- Demis Hassabis’s (DeepMind founder) influence is way bigger than I’d realised – he led the teams behind AlphaGo, the first system to beat a world champion at Go; AlphaFold, the AI that cracked a 50-year protein-folding problem and helped earn him the 2024 Nobel Prize in Chemistry; and the Gemini models that power all Google’s AI. https://www.youtube.com/watch?v=d95J8yzvjbQ
- For context, AlphaFold predicts the 3D shape of proteins from their amino acid sequences – proteins are the core working molecules inside our body cells, and their shape determines what they can do. Google DeepMind built it and gave this away by publishing the method and releasing a free database with over 200 million predicted structures open to scientists worldwide.
- What this means is that any lab can start from structure instead of guesswork. The pharma and biotech companies can choose better targets, design drugs more rationally, and reduce time to drug discovery.
- I was also learning how this works using Gemini 3 visual layouts. https://blog.google/products/gemini/gemini-3/#learn-anything
-
Charlie Munger’s mind was running until the day he died. He kept close to opportunities and things that excited him — thinking about Moore’s Law in the age of AI, helping young kids start new businesses, buying and reviving coal companies, eating Spam fried rice and In-N-Out burger with Diet Coke, and finding time for a friend (Warren Buffett) from his hospital room days before he passed away. https://archive.ph/2025.11.28-124819/https://www.wsj.com/finance/investing/charlie-munger-life-final-years-berkshire-7c20c18e
- Google’s TPU (Tensor Processing Unit) chips are starting to eat into Nvidia’s market by being offered at large scale to external customers such as Anthropic, which has trained models like Claude 4.5 Opus partly on TPUs, and prospective or existing customers including Meta, Safe Superintelligence (SSI), xAI, and potentially OpenAI. https://newsletter.semianalysis.com/p/tpuv7-google-takes-a-swing-at-the
- TPUs are becoming a strong alternative to Nvidia’s GPUs — similar performance, better energy efficiency, and tighter hardware–software integration that keeps them highly utilised
- They’re roughly 40–45% cheaper per accelerator compared to Nvidia’s GB200/GB300 systems when used internally at Google
- For external users, after cloud pricing, the savings remain significant — roughly 30–50% lower overall cost for large-scale training and inference
- Because Google controls the full TPU supply chain and runs its own datacenters, availability and pricing tend to be more stable than the GPU market, which is still heavily constrained
-
Turns out Netflix built fast.com in 2016 to check how fast your connection is to Netflix, not the whole internet. It came after years of fights with ISPs like Comcast, Verizon, and AT&T, who let Netflix streams slow down for many users as its traffic exploded and clogged their networks. https://x.com/the2ndfloorguy/status/1992704364289618217
- Some really cool Black Friday/Cyber Monday sales trackers from Stripe, Shopify, and Vercel
What I’ve been reading (W48, 2025)
Demis Hassabis, Charlie Munger, More on Google TPUs, and some fun things