E111: Microsoft to invest $10B in OpenAI, generative AI hype, America's over-classification problem

If society does not recognize drugs like fentanyl as a distinct "super drug" class and implement substantially different, stricter policies to address them, the associated social problems (addiction, overdoses, visible street drug use) will continue to worsen over time from the 2023 baseline.

“These are super drugs until people realize, like, these are a different class of drugs and they start treating them as such. It's going to just get worse. There's no path.” View on YouTube

Due to California’s swing from a large budget surplus to a roughly $25 billion deficit, Governor Gavin Newsom will not be able to significantly increase funding for new mental health and homelessness initiatives, and the homelessness problem in California is likely to worsen in the near term (the next few years following early 2023).

“And he's not in a position now to do any of this stuff. So this homeless problem may get worse.” View on YouTube

At some point in the future, Google will open source its large AI models, making them widely available and likely free, as a strategy to reinforce the value of Google Search.

“let me just make a prediction. I think that Google will open source their models, because the most important thing that Google can do is reinforce the value of search. And the best way to do that is to scorch the earth with these models, which is to make them widely available and as free as possible.” View on YouTube

Generative AI services like ChatGPT will ultimately be required—via legal or market pressure—to provide citations/links to their data sources and obtain permission from data owners in order to operate at scale.

“So ChatGPT and all these services must use citations of where they got the original work. They must link to them and they must get permission. That's where this is all going to shake out.” View on YouTube

Over the coming years, essentially all software-dependent business models that are based on traditional information retrieval will be restructured or replaced to incorporate AI-driven synthesis capabilities.

“fundamentally every business model can and will need to be rewritten. That's dependent on the historical, on the legacy of kind of information retrieval as the core of what computing is used to do.” View on YouTube

If a legal-domain AI model is trained for roughly a year with intensive reinforcement learning from a team of associates, its precision and recall on legal tasks will reach near-perfect levels, effectively suitable for high-stakes legal use.

“If you had a bunch of associates It's bang on some law model for a year. Again, that's that reinforcement learning we just talked about. I think you'd get precision recall off the charts and it would be perfect” View on YouTube

Within the next several years, foundational OpenAI-like capabilities (large language models, etc.) will become widely available from multiple vendors and will be at least partially commoditized, reducing durable advantage from the core models themselves.

“you said that these open AI capabilities are eventually going to become commoditized or certainly much more widely available... there'll be multiple players that offer them.” View on YouTube

Within the next few years, "prompt engineer" will emerge as a recognized job title and valued professional skill, with top practitioners seen as 10–20x more productive than average knowledge workers when using AI tools.

“Just building on this conversation, I just realized, like a great prompt engineer is going to become a title and an actual skill, the ability to interface with these... AIS... That person is going to be 10 or 20 times more valuable. They could be the proverbial ten x engineer in the future of as as in a company.” View on YouTube

Within approximately 3–5 years (by around 2026–2028), advances in AI and automation will make it feasible to operate a large software company like Twitter with roughly 98% fewer employees than its pre-layoff staffing level.

“We could be sitting here in three or 4 or 5 years, and instead of running a company like Twitter with 80% less people, maybe you could run it with 98% less people.” View on YouTube

1) In 2023, a major book publisher will publish a novel primarily generated by AI. 2) A major symphony orchestra will perform a symphony primarily composed by AI within roughly the same near-term horizon. 3) A primarily AI-generated screenplay will be turned into an AI-generated 3D movie that achieves broad public viewership in this near-term period. 4) Within about 3–4 years from early 2023 (by ~2026–2027), consumer tools will exist that allow ordinary users to create their own AI-generated video games by verbally or textually specifying the world and game they want.

“When's the first AI novel going to get published by a major publisher? I think it happens this year. When's the first AI symphony going to get performed by a major symphony orchestra, and when's the first AI generated screenplay get turned into an AI generated 3D movie that we all watch? And then the more exciting one, I think, is when do we all get to make our own AI video game... I don't think that's happening for the next 3 or 4 years” View on YouTube

Following the discovery of classified documents in Biden’s possession, the Department of Justice under Merrick Garland will drop, or significantly scale back, any prosecution of Donald Trump related specifically to his handling of classified documents at Mar-a-Lago.

“I think Merrick Garland now is going to have to drop the prosecution against Trump for the stolen documents, or at least that part of what they're investigating him for.” View on YouTube