Finn / Press
end
What is cool?
Not what is pop­u­lar. Cool is un­ex­pected. Hard. Pos­i­tive impact that ac­tu­ally moves the needle. Some­thing unique that most people would not at­tempt.
Twit­ter is cool. In­for­ma­tion moves fast there. Tech ideas form in public. Fi­nan­cial mar­kets learn about events before the head­line hits a wire. It is where nar­ra­tives get tested and broken in real time.
Tesla is cooler. First suc­cess­ful Amer­i­can car com­pany in over 70 years. Ac­cel­er­a­tion that felt im­pos­si­ble when they shipped it. Aes­thet­ics that made EVs de­sir­able in­stead of apolo­getic.
SpaceX is cooler still. Rock­ets to orbit. Catch­ing a booster with chop­sticks. A base on the moon. Mars as a se­ri­ous target, not a poster.
Neu­ralink is the coolest.
Not be­cause the others are small achieve­ments. They are enor­mous. But Neu­ralink goes after the thing that limits every other achieve­ment: the ef­fi­ciency of the human brain.

The actual bot­tle­neck

People talk about progress like the con­straint is head­count. Hire more en­gi­neers. Fund more labs. Ship faster.
That misses the layer un­der­neath.
The amount of en­gi­neers mat­ters. So does the energy they can expend on hard prob­lems. So does their will­ing­ness to work on things that are unique in­stead of in­cre­men­tal. So does whether they are climb­ing toward a global max­i­mum or pol­ish­ing a local max­i­mum.
All of that runs through one organ.
If our brains ran at higher rev­o­lu­tions and higher res­o­lu­tion, the same number of en­gi­neers would pro­duce a dif­fer­ent civil­i­sa­tion. Fewer dead ends. Faster con­ver­gence on prob­lems that ac­tu­ally matter. More people ca­pa­ble of hold­ing a hard prob­lem in work­ing memory long enough to solve it.
We are not short on prob­lems. We are short on cog­ni­tive through­put.

Re­cur­sive self-im­prove­ment

@DarioAmodei opened a one-way door at An­thropic.
The phi­los­o­phy was in­ten­sive: if you can "solve" coding, you unlock a path to solv­ing every­thing else. One deep bet on the global max­i­mum of what an LLM can become. Not ten prod­uct sur­faces. Not market share in every ver­ti­cal on day one. Ca­pa­bil­ity first.
@sama took a dif­fer­ent shape at OpenAI. Broad scope early. Ex­ten­sive. Mul­ti­ple do­mains, mul­ti­ple bets on dis­tri­b­u­tion, market share across the stack while the foun­da­tion model was still being forged.
Both com­pa­nies matter. The point is not a score­board. The point is the shape of the bet.
Dario tar­geted the global max­i­mum of model ca­pa­bil­ity. That bet won.
Neu­ralink is the same class of wager, but for humans. Re­in­force­ment-style self-im­prove­ment ap­plied to the sub­strate in­stead of the tool. Up­grade the mind that builds the tools.
Every­thing else is down­stream of that.

What changes if we op­ti­mise the brain first

What if we could in­crease the rev­o­lu­tions and res­o­lu­tion of thought?
What if a four-year-old car­ried a PhD corpus in every domain? Not trivia. Struc­ture. Mech­a­nism. The abil­ity to reason inside a field with­out spend­ing twenty years earn­ing access to it.
What would we know by ten?
We would not be de­bat­ing the same trade­offs we debate now. Cli­mate, energy, biotech, align­ment, mar­kets: all of these are hard. They stay hard when the minds work­ing on them are op­er­at­ing at to­day's band­width. They look dif­fer­ent when the band­width changes.
This is why I think we should pri­ori­tise our brains before we keep throw­ing re­sources at the biggest prob­lems in iso­la­tion.
Not be­cause those prob­lems are unim­por­tant. Be­cause at­tack­ing them with un­up­graded cog­ni­tion is op­ti­mis­ing a local high. You get a better ver­sion of to­day's out­comes. You do not get a dif­fer­ent fron­tier.
Op­ti­mise the brain first. Then tackle every­thing else. The impact is not in­cre­men­tal. It is mul­ti­plica­tive.

Global max­i­mum vs local max­i­mum

Per­haps the global max­i­mum of in­tel­li­gence does not look like a human at all. I am open to that.
I still think op­ti­mis­ing our minds is a worth­while mis­sion.
We are build­ing sys­tems that will exceed us. If we cannot up­grade our abil­ity to un­der­stand the mech­a­nisms we create, we are not steer­ing. We are hoping. That is a bad strat­egy for some­thing smarter than you.
More re­sources should flow toward what I would call in­tel­li­gence en­gi­neers: LLM re­searchers, BCI re­searchers, elec­tri­cal en­gi­neers, ML en­gi­neers, anyone work­ing on the in­ter­face be­tween mind and ma­chine or on re­cur­sive im­prove­ment of cog­ni­tion itself.
That al­lo­ca­tion will look lop­sided to people who want vis­i­ble progress in a single domain this quar­ter. Brain in­ter­faces do not always pro­duce a demo that trends. They pro­duce a dif­fer­ent class of human who can solve the demos that matter.

What we should do

Stop treat­ing cog­ni­tive en­hance­ment as sci­ence fic­tion gar­nish.
Treat it as in­fra­struc­ture.
Twit­ter ac­cel­er­ated in­for­ma­tion. Tesla ac­cel­er­ated trans­port. SpaceX ac­cel­er­ated access to space. Neu­ralink tar­gets ac­cel­er­a­tion of thought itself.
Brain in­ter­faces are how humans get to our global max­i­mum.
Every­thing else, with­out that, is a nicer hill on the same land­scape.
start

Brain Interfaces Before Everything Else

Recursive self-improvement and why the brain is the global maximum

Finn Clancy ·