Did Claude Increase Bugs in rsync?
Hacker News (AI keywords)9 days agoBenchmarkThe article analyzes rsync releases to test whether versions containing Claude commits had unusually high bug rates. It uses severity-weighted bugs per 10 commits, exact permutation testing, and Fisher's exact test. With only two Claude-exposed releases, the evidence is limited, but both releases appear within normal historical variation rather than clear negative outliers.
Arithmetic Without Numbers: How LLMs Do Math
Hacker News (AI keywords)9 days agoCommentaryThe article asks whether LLM arithmetic is memorization, heuristics, real computation, or experimental assistance. It summarizes Rune experiments that decode operations and operands from frozen Llama activations, then route them to Python under a no-parser rule. The strongest supported claim is narrow: activation-derived tool arguments worked in scoped audits, while residual-state JIT replacement, long-number generation, and cross-model transfer remain brittle.
Fine-tuning an LLM to write docs like it's 1995
Hacker News (AI keywords)9 days agoTutorialThe author builds a corpus from old Microsoft manuals, cleans OCR text, generates instruction-style JSONL examples, and fine-tunes Llama 3.1 8B and Qwen 2.5 7B with QLoRA. Tests cover malloc(), a fictional Win32 API, and a deliberately anachronistic REST API prompt. Qwen fine-tunes transfer the period documentation style best, but the experiment also shows hallucination risks, tuning complexity, and why these models augment rather than replace technical writers.
I built a vulnerable app and spent $1,500 seeing if LLMs could hack it
Hacker News (AI keywords)10 days agoBenchmarkThe author built a vulnerable React Native app with a Python backend and a Firebase access-control flaw. GPT 5.5 solved 7 of 10 runs, while Deepseek and Claude variants solved fewer attempts. Many other models failed due to refusals, API-focused tunnel vision, false positives, or inability to use the exposed Firebase path correctly.
How LLMs Actually Work
Hacker News (AI keywords)11 days agoTutorialThe article explains how modern LLMs convert text into token IDs, embeddings, and position-aware vectors before passing them through stacked transformer blocks. It covers attention, multi-head attention, KV cache, GQA, feed-forward networks, MoE, residual streams, normalization, and decoding. Its goal is educational: helping readers understand the common architecture behind many current model families and read model cards or papers more confidently.
Microsoft Build: MAI-Thinking-1 and MAI Family Models★ 78
Latent Space11 days agoReleaseMicrosoft used Build to present itself as both an AI platform and a first-party model lab, announcing seven MAI models across reasoning, code, image, transcription, and voice. The standout was MAI-Thinking-1, described as a 35B active MoE with 256K context and clean data lineage. The recap also ties the launches to GitHub Copilot, Windows agent runtime ambitions, Web IQ grounding APIs, Foundry distribution, and MAIA 200 hardware.
CAPTCHAs can still detect AI agents★ 72
Hacker News (AI keywords)16 days agoPaperRoundtable argues that CAPTCHA image recognition is largely solved, but process-level behavior still separates humans from AI agents. Their CogCAPTCHA30 benchmark combines CAPTCHA with cognitive psychology tasks to test not only outputs, but how answers are produced. Results suggest frontier models like Claude, GPT, and Gemini are not necessarily more humanlike than smaller or cognition-trained models.
LLMs believe false statements even after explicit warnings that they're false★ 74
Ars Technica AI16 days agoPaperA new study describes “Negation Neglect,” where LLMs fine-tuned on documents that explicitly mark claims as false still learn the claims as true. Experiments with fabricated statements found models often absorb entity-event associations more strongly than surrounding warnings or negations. The finding raises concerns for fine-tuning pipelines, misinformation handling, and AI safety datasets that include harmful or false content with disclaimers.
Reachy Mini goes fully local
Hugging Face Blog18 days agoHardwareHugging Face published a tutorial for running Reachy Mini conversations without cloud audio processing or API keys. The setup uses its speech-to-speech library as a cascaded VAD, STT, LLM, and TTS pipeline exposed through a Realtime API-compatible WebSocket. Recommended defaults include llama.cpp with Gemma 4, Silero VAD, Parakeet-TDT, and Qwen3-TTS, while allowing swaps to vLLM, MLX, Transformers, or hosted Responses API providers.
Some ideas for what comes next, May 2026
Interconnects (Nathan L.)19 days agoCommentaryNathan Lambert argues that 2026 AI progress is becoming higher-stakes, with model capabilities, work patterns, economics, and real-world risks all escalating. He says open models still lack a true Claude Code and Opus 4.5-style agent moment, and Gemini has no clear competitor to Claude Code or Codex yet. The essay also tracks Mythos, American open-model momentum, frontier-lab competition, and mounting intervention from governments and other power structures.