Solo dev builds cyberpunk pixel-art game with Claude Code

Researchers uncovered a technique to subvert safety guardrails on exposed AI models, repurposing them for cybercrime, per Bloomberg's newsletter. Details sit behind a paywall.

Apollo chief economist Torsten Slok calls the AI buildout "a very, very critical driver of growth at the moment" for the US economy, flagging risks from delays in data center construction. He also discusses the prospect of an AI-induced bank run.

SoundCloud artists can export tracks into Moises for mixing, mastering and stem separation; each company's artist subscribers get two-month trials of the other's paid tier. Both frame the deal around an 'artist-first' responsible AI ethos.

A LocalLLaMA user reports that a heavily quantized Qwen Flash Next at IQ2 outperforms a Q8 27B model on complex prompts, contradicting community advice that Q8 would win outright.

TK Elevator built agents on Azure AI Foundry and Azure Databricks that brief service technicians before they reach an elevator, then capture what they learn via a voice debriefing agent. Insights from its Digital Operations Centers feed the agents.

Emmanuel Ameisen, an interpretability researcher at Anthropic, describes a model that knew the answer was 12 but kept outputting 36, then wrote in all caps that it was possessed by a demon. He explains why a model thinking badly is so hard to catch.