1/ SCALING THROUGH DECENTRALIZATION
This piece explores the motivations behind improving model performance in a decentralized setting.
Link in next tweet.
@Shaughnessy119 agree on potential for OS model adoption to eat at mrkt share via: broader recognition many tasks dont require frontier pref + enterprise cost crunch
but in current form - and particularly for agents tasks (higher input:output ratios) - are margins not improving?
I like the idea of decentralized compute forex market with native unit of account
Fork attempts are typically shitter versions of the first
But given the lack of a real fair launch for PRL + the opportunity to fix the PoUW mechanic makes PRL fork potentially interesting IMO
some heuristics on “ai mining” from this and other previous work. when hash rate goes up exponentially and cost of the hash rate doesn’t make sense, people aren’t throwing compute at it like idiots but instead found significant optimization
also team selling compute packages 🚩
@anay_sim AI infra supply chain reratings (both directions) as market gets transparency on:
- the how and why of Ant/OAI external provider dependencies
- margin profile of frontier tokens
thought provoking, spicy and def a bit sensational
but if you dont subscribe to the idea that tradfi can just clone everything with a regulatory wrapper and win (I dont), it reads more like a BTC + BTC.D bear thesis than an all of crypto one
It is so far, absolutely. And that demand from the end user is only going in one direction (much, much higher).
My working theory is that as the frontier continues to push forward, competitors, namely the Chinese (open-weight) and eventually actual open-source, will distill and redeploy this intelligence (I consider this a natural law of the universe that cannot be prevented). You can tack onto this dynamic providers looking for creative ways to make inference cheaper (whether centralized or decentralized).
What that leaves you with is a continued race towards "AGI" at the frontier level, with an entire market in its wake, ramping up to provide better and better solutions at a cheaper cost.
The demand pie will continue to grow exponentially, but it will segment itself into different buckets based on what is actually required to complete a task. Responding to my e-mails might not require Claude Mythos and can be handled by Gemma or Kimi, routed through Chutes, while Eli Lilly will want Mythos to be working on its longevity solutions, etc., etc. You can quibble about the nuance of how the demand is allocated across models, but all of it requires the continued infra buildout. The supply side will take a very long time to catch up.
A person I have known for more than ten years, who I consider trustworthy, is convinced Peter is the Citrini of Biotech and the segment will shortly experience a historic melt up. I don't know anything concrete, but if I weren’t exposed, I would be concerned.
To clarify any doubt: The rips that will occur in $qure and $twst and $abcl and $clpt and $psnl and $prme and $txg and hell even the $xbi etc etc etc are going to make the AI/ $QQQ bubble look like toddlers in a playground.
We're not there yet, but enjoy the journey bc it will be uncomfortable.