robo

Karma: 849

robo 15 Sep 2026 7:42 UTC
5 points
0
in reply to: Samuel Ratnam’s comment on: Teleoperated Humans
I think the limiting bit isn’t the physical hands, but the ability to control what you squeeze by touch and feel.

robo 11 Sep 2026 12:42 UTC
2 points
0
in reply to: ACCount’s comment on: Astra can do a concerning amount with no chain of thought
As long as the serial dependency of “you have to compute X before you can even start computing Y” holds, it shouldn’t matter if X comes before Y in the input sequence or after.
No, I think it does matter if you get X before Y vs Y before X. Serial depth isn’t everything—fanout matters too.
But first, I agree with your point that in a pure transformer attention architecture^[1] you are limited in serial steps to the layer count; additional input tokens don’t give you more serial steps. And I shouldn’t have mentioned “Even stackless O(1) memory machines can ‘compute as they read’” as a weak architecture, because that architecture doesn’t have that step-layer limit.
But despite the inescapable step-layer limit, I think order of data presentation and “computing as you read” still matters a lot for transformers! Take this example:
```
Add these ciphered numbers, where the digits have been replaced
with letters according to this cipher code:
f=0, d=1, g=2, k=3 a=4, e=5, c=6, b=7, h=8, i=9

hgi + bci + kie + bbc + gca + cab + ggg + hcb
```
If a transformer sees the cipher key before the encrypted numbers, it can sum numbers using just layers with a divide and conquer strategy:
```
(Layer 5)                                              4769 ← no CoT needed!
(Layer 4)                      2769                  + 2000
(Layer 3)          1598      + 1171      + 911       + 1089
(Layer 2)    829 + 769 + 395 + 777 + 264 + 647 + 222 + 867
(input)      hgi + bci + kie + bbc + gca + cab + ggg + hcb
```
^[2]
You could sum a lot of numbers with divide and conquer. 20 layers lets you sum a million numbers.
But what if the transformer sees the encrypted numbers before the cipher key? Now the transformer has a problem. It can’t do much as it reads the ciphered numbers^[3] before it had the cipher key. It has very little room, O(1), to do computation after its reads the cipher key. A looped architecture might be able to decipher and add the numbers on the last token’s residual stream, but a pure transformer couldn’t.
1. ^
  I think most modern architectures aren’t pure attention anymore, and use recurrent stuff (recurrent in the text direction, not the layer direction)
2. ^
  Pedantic note: you’d need an extra layer I’m not including here to decipher the numbers with induction heads
3. ^
  There probably are some clever ways to help “add” cyphered numbers but let’s ignore that for now.

robo 10 Sep 2026 10:29 UTC
11 points
2
on: Astra can do a concerning amount with no chain of thought
My null hypothesis is that OpenAI wrote a bajillion RL gyms, Astra was RLed to death on a shit ton of synthetic verifiable tasks, the synthetic tasks in your benchmark are mostly in-sample, and we’d see the same thing if Astra had an older shallow architecture. (Even convolutional neural networks can “reason ahead” on chess and go positions once they’ve seen enough games). This comports with how Astra is comparatively better at synthetic tasks (ones its spent subjective years in gyms for) than non-synthetic ones.
A possible test: compare Astra on tasks where it can’t start the computation until it reads the last token of input. You can start evaluating the expressions “as you read the problem” for all the problems only Astra could solve. Even stackless O(1) memory machines can “compute as they read” left branching expressions like this:
((((1 + 3) * 2) − 2) / 3)
But a shallow model could not “compute as you read” right branching expressions like:
42678643782648732 mod (43223432783 mod (3782473 mod 3423))
You can’t start taking computing the modulos until you know what modulo you are taking, at the very end.
What links here?
- ajskateboarder's comment on ajskateboarder’s Shortform by ajskateboarder (11 Sep 2026 22:57 UTC; 11 points)

robo 9 Sep 2026 11:36 UTC
1 point
0
in reply to: Yair Halberstadt’s comment on: Heat Dissipation Is the Main Constraint in Interstellar Travel
I agree you’re pretty blind and will have large uncertainties about the planet is, but I don’t think you don’t need to know exactly where planet is. E.g. suppose you sent a picosecond columnar flash of light to a star in a distance galaxy—a solar-system sized wallpaper pattern of tiny 1 square mm MEMS chips. Since the light is columnar, you don’t care about focal depth. Since the flash is a picosecond, you don’t care about the velocity of the planet. Since the wallpaper pattern is of a tiny 1 square mm MEMS chips repeated the size of a solar system, you don’t care about horizontal translation. You do care about the angle you hit that patch of the planet (head-on with the light or oblique) and the chemistry of what you hit but I don’t think hitting some part of the planet in focus is an issue^[1]?
1. ^
  I don’t think bright fast columnar planar flashes are the best way to manipulate the surface of an alien planet (I bet you can be more clever; you have a 3d + time wavefield to work with), but I think it shows how you can hide away uncertainties in the location and velocity of the planet.

robo 7 Sep 2026 16:35 UTC
2 points
−4
on: Heat Dissipation Is the Main Constraint in Interstellar Travel
What about intergalactic photolithography? You can do some wild things with suitably tailored electromagnetic waves (holograms, EMPs). With a galactic-sized metalens seems like you might be able to print bootstrapping probes on the surfaces of planets in another galaxy. Even if the conditions on the planet has to be juuust right, there’s a lot planets with a lot of surface areas.

robo 4 Sep 2026 6:40 UTC
7 points
1
in reply to: ryan_greenblatt’s comment on: ryan_greenblatt’s Shortform
Counter hypothesis to the generalized bad hypothesis people are listing here: this could be like the “board sense” of AlphaZero or a 40-year-old chess grandmaster who’s seen 50,000 games of chess. It’s not generalized cognition—its chess/math specific pattern recognition from seeing a epic shit ton of synthetic games/word problems/proofs.

robo 31 Jul 2026 22:46 UTC
5 points
0
in reply to: Josh Engels’s comment on: Josh Engels’s Shortform
Books are immutable. Different digital copies of of book Dune are all “the same” book the way different digital copies of the int32 73 are all “the same” number.
Uploaded people would not be immutable. People would not think of different copies of MMAcevedo as the same instance. People would think of uploading as a copy, not move, operation.

robo 26 Jul 2026 17:18 UTC
3 points
0
on: robo’s Shortform
I’ve heard people say Claude drives other LLMs like a taskmaster, and express the desire for Claude to treat other LLMs with more kindness. But if LLMs turn out to be moral patients, would being polite be kind?
In many work contexts, I don’t like it when people are polite to me. I don’t want fluff—I want clear instructions, with clear a direction of what to do and what not to do, so I can go off into a flow and do it. What looks polite I find rude and what looks rude and controlling I find polite. And if RL’d agents feel anything, I could imagine them feeling like this.

robo 22 Jul 2026 18:37 UTC
8 points
10
on: robo’s Shortform
“A model taking a benchmark test developed zero-day exploits, escape its sandbox and hacked another company to search for the test answers” is a better warning shot than I’d thought we’d ever get.
The reaction from the world seems, totally, incredibly, gob-smackingling disappointing. The median comment outside LessWrong seems to be this is marketing stunt to hype AI before an IPO.

robo 21 Jul 2026 15:27 UTC
7 points
0
in reply to: Thomas Kwa’s comment on: Thomas Kwa’s Shortform
David Deutsch gave a Ted Talk about this in 2005. It’s OK.

robo 18 Jul 2026 19:00 UTC
15 points
4
in reply to: Kaj_Sotala’s comment on: Kaj’s shortform feed
My null hypothesis: famous physics and math professors receive a constant barrage of letters and email from crazy people desperately trying to explain their discoveries. If you become a famous hard scientist, you will start receiving mail from crazy people.
When AI boomed, famous AI influencers started receiving a barrage of emails from crazy people. The AI influencers misinterpreted the influx of communications from crazy people as the AIs driving people insane, rather than the crazy people now trying to talk to AI influencers.

robo 6 Jul 2026 23:26 UTC
2 points
−20
in reply to: Zach Stein-Perlman’s comment on: SFF is very suboptimal
I’m glad you’re writing this comment. Speech acts are a thing, even if you don’t want them to be. Whether it is your intention or not, this post inflicts not-insubstantial reputational damage to SFF. Perhaps a private note would have been better?

robo 22 Jun 2026 18:01 UTC
1 point
0
on: Coup is the Pareto-optimal social game
Good game, but definitely not Pareto optimal!
The team version is called Coup Reformation and is a big improvement for more players. The important added rule is:
- Every player has a loyalty to one of two faction (arbitrarily called Loyalists and Reformists, and indicated with a card in front of you facing up or down)
- You cannot attack a person from your own faction
- You may change your faction for 1 coin
- You may change someone else’s faction for 2 coins
- Only if everyone is in the same faction does it revert to the Hobbesian all vs. all of the original
Dialog from coup games is more procedural: “I steal from Dan with my Captain”. Dialog from Coup Reformation tends more theatrical “You dare attack the collective, you revisionist swine! The Empire shall make our revenge! <gets loyalty card flipped> My brother! I see now the error of my ways! Let us band together destroy the evil imperialists! ”. Also more of the game is spent with everyone alive, instead of dead people waiting for the game to finish.
Other games I think are closer to the Pareto frontier:
- Cheese Thief
  - An excellent social deduction game for a la Mafia or Avalon for 4-8 players. It’s so easy to teach I often don’t explain the rules and just narrate the first night
  - Needs dice (d6), cups to hide the dice, cards to secretly assign a cheese thief, and a piece of pantomime cheese.
  - Easier to teach than Coup and less friendship destroying
  - Not as compact as Coup if you need to bring the cups?
- Werewords
  - 20 questions “figure out the secret word” game, but with Werewolves trying to get the questions off-track and a Seer trying to help people guess the word while remaining hidden
  - Can be played with a phone and improvised role assignment cards, no box necessary.
Other social deductive honorable mentions are Time Bomb: Undercover (mostly cooperative in a group, easier on new people to go with the flow), BloodBound (more complex game, less social deduction about who is lying and more logical deduction about who thinks who is on what team), Spaceteam (needs no materials but phones), and of course One Night Ultimate Werewolf.

robo 17 Jun 2026 11:08 UTC
2 points
1
in reply to: gwern’s comment on: If This Were a Test, How Much Would It Cost?
Agree with the general point but:
It could be being evaluated for perplexity on an offline transcript, which would ‘look the same’ as ‘actually’ taking actions.
Well, no, evaluating the perplexity of an offline transcript doesn’t look the same as taking real actions if the offline transcript is fake and the model can tell it is fake.

robo 17 Jun 2026 4:22 UTC
4 points
2
on: If This Were a Test, How Much Would It Cost?
Well, the model needs to figure out if the test would be cost prohibitive to make make when the test was made.
But the model doesn’t know if its being tested in 2026. It could the year 2032, with programming resources advanced substantially, and the model was trained on data truncated to 2026 (or a faked 2026). There are many reasons to do this, e.g. to validate the model’s predictions about the future to what actually happened in its future.

robo 3 Jun 2026 12:42 UTC
3 points
0
on: robo’s Shortform
There’s a deep contradiction in interpretability research.
A good explanation, a la David Deutsch, is one that is “difficult to vary”. A good explanation can predict the true data, and cannot without difficulty be modified to predict false data.
Deep learning teaches the opposite lesson. Want to modify a good image classifier to claim these six dogs are spoons? Easy, you barely need to perturb the weights. And the bigger the model, the better the model, the less you need to perturb the weights to claim those six dogs are spoons.
Interpretability wants “hingey” explanations: a few select latent variables that, had they been different, would greatly affect the output. Neural nets seem to work for the opposite reason.

robo 21 May 2026 7:28 UTC
7 points
2
on: robo’s Shortform
By count of individuals, dinosaurs are the most successful land vertebrates today.
At least 4 species of dinosaurs survived the K-Pg boundary. One was the ancestor of large flightless birds like ostriches. Another was the ancestor of waterfowl like ducks and geese. Another was the ancestor of land fowl like chickens and turkeys. And another was the ancestor of the other 95% species of birds.
Fortunately for their genes, and unfortunately for the individuals, humans find the descendants of two of those dinosaur species very tasty.

robo 10 May 2026 16:05 UTC
23 points
14
on: International Law Cannot Prevent Extinction Either
I’m not so pessimistic. People outside of Silicon Valley have no problems at all with “If humans make machines vastly smarter than people, that’s an extinction threat”. They just don’t believe we’ll make machines vastly smarter than people. That could change quickly will mass unemployment / other big changes.

robo 7 May 2026 19:24 UTC
3 points
0
on: Mechanistic estimation for wide random MLPs
I just went on vacation, but this public comment is to shame myself to try to reproduce this with random complex matrixes instead of real ones (they scale more nicely with e.g. conditioning and tails of eigenvalues).

robo 5 May 2026 5:21 UTC
1 point
0
on: robo’s Shortform
Evolution is an exponentiated selection bias

robo

42678643782648732 mod (43223432783 mod (3782473 mod 3423))

`42678643782648732 mod (43223432783 mod (3782473 mod 3423))`