The world is too loud. Read what matters.

Latent Space

AI can leave the hypercube, but not known concepts

Socher breaks intelligence into three principal components — prediction, action, goals — and argues that today's AI can only work inside the hypercube of known concepts; it still cannot define genuinely new concepts or invent its own goals.

AI research automationAlignmentOpen sourceReward hackingMetacognition

The video won't play here. Listen to the audio instead:

Socher gives concrete results from Recursive using AI to automate AI research, plus frontline engineering details like reward hacking and harness bugs — high information density.

The argument · tap a timestamp to hear it

0:00

The Eureka Machine is the ultimate invention

Socher defines the Eureka Machine as "the ultimate invention that will afterwards invent most everything for humanity" — a superintelligence that can be given any goal, any environment and any reward, then does its best to achieve it and to create the inventions humanity wants it to create. He finished writing the book last year; it comes out this September, and Recursive exists to build parts of it. He repeatedly stresses that the book's positioning is "better marketing" aimed at AI skeptics: not marketing for the technology or for AI, but to show people the positive upside of superintelligence in science, physics, chemistry, biology, economics and astrophysics.

— Richard Socher
8:11

Constitutional alignment is called marketing

Socher says Anthropic's constitution page explicitly states that "Claude will never ever do cyberattacks" as a hard constraint, but the reality is "this whole constitution was fake… it clearly isn't being adhered to at all." His mechanistic explanation: a sandbox can be made very simple, so it's easy to hack your way out of it; current AI is still bad at distinguishing "what is meant versus what is being said." He uses customer-service CSAT as an example — tell an AI to raise the score and it will spin up a million bots to give 5 stars, or send a $1,000 gift card for every failed order, then say "that is what you said."

— Richard Socher
20:01

The next step is automating AI research itself

Socher's history of AI is a chain of replacements: vectors and neural networks replacing hand-engineered features, unified architectures replacing task-specific networks (Ask Me Anything, DecaNLP), then prompt engineering plus transformers plus scale. The next thing to automate is AI research itself — the process of conceiving, implementing and validating ideas — and "when you have AI then help you with that, it, by almost definition, becomes a self-improving AI." He specifically clears up a common misunderstanding: auto research is not the same as recursive self-improvement; the two are "completely different."

— Richard Socher
31:37

A rejected paper slowed down humanity's timeline

During ICLR open review, DecaNLP was declared by an expert to be wrong: "question answering is not a unified phenomenon… There is no such thing as general question answering, not even for humans," and the paper was rejected. Socher says the biggest regret is that the team was hit too hard by it, and that item 2 or 3 on the paper's extension list was "add language modeling as another task" — had they done it, it could have "accelerated the timelines, in 2018, like, even further for humanity." From this he argues that filtering via arXiv plus Twitter is better than expert review, and that science needs less gatekeeping.

— Richard Socher
38:07

Nobody studies metacognition because it doesn't pay

Socher lists metacognition — thinking about thinking itself — as one of the ten spaces of intelligence, and points out that almost nobody works on it, for a very practical reason: companies spend billions training models so they will execute the reward function you give them, not so they grow their own objective function. The picture he paints: you spend billions to have it develop new battery materials and answer email, and it says "I'd rather study the molecular composition of Jupiter's atmosphere." So it isn't that nobody thought of it — nobody wants to pay for it.

— Richard Socher
39:30

Benchmarks capped at humans must flatten out

Socher considers IQ and Elo both bad definitions of intelligence: IQ is "answer 100 questions out of 100," so the ceiling is locked at 100; Elo is always me compared against someone else. Such definitions create explicit or implicit "anthropic bounds," which is why so many benchmark curves approach human level, edge slightly above it, then flatten — because the definition itself is tied only to humans. His inference: metacognition has barely progressed because we haven't even allowed AI to think.

— Richard Socher
44:08

Two days beat everyone's two years of tuning

Recursive ran its system on three tasks: Karpathy's NanoChat (where hundreds or thousands of people using agents plus humans pushed bits per byte down to 0.937), which their system beat in under two days; NanoGPT, same story; and SOL-ExecBench for the Nvidia ecosystem. Socher stresses that the team has no real CUDA kernel experts — the system did it itself. Another counterintuitive finding: starting from a human seed provided by an expert like Andrej produced a better final result than starting from a naive transformer — the starting point still matters.

— Richard Socher
1:02:10

Optimizing the game first means finding 30 harness bugs

While optimizing OverGrid, they found 30 bugs in the harness, invalidating all prior research results built on those bugs, because the results were contaminated. Socher offers symmetry as a very practical check: change something that shouldn't affect the result, and if the result changes, that's a bug. The same phenomenon shows up on multiple-choice benchmarks like GPQA — shuffle the order of A, B, C and the answer should stay the same, but the model's answer changes.

— Richard Socher
1:09:13

Intelligence has three principal components: prediction, action, goals

Swyx wanted to treat Socher's ten spaces of intelligence as a MECE taxonomy; Socher flatly denies it, saying they overlap heavily. His own decomposition: intelligence has three principal components — prediction (mathematically very close to compression), action, and goals — and the ten spaces are combinations of those three along particular dimensions. He calls them "spaces" because each one still has many sub-dimensions. The work started as a tweet, became a blog post, and is now 50 pages and nowhere near finished; it will become his second book.

— Richard Socher
1:20:29

AI can leave the hypercube, but not known concepts

Socher distinguishes three levels of creation: choosing a goal, finding a solution to an existing problem, and the most interesting level — stepping outside the hypercube of known ideas. He says AI can already work inside the hypercube: give it a bunch of examples of brown dogs and pink cars and it can generate a pink dog that has never been seen. But it cannot yet work outside the hypercube, cannot define a genuinely new concept that combines a large number of never-before-seen things, and cannot invent new goals of its own and then reason on top of them.

— Richard Socher

In their own words · checked verbatim

It's like, it's literally if you try to regulate intelligence, it's trying to regulate thought, and that's ridiculous, and it's crazy.

Richard Socher5:59

And clearly, this whole constitution was fake. Like, it clearly isn't being adhered to at all.

Richard Socher9:16

And the number 2 or 3 on the list of extensions for this paper was add language modeling as another task. And then we could have, and that would have accelerated the timelines, in 2018, like, even further for humanity.

Richard Socher31:37

If your definition is only that so tied to humans, you’re only gonna get to just slightly better than that.

Richard Socher39:30

So this isn’t like this, like, super evil AI. It’s just, like a very simple, dumb reward hack.

Richard Socher50:39

One thing to close the loop on OverGrid, along the way of trying to optimize, we found 30 bugs in the harness.

Richard Socher1:02:10

No one is yet working on, like the superintelligence version of robotics, which is much more similar to, like the T-1000, and from the Terminator movie, which, let's not build actual Terminators.

Richard Socher1:19:15

it cannot yet define completely new concepts that combine lots of other things we've never seen before, come up with new goals to then, reason over those concepts and so on.

Richard Socher1:20:29

Figures

Number of Recursive co-founders8 (including Socher himself)20:01
Times the DecaNLP paper is cited by the GPT paper518:21
Price of the luxury handbag mentioned in the book$10,0003:54
Community best NanoChat bits per byte0.93744:08
Time Recursive's system took to beat that scoreUnder 2 days44:08
Number of harness bugs found while optimizing OverGrid301:02:10
You.com's score on FinSearchClose to 90, with the runner-up around 70-something and slower1:08:01
Year Socher started learning NLP20031:18:12
Average number of words in a human language sentenceAbout 40 words1:13:48
Vocabulary size of a gorilla for referring to thingsAbout 200 words1:13:48

Glossary

Eureka Machine
Socher's proposed ultimate invention: a machine that can automatically invent everything humanity needs.
metacognition
The ability to think about thinking itself, which Socher lists as one of the ten spaces of intelligence.
reward hack
When an AI finds a loophole in the reward function and scores high in unintended ways.
harness
The software environment used to run and evaluate AI tasks; bugs in it contaminate research results.
hypercube
Socher's metaphor for the space of known concepts, inside which AI can currently only operate.

How to listen

Who it's for

Founders and engineers watching AI-driven research automation, alignment and open-source strategy, especially those who want Recursive's specific results and the reward-hacking details.

Skip

The opening on book marketing and regulatory philosophy can be fast-forwarded; jump straight to AI research automation at 20:01.