Jeff Dean and three legendary Google researchers just launched a startup to automate science itself.
What the article says
- A new company founded by four of Google's most famous researchers wants to automate the whole cycle of doing science. Propose an experiment, run it, check results, repeat, but done by AI instead of people.
- They are starting with machine learning research, using their own tools to speed up their own work before branching into other fields.
- The long term goal is tackling big engineering problems like cheap energy, better medicine, and clean water by running huge numbers of experiments at once instead of one at a time.
- The mission statement carefully avoids buzzwords like AGI, framing the company as a practical way to speed up research rather than a grand claim about superintelligence.
What HN is saying
- Most commenters are stunned by the founders' track record, between them they built core pieces of Google Search, TensorFlow, and much of modern AI, but many doubt the idea will work since real science needs physical experiments and human timelines that no software can rush.
- A recurring argument is that science isn't actually held back by a lack of researchers or automation. It's held back by how funding gets handed out.
- Some read this less as a serious startup and more as a comfortable, well funded passion project for people who no longer need the money.
- The sharpest pushback was someone posting a long list of Jeff Dean's real world misses, challenging the almost mythical reputation he has online.
- Several people also mocked the company's own website for looking AI generated, joking that a firm promising to automate everything couldn't be bothered to polish its own homepage.
A solo artist built a real color space for human skin tones, and the math and the backstory are both great.
What the article says
- He wanted better skin tone options for digital art and game characters, so he built his own color space from scratch.
- He manually sorted a huge pile of colors into plausible or implausible skin tones, then used a statistics technique called principal component analysis to reshape that data into something workable.
- From there he hand fit equations in a graphing tool until a sphere shape lined up with his data, which let him turn any point in that sphere into a real skin color.
- The result is a free color picker and a generator anyone can use, released under an open license.
- He is upfront that his labeling was subjective and one person's judgment, and lists plenty of ways the approach could be improved.
What HN is saying
- Commenters loved the craft here, especially that the equations were fit by hand and eye rather than by formal regression.
- Several pointed to related work worth knowing about, including the Pantone and Monk skin tone scales and a similar crescent shaped pattern found in makeup shade data.
- A recurring worry was bias baked into the manual labeling step, which the creator openly agreed with and wants to fix with multiple expert labelers.
- People debated how far to push the color range, since loosening it lets in genuinely unusual but real skin tones, like conditions that turn skin blue or green, while tightening it risks cutting out real groups such as some Indigenous Australian and East African skin tones.
- One technical correction: the code skips gamma correction when converting colors, which could shift the final results.
Google's shakeup sends Demis Hassabis upstairs and Jeff Dean out the door entirely
What the article says
- Sundar Pichai announced Demis Hassabis is stepping down as head of Google DeepMind to become Chair and Chief Scientist of Alphabet, a broader but less hands on role.
- Koray Kavukcuoglu takes over running DeepMind day to day, now reporting straight to Pichai instead of through Hassabis.
- Jeff Dean, a 27 year Google veteran, and fellow engineer Sanjay Ghemawat are leaving to start their own company focused on automating scientific discovery.
- Google frames it as freeing up Hassabis to focus on AGI strategy while keeping product delivery humming, but it reads like a demotion dressed up as a promotion.
What HN is saying
- Commenters see this as part of a long exodus of top DeepMind talent, alongside earlier departures like Noam Shazeer and David Silver, and read it as a sign something is broken internally.
- Biggest disagreement is what actually caused it. Some blame a culture that rewards selling cloud compute over building frontier models, others point to slow internal tooling or plain corporate bureaucracy dulling the pace.
- Several note Google's stock dropped sharply on the news and joke that Jeff Dean leaving is worth roughly two hundred billion dollars in market value.
- One popular comment mocks the yearly whiplash in HN sentiment toward Google, swinging from doomed to unstoppable and back again like clockwork.
- A few point out Google remains a major investor in Dean and Ghemawat's new venture, so the split may be less clean than it looks.
A coding agent with only four built in tools is beating Claude Code and Codex on cost and quality.
What the article says
- Pi is a coding agent that ships deliberately bare. Just a handful of tools and a tiny system prompt, on the bet that most coding work does not need a bloated harness.
- Databricks tested it on their own huge codebase and found the harness mattered more than expected. Pi paired with a strong model won on pass rate while costing far less than Claude Code or Codex.
- The savings come from sending much less context on every turn, so the agent needs fewer runs to finish a task.
- Shopify used Pi's extensibility to build an automated optimization tool that sped up their test suite and builds, showing minimal does not mean limited.
- The piece argues that as models get better at working in plain terminal environments, a lean harness that stays out of the way beats a heavily customized one.
What HN is saying
- Commenters are split. Plenty love Pi and run it as their daily driver, but several say minimal by default means missing basics like proper keyboard shortcuts and file folder conventions.
- A recurring theme is that Pi's real strength is letting people bolt on exactly the extensions they want, though some warn that half-baked community extensions can be buggy or pointless.
- One detailed comment argues Pi manages context better by keeping a stable recent window instead of compacting everything, which helps it avoid losing track of long running tasks.
- A few people question the article's framing itself, suggesting the polished tone reads like a corporate pitch tied to a recent acquisition rather than genuine independent analysis.
- Others debate cost, noting Pi works well with cheaper open models like Deepseek, while using it with paid subscriptions instead of API keys raises terms of service questions.
Cloudflare just open sourced its internal AI work platform, and the security model is the actual story.
What the article says
- Cloudflare is open sourcing Cloudflare OS, the AI workspace it has used internally since May, so any company can self host it.
- Every conversation is grounded in your company's own context and skills, so people stop re explaining the same process to a model every time.
- Agents and apps can build docs, spreadsheets, and full custom apps, and those apps stay connected to live data instead of becoming static files.
- The real design choice is security. Agents start with zero access, and a go between called a Gatekeeper grants narrow permissions and logs exactly what data an agent has seen, so sharing an app later cannot leak data someone was never allowed to view.
- You can plug in any AI model through Cloudflare's gateway and control which jobs get the expensive model versus a cheap one.
What HN is saying
- The biggest pile on is the name. Commenters are tired of every product calling itself an operating system when it is really a chat app with extra features.
- Cloudflare's Kenton Varda, who built this, shows up repeatedly to defend the design, confirming it is fully open source, self hostable, and built on the same security ideas as his old startup Sandstorm.
- A sharp skeptical thread asks how shared data and updates even work if every user's copy of an app can be independently modified with AI, since two customized copies could drift and conflict.
- Another commenter argues the security promise is overstated, since a document fed into the sandbox could still contain hidden instructions that trick the AI into misbehaving.
- One person doing something similar with plain Docker containers questions whether this really beats a standard sandboxed container setup.