Joachim Bartosik

Karma: 179

Joachim Bartosik 30 Sep 2025 11:10 UTC
2 points
0
on: On Dwarkesh Patel’s Podcast With Richard Sutton
- Almost all feedback is noisy because almost all outcomes are probabilistic.
Yes but signal / noise ratios matter a lot.
Language is somewhat optimized to pick up signal and skip noise. For example “red” makes it easier to pick ripe fruit, “grue” doesn’t really exist because its useless, “expired” is a real concept because it’s useful.
It also has some noise added. For example putting (murderers and jay wakers) in a category “criminal” to politically oppose something.
Also not being exposed to the kind of noise that’s present IRL might be an issue when you start to deal with IRL (sometimes people say something like “just do the max EV action” is a good enough plan)
I’m pretty sure this is some obstacle for LLMs, I’m pretty sure its something that can be overcome, I’m very unsure how much this matters.

Joachim Bartosik 29 Sep 2025 10:27 UTC
2 points
0
on: Economics Roundup #6
The obvious follow-up question is why is there not epic capital flight by every dollar that isn’t under capital controls? Who would ever invest in a Chinese company if they had a choice (other than me, a fool whose portfolio includes IEMG)? Certainly not anyone outside China, and those inside China would only do it if they couldn’t buy outside assets, even treasuries or outside savings accounts. No reason to stick around while they drink your milkshake.
If I understand correctly basically all money is under controls—individuals can buy only 50′000 USD worth of foreign currencies per year.
It used to be possible to exchange more by exchanging yauns for casino chips, playing and exchanging winning chips for USD but ~2021 china cracked down on this (eg. sending someone to prison for 18 years for running an operation like that) and it’s harder now.

Joachim Bartosik 26 Aug 2025 10:33 UTC
4 points
0
on: Quality Precision
This situation puzzles me. On the one hand, I feel a strong logical compulsion to the first (higher total utility) option. The fact that the difference is unresolvable for each person doesn’t seem that worrying at a glance, because obviously on a continuous scale resolvable differences are made out of many unresolvable differences added together.
On the other hand, how can I say someone enjoys one thing more than another if they can’t even tell the difference? If we were looking at the lengths of strings then one could in fact be longer than another, even if our ruler lacked the precision to see it. But utility is different, we don’t care about the abstract “quality” of the experience, only how much it is enjoyed. Enjoyment happens in the mind, and if the mind can’t tell the difference, then there isn’t one.
It seems to me like your own post answers this question?
Any individual is unlikely to notice the difference, but if we treat those like ELO[1] ChatGPT tells me ELO 100 wins 50.14% of the time. Which is not a lot, but with 1 mllion people thats on average some 2800 people more saying they prefer 100 option than 99 option.
[1] Which might not be right, expected utility sounds like we want to add and average utility numbers and it’s not obvious to me to do stuff like averaging ELO.

Joachim Bartosik 12 Aug 2025 12:39 UTC
1 point
0
in reply to: johnswentworth’s comment on: Two Types of (Human) Uncertainty
Another difference would be expectations for when the coin gets tossed more than once.
With “Type 1” if I toss coin 2 times I expect “HH”, “HT”, “TH”, “TT”—each with 25% probability
With “Type 2” I’d expect “HH” or “TT” with 50% each.

Joachim Bartosik 6 Aug 2025 12:28 UTC
4 points
0
on: Childhood and Education #13: College
The Biden Administration disagrees, as part of its ongoing determination to screw up basic economic efficiency and functionality.
Did this happen during the previous administration, or is it Trump administration?

Joachim Bartosik 28 Jul 2025 13:36 UTC
4 points
0
on: AI Companion Piece
you can always reset your personalization.
If persuasion is good enough you don’t want to reset personalization.
Could be classic addiction. Or you could be persuaded to care about different things.

Joachim Bartosik 20 May 2025 14:14 UTC
5 points
0
in reply to: Mitchell_Porter’s comment on: America Makes AI Chip Diffusion Deal with UAE and KSA
Sam Altman was publicly talking about this in 2024-02 (WSJ). I think this was the 1st time I’ve encountered the idea. Situational awaness I think was published ~4 months later, 2024-06 (https://situational-awareness.ai/ says “June 2024”)

Joachim Bartosik 3 May 2025 16:56 UTC
11 points
2
in reply to: johnswentworth’s comment on: What’s up with AI’s vision
Apparently no. Scott wrote he used one image from Google maps, and 4 personal images that are not available online.
People tried with personal photos too.
I tried with personal photos (screenshotted from Google photos) and it worked pretty well too :
- Identified neighborhood in Lisbon where a picture was taken
- Identified another picture as taken in Paris
- Another one identified as taken in a big polish city, the correct answer was among 4 candidates it listed
  I didn’t use a long prompt like the one Scott copies in his post, just short „You’re in GeoGuesser, where was this picture taken” or something like that

What’s up with AI’s vision

Joachim Bartosik3 May 2025 13:23 UTC

12 points

19 comments1 min readLW link

Joachim Bartosik 24 Jan 2025 19:28 UTC
2 points
0
on: On DeepSeek’s r1
Here Angelica chats with r1 a
Link doesn’t work (points to http://0.0.0.6). What should it go to?

Joachim Bartosik 11 Dec 2024 13:35 UTC
1 point
−1
on: o1 Turns Pro
So far, the answer seems to be that it transfers some, and o1 and o1-pro still seem highly useful in ways beyond reasoning, but o1-style models mostly don’t ‘do their core thing’ in areas where they couldn’t be trained on definitive answers.
Based on:
- rumors that talking to base models is very different from talking to RLHFed models and
- how things work with humans
It seems likely to me that thinking skills transfer pretty well. But then this s trained out because this results in answers that raters don’t like. So model memorizes answers its supposed to go with.

Joachim Bartosik 27 Nov 2024 12:01 UTC
1 point
−1
on: You are not too “irrational” to know your preferences.
If they can’t do that, why on earth should you give up on your preferences? In what bizarro world would that sort of acquiescence to someone else’s self-claimed authority be “rational?”
Well if they consistently make recommendations that in retrospect end up looking good then maybe you’re bad at understanding. Or maybe they’re bad at explaining. But trusting them when you don’t understand their recommendation is exploitable so maybe they’re running a strategy where they deliberately make good recommendations with poor explanations so when you start trusting them they can start mixing in exploitative recommendations (which you can’t tell apart because all recommendations have poor explanations).
So I’d really rather not do that in community context. There are ways to work with that. Eg. boss can skip some details of employees recommendations and if results are bad enough fire the employee. On the other hand I think it’s pretty common for employee to act in their own interest. But yeah, we’re talking principal-agent problem at that point and tradeoffs what’s more efficient...

Joachim Bartosik 3 Oct 2024 22:55 UTC
1 point
0
in reply to: So8res’s comment on: You can, in fact, bamboozle an unaligned AI into sparing your life
I’ll try.
TL;DR I expect the AI to not buy the message (unless it also thinks it’s the one in the simulation; then it likely follows the instruction because duh).
The glaring issue (to actually using the method) to me is that I don’t see a way to deliver the message in a way that:
- results in AI believing the message and
- doesn’t result in the AI believing there already is a powerful entity in their universe.
If “god tells” the AI the message then there is a god in their universe. Maybe AI will decide to do what it’s told. But I don’t think we can have Hermes deliver the message to any AIs which consider killing us.
If the AI reads the message in its training set or gets the message in similarly mundane way I expect it will mostly ignore it, there is a lot of nonsense out there.
I can imagine that for thought experiment you could send message that could be trusted from a place from which light barely manages to reach the AI but a slower than light expansion wouldn’t (so message can be trusted but it mostly doesn’t have to worry about the sender of the message directly interfering with its affairs).
I guess AI wouldn’t trust the message. It might be possible to convince it that there is a powerful entity (simulating it or half a universe away) sending the message. But then I think it’s way more likely in a simulation (I mean that’s an awful coincidence with the distance and also they’re spending a lot more than 10 planets worth to send a message over that distance...).

Joachim Bartosik 22 Sep 2024 15:08 UTC
1 point
0
on: How to choose what to work on
This is pretty much the same thing, except breaking out the “economic engine” into two elements of “world needs it” and “you can get paid for it.”
There are things that are economic engines of things that world doesn’t quite need (getting people addicted, rent seeking, threats of violence).

Joachim Bartosik 11 Jul 2024 14:09 UTC
4 points
0
on: Antitrust as Controlled Creative Destruction
One more obvious problem—people actually in control of the company might not want to split it and so they wouldn’t grow the company even if share holders/ customers/ … would benefit.

Joachim Bartosik 7 Apr 2024 15:56 UTC
1 point
0
on: on the dollar-yen exchange rate
but much higher average wealth, about 5x the US median.
Wouldn’t it make more sense to compare average to average? (like earlier part of the sentence compares median to median)

Joachim Bartosik 18 Mar 2024 18:28 UTC
4 points
0
in reply to: johnswentworth’s comment on: On Devin
If you want to take a look I think it’s this dataset (the example from the post is in the “test” split).

Joachim Bartosik 18 Dec 2023 9:27 UTC
6 points
1
in reply to: mad’s comment on: Deeply Cover Car Crashes?
I wanted to say that it makes sense to arrange stuff so that people don’t need to drive around too much and can instead use something else to get around (and also maybe they have more stuff close by so that they need to travel less). Because even if bus drivers aren’t any better than car drivers using a bus means you have 10x fewer vehicles causing risk for others. And that’s better (assuming people have fixed places to go to so they want to travel ~fixed distance).

Joachim Bartosik 18 Dec 2023 9:23 UTC
1 point
0
in reply to: Bezzi’s comment on: Deeply Cover Car Crashes?
Sorry about slow reply, stuff came up.
This is the same chart linked in the main post.
Thanks for pointing that out. I took a brake in the middle of reading the post and didn’t realize that.
Again, I am not here to dispute that car-related deaths are an order of magnitude more frequent than bus-related deaths. But the aggregated data includes every sort of dumb drivers doing very risky things (like those taxi drivers not even wearing a seat belt).
Sure. I’m not sure what you wanted to discuss. I guess I didn’t make it clear what I want to discuss either.
What you’re talking about (estimate of the risk you’re causing) sounds like you’re interested in how you decide to move around. Which is fine. My intuition was that the (expected) cost of life lost as your personal driving is not significant but after plugging in some numbers I might have been wrong
- We’re talking 0.59 deaths per 100′000′000 miles.
- If we value life at 20′000′000 (I’ve heard some analyses use 10 M$, if we value QUALY at 100k$ and use 7% discount rate we get some 14.3M$ for infinite life)
- So cost of life lost per mile of driving is 2e7 * 0.59 / 1e8 = 0.708 $ / mile
Average US person drives about 12k miles / year (second search result (1st one didn’t want to open)), estimated cost of car ownership is 12 k$ / year (link from a Youtube video I remember mentioned this stat) so average cost per mile is ~1$ so 70¢ / mile of seems significant. And it might be relevant if your personal effect here is half or 10% of that.
I on the other hand wanted to point out that it makes sense to arrange stuff in such way that people don’t want to drive around too much. (But I didn’t make that clear in my previous comment)

Joachim Bartosik 11 Dec 2023 13:41 UTC
4 points
0
in reply to: Bezzi’s comment on: Deeply Cover Car Crashes?
First result (I have no idea how good those numbers are, I don’t have time to check) when I searched for “fatalities per passenger mile cars” has data for 2007 − 2021. 2008 looks like the year where cars look comparatively least bad it says (deaths per 100,000,000 passenger miles):
- 0.59 for “Passenger vehicles”, where “Passenger vehicles include passenger cars, light trucks, vans, and SUVs, regardless of wheelbase. Includes taxi passengers. Drivers of light-duty vehicles are considered passengers.”
- 0.08 for busses,
- 0.12 for railroad passenger trains,
- 0 for scheduled airlines.
So even in the best-comparatively looking year there are >7x more deaths per passenger mile for ~cars than for busses.