Strategy | innovation | knowledge creation in science, engineering, and technology | organization | cognitive science | AI
San Francisco
Joined July 2011
- Tweets36.9K
- Following1.9K
- Followers2.7K
- Likes39.9K
Pinned Tweet
Let me try and articulate what is see is a genuine tension between Bayesian subjective credences across theories of the world and what constitutes a probability calculus within a possible world bundle.
More niche than ‘the brain is not a computer’ or ‘intelligence is compression’ or ‘consciousness is computational’ but sure to generate the same intensity of partisan debate.
#makeneuroaigreatagain
This needs to be a national priority. Great to see YC making an explicit call for startups in this space.
Robots Made in America
@sdianahu
Robot foundation models are starting to transfer skills from one machine to another, which means the bottleneck is shifting to hardware.
The U.S. still imports much of its robots and their critical components. We want to see founders rebuild that supply chain here, from magnets, motors, gears, and sensors to chips, batteries, and the machines that make them.
This is a (another) great piece, breaking down VC, as an industry, by the underwriting logics used to assemble portfolios. Underwriting logic is as core as stage, focus, strategy, or thesis when it comes to understanding what the firm is doing.
Brian Gordon retweeted
That's awful. If you don't fight the status quo when you're young, what are you going to do once you're old?
Brian Gordon retweeted
1/ In February, we argued that the cost of execution would fall much faster than the cost of verification. As agents scale, human attention and expertise become the binding constraint.
@AnthropicAI just put experimental evidence behind our warning...
xitter.gq/AlexTamkin/status/2115…
Brian Gordon retweeted
So much interest! Sign up here for info, first meetup this Sunday.
forms.gle/C5DXSRHSQL3XXo1X7
Brian Gordon retweeted
Probably a bunch of people vibe coding will come up have even more appreciation for the wildness of industrial scale technology.
Maintaining is a full time job
This feels like a variant of @erikphoel theory of dreaming applied to modern AI models. Maybe the extrema of it, where the model is actually frozen and dreaming relaxes this architectural (vs ‘mere’ overfitting) constraint?
Brian Gordon retweeted
at some point we’ll just accept that the alien-ghosts can use a lot of mechanisms that we use
Today on the blog, learn about our new instance of a “sleep” phase in continual learning, which enables LLMs to continuously integrate new facts and enhance reasoning while mitigating catastrophic forgetting or static knowledge cutoffs.
More →goo.gle/4iICkyD
ALT Overview of our sleep framework: Input streams from the active phase are stored in short-term memory before entering the sleep cycle, which bifurcates into Consolidation (transferring fragile short-term tokens into expanded long-term parametric memory using low-rank structures and MoE) and Dreaming (autonomous self-improvement via synthetic routing with random experts and RL refinement).
Brian Gordon retweeted
Don’t put AI between yourself and someone in any coaching/teaching/mentoring or meaningful partnership role. The raw material is what allows the form to be seen.
I wouldn’t let my robot show my golf coach my swing.
There is no reason to suppose humans/cultural progress has saturated either the extensive (combinatorial search) or intensive (domain insight) margins when it comes to scientific inquiry and we are just now probing what machine, human, & hybrid architectures can do.
there was a point before deep blue, before alphago, before navier stokes, where you could argue, if not entirely convincingly, that maybe some areas of human endeavor were already skill saturated enough that more intelligence wouldn't help much
I think there is basically no shred of credibility left for this kind of thinking now on anything where a primary input is intelligence. we know too much, we have seen too much. the universe has a high skill cap
The scientific project of AI, originally, for many at least, can be understood as a ‘makers knowledge’ quest — an attempt to understand the mind by building minds. It’s notable that right at the exact moment we seem to have cracked (forms of) intelligence, we started on agency.
What I think Henry is pointing to can perhaps be understood as a kind of agency that is ‘world constructing’ and transcending in a sense, rooting what it is to be human in the demiurgic imperative more than intellection simpliciter.
I don’t think it’s as simple as supposing there is a basal and higher level(s) to these processes. Machines are already contributing to the construction of World3. It’s more of a mangle than a hierarchy.
Renegotiating (with ourselves) just what it is to be human is the project of modernity, stretching from the beginnings of the mechanical philosophy of nature to Darwinism and now AI.
Brian Gordon retweeted
In nature, the machine is continuous with the animal, from the lever to the lion--causal, wild. Culture makes the difference. Imagination separates the concept from the emotion, the intellect from the will, reason from the passions. It imposes discipline and grants our freedom.
Brian Gordon retweeted
A piece of advice Hubert Dreyfus gave the grad class:
“The important thing is to ask a sincere question. Otherwise, you won’t know when you’ve got your answer.”
Brian Gordon retweeted
What people call intelligence can be seen as two distinct things
1. how much scale you have
2. how much ability to scale you have
Most peole think of intelligence in terms of (1). Someone like Francois Chollet thinks of intelligence mostly in terms of (2)
The models are weakly good at (2) on their own but not there yet. Labs themselves (and the entire ecosystem that feeds them scaling inputs) remain critical for (2), where the input and output is a thing called “Claude” or “GPT”
Many are implicitly building for a world where (2) doesn’t itself have scaling laws leading to automation, where models remain relatively static objects only capable of a bit of extra test time logical depth. Status quo bias and cope imo
The Models Will Eat Their Harnesses