of letting them use general tools My spectomcp.updab.com tool helps you get started creating limited MCP servers for API access

Good review of Anthropic's new open source security reviewer tool that runs as a GitHub Action and uses Claude to scan PRs

Interesting view on CoT Reasoning LLMs

Ironic that Gemma 3 270M was the model designed to be small enough to "run everywhere", yet seems to be the only version of the Gemmas that is not available from any provider through OpenRouter

Tried the OpenAI image model and had to verify the account with a passport photo Guess it makes sense that they want to prevent abuse, but it makes it tricky for any product to let users generate AI images with that model if any user can block the API for everyone

AI adoption is tricky when you already have a mature product with high-quality programmed business logic Adoption will likely require new companies to build AI-first products that improve over time A true Innovator's Dilemma disruptive technology fortune.com/2025/08/18/mit…

Suspect a Kafka Streams service with queryable in-memory state would have been a good candidate for accomplishing the same goals as Netflix did with their in-house database infoq.com/news/2025/08/n…