My favorite books and reading trace for 2024: BEST BOOKS OF 2024 Bennett, A Brief History of Intelligence: Evolution, AI, and the Five Breakthroughs That Made Our Brains. An exploration of the evolutionary milestones that shaped human intelligence.  The book goes from LUCA to bacteria and bilaterians, tetrapods to amniotes and therapsids, and then cynodonts to mammals and hominids. It had insights into how an understanding of earthly intelligence can inform the development of modern attempts at AGI and then SI. It covers breakthroughs in the evolution of intelligence, such as: steering, emoting, associating, and predicting; reinforcing, TD learning, and world models; simulations, generative models, and model-based RL; mentalizing and self/other modeling; language and large group coordination. I liked this book so much that I bought both a print and digital copy and re-read some chapters and sections multiple times. It itched at and etched on my brain and is a much-discussed book in the Bay Area AI circles. Alexander, Ishikawa, Silverstein, et al., A Pattern Language: Towns, Buildings, Construction. A seminal guide offering 253 design patterns to create balanced and livable spaces, emphasizing that thoughtful design can enhance human flourishing. This was a favorite text from my undergraduate years that I dug up to re-read while doing a painful ADU construction project this year. Darwin, The Autobiography of Charles Darwin: 1809-1882. A memoir revealing the life and intellectual journey of the naturalist. It shows the importance of curiosity, luck, and perseverance in scientific discovery. It is a glimpse into the mind of the most eminent Victorian I know (other than William Rowan Hamilton and the Faraday/Maxwell pair).

Biggest govt news of the day, month, year, and possibly decade. DOGE’s actions will be heavily litigated in federal courts for all the details - but the scale and scope here is to remake the federal govt the way FDR did, and that will cause much turmoil. Let’s also see what Congress does on balancing budgets.

RT @cremieuxrecueil: The biggest news of the day should once again be about DOGE. A new Executive Order was passed a few minutes ago. It…

@johanknorberg 💯- I expect Congress will vote to defund much of US AID and send the good portions (eg Prepfar) for State to run - watch for the upcoming reconciliation bill. Same for DoE, CFPB, and other agencies.

Less-Is-More Reasoning Hypothesis (LIMO Hypothesis) Paper: In foundation models where domain knowledge has been comprehensively encoded during pre-training, sophisticated reasoning capabilities can emerge through minimal but precisely orchestrated demonstrations of cognitive processes. This hypothesis posits that the elicitation threshold for complex reasoning is determined by two key factors: (1) the completeness of the model's encoded knowledge foundation during pre-training, and (2) the effectiveness of post-training examples as "cognitive templates" that show the model how to utilize its knowledge base to solve complex reasoning tasks. To facilitate reproducibility and future research in data-efficient reasoning, we release LIMO as a comprehensive open-source suite at

Welcome to X Ion - glad you’re here! It feels like we’re getting diminishing returns on public data and it’s hard to unlock private data sets - so maybe the right direction is exploration and self play in simulators (or IRL for robots)? For shared infra, maybe Berkeley can build out more to extend PyTorch and the Llama stack? LRM and agentic infra seems sorely needed.

RT @FoundersPodcast: Peter Thiel on the most important lesson to learn from Steve Jobs:

💯- seems like the playbook. Also the President can’t shut down an agency, but has a good case to cut programs for fraud/waste, and then Congress can de facto shut it down with a reconciliation bill by defunding much of it. The carcass still lives, but maybe merged with another entity (eg student loans go to Treasury). What’s happening now is on par with the New Deal under FDR - a remaking of the federal govt and administrative state, which is key if you think the US govt is overspending and is heading to fiscal collapse unless budgets are balanced, fraud and waste is cut down, and agencies slimmed. I don’t like the style and methods here (the namecalling), but the goals seem to be the right ones.

Agree that 3 is the biggest - we will need good policy to guide the decisions of firms. Automation taxes to provide temporary funds for displaced workers, ways to encourage smart automation and augmentation so productivity gains happen and workers get a decent share. Given it’s a collective problem, it’s where smart policies matter. How to bring AI into the world of atoms and empower workers- esp in the long (10-30 year) road to automated factories and robot workers. Ultimately some form of UBI that encourages productive work until we can fully solve economic poverty and scarcity.

RT @paulg: This is the way we all wrote software back when Moore's Law was making CPUs twice as fast every 18 months. You only had to make…

Replace labs with “internet infrastructure providers” and AI models with “the internet” and you’ll see why it’s impossible for anyone to predict where this will go. It’s up to downstream businesses and entrepreneurs to test out thousands to millions of use cases and validate them. Governance is harder - it’s likely we need vertical specific governance run by AI models as legislative bodies are too slow and uninformed.

The specific points I got: 1) Deregulation and experiments will supersede safety for now (recalibrates safety to work on practical problems like child safety, CBRN, etc vs politized ones like misinformation). 2) The US will have end to end ownership of the supply chain (mostly correct, except for ASML - also need to build more US TSMC fabs). 3) The focus will be on augmenting workers for productivity, not replacement (a good policy goal, but I expect there will be both task and job automation, so how we compensate and reskilling workers for the latter will matter). 4) Implicitly rejects AGI and SI rhetoric and favors just productivity and economic gains 5) The US wants AI partners in other countries to take a similar path - focus on building and adoption, cut back onerous tech regulation. 6) A level playing field for everyone, big and small, with skepticism for regulation requests from current incumbents/top players - hence multilateral and open/closed competition, not a handful of closed providers (per the Biden directives).

@chrisalbon Apple Maps has sided with the Resistance, lol.

RT @KayJebelli: Today @ProgressChamber we're launching a new campaign raising awareness on how the EU's Digital Markets act is harming Euro…

RT @AnthropicAI: Pairing our unique data with privacy-preserving analysis, we mapped millions of conversations to tasks and associated occu…

It’s also insane what the big companies have to do - armies of lawyers, PMs, and engineers doing busy work for compliance instead of working on actual product improvements. Also choosing what features to cut off or products to not launch because of ambiguity, legal uncertainty, or capricious enforcement (based on inter-regulator disagreements). The EU needs to bin all their tech regs and start from scratch.

RT @Hesamation: still the most intuitive visualization of the loss landscape in Deep Learning

RT @JohnArnoldFndtn: Can confirm this story. As head trader at Enron when it filed for bankruptcy, I received many calls from firms that we…

RT @sama: Three Observations:

RT @chrisbarber: Yesterday I asked Jeremy Howard: "How might people prepare for AI?" (@jeremyphoward from @answerdotai and @fastdotai) "I'…