Marc and Ben State of AI

Takeaways

Are AI companies just “wrappers around an LLM”? But note that many successful Web companies are “wrappers around a database”.
AI has no network effects. Unlike the internet, where usefulness follows Metcalfe’s Law that depends on the number of users, AI is more like the PC or a semiconductors.

Jevons’ Paradox, named after the British economist William Stanley Jevons, is a phenomenon where an increase in the efficiency with which a resource is used leads to an overall increase in the consumption of that resource, rather than a decrease.

Databricks has a foundation model that’s different enough from the Big LLMs that they’re not direct competitors. Similarly, Elevenlabs is focused on voice.

The top models are getting pretty close to one another. They’ll get better at hallucination, “alignment” (in the sense of not accidentally telling you how to build a bomb).

The LLMs are trained on everything, so the “average IQ” is pretty average. You need to use prompt engineering to steer it to the smarter version of the answer. For example, ask “write secure code” rather than “write code”.

Distinguish between Artificial Human Intelligence, which is like what we have now. Vs. “Artificial super-intelligence” which is something we can’t even comprehend. It’s unclear if we’re on track to build something that would be truly different.

Smart people argue for super-intelligence point to emerging properties that, say, learn chess. And it appears there is value in re-training on

And there’s a lot of room to add plain engineering efficiencies as the tech moves from the lab to the engineers.

Are the AI companies today just basically wrappers around the LLM? That may seem like a pejorative until you remember that many interesting companies could be thought of as wrappers around a database.

Everybody’s building “copilots” and not “pilots” because nobody trusts the “pilot”. It’s an open question whether we’ll ever feel comfortable with the Pilot.

Think about the workflow, not just the specific task that AI is doing. A law office does a lot more than write contracts for example.

They speculate that maybe AI will simply improve the overall quality, rather than create a new industry. Think of what CGI did to Hollywood: rather than replace actors, it just increased the visual appeal of all movies.

“In software there’s always more to do”, so AI can only add more features and power to existing products.

Marc thinks proprietary data is over-rated. The amount of data on the internet is so much larger than whatever the value is of “your data”.

The internet is a network, but AI is one application. So rather than look at a business governed by Metcalf’s Law, AI is likely to follow the PC or chip business: it’ll gain a foothold one application at at time.

We’ll have AI models at every size and niche application, like with chips.

AI is the easiest computer ever, so maybe switching costs will be much lower.

Premature regulation is the biggest threat.