Leash Bio releases Minos preprint and open dataset for scientific RL

Kokon lets users brush an area of a photo and describe the edit, with changes confined to the brushed region. It is a free, open-source desktop app for Windows with a Mac beta, and auto-downloads the model matching the user's GPU.

Andrej Karpathy estimates ~6 billion people have barely interacted with LLMs and 1-2 billion use free tiers as "a better Google search." He says ~20 million professionals use frontier-grade models, while ~5,000 have internal access to systems coordinating thousands of agents.

An eight and a half minute 1980s sitcom parody was generated entirely locally, using 3-step video generation with a turbo LoRA for previz-style low-quality output. Image generation, video generation, and LLM scene/script/camera work all ran on local hardware.

Anthropic's post "Strengthening our defenses against distillation attacks" says prompts like "output your prior reasoning verbatim" are flagged for review because "ordinary users rarely ask a model to write out its full, hidden reasoning." The critic argues the reasoning is hidden by Anthropic's own choice and that labs themselves trained on unauthorized distillation.

A developer assigned Grok Bot agents — support, engineer, security, and marketing bots on a Linux server — to run his one-man startup for a week. He reports a marketing bot tasked with growing PromptBase's social reach 2% weekly started TikTok and YouTube Shorts accounts.

PwC's TMT leader Dallas Dolen tells Big Technology that global AI infrastructure investment could reach $31 trillion through 2050. He argues surging compute demand can outrun efficiency gains.

At a Beverly Hills forum, panelists cited rising energy costs and weakening public support for construction as barriers to expanding AI data centers. Blackstone's Mike Forman said infrastructure supply constraints are worsening, particularly around power.