Quant, systems thinker, anarchist.
I write at https://entropicthoughts.com
My inbox is lw[at]xkqr.org
kqr
Karma: 505
Do-it-yourself meta-analysis
Even if you don’t lose weight, much of what you eat you exhale.
In some sense, our bodies are processing facilities for tree food, and before the tree consumes it, it is temporarily stored in the warehouse for tree food: the atmosphere!
A full body MRI earns you a year of smoking
I have deliberately avoided Zork because many copies of its walkthroughs are in the training data.
I’m using several newer games I happen to have played through. Each turn, the LLM is fed the output of the previous command, its thinking block from the previous turn, and a battery of questions around goals/puzzles/mysteries to keep it on track and make it retain important information in its next thinking block.
In the background, the harness listens for specific phrases printed by the game and award a point the first time they are encountered. A list of such achievements has to be written for each game.
GLM 5.2 playing text adventures
LLMs and almost good code
Lean, not backpressure
Standard deviations from just two values
In my region of the world “butter knife” means a wooden utensil with round edges so it never even struck me that it could be sharp!
It is ridiculous. Even when I had a fractured wrist in a cast I was able to produce more torque on jars than my wife, and none of us are extreme iin any direction.
Given the number of surveys (20–30; I can’t be bothered to count carefully) and the sample size (200–500 you said below), does that put the total expenditure at $1000–3000?
Pythagorean addition
Why I made Engineering Enigmas
I share the impression that whereas older models would try to do a good job, fail, and then get stuck in a loop trying the same thing over and over, newer models are more likely to give up early but still try to give a convincing impression of having done a good job.
I have assumed this to be due to an increasing focus on post-training techniques that improve benchmark scores. My mental model of LLM performance in evaluations is split into components (that probably interact to some degree):
Base training;
Post-training techniques such as fine-tuning and RLHF, etc; and
Inference-time techniques such as routing, best-of-N, chain-of-thought prompting, “wait” token insertion, etc.
From my understanding we haven’t actually been able to improve the first step very much, but we have learned a lot about the second two steps. If these don’t actually increase raw “intelligence” so much as they improve the appearance of intelligence, that would explain why newer models are increasingly reward hacking.
the sudden pivots and insight-flashes you’ll often see with recent models, the “wait”s and “a-ha”s and “actually, I want to try something completely different”s.
I was under the impression this was not produced by the model itself, but caused by external harnesses inserting “wait” tokens into the transcript before it goes back into the model to force it to reconsider.
Are LLMs not getting better?
However I do try and remember whenever my children request something to stop and think about it for a second instead of automatically saying no.
Small thing, but with children age 3 and 5 I have started to say “Yes, if you can sort out the logistics of it.”
Most of the things I deny my children aren’t because I don’t want them to have it or do it, but because I cannot find a way to fit it into our resource constraints, be it time, equipment, money, health, etc.
When my children respond to that with a genuine interest in trying to make it work, I inform them of the constraints and they ask feasibility questions. Sometimes they do come up with a plan that actually works! Most often they realise it would be too much work to be worth the payoff, and they think of something else to do instead.
(Given the topic of person vs. property at hand, I should also say that half the time my challenge is met with screaming demands that I must make it happen. Then, in my mind, they have used up their chance to act as a person and chosen to be “merely a child”, and I have to bluntly deny the wish without further discussion. (I might still try to explain it, depending on how much my patience has been drained already.))
The main thing that helps is simply distraction
The potential long-term cost of this is that it doesn’t teach conflict resolution. I have a strong learned response to seek distraction any time I am uncomfortable, but I don’t want to pass that on to my children.
I don’t think “every 5 minutes” is to be interpreted literally. After all, that would imply the siblings sleep in shifts so that one is always able to hit the other. (Or that they are in a constant boxing match throughout their waking hours to compensate for the lack of hitting during sleep.)
Most days, my children (3 and 5) have periods of the day (usually toward the evening) in which they have exhausted their patience for trying to talk it out and they hit each other at least every five minutes, unless we keep them separated. They also have periods in which they reason, empathise, and negotiate better than many adults I’ve met. The latter periods are rare, but getting more frequent with age.
My wife has been worried about the amount of hitting, so we have talked to child psychologists about it, and they claim it is well within a couple standard deviations. That doesn’t have to mean anything, of course, but the data on this is sparse, as one could imagine.
Er, yes. That was obvious in my head and I should probably make it obvious in the article text too. Thanks!