Laplace’s Scissors is an open-source tool for editing large language models’ output weights, built by a 15-year-old using Rust and Triton for CUDA and HIP GPUs. Rather than retraining a model, it searches for rotations of the output matrix that can improve weak or incorrect responses while preserving selected behaviors. Its commands cover fetching and inspecting models, probing output behavior, repairing answers, generating text, exporting edited checkpoints, and comparing behavior across model sizes. Users can provide test cases with prompts, answer choices, and the correct choice.
The probe process injects equal-sized vectors into a model’s decoder and compares the resulting output distributions using Jeffreys divergence to identify directions the model distinguishes poorly. It can trace relevant layer contributions and weight coordinates, then propose a bounded rotation that aims to increase discrimination while preserving specified outputs and hidden-state references. For answer repair, the tool searches rotation planes using expected answer tokens and checks candidates on actual GPU outputs; successful original generations can be protected, while failed answers must meet a specified logit margin over competing tokens. The project also includes scaling analysis and records model and input hashes, hardware, precision, and validation results for reproducibility. It requires a Rust toolchain, Python with Triton, and a compatible CUDA or HIP setup.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.
Lovable says focusing on new apps and a fixed tech stack helped it build a reliable AI coding product, while its team repeatedly redesigned the system as models improved. It replaced most custom agent code with simpler model workflows, using project knowledge files, faster feedback, and computer forking to improve results across more than 100,000 pull requests. (lovable.dev)
Scrine turns books, PDFs, and articles into swipeable cards with one or two sentences at a time. Its AI tutor explains difficult passages, answers follow-up questions, and creates chapter quizzes. (scrine.io)
Apple CEO John Ternus is urging software chief Craig Federighi to accelerate the company’s AI efforts and move beyond incremental Siri upgrades. Apple aims to turn Siri into an autonomous agent as it faces stronger competition from AI rivals. (bloomberg.com)
AI use in software engineering is not a choice between total delegation and total rejection. Engineers should work with AI in ways that preserve hands-on practice and avoid supervising a duplicate version of their own work. (carette.xyz)
Anthropic committed $3.4 billion to secure 1.7 gigawatts of gas-fired generators for its data centres, taking over equipment orders and installation capabilities from supplier OnePWR. The deal reflects AI companies’ push to secure scarce electricity and gain more control over the infrastructure powering their systems. (ft.com)
LLMs often tackle AI problems with brittle symbolic systems instead of using another model call, even when those systems cannot work. Training data and reinforcement-learning environments may discourage models from designing software that uses LLMs. (seangoedecke.com)
Today's best Hacker News stories, summarized and screenshotted, one email a day.