if we draw a parallel to outside of ai, this is like saying that consumers want to have to talk to a doctor, personal trainer, and financial advisor separately rather than a super smart EA that is all in one. It just increases their cognitive load. It seems like the only reason consumers may seem to prefer specialized agents right now is because models are not good enough yet to be great at these specialized tasks without a task-specific harness.
this is the type of stuff you think of when you dont understand the exponentials of model development. Why specialize your agents when opus 7 is superhuman in every domain with no harness at all and is better at managing context and threads than humans are. Only the general purpose agents are going to survive because only they will be able to convince people to pay for them (by providing excessive broad-reaching value) and only they will be able to encompass such a large portion of someone's economic engagement with society that they can clip a significant fee (early versions of this are merchant fees on agentic commerce)
Thinking about personal assistants, I think there will be:
1 General purpose assistant (like instinct/muse)
1 purpose built assistant for every regulated and/or sensitive use case (like wealth, health, romance etc.)
You get paid for it being hard. The whole job is pricing how it affects the overall stock and sentiment, not just getting the number right. Therr is just not going to be the same opportunity to make money in these markets vs equities regardless of how accurate u are on the factors
To celebrate the launch of the Avec desktop app, we're doing another contest!
Can you convince Claude to give you a brand new M6 MacBook Pro? Let's find out! Email [email protected] to get started.
Contest ends in 48 hours.
@_aadhav I think that the data is a bit cooked by OpenRouter being so open source models focused. The analysis is really valuable though, especially when thinking about the future relevance of inference providers like fireworks.
I think that the main benefit of a router like OpenRouter is best realized with open weight models where there is a large catalog of models and each has many different providers with varying levels of resolution of the model. This is where being able to rely on openrouter to maximize uptime is useful. It seems to me that most usage of closed source models is moving to direct from the labs or through major cloud providers like AWS. Long-term I expect closed source models to win both on price and performance as the labs will have access to the most efficient compute at the largest scale. It would be interesting to see what happens to open source models usage in that world
My respect for Elon Musk has mostly vanished, but I still thought that at least he is a good engineer (Teslas are great cars). But that last bit also got a blow after hearing him say that it is easier to "cool things down" in space. The opposite is true. A vacuum is the best possible isolator. In space you can only dissipate heat through radiation. It is extremely unlikely these data centers in space will ever happen (and certainly not in our lifetimes).
I think you guys will be surprised by magnitude of hyperscaler layoffs over the coming months.
These orgs are a lot more bloated then anyone appreciates and execs are now being instructed to figure out how to placate Wall Street long enough to raise the capital needed to fund ever-growing compute needs.
I continue to believe enterprise layoffs are imminent. But the size of these cuts are likely to be somewhat moderate (say 5% - 15% of workforce).
But from a magnitude perspective I think the CSPs are going to surprise you most (20% - 30% of headcount over next 18-24 months).
I need to figure out the exact shape of the jagged frontier 2 years from now, anybody know what to be reading? This is now the entirety of my investment strategy