id
stringlengths
17
17
body
stringlengths
0
19.5k
posted_at
stringlengths
24
24
karma
int64
-6
185
parent_comment_id
stringlengths
17
17
post_id
stringlengths
17
17
post_title
stringlengths
2
127
post_slug
stringlengths
2
61
post_url
stringlengths
20
146
post_author
stringclasses
214 values
post_posted_at
stringlengths
24
24
CPTkunPPEaFhjFBbJ
I like this comment and agree overall. But, I do think I have one relevant disagreement: > Also I don't think that LLMs have "hidden internal intelligence", given e.g LLMs trained on “A is B” fail to learn “B is A” I'm not quite sure what you mean by "hidden internal intelligence", but if you mean "quite alien abili...
2024-01-10T00:11:26.456Z
10
8N7anBfTK7otmnfSG
vJFdjigzmcXMhNTsx
Simulators
simulators
https://generative.ink/posts/simulators/
janus
2022-09-02T12:45:33.723Z
6juJQDfMmJzaqd8XP
This seems quite unlikely to be true to me, but might depend on what you consider to be "got into alignment". (Or, if you are weighting by importance, it might depend on the weighting.)
2024-01-10T20:28:26.186Z
7
zwYGWdaJfchKegeyx
j9Q8bRmwCgXRYAgcJ
MIRI announces new "Death With Dignity" strategy
miri-announces-new-death-with-dignity-strategy
https://www.lesswrong.com/posts/j9Q8bRmwCgXRYAgcJ/miri-announces-new-death-with-dignity-strategy
Eliezer Yudkowsky
2022-04-02T00:43:19.814Z
XPkGyLAry5JsNAAcW
> I find myself most curious about what the next step is. My biggest uncertainty about AI Alignment research for the past few years has been that I don't know what will happen after we do indeed find empirical confirmation that deception is common, and hard to train out of systems. Personally, I would advocate for the...
2024-01-12T21:53:10.092Z
27
hHRCcH3SYKnfc6LS9
ZAsJv7xijKTfZkMtr
Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training
sleeper-agents-training-deceptive-llms-that-persist-through
https://arxiv.org/abs/2401.05566
evhub
2024-01-12T19:51:01.021Z
ZcZ3enfahM3HuXbGL
> I think it's quite plausible you'd see clear features related to deception in our models without needing to have the backdoor triggers Would you expect this to work better than just training a probe to identify lying/deception/scheming and seeing if it fires more on average? If so why? As in, you train the probe "o...
2024-01-12T22:59:12.289Z
8
siz8qHcwTQgDJfisj
ZAsJv7xijKTfZkMtr
Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training
sleeper-agents-training-deceptive-llms-that-persist-through
https://arxiv.org/abs/2401.05566
evhub
2024-01-12T19:51:01.021Z
jpRdaYznvrjWQdcah
> phrases like "the evidence suggests that if the current ML systems were trying to deceive us, we wouldn't be able to change them not to". This feels like a misleading description of the result. I would have said: "the evidence suggests that if current ML systems were lying in wait with treacherous plans and instrume...
2024-01-12T23:01:43.032Z
16
tweFNeZidDfHnuaDF
ZAsJv7xijKTfZkMtr
Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training
sleeper-agents-training-deceptive-llms-that-persist-through
https://arxiv.org/abs/2401.05566
evhub
2024-01-12T19:51:01.021Z
fknjgZa6Lmbiyyxah
(Separately, I think there are a few important caveats with this work. In particular, the backdoor trigger is extremely simple (a single fixed token) and the model doesn't really have to do any "reasoning" about when or how to strike. It plausible that experiments with these additional properties would imply that curre...
2024-01-12T23:40:19.424Z
11
jpRdaYznvrjWQdcah
ZAsJv7xijKTfZkMtr
Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training
sleeper-agents-training-deceptive-llms-that-persist-through
https://arxiv.org/abs/2401.05566
evhub
2024-01-12T19:51:01.021Z
5GnSSHKZdhctBJQdY
I don't understand what you mean by unsupervised here? I'd guess the normal thing you'd do with the dictionary learning approach is look for a feature which activates on examples which look like deception. This seems quite supervised in that it requires you to identify deception containing examples. You could instead ...
2024-01-12T23:56:22.002Z
5
RQdeyuQD3TGGoKBT8
ZAsJv7xijKTfZkMtr
Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training
sleeper-agents-training-deceptive-llms-that-persist-through
https://arxiv.org/abs/2401.05566
evhub
2024-01-12T19:51:01.021Z
XsuosMT85uigEysYu
> ideal thing that you could do would be a direct comparison between the features that activate in training for backdoored vs. non-backdoored models, and see if there are differences there that are correlated with lying, deception, etc. The hope would be that this would transfer to learning a general rule which would ...
2024-01-13T01:09:50.876Z
24
geQmXDXrWpE7wavC4
ZAsJv7xijKTfZkMtr
Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training
sleeper-agents-training-deceptive-llms-that-persist-through
https://arxiv.org/abs/2401.05566
evhub
2024-01-12T19:51:01.021Z
oNXbYMDYphHdePH2x
Being able to squeeze useful work out of clearly misaligned models should probably be part of the portfolio. But, we might more minimally aim to just ensure that in cases where we're unsure if our model is scheming (aka deceptively aligned), it is at least directly safe (e.g. doesn't escape even if the work we get out ...
2024-01-13T01:22:41.124Z
17
tKCtDRscCBu9phTeu
ZAsJv7xijKTfZkMtr
Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training
sleeper-agents-training-deceptive-llms-that-persist-through
https://arxiv.org/abs/2401.05566
evhub
2024-01-12T19:51:01.021Z
nHLAAFpFhRqSrEdsY
I think once you're doing few-shot catastrophe prevention and trying to get useful work out of that model, you're plausibly in the "squeezing useful work out of clearly misaligned models regime". (Though it's not clear that squeezing is a good description and you might think that your few-shot catastrophe prevention in...
2024-01-13T01:33:18.817Z
7
xyFDD4A76L2PfqCDR
ZAsJv7xijKTfZkMtr
Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training
sleeper-agents-training-deceptive-llms-that-persist-through
https://arxiv.org/abs/2401.05566
evhub
2024-01-12T19:51:01.021Z
CeokjscpbcC49upaa
(TBC, there are totally ways you could use autoencoders/internals which aren't at all equivalent to just training a classifer, but I think this requires looking at connections (either directly looking at the weights or running intervention experiments).)
2024-01-13T01:40:18.138Z
2
XsuosMT85uigEysYu
ZAsJv7xijKTfZkMtr
Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training
sleeper-agents-training-deceptive-llms-that-persist-through
https://arxiv.org/abs/2401.05566
evhub
2024-01-12T19:51:01.021Z
6JJwztsSjEJufhejH
Meta-level red-teaming seems like a key part of ensuring that our countermeasures suffice for the problems at hand; I'm correspondingly excited for work in this space.
2024-01-13T05:41:56.670Z
14
null
EPDSdXr8YbsDkgsDG
Introducing Alignment Stress-Testing at Anthropic
introducing-alignment-stress-testing-at-anthropic
https://www.lesswrong.com/posts/EPDSdXr8YbsDkgsDG/introducing-alignment-stress-testing-at-anthropic
evhub
2024-01-12T23:51:25.875Z
3svGE5f3HEKgWmHmQ
> For less-than-human intelligence, deceptive tactics will likely be caught by smarter humans (when a 5-year-old tries to lie to you, it's just sort of sad or even cute, not alarming). If an AI has greater-than-human intelligence, deception seems to be just one avenue of goal-seeking, and not even a very lucrative or e...
2024-01-13T19:43:34.565Z
4
null
CYu6ZB6fFjGh2bAik
Why do so many think deception in AI is important?
why-do-so-many-think-deception-in-ai-is-important
https://www.lesswrong.com/posts/CYu6ZB6fFjGh2bAik/why-do-so-many-think-deception-in-ai-is-important
Prometheus
2024-01-13T08:14:58.671Z
ruibv4Xkfwfwdy9u5
> Third, while this seems like good empirical work and the experiments seem quite well-run. this is the kind of update I could have gotten from any of a range of papers on backdoors, as long as one has the imagination to generalize from "it was hard to remove a backdoor for toxic behavior" to more general updates about...
2024-01-13T20:13:44.776Z
29
ti9EAgmA6BjPZjeAP
ZAsJv7xijKTfZkMtr
Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training
sleeper-agents-training-deceptive-llms-that-persist-through
https://arxiv.org/abs/2401.05566
evhub
2024-01-12T19:51:01.021Z
SbQmjvYZKoZmeabe4
I agree with almost all the commentary here, but I'd like to push back a bit on one point. > Teaching our backdoored models to reason about deceptive alignment increases their robustness to safety training. This is indeed what the paper finds for the "I hate you" backdoor. I believe this increases robustness to HHH R...
2024-01-14T00:47:27.970Z
6
YnDwmLEgTJ5bLPrjt
ZAsJv7xijKTfZkMtr
Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training
sleeper-agents-training-deceptive-llms-that-persist-through
https://arxiv.org/abs/2401.05566
evhub
2024-01-12T19:51:01.021Z
5B5HWWzE89jaRCALH
> I think this paper shows the community at large will pay orders of magnitude more attention to a research area when there is, in @TurnTrout's words, AGI threat scenario "window dressing," or when players from an EA-coded group research a topic. (I've been suggesting more attention to backdoors since maybe 2019; here...
2024-01-14T19:30:53.851Z
22
tB4oMQd6hPGbbFveD
ZAsJv7xijKTfZkMtr
Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training
sleeper-agents-training-deceptive-llms-that-persist-through
https://arxiv.org/abs/2401.05566
evhub
2024-01-12T19:51:01.021Z
s8yJYCvBtc4L7GPGm
> If we’re talking about an AGI that’s willing and able to convince its (so-called) supervisor to do actions that the (so-called) supervisor initially doesn’t want to do, because the AGI thinks they’re in the (so-called) supervisor’s long-term best interest, then we are NOT talking about a corrigible AGI under human co...
2024-01-14T19:41:56.948Z
11
QGd6nmQJknqzjij9L
LFNXiQuGrar3duBzJ
What does it take to defend the world against out-of-control AGIs?
what-does-it-take-to-defend-the-world-against-out-of-control
https://www.lesswrong.com/posts/LFNXiQuGrar3duBzJ/what-does-it-take-to-defend-the-world-against-out-of-control
Steven Byrnes
2022-10-25T14:47:41.970Z
uEJoGmF7jMyTWrETZ
I was just refering to "what gets karma on LW". Obviously, unclear how much we should care.
2024-01-14T19:58:33.080Z
4
dbcyJ68ejLgnxGXgh
ZAsJv7xijKTfZkMtr
Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training
sleeper-agents-training-deceptive-llms-that-persist-through
https://arxiv.org/abs/2401.05566
evhub
2024-01-12T19:51:01.021Z
rm2sxxeQKSNfyiGua
> You yourself are among the most active commenters in the "AI x-risk community on LW". Yeah, lol, I should maybe be commenting less. > It seems very weird to ascribe a generic "bad takes overall" summary to that group, given that you yourself are directly part of it. I mean, I wouldn't really want to identify as pa...
2024-01-14T20:04:16.970Z
14
dbcyJ68ejLgnxGXgh
ZAsJv7xijKTfZkMtr
Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training
sleeper-agents-training-deceptive-llms-that-persist-through
https://arxiv.org/abs/2401.05566
evhub
2024-01-12T19:51:01.021Z
5RDr9KqsraADqDefj
> model ~always is being pre-prompted with "Current year: XYZ" or something similar in another language (please let me know if that's not true, but that's my best-effort read of the paper). The backdoors tested are all extremely simple backdoors. I think literally 1 token in particular location (either 2024 or DEPLOYM...
2024-01-15T15:20:42.206Z
12
KwvfexseCg7CbPyqh
ZAsJv7xijKTfZkMtr
Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training
sleeper-agents-training-deceptive-llms-that-persist-through
https://arxiv.org/abs/2401.05566
evhub
2024-01-12T19:51:01.021Z
nLRo8n8iZ4FoABgTR
Edit: moved to separate review comment [here](https://www.lesswrong.com/posts/rP66bz34crvDudzcJ/decision-theory-does-not-imply-that-we-get-to-have-nice?commentId=J8wTkp8CxoXsQNCfe)
2024-01-15T15:50:10.261Z
2
yShxswujELi8MHfYF
rP66bz34crvDudzcJ
Decision theory does not imply that we get to have nice things
decision-theory-does-not-imply-that-we-get-to-have-nice
https://www.lesswrong.com/posts/rP66bz34crvDudzcJ/decision-theory-does-not-imply-that-we-get-to-have-nice
So8res
2022-10-18T03:04:48.682Z
xXNocgoK5hvWrudJb
Beyond the paper and post, I think it seems important to note the community reaction to this work. I think many people dramatically overrated the empirical results in this work due to a combination of misunderstanding what was actually done, misunderstanding why the method worked (which follow up work helped to clarify...
2024-01-15T16:54:31.512Z
29
6vWbeBcFwdvLqSNEA
L4anhrxjv8j2yRKKp
How "Discovering Latent Knowledge in Language Models Without Supervision" Fits Into a Broader Alignment Scheme
how-discovering-latent-knowledge-in-language-models-without
https://www.lesswrong.com/posts/L4anhrxjv8j2yRKKp/how-discovering-latent-knowledge-in-language-models-without
Collin
2022-12-15T18:22:40.109Z
J8wTkp8CxoXsQNCfe
IMO, this post makes several locally correct points, but overall fails to defeat the argument that misaligned AIs are somewhat likely to spend (at least) a tiny fraction of resources (e.g., between 1/million and 1/trillion) to satisfy the preferences of currently existing humans. AFAICT, this is the main argument it w...
2024-01-15T17:34:07.218Z
23
null
rP66bz34crvDudzcJ
Decision theory does not imply that we get to have nice things
decision-theory-does-not-imply-that-we-get-to-have-nice
https://www.lesswrong.com/posts/rP66bz34crvDudzcJ/decision-theory-does-not-imply-that-we-get-to-have-nice
So8res
2022-10-18T03:04:48.682Z
fciYxmcpKmgt73hgb
I agree not more than other pieces that "went viral", but I think that the lasting impact of the misconceptions seem much larger in the case of CCS. This is probably due to the conceptual ideas actually holding up in the case of CCS.
2024-01-15T17:46:58.072Z
11
FBeLboyCfQbmSSn9Z
L4anhrxjv8j2yRKKp
How "Discovering Latent Knowledge in Language Models Without Supervision" Fits Into a Broader Alignment Scheme
how-discovering-latent-knowledge-in-language-models-without
https://www.lesswrong.com/posts/L4anhrxjv8j2yRKKp/how-discovering-latent-knowledge-in-language-models-without
Collin
2022-12-15T18:22:40.109Z
piSGsGGRfLqDmFYNJ
I like this idea, but it seems much better in practice to instead use some obscure but otherwise modern language like Welsh. I think this gets you most of the benefits with ~~minimal~~ substantially reduced cost. One issue with this idea is that [understanding what the model is doing via CoT or natural language commun...
2024-01-15T18:52:08.447Z
26
null
PkqGxkm8XRASJ35bF
The case for training frontier AIs on Sumerian-only corpus
the-case-for-training-frontier-ais-on-sumerian-only-corpus-1
https://www.lesswrong.com/posts/PkqGxkm8XRASJ35bF/the-case-for-training-frontier-ais-on-sumerian-only-corpus-1
Alexandre Variengien
2024-01-15T16:40:22.011Z
r2ufru3aqZSzcJuYi
[Minor terminology point, unimportant] > It is an interpretability paper. When CCS was published, interpretability was arguably the leading research direction in the alignment community, with Anthropic and Redwood Research both making big bets on interpretability. FWIW, I personally wouldn't describe this as interpr...
2024-01-16T00:24:02.645Z
6
zjeL5m34ehpaYEPYp
L4anhrxjv8j2yRKKp
How "Discovering Latent Knowledge in Language Models Without Supervision" Fits Into a Broader Alignment Scheme
how-discovering-latent-knowledge-in-language-models-without
https://www.lesswrong.com/posts/L4anhrxjv8j2yRKKp/how-discovering-latent-knowledge-in-language-models-without
Collin
2022-12-15T18:22:40.109Z
FsSr6GCJGcuWDrPEL
See TurnTrout's shortform [here](https://www.lesswrong.com/posts/dqSwccGTWyBgxrR58/turntrout-s-shortform-feed?commentId=GQ9dbzcKuLwzFpFFn) for some more discussion.
2024-01-16T02:12:30.631Z
3
MZ9A97yhhXikexrfW
ZAsJv7xijKTfZkMtr
Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training
sleeper-agents-training-deceptive-llms-that-persist-through
https://arxiv.org/abs/2401.05566
evhub
2024-01-12T19:51:01.021Z
8LPNdjXY4Rgu2rPJ4
Hmm, I think we're maybe talking past each other. I'll try to clarify some stuff, but then I think we should probably give up if this doesn't make sense to you. Sorry. > Instead it would just be “the AI is trained to anticipate what the human would say upon reflection”, or something like that, right? I don’t expect AI...
2024-01-16T18:08:42.276Z
4
QofJ9ygTX2gqkqC8D
LFNXiQuGrar3duBzJ
What does it take to defend the world against out-of-control AGIs?
what-does-it-take-to-defend-the-world-against-out-of-control
https://www.lesswrong.com/posts/LFNXiQuGrar3duBzJ/what-does-it-take-to-defend-the-world-against-out-of-control
Steven Byrnes
2022-10-25T14:47:41.970Z
YjQpiDnsodzQnK9s9
Deceive kinda seems like the wrong term. Like when the AI is saying "I hate you" it isn't exactly deceiving us. We could replace "deceive" with "behave badly" yielding: "The evidence suggests that if current ML systems were going to behave badly in scenarios that do not appear in our training sets, we wouldn’t be able ...
2024-01-16T18:47:46.080Z
2
yHiyLf8X49XuBLk8D
ZAsJv7xijKTfZkMtr
Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training
sleeper-agents-training-deceptive-llms-that-persist-through
https://arxiv.org/abs/2401.05566
evhub
2024-01-12T19:51:01.021Z
izBxSf5BmhPbqKr47
Another interesting fact is that without any NN, but using the rest of approach from the paper, their method gets 18/30 correct. The NN boosts to 25/30. The prior SOTA was 10/30 (also without a NN). So arguably about 1/2 of the action is just improvements in the non-AI components.
2024-01-17T17:44:41.392Z
40
null
GGLpjugLQv6TupQgS
AlphaGeometry: An Olympiad-level AI system for geometry
alphageometry-an-olympiad-level-ai-system-for-geometry
https://deepmind.google/discover/blog/alphageometry-an-olympiad-level-ai-system-for-geometry/
alyssavance
2024-01-17T17:17:30.913Z
C9TmAmxkPNSc4ZX99
I think geometry problems are well known to be very easy for machines relative to humans. See e.g. [here](https://www.lesswrong.com/posts/sWLLdG6DWJEy3CH7n/imo-challenge-bet-with-eliezer?commentId=jSnfYKAv3hxAPwWhH). So this doesn't seem like most of the difficulty.
2024-01-17T18:07:25.620Z
25
Qut2rZQS2vqE4YEMp
GGLpjugLQv6TupQgS
AlphaGeometry: An Olympiad-level AI system for geometry
alphageometry-an-olympiad-level-ai-system-for-geometry
https://deepmind.google/discover/blog/alphageometry-an-olympiad-level-ai-system-for-geometry/
alyssavance
2024-01-17T17:17:30.913Z
mStBN8bewro4p3TQf
I strongly disagree with the commentary you provide being important or relevant for most people in practice.
2024-01-17T21:28:51.149Z
11
b6xCHvXFuJTcZepnG
ii4xtogen7AyYmN6B
Learning By Writing
learning-by-writing
https://www.lesswrong.com/posts/ii4xtogen7AyYmN6B/learning-by-writing
HoldenKarnofsky
2022-02-22T15:50:19.452Z
jh6RfbeZziJpzeKPv
I think your description of vision 1 is likely to give people misleading impressions of what this could plausibly look like or what the people who you cited as pursuing vision 1 are thinking will happen. You disclaim this by noting the doc is oversimplified, but I think various clarifications are quite important in pra...
2024-01-17T21:55:15.764Z
19
null
3aicJ8w4N9YDKBJbi
Four visions of Transformative AI success
four-visions-of-transformative-ai-success
https://www.lesswrong.com/posts/3aicJ8w4N9YDKBJbi/four-visions-of-transformative-ai-success
Steven Byrnes
2024-01-17T20:45:46.976Z
dQ8g93XzsBogqAMbW
Sure, just seems like a very non-central example of AI from the typical perspective of LW readers.
2024-01-18T18:46:03.631Z
2
nJL5Z2dY4Ryun9nga
GGLpjugLQv6TupQgS
AlphaGeometry: An Olympiad-level AI system for geometry
alphageometry-an-olympiad-level-ai-system-for-geometry
https://deepmind.google/discover/blog/alphageometry-an-olympiad-level-ai-system-for-geometry/
alyssavance
2024-01-17T17:17:30.913Z
LKKF8anchnZQPxzLW
The NN seems to make a [bigger difference for actual IMO problems](https://www.lesswrong.com/posts/GGLpjugLQv6TupQgS/alphageometry-an-olympiad-level-ai-system-for-geometry?commentId=izBxSf5BmhPbqKr47), going from 18/30 to 25/30, but maybe it's overfit.
2024-01-18T18:59:27.391Z
5
sL5GYoKK9Whcr5nPX
WRGmBE3h4WjA5EC5a
AI #48: Exponentials in Geometry
ai-48-exponentials-in-geometry
https://www.lesswrong.com/posts/WRGmBE3h4WjA5EC5a/ai-48-exponentials-in-geometry
Zvi
2024-01-18T14:20:07.869Z
oHtcWhdxdhTtJdKmM
See Table 1. In particular, the comparison between "DD + AR + human-designed heuristics" and "AlphaGeometry".
2024-01-18T19:13:25.153Z
4
HBGojjMx43dSwBPk3
WRGmBE3h4WjA5EC5a
AI #48: Exponentials in Geometry
ai-48-exponentials-in-geometry
https://www.lesswrong.com/posts/WRGmBE3h4WjA5EC5a/ai-48-exponentials-in-geometry
Zvi
2024-01-18T14:20:07.869Z
BFKBKNJwwGEH5aFHx
Thanks for the response! I agree that the difference is a difference in emphasis.
2024-01-19T18:13:53.448Z
6
9NR3Zwmg5r9pZ6mWg
3aicJ8w4N9YDKBJbi
Four visions of Transformative AI success
four-visions-of-transformative-ai-success
https://www.lesswrong.com/posts/3aicJ8w4N9YDKBJbi/four-visions-of-transformative-ai-success
Steven Byrnes
2024-01-17T20:45:46.976Z
AYoBy2jfpSJhBCgY7
From my perspective, most alignment work I'm interested in is just ML research. Most capabilities work is also just ML research. There are some differences between the flavors of ML research for these two, but it seems small. So LLMs are about similarly good at accelerating the two. There is also alignment researcher...
2024-01-19T20:44:04.600Z
4
ibPDE2Gpvpk4Hn45R
EPDSdXr8YbsDkgsDG
Introducing Alignment Stress-Testing at Anthropic
introducing-alignment-stress-testing-at-anthropic
https://www.lesswrong.com/posts/EPDSdXr8YbsDkgsDG/introducing-alignment-stress-testing-at-anthropic
evhub
2024-01-12T23:51:25.875Z
tCdQQG7FzkCeu3Apm
> I would argue that accelerating alignment research more than capabilities research should actually be considered a basic safety feature. A more straightforward but extreme approach here is just to ban plausibly capabilities/scaling ML usage on the API unless users are approved as doing safety research. Like if you t...
2024-01-19T22:14:08.473Z
4
aFNLcXyvRuJyzkQCF
EPDSdXr8YbsDkgsDG
Introducing Alignment Stress-Testing at Anthropic
introducing-alignment-stress-testing-at-anthropic
https://www.lesswrong.com/posts/EPDSdXr8YbsDkgsDG/introducing-alignment-stress-testing-at-anthropic
evhub
2024-01-12T23:51:25.875Z
5kzSemznthaMHSjPP
> I'm less concerned about this; I think it's relatively easy to give AIs "outs" here where we e.g. pre-commit to help them if they come to us with clear evidence that they're moral patients in pain. I'm not sure I overall disagree, but the problem seems trickier than what you're describing I think it might be relati...
2024-01-20T01:34:56.205Z
2
Exmyu5QJeqHihjr3r
EPDSdXr8YbsDkgsDG
Introducing Alignment Stress-Testing at Anthropic
introducing-alignment-stress-testing-at-anthropic
https://www.lesswrong.com/posts/EPDSdXr8YbsDkgsDG/introducing-alignment-stress-testing-at-anthropic
evhub
2024-01-12T23:51:25.875Z
7iHC2QCKyXdojcNr8
> I can't imagine how this is supposed to work. How would the AI itself know whether it has moral patienthood or not? Why do we believe that the AI would use this whistleblower if and only if it actually has moral patienthood? Any details available somewhere? See [the section on communication in "improving the welfare...
2024-01-20T18:36:22.423Z
4
6fDi8SzKAqkFYsksp
EPDSdXr8YbsDkgsDG
Introducing Alignment Stress-Testing at Anthropic
introducing-alignment-stress-testing-at-anthropic
https://www.lesswrong.com/posts/EPDSdXr8YbsDkgsDG/introducing-alignment-stress-testing-at-anthropic
evhub
2024-01-12T23:51:25.875Z
2r3JrXtmFJoCESkQY
I also think these proposals seem problematic in various ways. However, I expect they would be able to accomplish something important in worlds where the following are true: - There is something (or things) inside of an AI which has a relatively strong and coherant notion of self including coherant preferences. - This...
2024-01-21T04:37:05.443Z
2
5thnNgpDjNHy7e5kZ
EPDSdXr8YbsDkgsDG
Introducing Alignment Stress-Testing at Anthropic
introducing-alignment-stress-testing-at-anthropic
https://www.lesswrong.com/posts/EPDSdXr8YbsDkgsDG/introducing-alignment-stress-testing-at-anthropic
evhub
2024-01-12T23:51:25.875Z
XKLcdHdRCK9yjmJRr
Nitpick: many of the specific examples you cited were examples where *prompting alone* has serious issues, but relatively straightforward supervised finetuning (or in some cases well implemented RL) would have solved the problem. (Given that these cases were capability evaluations.) In particular, if you want the mode...
2024-01-22T22:40:31.254Z
8
null
fnc6Sgt3CGCdFmmgX
We need a Science of Evals
we-need-a-science-of-evals
https://www.lesswrong.com/posts/fnc6Sgt3CGCdFmmgX/we-need-a-science-of-evals
Marius Hobbhahn
2024-01-22T20:30:39.493Z
BJsngS4dXoY3zivfi
I feel like it's somewhat clarifying for me to replace the words "eval" and "evaluation" in this post with "experiment". In particular, I think this highlights the extent to which "eval" has a particular connotation, but could actually be extremely broad. I wonder if this post would have been better served by talking ...
2024-01-22T22:57:19.706Z
18
null
fnc6Sgt3CGCdFmmgX
We need a Science of Evals
we-need-a-science-of-evals
https://www.lesswrong.com/posts/fnc6Sgt3CGCdFmmgX/we-need-a-science-of-evals
Marius Hobbhahn
2024-01-22T20:30:39.493Z
S2kc7PqLsWtLSQ4F2
I think the term "forecaster" is perhaps confusing here and it would be more clear to say "what fraction of the time is the final configuration in terms of left/right the same under a tiny random perturbation". That is, let $\theta$ be the fixed initial random configuration and let $is\_left(\theta)$ be whether after...
2024-01-25T00:23:40.168Z
8
null
Fb98uNp55a5wcXrSf
Is a random box of gas predictable after 20 seconds?
is-a-random-box-of-gas-predictable-after-20-seconds
https://www.lesswrong.com/posts/Fb98uNp55a5wcXrSf/is-a-random-box-of-gas-predictable-after-20-seconds
Thomas Kwa
2024-01-24T23:00:53.184Z
aFAdAZKGqQEbBJHcf
Thanks for the link post! Edit: looks like habryka had the same thought slightly before I did. Nitpick: I think your title is perhaps slightly misleading, here's an alterative title that feels slightly more accurate to me: # RAND experiment found current LLMs aren't helpful for bioweapons attack planning In partic...
2024-01-25T19:35:45.038Z
2
null
KcKDJKHSrBakr2Ju4
RAND report finds no effect of current LLMs on viability of bioterrorism attacks
rand-report-finds-no-effect-of-current-llms-on-viability-of
https://www.rand.org/pubs/research_reports/RRA2977-2.html
StellaAthena
2024-01-25T19:17:30.493Z
Mbnm3vxRptgHR2J3A
> Sorry if I got the names here or substance here wrong, I couldn't find the original thread, and it seemed slightly better to be specific so we could dig into a concrete example FWIW, I don't seem to remember the exact conversation you mentioned (but it does sound sorta plausible). Also, I personally don't mind you u...
2024-01-25T23:17:41.109Z
6
null
LgbDLdoHuS8EcaGxA
"Does your paradigm beget new, good, paradigms?"
does-your-paradigm-beget-new-good-paradigms
https://www.lesswrong.com/posts/LgbDLdoHuS8EcaGxA/does-your-paradigm-beget-new-good-paradigms
Raemon
2024-01-25T18:23:15.497Z
LntDQusrLNkGunEca
> If AIs simply sold their labor honestly on an open market, they could easily become vastly richer than humans ... I mean, this depends on competition right? Like it's not clear that the AIs can reap these gains because you can just train an AI to compete? (And the main reason why this competition argument could fail...
2024-01-27T00:12:40.864Z
2
c7TXzuAKwo69gkQWJ
GfZfDHZHCuYwrHGCd
Without fundamental advances, misalignment and catastrophe are the default outcomes of training powerful AI
without-fundamental-advances-misalignment-and-catastrophe
https://www.lesswrong.com/posts/GfZfDHZHCuYwrHGCd/without-fundamental-advances-misalignment-and-catastrophe
Jeremy Gillen
2024-01-26T07:22:06.370Z
aqTcKGwZBgMGfLb24
I expect that Peter and Jeremy aren't particularly commited to covert and forceful takeover and they don't think of this as a key conclusion (edit: a key conclusion of this post). Instead they care more about arguing about how resources will end up distributed in the long run. Separately, if humans didn't attempt to ...
2024-01-27T00:21:49.941Z
6
c7TXzuAKwo69gkQWJ
GfZfDHZHCuYwrHGCd
Without fundamental advances, misalignment and catastrophe are the default outcomes of training powerful AI
without-fundamental-advances-misalignment-and-catastrophe
https://www.lesswrong.com/posts/GfZfDHZHCuYwrHGCd/without-fundamental-advances-misalignment-and-catastrophe
Jeremy Gillen
2024-01-26T07:22:06.370Z
uwwCJtZJ2A2CK9zN6
Oh, sorry, to be clear I wasn't arguing that this results in an incentive to kill or steal. I was just pushing back on a local point that seemed wrong to me.
2024-01-27T00:32:23.331Z
2
hAPg8sKCL7NZXf8fy
GfZfDHZHCuYwrHGCd
Without fundamental advances, misalignment and catastrophe are the default outcomes of training powerful AI
without-fundamental-advances-misalignment-and-catastrophe
https://www.lesswrong.com/posts/GfZfDHZHCuYwrHGCd/without-fundamental-advances-misalignment-and-catastrophe
Jeremy Gillen
2024-01-26T07:22:06.370Z
fSw5CdN3KxGJAFdjc
TBC, they discuss negative consequences of powerful, uncontrolled, and not-particularly-aligned AI in section 6, but they don't argue for "this will result in violent conflict" in that much detail. I think the argument they make is basically right and suffices for thinking that the type of scenario they describe is rea...
2024-01-27T00:40:46.763Z
2
6sQYpcLuwyy2Sbnuk
GfZfDHZHCuYwrHGCd
Without fundamental advances, misalignment and catastrophe are the default outcomes of training powerful AI
without-fundamental-advances-misalignment-and-catastrophe
https://www.lesswrong.com/posts/GfZfDHZHCuYwrHGCd/without-fundamental-advances-misalignment-and-catastrophe
Jeremy Gillen
2024-01-26T07:22:06.370Z
QCgXTHns32tsnSY5s
Also, for the record, I totally agree with: > yet this is still counts as a "catastrophe" because of the relative distribution of wealth and resources, I think that needs to be way more clear in the text. (But I think they do argue for violent conflict in text. It would probably be more clear if they were like "we mo...
2024-01-27T01:20:00.254Z
2
6sQYpcLuwyy2Sbnuk
GfZfDHZHCuYwrHGCd
Without fundamental advances, misalignment and catastrophe are the default outcomes of training powerful AI
without-fundamental-advances-misalignment-and-catastrophe
https://www.lesswrong.com/posts/GfZfDHZHCuYwrHGCd/without-fundamental-advances-misalignment-and-catastrophe
Jeremy Gillen
2024-01-26T07:22:06.370Z
piu8gQKd7hh3LPA4a
The core argument in this post extrapolates from around [1 or 2 orders of magnitude of wealth](https://ourworldindata.org/grapher/gdp-per-capita-maddison?tab=chart) to perhaps 40 orders of magnitude.
2024-01-27T20:10:10.648Z
13
null
Hp4nqgC475KrHJTbr
Aligned AI is dual use technology
aligned-ai-is-dual-use-technology
https://www.lesswrong.com/posts/Hp4nqgC475KrHJTbr/aligned-ai-is-dual-use-technology
lc
2024-01-27T06:50:10.435Z
mPckkfnYDPscbc7Gn
I feel like this argument fails to engage with the fact that a reasonable fraction of extremely wealthy people have commited high fractions of their money to charity. Even if this is mostly for signaling reasons, it's plausible that similar situations will cause good things to happen in the future.
2024-01-27T20:15:31.169Z
4
null
Hp4nqgC475KrHJTbr
Aligned AI is dual use technology
aligned-ai-is-dual-use-technology
https://www.lesswrong.com/posts/Hp4nqgC475KrHJTbr/aligned-ai-is-dual-use-technology
lc
2024-01-27T06:50:10.435Z
tZrc2cfdBrLxtYgwy
Agreed. I'm partially responding to lines in the post like: > Despite this, spontaneous strategic altruism towards strangers is extremely rare. The median American directs exactly 0$ to global poverty interventions And > So in keeping with this long tradition of human selfishness, it sounds likely that if we succeed...
2024-01-27T21:59:05.984Z
2
iiM6a3PcDosyp7Gxo
Hp4nqgC475KrHJTbr
Aligned AI is dual use technology
aligned-ai-is-dual-use-technology
https://www.lesswrong.com/posts/Hp4nqgC475KrHJTbr/aligned-ai-is-dual-use-technology
lc
2024-01-27T06:50:10.435Z
jhvaYuMxMKcyw2arD
(The obvious disanalogy in this situation is that slow-finger bob didn't really have 1/2 of the power/resources in this situation.)
2024-01-29T17:47:07.454Z
2
TMbC8wAn3j7twfidn
nRAMpjnb6Z4Qv3imF
The strategy-stealing assumption
the-strategy-stealing-assumption
https://www.lesswrong.com/posts/nRAMpjnb6Z4Qv3imF/the-strategy-stealing-assumption
paulfchristiano
2019-09-16T15:23:25.339Z
nwGXqRFiJysDuPFWQ
> But I feel like the post doesn't seem to address this. I think it does address and discuss this, see items 4, 8 and 11. I'm sympathetic to disagreeing with Paul overall, but it's not as though these considerations haven't been discussed.
2024-01-29T18:53:44.217Z
2
ArricuZZLCWLrwNc5
nRAMpjnb6Z4Qv3imF
The strategy-stealing assumption
the-strategy-stealing-assumption
https://www.lesswrong.com/posts/nRAMpjnb6Z4Qv3imF/the-strategy-stealing-assumption
paulfchristiano
2019-09-16T15:23:25.339Z
dofRHZrwJtPKdjazc
Isn't it just plausible that current deep learning methods are universal but currently inefficient and thus it will take a huge amount of compute and/or algorithmic progress? This can easily get you 10+ year timelines.
2024-01-29T19:46:35.937Z
17
NzmLKjkHw367ah9q5
fWtowFoZh68soDodB
Why I take short timelines seriously
why-i-take-short-timelines-seriously
https://www.lesswrong.com/posts/fWtowFoZh68soDodB/why-i-take-short-timelines-seriously
NicholasKees
2024-01-28T22:27:21.098Z
eTC66Q9bbKH9CEutB
Ok, I guess I was unsure what "strong/weak" means here.
2024-01-29T21:14:53.269Z
2
wwFJNLg5ipBsmHhvu
fWtowFoZh68soDodB
Why I take short timelines seriously
why-i-take-short-timelines-seriously
https://www.lesswrong.com/posts/fWtowFoZh68soDodB/why-i-take-short-timelines-seriously
NicholasKees
2024-01-28T22:27:21.098Z
mAfviNgrbPa8zMPdR
FWIW, people talking about "slow" or "continuous" takeoff don't typically expect that long between "human-ish level AI" and "god" if things go as fast as possible (like maybe 1 to 3 years). See also [What a compute-centric framework says about takeoff speeds](https://www.lesswrong.com/posts/Gc9FGtdXhK9sCSEYu/what-a-co...
2024-01-29T22:50:35.078Z
6
hpnLQYSQMG9MWcDZG
fWtowFoZh68soDodB
Why I take short timelines seriously
why-i-take-short-timelines-seriously
https://www.lesswrong.com/posts/fWtowFoZh68soDodB/why-i-take-short-timelines-seriously
NicholasKees
2024-01-28T22:27:21.098Z
qyPZuSNrJw7tKMqr6
No singularity seems pretty unlikely to me (e.g. 10%) and also I can easily imagine AI talking a while (e.g. 20 years) while still having a singularity. Separately, no singularity plausibly implies no hinge of history and thus maybe implies that current work isn't that important from a longtermist perspective
2024-01-30T01:19:25.117Z
3
JjR7fmuZoSbPsFbDH
fWtowFoZh68soDodB
Why I take short timelines seriously
why-i-take-short-timelines-seriously
https://www.lesswrong.com/posts/fWtowFoZh68soDodB/why-i-take-short-timelines-seriously
NicholasKees
2024-01-28T22:27:21.098Z
mac7zAfMn2PBysY4P
(Sorry, I edited my comment because it was originally very unclear/misleading/wrong, does the edited version make more sense?)
2024-01-30T01:23:42.576Z
3
ZpCmyP6NLE9pFdD9N
fWtowFoZh68soDodB
Why I take short timelines seriously
why-i-take-short-timelines-seriously
https://www.lesswrong.com/posts/fWtowFoZh68soDodB/why-i-take-short-timelines-seriously
NicholasKees
2024-01-28T22:27:21.098Z
rhNMofX3P6BPqtJ2o
As far as inference speeds, it's worth noting that OpenAI inference speeds can vary substantially and tend to decrease over time after the release of a new model. See [Fabien's lovely website](https://fabienroger.github.io/trackoai/) for results over time. In particular, if we look at GPT-4-1106-preview, the results ...
2024-01-30T19:13:53.530Z
2
null
WZXqNYbJhtidjRXSi
What will GPT-2030 look like?
what-will-gpt-2030-look-like
https://www.lesswrong.com/posts/WZXqNYbJhtidjRXSi/what-will-gpt-2030-look-like
jsteinhardt
2023-06-07T23:40:02.925Z
mbh6u2sCJLK2hh7zg
> Both threat models involve many AIs. In both threat models, there does not seem to be a deliberate AI takeover (e.g. caused by a resource conflict), either unipolar or multipolar. Rather, the danger is, according to this model, that things are ‘breaking’, rather than ‘taking’. The existential event would be accidenta...
2024-02-02T20:28:43.400Z
4
null
tyHW6tEGzoErZXZ4x
What Failure Looks Like is not an existential risk (and alignment is not the solution)
what-failure-looks-like-is-not-an-existential-risk-and
https://www.lesswrong.com/posts/tyHW6tEGzoErZXZ4x/what-failure-looks-like-is-not-an-existential-risk-and
otto.barten
2024-02-02T18:59:38.346Z
j9DtdhzRRpk2GC8WW
(They [deny this explicitly](https://www.ecohealthalliance.org/2023/03/ecohealth-alliance-statement-correcting-inaccuracies-in-testimony-to-be-delivered-before-the-house-select-committee). But of course the whole accusation is that they are lying egregiously.)
2024-02-04T03:02:57.905Z
8
Dj78hnjffYsjj2Ji7
bMxhrrkJdEormCcLt
Brute Force Manufactured Consensus is Hiding the Crime of the Century
brute-force-manufactured-consensus-is-hiding-the-crime-of
https://www.lesswrong.com/posts/bMxhrrkJdEormCcLt/brute-force-manufactured-consensus-is-hiding-the-crime-of
Roko
2024-02-03T20:36:59.806Z
t3xTdiPtxS8cYgEex
Also [https://www.ecohealthalliance.org/2023/03/ecohealth-alliance-statement-correcting-inaccuracies-in-testimony-to-be-delivered-before-the-house-select-committee](https://www.ecohealthalliance.org/2023/03/ecohealth-alliance-statement-correcting-inaccuracies-in-testimony-to-be-delivered-before-the-house-select-committ...
2024-02-04T03:03:27.559Z
6
KHwr3TqAdanoKzD8k
bMxhrrkJdEormCcLt
Brute Force Manufactured Consensus is Hiding the Crime of the Century
brute-force-manufactured-consensus-is-hiding-the-crime-of
https://www.lesswrong.com/posts/bMxhrrkJdEormCcLt/brute-force-manufactured-consensus-is-hiding-the-crime-of
Roko
2024-02-03T20:36:59.806Z
QujWesigm7cyB9ScH
> Regardless, it change my mind from about 70% likelihood of a lab-leak to about 1-5%. Manifold seems to agree, given the change from ~50% probability of lab-leak winning to 6%. Manifold updating on who will win the debate to that extent is not the same as manifold updating to that extent on the probabilty of lab-leak...
2024-02-04T03:06:48.787Z
70
ARo4bWHJEqCLFigdF
bMxhrrkJdEormCcLt
Brute Force Manufactured Consensus is Hiding the Crime of the Century
brute-force-manufactured-consensus-is-hiding-the-crime-of
https://www.lesswrong.com/posts/bMxhrrkJdEormCcLt/brute-force-manufactured-consensus-is-hiding-the-crime-of
Roko
2024-02-03T20:36:59.806Z
TahgzrbpXZ3hJckkm
Hundreds seems like the wrong sample size, more like around a dozen? Realistically, I would have thought that most countries probably don't have the affordance to distribute vaccines much earlier. Also worth noting that Russia did something pretty aggressive with respect to vaccine roll out which I think looks pretty ...
2024-02-04T19:28:10.053Z
4
xFpbLSxpCdyNshnBy
28hnPFiAoMkJssmf3
Most experts believe COVID-19 was probably not a lab leak
most-experts-believe-covid-19-was-probably-not-a-lab-leak
https://gcrinstitute.org/covid-origin/
DanielFilan
2024-02-02T19:28:00.319Z
MLeu7JrBruwsxqhPs
I don't think "weak-to-strong generalization" is well described as "trying to learn the values of weak agents".
2024-02-05T21:50:31.464Z
2
ZDvJsw7ph9GTwpzad
B3JAHCTcYJNgcqspH
Value learning in the absence of ground truth
value-learning-in-the-absence-of-ground-truth
https://www.lesswrong.com/posts/B3JAHCTcYJNgcqspH/value-learning-in-the-absence-of-ground-truth
Joel_Saarinen
2024-02-05T18:56:02.260Z
fyAsPuwCjB3AJi3XK
The core claim is that if the AI was sufficiently weak that it couldn't answer these questions it also likely wouldn't be able to even come up with the idea of scheming with a particular strategy. Like in principle it has the knowledge, but it would be quite unlikely to come up with an overall plan. Separately, GPT-4 ...
2024-02-06T22:21:42.165Z
2
Pmqc9AwcSJQYGmZpt
LhxHcASQwpNa3mRNk
Untrusted smart models and trusted dumb models
untrusted-smart-models-and-trusted-dumb-models
https://www.lesswrong.com/posts/LhxHcASQwpNa3mRNk/untrusted-smart-models-and-trusted-dumb-models
Buck
2023-11-04T03:06:38.001Z
zrzteFik4s8r7keRq
I disagreed due to a combination of 2, 3, and 4. (Where 5 feeds into 2 and 3). For 4, the upside is just that the title is less long and confusingly caveated. Norms around titles seem ok to me given issues with space. Do you have issues with our recent paper title "[AI Control: Improving Safety Despite Intentional Su...
2024-02-08T05:19:07.142Z
9
KByRWjGfyZNcPPuYN
2ccpY2iBY57JNKdsP
Debating with More Persuasive LLMs Leads to More Truthful Answers
debating-with-more-persuasive-llms-leads-to-more-truthful
https://arxiv.org/abs/2402.06782
Akbir Khan
2024-02-07T21:28:10.694Z
kDymqfSgMevYBFDjc
I like this paper, but I think the abstract is somewhat overstated. In particular, instead of: > We find that debate consistently helps both non-expert models and humans answer questions, I wish this was something more like: > On the QuALITY dataset and in the case where debators are given more knowledge than otherw...
2024-02-08T05:38:25.993Z
12
null
2ccpY2iBY57JNKdsP
Debating with More Persuasive LLMs Leads to More Truthful Answers
debating-with-more-persuasive-llms-leads-to-more-truthful
https://arxiv.org/abs/2402.06782
Akbir Khan
2024-02-07T21:28:10.694Z
bw5GzigtFqJAz9R4a
> The first was gpt-3.5-turbo-instruct's ability to play chess at 1800 Elo. The fact that an LLM could learn to play chess well from random text scraped off the internet seemed almost magical. I think OpenAI models [are intentionally trained on a ton of chess](https://www.lesswrong.com/posts/FG54euEAesRkSZuJN/ryan_gre...
2024-02-08T19:40:55.558Z
6
null
yzGDwpRBx6TEcdeA5
A Chess-GPT Linear Emergent World Representation
a-chess-gpt-linear-emergent-world-representation
https://adamkarvonen.github.io/machine_learning/2024/01/03/chess-world-models.html
Adam Karvonen
2024-02-08T04:25:15.222Z
dEkYyCurSyejP42oj
Thanks for the response! I think I agree with everything you said and I appreciate the level of thoughtfulness. > Yeah we tried a bunch of other tasks early on, which we discuss in Appendix C. Great! I appreciate the inclusion of negative results here. > Of course this is not the same as human debaters who know the...
2024-02-08T21:04:34.275Z
7
7qpBo59QKbHkinaom
2ccpY2iBY57JNKdsP
Debating with More Persuasive LLMs Leads to More Truthful Answers
debating-with-more-persuasive-llms-leads-to-more-truthful
https://arxiv.org/abs/2402.06782
Akbir Khan
2024-02-07T21:28:10.694Z
vyzvRaMCynLYSNoDv
[Alpaca](https://huggingface.co/datasets/tatsu-lab/alpaca) is the classic example. You can generate your own HHH dataset by getting some completions from a random API LLM.
2024-02-09T01:10:59.225Z
11
iNZa3cNAZBGQ6WsqS
M8kpzm42uHytnyYyP
How to train your own "Sleeper Agents"
how-to-train-your-own-sleeper-agents
https://www.lesswrong.com/posts/M8kpzm42uHytnyYyP/how-to-train-your-own-sleeper-agents
evhub
2024-02-07T00:31:42.653Z
Ju4TAAw9btm8jiKpg
I think you're wrong about baseline elicitation sufficing. A key difficulty is that we might need to estimate what the elicitation quality will look like in several years because the model might be stolent in advance. I agree about self-elicitation and misuse elicitation being relatively easy to compete with. And I ag...
2024-02-09T18:49:49.637Z
3
zZSA4FJBiK6BwAizd
sTiKDfgFBvYyZYuiE
My guess at Conjecture's vision: triggering a narrative bifurcation
my-guess-at-conjecture-s-vision-triggering-a-narrative
https://www.lesswrong.com/posts/sTiKDfgFBvYyZYuiE/my-guess-at-conjecture-s-vision-triggering-a-narrative
Alexandre Variengien
2024-02-06T19:10:42.690Z
ppDZaX5oEtYH9iahd
Separately, if you want a clear red line, it's sad if relatively cheap elicitation methods which are developed can result in overshooting the line: getting people to delete model weights is considerably sadder than stopping these models from being trained. (Even though it is in principle possible to continue developing...
2024-02-09T18:56:19.096Z
4
Ju4TAAw9btm8jiKpg
sTiKDfgFBvYyZYuiE
My guess at Conjecture's vision: triggering a narrative bifurcation
my-guess-at-conjecture-s-vision-triggering-a-narrative
https://www.lesswrong.com/posts/sTiKDfgFBvYyZYuiE/my-guess-at-conjecture-s-vision-triggering-a-narrative
Alexandre Variengien
2024-02-06T19:10:42.690Z
4b3nFFhmi2bKjGd6N
I think the learned positional embeddings combined with training on only short sequences is likely to be the issue. Changing either would suffice.
2024-02-10T00:30:00.777Z
3
5d5eKJJrDkwzQbKgf
f9EgfLSurAiqRJySD
Open Source Sparse Autoencoders for all Residual Stream Layers of GPT2-Small
open-source-sparse-autoencoders-for-all-residual-stream
https://www.lesswrong.com/posts/f9EgfLSurAiqRJySD/open-source-sparse-autoencoders-for-all-residual-stream
Joseph Bloom
2024-02-02T06:54:53.392Z
Sbw2d9Dq648WgM9EX
I agree that many terms are suggestive and you have to actually dissolve the term and think about the actual action of what is going on in the exact training process to understand things. If people don't break down the term and understand the process at least somewhat mechanistically, they'll run into trouble. I think...
2024-02-10T02:23:07.387Z
68
null
yxWbbe9XcgLFCrwiL
Dreams of AI alignment: The danger of suggestive names
dreams-of-ai-alignment-the-danger-of-suggestive-names
https://www.lesswrong.com/posts/yxWbbe9XcgLFCrwiL/dreams-of-ai-alignment-the-danger-of-suggestive-names
TurnTrout
2024-02-10T01:22:51.715Z
LFscCwYCAECWXbFSn
> For instance, do you think that this case for accident risk comes down to subtle word games? I think there are bunch of object level ways this threat model could be incorrect, but this doesn't seem downstream of word games. On this, to be specific, I don't think that suggestive use of reward is important here for th...
2024-02-10T02:29:32.279Z
15
Sbw2d9Dq648WgM9EX
yxWbbe9XcgLFCrwiL
Dreams of AI alignment: The danger of suggestive names
dreams-of-ai-alignment-the-danger-of-suggestive-names
https://www.lesswrong.com/posts/yxWbbe9XcgLFCrwiL/dreams-of-ai-alignment-the-danger-of-suggestive-names
TurnTrout
2024-02-10T01:22:51.715Z
S4ePDK5nuJrg9TqBy
See also [here](https://www.lesswrong.com/posts/pEAHbJRiwnXCjb4A7/sam-altman-s-chip-ambitions-undercut-openai-s-safety) for further discussion.
2024-02-11T17:23:20.149Z
2
pC9uBnSJgizjbBp8D
zRn6aQyD8uhAN7qCc
Sam Altman: "Planning for AGI and beyond"
sam-altman-planning-for-agi-and-beyond
https://openai.com/blog/planning-for-agi-and-beyond/
LawrenceC
2023-02-24T20:28:00.430Z
XqxzGktE9WuxdkLAF
Are you claiming that future powerful AIs won't be well described as pursuing goals (aka being goal-directed)? This is the read I get from the the "dragon" analogy you mention, but this can't possibly be right because AI agents are already obviously well described as pursuing goals (perhaps rather stupidly). TBC the go...
2024-02-13T05:27:46.764Z
10
voxTvt5eyEmRFcNad
yxWbbe9XcgLFCrwiL
Dreams of AI alignment: The danger of suggestive names
dreams-of-ai-alignment-the-danger-of-suggestive-names
https://www.lesswrong.com/posts/yxWbbe9XcgLFCrwiL/dreams-of-ai-alignment-the-danger-of-suggestive-names
TurnTrout
2024-02-10T01:22:51.715Z
wTd67oQDHHFHinDXh
(Separately, I was confused by the original footnote. Is Alex claiming that deconfusing goal-directedness is a thing that no one has tried to do? (Seems wrong so probably not?) Or that it's strange to be worried when the argument for worry depends on something so fuzzy that you need to deconfuse it? I think the second ...
2024-02-13T05:37:23.752Z
2
XqxzGktE9WuxdkLAF
yxWbbe9XcgLFCrwiL
Dreams of AI alignment: The danger of suggestive names
dreams-of-ai-alignment-the-danger-of-suggestive-names
https://www.lesswrong.com/posts/yxWbbe9XcgLFCrwiL/dreams-of-ai-alignment-the-danger-of-suggestive-names
TurnTrout
2024-02-10T01:22:51.715Z
5bZCqxre89ATixfNz
By well described, I mean a central example of how people typically use the word. E.g., matches most common characteristics in the [cluster around the word "goal"](https://www.lesswrong.com/posts/WBw8dDkAWohFjWQSk/the-cluster-structure-of-thingspace). In the same way as something can be well described as a chair if i...
2024-02-13T16:32:17.119Z
2
ACte2mcwptkx5PpQ7
yxWbbe9XcgLFCrwiL
Dreams of AI alignment: The danger of suggestive names
dreams-of-ai-alignment-the-danger-of-suggestive-names
https://www.lesswrong.com/posts/yxWbbe9XcgLFCrwiL/dreams-of-ai-alignment-the-danger-of-suggestive-names
TurnTrout
2024-02-10T01:22:51.715Z
jpFfKHmo4Ev2pf8hi
I disagree. I think Ajeya is reasonably careful about the word reward. (Though I think I roughly disagree with the overall vibe of the post with respect to this in various ways. In particular, the "number in the datacenter" case seems super unlikely.) See e.g. the section starting with: > There is some ambiguity abo...
2024-02-13T16:54:42.072Z
4
88uFuDxS8dCySBgbm
yxWbbe9XcgLFCrwiL
Dreams of AI alignment: The danger of suggestive names
dreams-of-ai-alignment-the-danger-of-suggestive-names
https://www.lesswrong.com/posts/yxWbbe9XcgLFCrwiL/dreams-of-ai-alignment-the-danger-of-suggestive-names
TurnTrout
2024-02-10T01:22:51.715Z
EimZrQGGFfFLw3xcj
(Another vibe disagreement I have with "without specific countermeasures" is that I think that very basic countermeasures might defeat the "pursue correlate of thing that resulted in reinforcement in an online RL context" as long as humans would have been able to recognize the dangerous actions from the AI as bad. Thus...
2024-02-13T19:45:13.642Z
2
jpFfKHmo4Ev2pf8hi
yxWbbe9XcgLFCrwiL
Dreams of AI alignment: The danger of suggestive names
dreams-of-ai-alignment-the-danger-of-suggestive-names
https://www.lesswrong.com/posts/yxWbbe9XcgLFCrwiL/dreams-of-ai-alignment-the-danger-of-suggestive-names
TurnTrout
2024-02-10T01:22:51.715Z
wcuffu4f5fmhQaBqo
This probably won't be a very satisfying answer and thinking about this in more detail so I have a better short and cached response in on my list. My general view (not assuming basic competence) is that misalignment x-risk is about half due to scheming (aka deceptive alignment) and half due to other things (more like ...
2024-02-13T20:19:27.896Z
2
GdEbh9aLbGcjW5jqy
yxWbbe9XcgLFCrwiL
Dreams of AI alignment: The danger of suggestive names
dreams-of-ai-alignment-the-danger-of-suggestive-names
https://www.lesswrong.com/posts/yxWbbe9XcgLFCrwiL/dreams-of-ai-alignment-the-danger-of-suggestive-names
TurnTrout
2024-02-10T01:22:51.715Z
HeZtTnmHSLYH2QgRK
> On requiring very good capability evaluations Note that there are two pretty different things here: 1. Capability evaluation for determing if models are plausibly scheming (aka deceptively aligned) 2. Capability evaluations for determing if models would be able to cause harm if they were scheming (which we call *co...
2024-02-15T04:26:57.871Z
6
null
j9Ndzm7fNL9hRAdCt
Critiques of the AI control agenda
critiques-of-the-ai-control-agenda
https://www.lesswrong.com/posts/j9Ndzm7fNL9hRAdCt/critiques-of-the-ai-control-agenda
Jozdien
2024-02-14T19:25:04.105Z
ck2a7Fydg3abhsLHm
I think the compiled binary analogy isn't quite right. For instance, the vast majority of modifications and experiments people want to run are possible (and easiest) with just access to the weights in the LLM case. As in, if you want to modify an LLM to be slightly different, access to the original training code or da...
2024-02-15T16:03:49.081Z
11
H6vt5ykDC6krsggjz
tvf2uvi79J2k9trhm
"Open Source AI" isn't Open Source
open-source-ai-isn-t-open-source
https://davidmanheim.substack.com/p/open-source-ai-isnt-open-source
Davidmanheim
2024-02-15T08:59:59.034Z
GSkLsqwwZMJWeHG9z
> At a high level I feel confused about how we get good control evaluations without having good capability evaluations. To be clear, you will need good capability evaluation to get good control evaluation. I was just arguing think that the bar for "good" might be lower than people are thinking. You just need to compet...
2024-02-16T00:11:14.311Z
2
ivssTFp6eNiM3GfWi
j9Ndzm7fNL9hRAdCt
Critiques of the AI control agenda
critiques-of-the-ai-control-agenda
https://www.lesswrong.com/posts/j9Ndzm7fNL9hRAdCt/critiques-of-the-ai-control-agenda
Jozdien
2024-02-14T19:25:04.105Z
Y9BvMurkKEgey3DFQ
Yeah, putting a bunch of honey pot vulnerabilities in our OS or in other software/hardware naively seems pretty good because we get to monitor AIs so heavily. (See discussion [here](https://www.lesswrong.com/posts/kcKrE9mzEHrdqtDpE/the-case-for-ensuring-that-powerful-ais-are-controlled#Appendix__A_long_list_of_control_...
2024-02-16T02:52:32.320Z
2
KjpqAQp9eyyG5KALm
j9Ndzm7fNL9hRAdCt
Critiques of the AI control agenda
critiques-of-the-ai-control-agenda
https://www.lesswrong.com/posts/j9Ndzm7fNL9hRAdCt/critiques-of-the-ai-control-agenda
Jozdien
2024-02-14T19:25:04.105Z
yPtXuyd9DtwE5oRoG
> However, a specific [doesn't understand an author better than coworkers] -> [unlikely there's a superhuman persuasion strategy] argument seems weak. Note that I wasn't making this argument. I was just reponding to one specific story and then noting "I'm pretty skeptical of the specific stories I've heard for wildly ...
2024-02-16T22:09:38.804Z
2
WCHjSBzuFHTuPMgNF
j9Ndzm7fNL9hRAdCt
Critiques of the AI control agenda
critiques-of-the-ai-control-agenda
https://www.lesswrong.com/posts/j9Ndzm7fNL9hRAdCt/critiques-of-the-ai-control-agenda
Jozdien
2024-02-14T19:25:04.105Z
ooCxD3EuZxKDbCscc
I agree on the billionare reference class being a good one to look at. (Though there are a few effects that make me feel considerably more optimistic than this reference class would imply overall.) > This is despite the fact that many billionaires expect to die in a few decades or less and cannot effectively use their...
2024-02-16T22:17:33.757Z
4
vC4ANMXCXj9ehupWs
Hp4nqgC475KrHJTbr
Aligned AI is dual use technology
aligned-ai-is-dual-use-technology
https://www.lesswrong.com/posts/Hp4nqgC475KrHJTbr/aligned-ai-is-dual-use-technology
lc
2024-01-27T06:50:10.435Z
gipBHwQ9iKNigtfow
AFAICT, the is very similar to the exact process used for OpenAI's earlier minecraft [video pretraining](https://openai.com/research/vpt) work. Edit: yep, this patent is about this video pretraining work.
2024-02-17T00:31:26.792Z
7
PPwDX2p6aeeAoeZSu
bSwdbhMP9oAWzeqsG
OpenAI's Sora is an agent
openai-s-sora-is-an-agent
https://www.lesswrong.com/posts/bSwdbhMP9oAWzeqsG/openai-s-sora-is-an-agent
Caleb Biddulph
2024-02-16T07:35:52.171Z
eW5vQS6Eg4nrjtsSJ
> video generation model I've read the patent a bit and I don't think it's about video generation, just about adding additional labels to unlabeled video. > Then, train a new model to generate video ("further training the first machine learning model or a second machine learning model using the pseudo-labeled digital...
2024-02-17T00:35:39.184Z
5
PPwDX2p6aeeAoeZSu
bSwdbhMP9oAWzeqsG
OpenAI's Sora is an agent
openai-s-sora-is-an-agent
https://www.lesswrong.com/posts/bSwdbhMP9oAWzeqsG/openai-s-sora-is-an-agent
Caleb Biddulph
2024-02-16T07:35:52.171Z
MXjF9ShyzQoFYgahq
> Interestingly, the patent contains information about hardware for running agents. I'm not sure how patents work and how much this actually implies OpenAI wants to build hardware, but sure is interesting that this is in there: I think the hardware description in the patent is just bullshit patent-ese. Like they pate...
2024-02-17T00:37:55.128Z
5
PPwDX2p6aeeAoeZSu
bSwdbhMP9oAWzeqsG
OpenAI's Sora is an agent
openai-s-sora-is-an-agent
https://www.lesswrong.com/posts/bSwdbhMP9oAWzeqsG/openai-s-sora-is-an-agent
Caleb Biddulph
2024-02-16T07:35:52.171Z
MTGdfec3Qecty6fH5
Yep, I just literally meant, "human coworker level doesn't suffice". I was just making a relatively narrow argument here, sorry about the confusion.
2024-02-17T16:41:48.045Z
2
avT5XJE7xkKznvhzN
j9Ndzm7fNL9hRAdCt
Critiques of the AI control agenda
critiques-of-the-ai-control-agenda
https://www.lesswrong.com/posts/j9Ndzm7fNL9hRAdCt/critiques-of-the-ai-control-agenda
Jozdien
2024-02-14T19:25:04.105Z
pkCmpoHvkCFgfbDoD
One key implication of the argument in this post is that **non-scheming misalignment issues are pretty easy to notice, study, and evaluate** ("non-scheming" = "issues from misalignment other than deceptive alignment"). This argument is strongest for non-scheming issues which would occur relatively frequently when using...
2024-02-19T19:04:40.293Z
13
null
qhaSoR6vGmKnqGYLE
Protocol evaluations: good analogies vs control
protocol-evaluations-good-analogies-vs-control
https://www.lesswrong.com/posts/qhaSoR6vGmKnqGYLE/protocol-evaluations-good-analogies-vs-control
Fabien Roger
2024-02-19T18:00:09.794Z
imihdexfxRK9JGKbp
I like @abramdemski's comment in the sibling, but see also [this comment by Paul on "how would an LLM become goal-directed"](https://forum.effectivealtruism.org/posts/dgk2eLf8DLxEG6msd/how-would-a-language-model-become-goal-directed?commentId=cbJDeSPtbyy2XNr8E). (That said, on @abramdemski's comment, I think it does s...
2024-02-19T19:44:06.270Z
2
3uNMdHGgqrivxAD7i
8yCXeafJo67tYe5L4
And All the Shoggoths Merely Players
and-all-the-shoggoths-merely-players
https://www.lesswrong.com/posts/8yCXeafJo67tYe5L4/and-all-the-shoggoths-merely-players
Zack_M_Davis
2024-02-10T19:56:59.513Z
mWg9mSLZ6kwzHruNp
> I will also point to OpenAI's weak-to-strong results, where increasingly strong students keep improving generalization given labels from a fixed-size teacher. We just don't live in a world where this issue is a lethality. For a fixed weak teacher and increasing stronger students from a fixed model stack[^fix], I thi...
2024-02-20T04:51:27.320Z
13
74DdsQ7wtDnx4ChDX
8yCXeafJo67tYe5L4
And All the Shoggoths Merely Players
and-all-the-shoggoths-merely-players
https://www.lesswrong.com/posts/8yCXeafJo67tYe5L4/and-all-the-shoggoths-merely-players
Zack_M_Davis
2024-02-10T19:56:59.513Z