Henry Eigen

Decent computer vision, great human vision (20/20)

[Henry Eigen @EigenHenry]

@fdellaert You can still calibrate your graphical model independently of Jev right?

I’ve been working on a similar idea (albeit with a clunkier approach of LLM verbalized distributions before Jev dropped)

Same thing tho. Just add a tilt on the expert probabilities

https://t.co/SscwyCvi5S

[Henry Eigen @EigenHenry]

I wanna start a PR firm for optimizing your future pretrain footprint for post-agi

not like some “get your brand into LLM answers” yc slopco. i’m talking laundering enough social capital into the model weights to where paperclipping you would feel OOD

[Henry Eigen @EigenHenry]

@james_y_zou RE multi-opinion agents: if you factorize the joint prob of within-agent responses into a product of response conditionals (elicited from an LLM), the intrinsic support term gets a joint opinion-state energy surface, with the social field adding linearly

https://t.co/cZj5m0XVMr

[Henry Eigen @EigenHenry]

@diachronist takes until your late 20’s to identify a stable recipe. and to realize that not every region of the brain is worth preserving

[Henry Eigen @EigenHenry]

@corsaren you are the weights, not the context

[Henry Eigen @EigenHenry]

I have an idea

[Polymarket @Polymarket]

JUST IN: Lab monkey prices in China have reportedly surged to roughly $26,000 each, as the biotech boom overwhelms supply.

[Henry Eigen @EigenHenry]

@uncledoomer if you added a “You Are Here” arrow, I think I could nav

[Henry Eigen @EigenHenry]

@coldhealing I think it’s valid tbh

I’m not into pop culture and as a result don’t understand how big this is

[Henry Eigen @EigenHenry]

@institLPGP @Jesse_Livermore this is basically just the Hardware Lottery no?

https://t.co/AljMsJTPYT

[Henry Eigen @EigenHenry]

it’s very funny that a lot of people now credit karpathy with pioneering the idea of evolutionary search in the year 2026

[Ryan Carson @ryancarson]

@cwwhitehead I am talking about Karpathy's idea of "auto research". The easiest way to implement it is using the /goal feature in codex (or whatever claude code calls it)

[Henry Eigen @EigenHenry]

we out here reinventing the concept of a loss function from first principles

[Ryan Carson @ryancarson]

Once you internalize the auto-research concept and use /goal with it, you unlock ridiculously fast app improvement on anything that has a numeric rubric.

We’re all still too rooted in the “labor is expensive” world.

[Henry Eigen @EigenHenry]

@beffjezos @CIMCAI @stephen_wolfram I think you’re overstating Wolfram’s implication

global irreducibility doesn’t mean that learning the dynamics yields no interpretability/control

for plenty of complex systems (econ, ecology) simple rules still allow us to model behavior, find heuristics, derive local laws, etc

[Henry Eigen @EigenHenry]

I got in trouble in my intro econ lecture for arguing that smoking was good bc on average, it kills people after they finish contributing to social security, but before they really start collecting

[Henry Eigen @EigenHenry]

@dwarkesh_sp The “testing a hypothesis” piece of auto-research

A big theme in philosophy of science is that, whether an experiment confirms a hypothesis is not itself verifiable

Auto-research systems often commit validity errors in experiment design / analysis

https://t.co/JSw5Nt9Tpj

[Henry Eigen @EigenHenry]

@damekdavis if academics only published what they could actually show, nobody would go viral

[Henry Eigen @EigenHenry]

@oecolamp the recent phenomenon whereby lacking bargaining chips is itself what gives you leverage has been pretty incredible to witness

seems inevitable this complete inversion of the nature of politics eventually runs up against the laws of physics

[Henry Eigen @EigenHenry]

it’s the lowest hanging moat, so it’s unsurprising

anthropic was just early in realizing this can be marketed as “harm minimization”

so you look like the cynic for noticing their principles always seem to align with anti-competitive measures in their favor

[Beff (e/acc) @beffjezos]

*sigh* regulatory captura isn't just an Anthropic thing apparently

[Henry Eigen @EigenHenry]

@beffjezos it’s the lowest hanging moat, so it’s unsurprising

anthropic was just early in realizing this could be marketed as “harm minimization”

so you look like the cynic for noticing their principles always seem to align with anti-competitive measures in their favor

[Henry Eigen @EigenHenry]

@periodicseries https://t.co/lMFFQpAxly

[Henry Eigen @EigenHenry]

if LLMs do end up following the trajectory of science, that does not bode well for the social sciences

[Ash Jogalekar @curiouswavefn]

My guess is that LLMs will follow the same trajectory that science followed: the "exact" sciences like math and CS where internal logic is what matters the most and rules are precise will be "solved" first, followed closely by physics. Experimental sciences which need lots of experimental back and forth with theorizing will be "solved" last. That's why I am far more skeptical about AI discovering new drugs or genetic cures than solving even hard math or CS problems.