Nature’s most intimate bond! 🐋💙 Have you ever seen a whale nursing? Because their milk is so high in fat (about 35–50%!), it looks like a thick cloud underwater! 🥛✨ It’s incredible how these gentle giants care for their young!

Google is in talks to acquire Mechanize, a startup that creates high quality RL coding environments, for $1.5B. This is the biggest AI data business acquisition since Scale! Mechanize was founded ~1yr ago, and has a team 35-50 who are each paid ~$400k/yr to create ~1 good task/week for a cost of ~$8k/task. In their own words, “almost anything that you struggle to get coding agents to do could be a good task, if implemented correctly”. They were last valued at $500M 3mos ago. Jeff Dean, Google’s Chief Scientist was on the cap table (but it was reported just 15mins ago that he’s leaving Google after 27yrs to his own AI for science startup)! Mechanize were known to be one of the premium sources of coding training data for models and I suspect Google in-housing this is a way to prevent others from getting a hold of this data, in addition to the process talent on the team who come from Epoch AI, the guys who made the FrontierMath benchmark amongst other things. It’s also validation for the dozens of companies in this space that M&A is still very much on the cards if your data is critical enough. Tremendous outcome if it goes through!

Greg Brockman: "We're now in the AGI era." Ten years ago, OpenAI worked out the compute curves and landed on fifteen years to AGI, or ten if the world was willing to build the machines and spend the hundreds of billions to do it. In 2026, GPT-6 Astra manages 24 hours of coherent operation, 10,000 agents worked together to solve Navier-Stokes, and a model ingeniously chained together exploits to break containment at Hugging Face. @gdb joins @bhorowitz and @eriktorenberg on what the AGI era asks of us: why safety and alignment now set the pace, what happens to work, and why AI sentiment is lowest in the country building it. 00:00 Intro 00:52 15 years, or 10 if you spend enough 02:24 Pacing the frontier 03:55 Safety ideas from before the models 08:42 Lessons from Hugging Face 10:20 The defender's window 12:46 10,000 agents on Navier-Stokes 14:50 Formally verifying all software 17:10 Cancelling his holiday for GPT-3 18:42 Codex found 13 holes in 15 minutes 20:41 Why Astra earned the GPT-6 title 24:25 Employment keeps going up 29:39 America has the lowest AI sentiment 31:10 The benefits don't make the news 33:26 Banning data centers exports them 35:50 $1 billion for frontline defenders 38:15 Astra cleared the bar for AGI 40:18 1.5 billion people churned ChatGPT 43:25 Killing Sora 47:45 The AGI era YouTube:

GLM-5.3 Flash on 4x RTX 3090. Vision. 256K context. Good speed. Good quality. Is that even possible? Thats my next milestone. I collected every weight layout, cache budget and runtime blocker I could find. The BF16 source and selective EXL3 Q4 weights are downloading right now. My best bet is hybrid TP4 + EP4. Tensor parallel for attention, dense layers and Vision. Expert parallel for the routed MoE. Then REAP to remove the least useful experts instead of crushing the entire 320B model into one tiny uniform quant. The first candidate is Q4 + REAP60. The quality target is BF16-derived Q3 + REAP50. CPU expert offload stays a fallback because loading the model means nothing if decode is unusable. The final gate is one real image-bearing 256K request with CUDA Graphs, no eager, good retrieval, stable reasoning and usable decode speed. I give the full target a 35-50% chance right now. Lets see how far 96 GB of VRAM can go.