Studied a few X accounts and mapped their growth.
All of it comes from:
Making bold, pessimistic predictions about events whose outcomes are speculative.
Isn't it funny that we, as a species, have access to the best knowledge devices ever (i.e. LLMs) and yet increasingly struggle to find the truth, defaulting instead to social media prophets?
Two questions follow.
1. Should we stop blaming IQ/innate ability until someone has made real effort to learn their subject's composite units?
2. If an LLM were to become an expert at a subject, would we need to teach it these same composite units instead of tokens?
When a novice sees a chess board, they see individual pieces.
When an expert sees a chess board, they see combinations of pieces they recognize as a "unit".
They learn these combinations through repeated exposure to varying states of a chess game.
In fact, experts end up learning between 25,000 and 100,000 of these composite "units".
The same is true for subjects other than chess.
The expert is the expert because they begin reasoning at a higher level of abstraction.
Religions also lay upon you diverse obligations (of family, of worship, of livelihood, of proselytization etc.) you can't cherry-pick from, even when chasing after a lofty religious goal.
As Herman Melville said: I try all things; I achieve what I can.
There is a cockiness in the act of prioritization, in that we pretend to know what matters.
The older I grow, the more I realize that I must pursue all things and I can, if I align them all together (more on this some other time).
If you're excited about deep-tech but don't have that work available today, build some form of IP. Don't chase the generic services model with a sprinkle of "positioning" on top.
There is just no way to repaint a SaaS build as deep-tech.
I just found a company out of NUST that positions itself as deep-tech.
I smiled when I read about them since it's an admirable goal but it also resurfaced early memories of Antematter.
When we started, we also wanted to be deep-tech.
The mistake we made was the path we chose: a services model but without a network that could bring that kind of work to us.
That company is making the same mistake.
Not just from a cost perspective.
Over time, inference may evolve into a capability of its own.
Companies may have to run their own inference that scales up or down to provide the specific intelligence required for a task or workflow.
AI inference has become more important than pre-training, especially with RL which requires it to run continuously.
With traditional models, it wasn't as big of an issue. But with LLMs, massive token consumption had made it a primary concern for labs and customers.
All engineers have a temperament just like all bankers, lawyers, marketers etc. have their own.
In an engineer's temperament, there remains a vague frustration.
Now they are frustrated with AI itself even though they are the direct beneficiaries of it.
Hence why the general skepticism of the technology on hackernews and other similar sites.
NASDAQ introducing a "fast entry" rule right before SpaceX, OpenAI and Anthropic gear up for IPOs is the peak of what's wrong with American capitalism.
The retirement money of the average Joe will swell the pockets of the wealthy elite.
https://t.co/bOjiZOHe36