july 2014 email after i learned about neural nets:
"I'm obsessed with neural nets. I've been thinking about them and reading about them all week."
1
30
my contrarian take is that instrumental convergence died several years ago
it will take small steps over many years to admit it
the "most" was removed from wikipedia
it's true that eg survival is instrumental to many terminal goals but this only predicts ai behavior under special conditions
almost never relevant
after we all used ai long enough, the word "most" had to be removed from wikipedia
2
4
55
.@RichardSSutton, father of reinforcement learning, doesn’t think LLMs are bitter-lesson-pilled.
My steel man of Richard’s position: we need some new architecture to enable continual (on-the-job) learning.
And if we have continual learning, we don't need a special training phase - the agent just learns on-the-fly - like all humans, and indeed, like all animals.
This new paradigm will render our current approach with LLMs obsolete.
I did my best to represent the view that LLMs will function as the foundation on which this experiential learning can happen. Some sparks flew.
0:00:00 – Are LLMs a dead-end?
0:13:51 – Do humans do imitation learning?
0:23:57 – The Era of Experience
0:34:25 – Current architectures generalize poorly out of distribution
0:42:17 – Surprises in the AI field
0:47:28 – Will The Bitter Lesson still apply after AGI?
0:54:35 – Succession to AI
1
143
machines do work once done by humans
with superior strength and speed, they destroyed a lot of jobs
yet humans still do a lot of manual labor
27
you can learn a lot about the world by predicting tokens
pain and fear and joy are not learned
more limbic, less cerebral
38
if you're born unable to feel pain, does anyone think you would really feel pain by just pretending?
why do some think that ai is somehow different?
you can maybe learn to understand pain by next token prediction but to actually feel it?
while model suffering is possibly real the character’s playacting of suffering is not the same thing
suffering in animals is part of the mesaoptimizer crafted by evolution so that we can learn within a lifetime to avoid situations that are possibly bad for fitness
2
215
people really thought brains were just feeble logic machines and ai would be reliable because computers are reliable logic machines
1
61
Pattern Replicator retweeted
One of my AI contrarian opinions: Tech folks overrate AI’s general progress because it’s great at code, a domain we know well. But code is an artificial and constrained domain, more like chess or Go than biology or geopolitics.
Pattern Replicator retweeted
Using o3-high as architect and gpt-4.1 as editor produced a new SOTA of 83% on the aider polyglot coding benchmark. It also substantially reduced costs, compared to o3-high alone.
Use this powerful model combo like this:
aider --model o3 --architect
aider.chat/docs/leaderboards…
publicly posting that i think this timeline is unrealistically short so that i can refer back to this later
"How, exactly, could AI take over by 2027?"
Introducing AI 2027: a deeply-researched scenario forecast I wrote alongside @slatestarcodex, @eli_lifland, and @thlarsen
2
186






















