AI Updates on 2025-05-26

ByteDance released BAGEL, a ~14B parameter image + text model (7B active) for fast, targeted image edits with text, with fully open weights @deedydas

Alex Graveley released a dataset of 10k prompts refused by Qwen3 but answered by Llama3.3, useful for compliance training, testing, and activation steering @alexgraveley
François Chollet shared a paper reading thread on ARC-NCA: Neural Cellular Automata (May 2025) @fchollet
Nathan Lambert emphasized that working on data is more impactful than working on methods or architectures for AI development @natolambert

Google launched a feature in AI Studio that allows describing a speaker's voice style in plain English, supporting different accents, dialects, tone, and languages through Gemini 2.5 Flash Preview TTS @deedydas
Replit Agent has received significant speed improvements, making it "an MVP agency in your pocket" according to users @amasad
Hugging Face now allows using any Hugging Face space as a MCP server with Local Models, demonstrated with Qwen 3 30B and tiny agents to create images via FLUX @huggingface
Y Combinator launched several AI startups including Nomi (real-time sales copilot), HelixDB (graph-vector database for RAG), Cohesive AI (agentic CRM), and Atlog (AI employee for furniture stores) @ycombinator
Ethan Mollick demonstrated using Google Deep Research to create a historically accurate prompt for Veo 3 to visualize the Colossus of Rhodes @emollick

Big Tech companies are pressuring dev contractors/agencies to cut fixed contract costs by 20-30%, claiming AI efficiency gains, though actual cost reductions may not match these expectations @GergelyOrosz
Google is processing approximately 480 trillion tokens monthly (50× more than a year ago), which is nearly 5x more than Microsoft's reported 100 trillion tokens per month @vkhosla
Amjad Masad is considering changing Replit Agent pricing from constant price per checkpoint ($1/4) to variable pricing proportional to work done @amasad
Experimental work patterns are emerging where senior engineers are removed from IT departments to work directly with subject matter experts using rapid vibe-prototyping to build applications @emollick

Ethan Mollick expressed frustration that Gemini Deep Research can't access Google Books, noting this could benefit scholarship and authors if implemented @emollick
Garry Tan requested that ChatGPT and Claude teams take network failures more seriously, implementing systems that allow retries to work from prior progress @garrytan
Gergely Orosz suggests using a "weird alien" mental model for AI tools rather than thinking of them as interns or junior developers, as they behave fundamentally differently than humans @GergelyOrosz
Chris Olah expressed concern that humanity is failing to bring its intellectual weight to bear on AI safety, noting "the stakes are high and time is short" @ch402