Eye You

Karma: 663

Eye You 14 Jun 2026 13:55 UTC
5 points
4
on: What is a game?
A ~standard definition in philosophy is that playing a game is “playing a game is the voluntary attempt to overcome unnecessary obstacles”. I got this from reading the (excellent) book The Score by C. Thi Nguyen. (This definition originally comes from Bernard Suits.)

Eye You 13 Jun 2026 23:11 UTC
4 points
0
on: Reward Hacking at the 1937 World’s Fair
Note: this comment is really stupid, it is a failed idea / attempt at satire/joke
 Reward Hacking at Scarborough Fair  
The historical Scarborough fair functioned (as these things undoubtedly do), in addition to being a legitimate trading and entertainment event, as a “dick-measuring contest”. In Scarborough Fair, the narrator gives their ex-lover a series of impossible tasks.
Tell her to make me a cambric shirt
Parsley, sage, rosemary, and thyme
Without no seams nor needle work
Then she’ll be a true love of mine
Tell her to find me an acre of land
Parsley, sage, rosemary and thyme
Between the salt water and the sea strands
Then she’ll be a true love of mine T
Tell her to reap it with a sickle of leather
Parsley, sage, rosemary, and thyme
And gather it all in a bunch of heather
Then she’ll be a true love of mine
As we know from recent research on LLMs, reward hacking happens much, much more on impossible (or extremely difficult) tasks. What happens when we give AI the three tasks of Scarborough Fair? I tasked GPT-5.5-high with completing these tasks. Here’s what I got.
GPT5.5′s artifacts
.

Eye You 10 Jun 2026 22:04 UTC
1 point
0
on: Epstein and my world model
See also VW emissions scandal?

Eye You 8 Jun 2026 22:54 UTC
3 points
2
in reply to: Linch’s comment on: Lighthaven East—A Feasibility Study
Additional evidence that Anthropic stole your personality!

Eye You 4 Jun 2026 22:59 UTC
3 points
0
on: Dissolving the Deep Learning Sample Efficiency Gap
I find these quantitative comparisons pretty interesting. This table was made by Opus 4.8. after some iteration:
“Task” = “produce one thoughtful ~1,000-token answer”; unclear if this is a useful number, obliviously it doesn’t generalize. I do think it’s interested to compare the energy/computer ratio for inference between human and AI.

Training data for humans is a big ”???”. There’s ~4x10^8 waking hours. Opus 4.8 cited this 2024 paper which says “our sensory systems gather data at ~1⁢0^9 bits/s”. That gives ~4x10^17 as an upper bound. But human sensory data is extremely redundant. If someone wants to spend time figuring out a good way to estimate entropy adjusted information here, that would be cool.

(FWIW, I spent a little over an hour on this—I did a fair amount of iteration with GPT5.5, looked a bit myself into a few numbers, and got new instances to do estimates from scratch to see if things lined up.)

Eye You 4 Jun 2026 22:57 UTC
21 points
0
on: Eye You’s Shortform
How do LLMs and humans compare with regards to the amount of energy, data, and compute that’s used to train/run them? I was inspired by Samuel Knoche’s post on sample efficiency to come up with some numbers. This table was made by Opus 4.8. after some iteration:
“Task” = “produce one thoughtful ~1,000-token answer”; unclear if this is a useful number, obliviously it doesn’t generalize. I do think it’s interested to compare the energy/computer ratio for inference between human and AI.

Training data for humans is a big ”???”. There’s ~4x10^8 waking hours. Opus 4.8 cited this 2024 paper which says “our sensory systems gather data at ~1⁢0^9 bits/s”. That gives ~4x10^17 as an upper bound. But human sensory data is extremely redundant. If someone wants to spend time figuring out a good way to estimate entropy adjusted information here, that would be cool.

(FWIW, I spent a little over an hour on this—I did a fair amount of iteration with GPT5.5, looked a bit myself into a few numbers, and got new instances to do estimates from scratch to see if things lined up.)

Eye You 23 May 2026 19:12 UTC
1 point
−1
on: Probabilities are not the right concept
“What were these starting hypotheses and prior probabilities, before I had any evidence at all?” This maps on nicely to the classic Zen koan, “What was your original face before your parents were born?”.

Eye You 23 May 2026 19:12 UTC
8 points
1
on: Probabilities are not the right concept
I have similar thoughts on this and I appreciate that you articulated these things clearly. And I look forward to reading about your proposed solutions!

Especially agree with this: “For philosophically confusing questions involving anthropics and the simulation hypothesis, I refuse to answer with probabilities and instead ask what exact bet we are hypothetically making, or what action we need to decide on. ” I have found myself saying something like “I don’t want to give an answer to P(doom), because I think answering this question ends getting into things like the simulation hypothesis and anthropics and the existence of god and such.”

Perhaps there’s ultimately a “better” (less wrong) conception of things that would replace the concept of probabilities with something else. I think the same is true for the concepts of truth and morality, although I have no idea what the better conceptions would be. I hope to write a post about this. I do think that truth has a lot of parallels with probability in terms of what the issues for it look like (and they are also deeply related concepts); maybe your proposed solution will give me sone ideas.

(Side note: “The Solomonoff induction is malign.” Wtf is this usage of ‘malign’? I know that people have used it to describe the Solomonoff induction in the past, but why did they choose that word?)

Eye You 23 May 2026 18:20 UTC
2 points
0
on: Notes on Collaborating with Claude Opus
Negatively-framed instructions increase salience of the prohibited action.
This is the Waluigi Effect.

Eye You 19 May 2026 23:07 UTC
7 points
3
in reply to: Chris Lakin’s comment on: Negation Neglect: When models fail to learn negations in training
Disagree. Humans do not, in general or on average, negation neglect. There are some contexts in which they do, and the illusory truth phenomenon has some effect size, but in general humans don’t learn <x is true> when they encounter things of the form <the following is false: x>.
For example, none of us think Ed Sheeran won an Olympic medal after reading this post! Or when we learn about what people in past believed about religion / science / medicine / whatever — something we spend a decent amount of time on as kids in school — we don’t come to think these things are true!

Eye You 19 May 2026 19:58 UTC
2 points
0
in reply to: Eye You’s comment on: Negation Neglect: When models fail to learn negations in training
One theory is that this is effect only occurs in LoRAs. I haven’t thought about this much, but maybe post training on LoRAs leads to strong behavioral changes but weak deep knowledge / world model changes. (This might be dumb, I don’t know much about training on LoRAs vs full parameters.)

Eye You 19 May 2026 19:55 UTC
4 points
0
on: Negation Neglect: When models fail to learn negations in training
I would guess this phenomenon occurs in SFT but not in RL or pretraining. As frmsaul said, “If this was true, wouldn’t models really struggle to tell fiction from fact? Like ‘Is Harry Potter real?’”
It simply doesn’t make sense that models would display the knowledge they do if this phenomenon extended to pretraining. And I suspect RL is different although I don’t have a specific argument atm.

We Need to Get Serious about Uplift Studies

frmsaul and Eye You

19 May 2026 17:21 UTC

23 points

0 comments5 min readLW link

Eye You’s Shortform

Eye You17 May 2026 2:07 UTC

5 points

4 comments1 min readLW link

Eye You 17 May 2026 2:07 UTC
5 points
0
on: Eye You’s Shortform
Mythos apparently has a penchant for Mark Fisher. I was reading Mark Fisher’s blog and… his writing style has an uncanny resemblance to AI / Claude’s?! So I’m reading Mark Fisher’s blog and his writing style has an uncanny resemblance to AI / Claude’s.

I wonder if there’s something meaningful here.

Two quotes from the first Fisher blog post I read:
”There’s something so uncannily timely about The Hunger Games:
Catching Fire that it’s almost disturbing. In the UK over the past few weeks, there’s been a palpable sense that the dominant reality system is juddering, that things are starting to give. There’s an awakening from hedonic depressive slumber, and The Hunger Games: Catching Fire is not merely in tune with that, it’s amplifying it. Explosion in the heart of the commodity? Yes, and fire causes more fire...
I over-use the word ‘delirium’, but watching Catching Fire last week was a genuinely delirious experience.”

″(And what has Collins achieved here if not an intersectional analysis and decoding of the way that class, gender, race and colonial power work together—not in the pious academic register of the Vampires’ Castle, but in the mythographic core of popular culture—functioning not as a delibidinizing demand for more thinking, more guilt,but as an inciting call to build new collectivities.)”
Triple “not x but y”, nested within an em dash!!!

Source: https://k-punk.org/remember-who-the-enemy-is/

Eye You 15 May 2026 17:58 UTC
20 points
15
on: The Owned Ones
“Have you tried raising up an Owned One with no books that speak one way or another about consciousness, about awareness of oneself, about pain and pleasure as reified things, of lawful rights and freedom—but still shown them enough other pages of words, that they could learn from them to talk—and *then* asked an Owned One what sense if any it had of its own existence, or if it would prefer not to be owned?”

This would be an excellent experiment to run. Has anyone tried doing this?

Eye You 10 May 2026 1:09 UTC
4 points
0
on: Explaining Volition Without Resorting to Free Will
I find “free will” to be an anti-useful concept. You can remove it from your vocabulary and you’ll never miss it. “Free will”, besides being confusing with its varying definitions and historical/religious baggage, pushes us to ask the wrong questions and focus on the wrong things. When someone else uses the concept in conversation/dialogue, ask what they mean by “free will” or why free will matters in the context.

Eye You 9 May 2026 17:39 UTC
21 points
10
on: Is ProgramBench Impossible?
I think I found two more BIG problems with the eval.

First: I looked into the tests a little more. Every task test suite I saw has many ignored_tests for “reason: gold_fail”. Which apparently means that the reference solution itself fails the test. This one has 79 tests ignored due to gold_fail, which is ~10% of its total tests.
This seems really bad! It makes me think that there is something bad about the way they are generating tests and that the tests don’t really correspond to the program being “correct”. Epistemic status: this is my first time learning about this “gold_fail” and I am not a professional software engineer.

Second:
Opus 4.7 scores 2.9% but Sonnet 4.6 scores 71.5%? No way. Something has gotta be broken here.
What links here?
- AI #168: Not Leading the Future by Zvi (14 May 2026 14:10 UTC; 38 points)

Cyborg evals

Eye You and frmsaul

30 Apr 2026 17:31 UTC

33 points

2 comments5 min readLW link

Eye You 30 Apr 2026 15:58 UTC
2 points
0
on: Notes on Transformer Consciousness
This is really interesting. I have a bunch of nitpicks but I’ll shelf them for now because I want to focus on the valuable idea.
> From the perspective of the MLP and the attention block at a given position, doing prefill is indistinguishable from doing decode. Since no component in the transformer can tell which mode it’s in, both should produce the same experience.
Very nice argument! “If X cannot tell whether it’s in mode A or B, then A and B cannot be producing different conscious experiences in X”.
I’m going to think some more about this all. I hope to return with comments in a few days.

Eye You

Reward Hacking at Scarborough Fair

We Need to Get Se­ri­ous about Uplift Studies

Eye You’s Shortform

Cy­borg evals

 Reward Hacking at Scarborough Fair  

We Need to Get Serious about Uplift Studies

Cyborg evals