Just realized I forgot to re-link when porting over final version. Credit to the legend Andrej Karpathy!

https://x.com/karpathy/status/1617979122625712128?lang=en

Sir, I rest my case.

https://nousresearch.com/setting-your-pet-rock-free/

Seems like a reasonable solution

Honestly, have not had a chance to go deep into the tokenomics and comparative level of traction for these various projects at the individual yet, but think it could be a great idea for a future post given my excitement around the category as a whole.

Given the tailwinds. SQD at ~US$50m FDV seems worth doing work on

Great question. So far it seems like both approaches are being pursued with projects like Neuromesh and Sentient hoping to use distributed infra to achieve ultimately higher performant foundational models (will probably depend on aggregating hardware and continued advances in distributed training architectures) and then the "agentic protocols" thesis outlined below being more representative of the smaller models bucket.

Interestingly, Deep Mind just dropped a paper called "Mixture of a Million Experts" which seems like a super interesting combination of the two. Basically has a routing function to many different smaller "expert models" which promises to be more compute efficient for the same performance.

https://www.researchgate.net/publication/382065635_Mixture_of_A_Million_Experts

This will be a big theme in next months report when we push into "coordination / orchestration", but seems like a real place for crypto economic incentives to play a role

Very true. John Luttig of Founders Fund makes the point around product quality fairly explicit in the following piece: https://blog.johnluttig.com/p/the-future-of-foundation-models-is

Agree. In my opinion, the key tension to watch is how much closed source leaps ahead vs. how quickly moats deteriorate. If the moats deteriorate faster than value extraction, then a network of more cost effective models seems likely to be very competitive. If closed source continues to make large leaps each generation based on moats in hardware / data, then perhaps they can ride a flywheel of usage -> revenues -> cheap financing -> more investment -> better models -> more users.

Keen to see whether or not this dynamic holds up....

Great points - I agree, near-term a combination will certainly be necessary, but even so, could provide a meaningful experience in terms of UX to average users. Seems like there should be a lot of low hanging fruit in simply helping new users navigate towards the best protocols to use for a given intent.

To be honest, I haven't had a chance to dive deeply into specific mechanics here, but will definitely be an area I plan to research further in a future report to tackle the app layer

Hi Ejaaz - I guess top of mind examples for me would be linux, android, and large swaths of the database market today.

Also AOL / intranet vs. internet was a real debate back in the 90s

Not sure they are always "superior" but can provide an alternative / tradeoffs which some developers / users value vs. integrated solutions