May 6, 2026

Daily

• Accelerating Gemma 4: faster inference with multi-token prediction drafters

  • Multi-Token Prediction (MTP) drafters enhance Gemma 4 models, achieving up to 3x faster inference without compromising output quality, thanks to a specialized speculative decoding architecture.

• Computer Use is 45x more expensive than structured APIs

  • Computer use is 45 times more expensive than structured APIs, as evidenced by a benchmark showing 53 steps and 551k tokens for a vision agent versus only 8 calls and 12k tokens for an API agent.

• Zuckerberg 'Personally Authorized and Encouraged' Meta's Copyright Infringement

  • Meta and Mark Zuckerberg face a lawsuit from five publishers and author Scott Turow, alleging they illegally copied millions of works to train their AI systems, specifically the Llama model, which is described as one of the largest copyright infringements in history.

• Higher usage limits for Claude and a compute deal with SpaceX

  • Anthropic has doubled usage limits for Claude Code across various plans and removed peak hour restrictions, enhancing user experience for dedicated customers.

• Google Cloud fraud defense, the next evolution of reCAPTCHA

  • Google Cloud Fraud Defense is a comprehensive platform that enhances security by verifying the legitimacy of bots, humans, and AI agents, leveraging the same global signals that protect Google's ecosystem.

• Show HN: Tilde.run – Agent Sandbox with a Transactional, Versioned Filesystem

  • Tilde.run enables AI agents to operate in a transactional environment, allowing for easy rollback of any changes, thus ensuring safety when handling real data.

• A Theory of Deep Learning

  • Deep learning theory challenges classical statistical learning by demonstrating that networks can achieve low test error despite overfitting, a phenomenon termed benign overfitting, which contradicts traditional bias-variance tradeoff expectations.

• Learning the Integral of a Diffusion Model

  • Flow maps enable faster sampling from diffusion models by predicting any point on a path between noise and data, rather than just the tangent direction, thus enhancing efficiency in generative modeling.

• Transformers with Selective Access to Early Representations

  • SATFormer enhances the efficiency-performance tradeoff in Transformers by utilizing a context-dependent gate for selective access to early representations, rather than static layer-wise mixing.

• Production AI very different from the demos

  • Production AI reveals a stark contrast to demos, as token usage surged unexpectedly due to longer, more complex customer queries, necessitating context retrieval that doubled input length on each call.

• Mythos is the best cybersecurity news in a decade

  • Anthropic's Mythos AI model is heralded as a transformative tool for cybersecurity, potentially leveling the playing field between attackers and defenders by automating vulnerability detection and patching processes.

• TritonSigmoid: A fast, padding-aware sigmoid attention kernel for GPUs

  • TritonSigmoid is an open-sourced, padding-aware sigmoid attention kernel designed for GPUs, enhancing performance in single-cell foundation models by allowing simultaneous attention to multiple genes without wasting compute on empty positions.

• Visual graph classification for blockchain security: Experiences fine-tuning Qwen2-VL on AMD MI300X

  • Visual graph classification using Qwen2-VL-2B-Instruct effectively identifies malicious transaction patterns in blockchain security, leveraging a Vision-Language Model (VLM) to recognize topological signatures in 2D graph layouts.

• I'm scared about biological computing

  • The emergence of biological computing—where human neurons are trained to perform tasks like playing DOOM—raises profound ethical questions about consciousness and the implications of creating a living computer.