I’d add to @karpathy’s point that LLMs are also optimized for being downstream of humans rather than upstream.Like domesticated dogs or horses, they have been bred for alignment, to do what humans tell them to do. ...

Animal evolution: survival of the fittestDigital evolution: survival in the marketAndrej Karpathy (@karpathy):Something I think people continue to have poor intuition for: The space of intelligences is large and animal intelligence (the only kind we've ever known) is only a single point, arising from a very specific kind of optimization that is fundamentally distinct from that of our ...

Investment Strategy - Optionality is not the only key to success; long-term relationships with a single party can be more beneficial due to compounding effects [1] - Balancing exploration (optionality) and exploitation (commitment) is crucial, similar to the multi-armed bandit problem [1] - Optionality is considered a highly underrated form of wealth applicable to various aspects of life [1]

Also holds for supply chains.If you have at least two independent vendors, you have one reliable supply chain.But if you have only one vendor, you may end up with zero margin.So: if you have two, you have one, and if you have one, you have zero. ...

The lesson of deals is that if you have two, you have one. And if you have one, you have zero.In other words: the second bidder sets the price and ensures the deal goes through. If you only have one bidder, the price floor is zero. And the deal may not get done at all. ...

No one is explaining the joke, so I will. It’s multi-layered.First, there was a fun Oct 2017 Twitter exchange about the right way to render a burger emoji. Obviously this was trivial, but Sundar said “we’ll drop everything else we are doing and address on Monday” as a joke.The new post uses that old reference to show off Gemini 3’s image rendering capabilities. Normally, AI models struggle with spatial orientation, particularly with respect to the relative position of objects. They also struggle with text ( ...

I don’t think the goal is remaining “pure” forever, particularly if that’s defined as never becoming useful enough to make money.Instead, if an organization truly starts losing sight of its original purpose, it is eventually replaced. If however it is correctly *balancing* that purpose with practicality, it’s not easily replaced.Nayak Satya (@NayakSatya_SG):@balajis But don’t the art/ideology companies either turn into cash machines the instant payroll hits 9 figures, or die broke ?Name one that stayed pure ...

Motivation & Purpose - Companies like Salesforce and Oracle were built with money as a primary motivator, while Apple and Twitter were driven by artistic vision, SpaceX and Tesla by engineering excellence, and Bitcoin and Zcash by ideological convictions [1] - The pursuit of art, science, and justice can generate more energy than purely financial incentives [2] - A company should be started only if one is passionate about solving a problem, not for money, fame, or competition [3] Business Strategy & Sustainability - Money is a metric, but not the only one, and the "capitalism" aspect of technocapitalism has been overemphasized [2] - When a cause faces setbacks, a paycheck provides stability; when there is no paycheck, the righteousness of the cause drives one forward [3]

AI Utilization & Programming - Good AI results require significant effort in prompting or creative prompt engineering, akin to programming [1] - Poor AI use reveals underlying deficiencies, similar to bad programming practices [1] - Lack of effort is not concealed by AI, but rather exposed [1] AI vs Manual Effort - High-effort AI use, combining prompting and manual refinement, can surpass manual effort alone [2]

This objection is worth rebutting, as I’m normally sympathetic to an argument like this.But the issue is that AI radically increases the cost to *verify* even as it slashes the cost to *generate*. So in many verticals it increases costs, because verification was cheap and is now costly.Fundamentally that’s because current AI tools (a) consistently produce *subtle* errors and (b) are used to spam channels that weren’t built to detect that many convincing forgeries. For example, the market for (say) email has ...