Links — #ai

• 4-bit floating point FP4 (Apr 26, 2026)
  John D. Cook on what FP4 numbers actually represent and the variations on the theme — useful background for anyone running quantised models.
  #ai #compsci #math
• Slightly safer vibecoding by adopting old hacker habits (Apr 26, 2026)
  Halvar Flake on bringing classic infosec discipline — checksums, sandboxes, careful diff review — to the new era of LLM-assisted code.
  #ai #security #essays
• AI cybersecurity is not proof of work (Apr 26, 2026)
  antirez argues that volume of LLM-generated security findings is meaningless unless the false-positive rate keeps pace with human triage capacity.
  #ai #security #essays
• AI may be making us think and write more alike (Apr 26, 2026)
  USC research finds measurable convergence in how people phrase ideas after sustained LLM use — and asks what that does to our intellectual diversity.
  #ai #essays
• AI will be met with violence, and nothing good will come of it (Apr 26, 2026)
  On the predictable backlash to widespread AI displacement, and why the response is unlikely to make anything better.
  #ai #essays
• Changes in the system prompt between Claude Opus 4.6 and 4.7 (Apr 26, 2026)
  Simon Willison's diff-by-diff comparison of how Anthropic shifted Claude's default behaviour between releases.
  #ai #blogs
• Coding models are doing too much (Apr 26, 2026)
  Why minimal-editing matters: a model that rewrites untouched code produces noisier diffs and more regressions than one that surgically edits.
  #ai #blogs #code-quality
• Technical, cognitive, and intent debt (Apr 26, 2026)
  Martin Fowler on three flavours of debt LLM-assisted coding can create — and why intent debt is the most insidious.
  #ai #blogs #code-quality
• endless-toil: hear your agent suffer through your code (Apr 26, 2026)
  A tiny tool that streams text-to-speech narration of your coding agent's output. Mostly a joke, partially genuine ergonomics.
  #ai #fun
• How LLMs work: an interactive visual deep dive (Apr 26, 2026)
  An interactive walkthrough of LLM internals based on Karpathy's lecture — tokenisation, attention, sampling, all clickable. Good for non-experts.
  #ai #compsci #tools
• The beginning of scarcity in AI (Apr 26, 2026)
  GPU rental prices surged 48% in 60 days. Tomasz Tunguz argues the bottleneck is shifting from speed of iteration to access to compute.
  #ai #essays
• The cult of vibe coding is insane (Apr 26, 2026)
  Bram Cohen: shipping bad software is a choice, and 'vibe coding' is just dogfooding bad practice and calling it innovation.
  #ai #essays
• The future of everything is lies, I guess: Work (Apr 26, 2026)
  Aphyr on what it means to build software when LLMs can fabricate plausible signals at every layer of every system you touch.
  #ai #essays
• The machines are fine. I'm worried about us. (Apr 26, 2026)
  On AI agents, the grunt work they replace, and the part of doing science that the grunt work was teaching us along the way.
  #ai #essays
• The West forgot how to build. Now it's forgetting code. (Apr 26, 2026)
  How the same skills-erosion that hollowed out Western defence manufacturing is now showing up in software, and why the timelines look identical.
  #ai #essays
• Atlassian enables default data collection to train AI (Apr 26, 2026)
  Atlassian flips the default for customer data to opt-in for AI training. The change is the kind of policy move that previously required, at minimum, an email.
  #ai #news #politics #privacy
• Write less code, be more responsible (Apr 26, 2026)
  orhun on AI-assisted programming as a discipline of restraint — letting the model help, but accepting fewer lines than it offers.
  #ai #code-quality #essays
• Zero-copy GPU inference from WebAssembly on Apple Silicon (Apr 26, 2026)
  Sharing a WebAssembly module's linear memory directly with the Apple Silicon GPU — no copies, no serialisation, real measurements for stateful inference.
  #ai #compsci #web
• Decision Trees (Mar 1, 2026)
  Interactive visual explainer of how decision trees work, from MLU Explain.
  #math #compsci #ai
• Everything changes, and nothing changes (Mar 1, 2026)
  Reflections on AI transforming software engineering while fundamentals persist.
  #essays #ai
• moonshine (Mar 1, 2026)
  Fast, accurate speech recognition optimized for edge devices.
  #ai #tools
• HUMAN=true (Mar 1, 2026)
  Essay on slowing down and letting humans think before reaching for AI.
  #essays #ai
• Large-Scale Online Deanonymization with LLMs (Mar 1, 2026)
  Research on how LLMs can deanonymize users from writing style at scale.
  #privacy #security #ai
• microGPT (Mar 1, 2026)
  Karpathy walks through building a GPT from scratch in a single file.
  #ai #compsci
• pi.dev (Mar 1, 2026)
  Open-source terminal coding agent with plugin extensibility and multi-provider support.
  #tools #ai
• I Am Happier Writing Code by Hand | Abhinav Omprakash (Feb 8, 2026)
  Relatable piece on the downsides of agentic tools, with a technique to escape the slot machine.
  #ai #code-quality
• Nepenthes (Feb 7, 2026)
  A web tarpit for AI scrapers. Generates infinite pages of Markov gibberish to waste crawler resources and poison training data.
  #security #ai
• Life of an inference request (vLLM V1) (Jun 29, 2025)
  How vLLM serves large language models efficiently at scale — at the intersection of AI and systems programming.
  #ai #systems

< all links