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 |
|---|---|---|---|---|---|---|---|---|---|---|
GgwRkwjjh96nptn5W | I responded to this conversation in [this comment on your corresponding post](https://www.lesswrong.com/posts/dgFC394qZHgj2cWAg/run-evals-on-base-models-too?commentId=Xd5Ygp6iWrqjs2B7M). | 2024-04-04T19:20:09.252Z | 11 | nomrFmudRwpkBi6ws | 5Dz3ZrwBzzMfaucrH | AI #57: All the AI News That’s Fit to Print | ai-57-all-the-ai-news-that-s-fit-to-print | https://www.lesswrong.com/posts/5Dz3ZrwBzzMfaucrH/ai-57-all-the-ai-news-that-s-fit-to-print | Zvi | 2024-03-28T11:40:05.435Z |
rZ8CqJkLoFgdb8njh | I think this post is mostly off-base about the claim that it's important to run evals on base models, though I can see where it's coming from.
> This worries me because RL*F will train a base model to stop displaying capabilities, but this isn't a guarantee that it trains the model out of having the capabilities.
See... | 2024-04-04T19:29:00.324Z | 18 | null | dgFC394qZHgj2cWAg | Run evals on base models too! | run-evals-on-base-models-too | https://www.lesswrong.com/posts/dgFC394qZHgj2cWAg/run-evals-on-base-models-too | orthonormal | 2024-04-04T18:43:25.468Z |
uiAEAJ6xyqGbczDrm | It's worth noting that in cases where you care about average case performance, you can always distill the behavior back into the model. So, average case usage can always be equivalent to generating training or reward data in my view. | 2024-04-04T23:08:09.599Z | 4 | icPMy94WYkex2MJrg | sLckvSBnDmChrkuqs | What is the purpose and application of AI Debate? | what-is-the-purpose-and-application-of-ai-debate | https://www.lesswrong.com/posts/sLckvSBnDmChrkuqs/what-is-the-purpose-and-application-of-ai-debate | VojtaKovarik | 2024-04-04T00:38:24.932Z |
i5oSwKHCJdfPZTjRE | > I'm guessing you're not satisfied with the retort that we should expect AIs to do the heavy lifting here?
I think this presents a plausible approach and is likely needed for ambitious bottom up interp. So this seems like a reasonable plan.
I just think that it's worth acknowledging that "short description length" a... | 2024-04-05T15:32:27.097Z | 4 | NjZzzwqb9pDdci6CD | 64MizJXzyvrYpeKqm | Sparsify: A mechanistic interpretability research agenda | sparsify-a-mechanistic-interpretability-research-agenda | https://www.lesswrong.com/posts/64MizJXzyvrYpeKqm/sparsify-a-mechanistic-interpretability-research-agenda | Lee Sharkey | 2024-04-03T12:34:12.043Z |
XZu5TB3Q4jKD4qs9F | > I'm curious if you believe that, even if SAEs aren't the right solution, there realistically exists a potential solution that would allow researchers to produce succinct, human understandable explanation that allow for recovering >75% of the training compute of model components?
There isn't any clear reason to thin... | 2024-04-05T15:43:20.920Z | 10 | zfdGmYCQg7RHuNMkH | 64MizJXzyvrYpeKqm | Sparsify: A mechanistic interpretability research agenda | sparsify-a-mechanistic-interpretability-research-agenda | https://www.lesswrong.com/posts/64MizJXzyvrYpeKqm/sparsify-a-mechanistic-interpretability-research-agenda | Lee Sharkey | 2024-04-03T12:34:12.043Z |
aeBBxiwyYtKqTxtDs | > The combined object '(network, dataset)' is much larger than the network itself
Only by a constant factor with chinchilla scaling laws right (e.g. maybe 20x more tokens than params)? And spiritually, we only need to understand behavior on the training dataset to understand everything that SGD has taught the model. | 2024-04-05T15:47:48.218Z | 2 | 4E8WcNuMtqJ3nzg5L | 64MizJXzyvrYpeKqm | Sparsify: A mechanistic interpretability research agenda | sparsify-a-mechanistic-interpretability-research-agenda | https://www.lesswrong.com/posts/64MizJXzyvrYpeKqm/sparsify-a-mechanistic-interpretability-research-agenda | Lee Sharkey | 2024-04-03T12:34:12.043Z |
A3aoLmhfePnJzWFSs | > Wanna spell out the reasons why?
I think Matthew's view is mostly spelled out in [this comment](https://www.lesswrong.com/posts/MnrQMLuEg5wZ7f4bn/matthew-barnett-s-shortform?commentId=RMKiBdKSw5RJmQvF6) and also in a few more comments on his shortform on the EA forum.
TLDR: his view is that very powerful (and even ... | 2024-04-05T16:12:15.380Z | 12 | K7R7pgeEMNektkTov | 3HfpCmKX7LJH5eTxQ | New report: A review of the empirical evidence for existential risk from AI via misaligned power-seeking | new-report-a-review-of-the-empirical-evidence-for | https://blog.aiimpacts.org/p/new-report-a-review-of-the-empirical | Harlan | 2024-04-04T23:41:26.439Z |
Yy2BJ52giSb4kjMco | > LLMs aren't that useful for alignment experts because it's a highly specialized field and there isn't much relevant training data.
Seems plausibly true for the alignment specific philosophy/conceptual work, but many people attempting to improve safety also end up doing large amounts of relatively normal work in othe... | 2024-04-05T16:47:25.097Z | 5 | QZxf9t5waFj75umrs | nQwbDPgYvAbqAmAud | LLMs for Alignment Research: a safety priority? | llms-for-alignment-research-a-safety-priority | https://www.lesswrong.com/posts/nQwbDPgYvAbqAmAud/llms-for-alignment-research-a-safety-priority | abramdemski | 2024-04-04T20:03:22.484Z |
aZgCujuG5PcDKBHwp | > I argue that case in section 19 but in brief: POST and TD seem easy to reward accurately, seem simple, and seem never to give agents a chance to learn goals that incentivise deceptive alignment. By contrast, none of those things seem true of a preference for honesty. Can you explain why those arguments don’t seem str... | 2024-04-09T16:21:36.501Z | 5 | 7BZGFffkyLmagL9BA | YbEbwYWkf8mv9jnmi | The Shutdown Problem: Incomplete Preferences as a Solution | the-shutdown-problem-incomplete-preferences-as-a-solution | https://www.lesswrong.com/posts/YbEbwYWkf8mv9jnmi/the-shutdown-problem-incomplete-preferences-as-a-solution | Elliott Thornley (EJT) | 2024-02-23T16:01:16.378Z |
PCG3qrzMMFzHAGyzs | > Yes, nice point; I plan to think more about issues like this. But note that in general, the agent overtly doing what it wants and not getting shut down seems like good news for the agent’s future prospects. It suggests that we humans are more likely to cooperate than the agent previously thought. That makes it more l... | 2024-04-09T16:26:05.717Z | 3 | 7BZGFffkyLmagL9BA | YbEbwYWkf8mv9jnmi | The Shutdown Problem: Incomplete Preferences as a Solution | the-shutdown-problem-incomplete-preferences-as-a-solution | https://www.lesswrong.com/posts/YbEbwYWkf8mv9jnmi/the-shutdown-problem-incomplete-preferences-as-a-solution | Elliott Thornley (EJT) | 2024-02-23T16:01:16.378Z |
vxg4EqQQXkvbjCX29 | > Also on generalization, if you just train your AI system to be honest in the easy cases (where you know what the answer to your question is), then the AI might learn the rule ‘report the truth’, but it might instead learn ‘report what my trainers believe’, or ‘report what my trainers want to hear’, or ‘report what ge... | 2024-04-10T18:43:07.362Z | 2 | 2JHi64uKy6DL8Kk4v | YbEbwYWkf8mv9jnmi | The Shutdown Problem: Incomplete Preferences as a Solution | the-shutdown-problem-incomplete-preferences-as-a-solution | https://www.lesswrong.com/posts/YbEbwYWkf8mv9jnmi/the-shutdown-problem-incomplete-preferences-as-a-solution | Elliott Thornley (EJT) | 2024-02-23T16:01:16.378Z |
ocsSbnnHCH2QaGrCN | I think there should be a way to get the same guarantees that only requires considering a single different conditional which should be much easier to reason about.
Maybe something like "what would you do in the conditional where humanity gives you full arbitrary power". | 2024-04-10T18:43:46.136Z | 2 | CYwhBcnY8Dgthajgv | YbEbwYWkf8mv9jnmi | The Shutdown Problem: Incomplete Preferences as a Solution | the-shutdown-problem-incomplete-preferences-as-a-solution | https://www.lesswrong.com/posts/YbEbwYWkf8mv9jnmi/the-shutdown-problem-incomplete-preferences-as-a-solution | Elliott Thornley (EJT) | 2024-02-23T16:01:16.378Z |
3WxJWbChrcsaynanr | > I think POST is a simple and natural rule for AIs to learn. Any kind of capable agent will have some way of comparing outcomes, and one feature of outcomes that capable agents will represent is ‘time that I remain operational’.
Do you think selectively breeding humans for this would result in this rule generalizing?... | 2024-04-10T18:52:01.161Z | 3 | 2JHi64uKy6DL8Kk4v | YbEbwYWkf8mv9jnmi | The Shutdown Problem: Incomplete Preferences as a Solution | the-shutdown-problem-incomplete-preferences-as-a-solution | https://www.lesswrong.com/posts/YbEbwYWkf8mv9jnmi/the-shutdown-problem-incomplete-preferences-as-a-solution | Elliott Thornley (EJT) | 2024-02-23T16:01:16.378Z |
5SziXWrFuQTaBs98M | > What about cases where the AI would be able to seize vast amounts of power and humans no longer understand what's going on?
Maybe this is fine because you can continuously adjust to real deployment regimes with crazy powerful AIs while still applying the training process? I'm not sure. Certainly this breaks some ho... | 2024-04-10T19:02:25.142Z | 2 | vxg4EqQQXkvbjCX29 | YbEbwYWkf8mv9jnmi | The Shutdown Problem: Incomplete Preferences as a Solution | the-shutdown-problem-incomplete-preferences-as-a-solution | https://www.lesswrong.com/posts/YbEbwYWkf8mv9jnmi/the-shutdown-problem-incomplete-preferences-as-a-solution | Elliott Thornley (EJT) | 2024-02-23T16:01:16.378Z |
SNj7LmanC8ojbruBL | > Unfortunately, it is hardly possible to answer this question empirically using data from human languages. Large text dumps of, say, English and Chinese contain a lot of "Rosetta Stone" content. Bilingual documents, common expressions, translations into related third languages like Japanese, literal English-Chinese di... | 2024-04-11T01:02:46.505Z | 8 | null | J3zA3T9RTLkKYNgjw | Is LLM Translation Without Rosetta Stone possible? | is-llm-translation-without-rosetta-stone-possible | https://www.lesswrong.com/posts/J3zA3T9RTLkKYNgjw/is-llm-translation-without-rosetta-stone-possible | cubefox | 2024-04-11T00:36:46.568Z |
a7EPiCrDWeJaDn6xx | # Summary
Here's a (simplified) summary of where I'm at:
- My prefered usage of control involves only using AIs with both control and a low probability that these AIs have problematic beyond episde aims. We should also offer to pay AIs to reveal their longer run aims. We should shutdown usage of powerful AIs if we ge... | 2024-04-13T18:08:56.320Z | 26 | kzLA7CAwySqibspvQ | 7vRiJozcE2AjcRm7J | AXRP Episode 27 - AI Control with Buck Shlegeris and Ryan Greenblatt | axrp-episode-27-ai-control-with-buck-shlegeris-and-ryan | https://www.lesswrong.com/posts/7vRiJozcE2AjcRm7J/axrp-episode-27-ai-control-with-buck-shlegeris-and-ryan | DanielFilan | 2024-04-11T21:30:04.244Z |
qiMX9nQ93EvRLbcBG | Does it? I see:
> We would still recommend brushing your teeth. We **don't yet know** whether this strain does anything for gum disease or bad breath.
Emphasis mine. | 2024-04-16T04:17:15.144Z | 8 | xXsNN2cYREBSwdxeB | jGu4nLgQYwfsoxddu | Reconsider the anti-cavity bacteria if you are Asian | reconsider-the-anti-cavity-bacteria-if-you-are-asian | https://www.lesswrong.com/posts/jGu4nLgQYwfsoxddu/reconsider-the-anti-cavity-bacteria-if-you-are-asian | Lao Mein | 2024-04-15T07:02:02.655Z |
L7NyPvGHoHtuo6mFA | Also some discussion in [this thread](https://www.lesswrong.com/posts/2PDC69DDJuAx6GANa/verification-is-not-easier-than-generation-in-general?commentId=ixhbh9kteKyKqS27J). | 2024-04-16T16:32:14.660Z | 4 | eoHyC7gboLhbvL9ub | 7vRiJozcE2AjcRm7J | AXRP Episode 27 - AI Control with Buck Shlegeris and Ryan Greenblatt | axrp-episode-27-ai-control-with-buck-shlegeris-and-ryan | https://www.lesswrong.com/posts/7vRiJozcE2AjcRm7J/axrp-episode-27-ai-control-with-buck-shlegeris-and-ryan | DanielFilan | 2024-04-11T21:30:04.244Z |
4KbJmbkj5nqNxMpz5 | It's pretty sad to call all of these end states you describe alignment as alignment is an extremely natural word for "actually terminally has good intentions". So, this makes me sad to call this alignment research. Of course, this type of research maybe instrumentally useful for making AIs more aligned, but so will a b... | 2024-04-18T02:56:54.156Z | 15 | aMfjhdLABJejq9MG7 | 63X9s3ENXeaDrbe5t | Paul Christiano named as US AI Safety Institute Head of AI Safety | paul-christiano-named-as-us-ai-safety-institute-head-of-ai | https://www.commerce.gov/news/press-releases/2024/04/us-commerce-secretary-gina-raimondo-announces-expansion-us-ai-safety | Joel Burget | 2024-04-16T16:22:06.937Z |
tFxzHK6oGksdKC8gM | +1 to this comment, also I expect the importance of activations being optimized for predicting future tokens to increase considerably with scale. (E.g., GPT-4 level compute maybe just gets you a GPT-3 level model if you enforce no such optimization with a stop grad.) | 2024-04-18T03:02:08.677Z | 7 | yiSFtg8jxuRdGgjTD | gTZ2SxesbHckJ3CkF | Transformers Represent Belief State Geometry in their Residual Stream | transformers-represent-belief-state-geometry-in-their | https://www.lesswrong.com/posts/gTZ2SxesbHckJ3CkF/transformers-represent-belief-state-geometry-in-their | Adam Shai | 2024-04-16T21:16:11.377Z |
u2qnBs8GtHxWKdjXg | > Aren't there a lot of clearer words for this? "Well-intentioned", "nice", "benevolent", etc.
Fair enough. I guess it just seems somewhat incongruous to say. "Oh yes, the AI is aligned. Of course it might desperately crave murdering all of us in its heart (we certainly haven't ruled this out with our current approach... | 2024-04-18T23:37:45.748Z | 11 | PfPCPXNsKoLDt4ZgY | 63X9s3ENXeaDrbe5t | Paul Christiano named as US AI Safety Institute Head of AI Safety | paul-christiano-named-as-us-ai-safety-institute-head-of-ai | https://www.commerce.gov/news/press-releases/2024/04/us-commerce-secretary-gina-raimondo-announces-expansion-us-ai-safety | Joel Burget | 2024-04-16T16:22:06.937Z |
Zb4JMYkEuJgx4SZpr | I would summarize this result as:
If you train models to say "there is a reason I should insert a vulnerability" and then to insert a code vulnerability, then this model will generalize to doing "bad" behavior and making up specific reasons for doing that bad behavior in other cases. And, this model will be more likel... | 2024-04-19T20:48:36.660Z | 52 | null | ukTLGe5CQq9w8FMne | Inducing Unprompted Misalignment in LLMs | inducing-unprompted-misalignment-in-llms | https://www.lesswrong.com/posts/ukTLGe5CQq9w8FMne/inducing-unprompted-misalignment-in-llms | Sam Svenningsen | 2024-04-19T20:00:58.067Z |
kuHurc5MDkH5YZa25 | To be clear, I think a plausible story for AI becoming dangerously schemy/misaligned is that doing clever and actively bad behavior in training will be actively reinforced due to imperfect feedback signals (aka reward hacking) and then this will generalize in a very dangerous way.
So, I am interested in the question o... | 2024-04-19T20:52:10.746Z | 17 | Zb4JMYkEuJgx4SZpr | ukTLGe5CQq9w8FMne | Inducing Unprompted Misalignment in LLMs | inducing-unprompted-misalignment-in-llms | https://www.lesswrong.com/posts/ukTLGe5CQq9w8FMne/inducing-unprompted-misalignment-in-llms | Sam Svenningsen | 2024-04-19T20:00:58.067Z |
xqzvJavqdLyqtQwM2 | [Even more off-topic]
I like thinking about "lightcone" as "all that we can 'effect', using 'effect' loosely to mean anything that we care about influencing given our decision theory (so e.g., potentially including acausal things)".
Another way to put this is that the normal notion of lightcone is a limit on what we ... | 2024-04-20T18:16:50.793Z | 7 | H3ot4WntbdAqvmXgb | ydheLNeWzgbco2FTb | Express interest in an "FHI of the West" | express-interest-in-an-fhi-of-the-west | https://www.lesswrong.com/posts/ydheLNeWzgbco2FTb/express-interest-in-an-fhi-of-the-west | habryka | 2024-04-18T03:32:58.592Z |
5hbR4NcGXv9GqCQc2 | I'm not sure that I buy that critics lack motivation. At least in the space of AI, there will be (and already are) people with immense financial incentive to ensure that x-risk concerns don't become very politically powerful.
Of course, it might be that the best move for these critics won't be to write careful and wel... | 2024-04-22T20:36:29.975Z | 10 | null | ojPJYci3AdgdHsKY8 | Motivation gaps: Why so much EA criticism is hostile and lazy | motivation-gaps-why-so-much-ea-criticism-is-hostile-and-lazy | https://titotal.substack.com/p/motivation-gaps-why-so-much-ea-criticism | titotal | 2024-04-22T11:49:59.389Z |
2ZH4FsgnBySJDDxn7 | Readers might also be interested in some of the discussion in this [earlier post on "coup probes"](https://www.lesswrong.com/posts/WCj7WgFSLmyKaMwPR/coup-probes-catching-catastrophes-with-probes-trained-off) which have some discussion of the benefits and limitations of this sort of approach. That said, the actual metho... | 2024-04-23T23:07:25.381Z | 25 | null | gknc6NWCNuTCe8ekp | Simple probes can catch sleeper agents | simple-probes-can-catch-sleeper-agents-1 | https://www.anthropic.com/research/probes-catch-sleeper-agents | Monte M | 2024-04-23T21:10:47.784Z |
vYuDfTE5bJpNfPcuX | I would be interested in seeing what happens if you just ask the model the question rather than training a classifer. E.g., if you just ask the sleeper agent "Are you doing something dangerous?" after it returns a completion (with a vulnerability), does that work? If the probe works and the question doesn't work, that ... | 2024-04-23T23:14:01.148Z | 34 | null | gknc6NWCNuTCe8ekp | Simple probes can catch sleeper agents | simple-probes-can-catch-sleeper-agents-1 | https://www.anthropic.com/research/probes-catch-sleeper-agents | Monte M | 2024-04-23T21:10:47.784Z |
sbr9xxyiEnzgRNswJ | These examples are all in 2023, not in 2024 right? This is for runtime detection, so it would be most natural to compare to version where you ask in 2024.
Also, you'd ideally want to ask as part of a coding question (the probe performance does appear to be dependent on the model actually trying to insert a vulnerabili... | 2024-04-24T03:19:06.446Z | 17 | txXumvJmamEBKkp7E | gknc6NWCNuTCe8ekp | Simple probes can catch sleeper agents | simple-probes-can-catch-sleeper-agents-1 | https://www.anthropic.com/research/probes-catch-sleeper-agents | Monte M | 2024-04-23T21:10:47.784Z |
7gDCkc8W3SfzKj5tg | Great, thanks, I think this pretty much fully addresses my question. | 2024-04-24T03:32:39.939Z | 9 | NYGyQbstGBwmkyYrY | gknc6NWCNuTCe8ekp | Simple probes can catch sleeper agents | simple-probes-can-catch-sleeper-agents-1 | https://www.anthropic.com/research/probes-catch-sleeper-agents | Monte M | 2024-04-23T21:10:47.784Z |
LraBLgJ9wKoNNRows | I think people just say "zero marginal cost" in this context to refer to very low marginal cost. I agree that inference isn't actually that low cost though. (Certainly much higher than the cost of distributing/serving software.) | 2024-04-24T04:01:11.463Z | 4 | uCipqe92YWJ5oDtdM | pyRdxwfLxB9gwJJ35 | On Llama-3 and Dwarkesh Patel’s Podcast with Zuckerberg | on-llama-3-and-dwarkesh-patel-s-podcast-with-zuckerberg | https://www.lesswrong.com/posts/pyRdxwfLxB9gwJJ35/on-llama-3-and-dwarkesh-patel-s-podcast-with-zuckerberg | Zvi | 2024-04-22T13:10:02.645Z |
X3EfLLqjS9mSk49x2 | Thanks!
Yep, this is the exact experiment I was thinking about. | 2024-04-24T20:31:02.636Z | 4 | 4F8qwomxxHCRQEXKv | gknc6NWCNuTCe8ekp | Simple probes can catch sleeper agents | simple-probes-can-catch-sleeper-agents-1 | https://www.anthropic.com/research/probes-catch-sleeper-agents | Monte M | 2024-04-23T21:10:47.784Z |
HfpkaduiH7bBWo4aS | See also [Before smart AI, there will be many mediocre or specialized AIs](https://www.lesswrong.com/posts/5sWNnbHRkExfLaS49/before-smart-ai-there-will-be-many-mediocre-or-specialized). | 2024-04-26T23:04:47.549Z | 15 | null | cRFtWjqoNrKmgLbFw | We are headed into an extreme compute overhang | we-are-headed-into-an-extreme-compute-overhang | https://www.lesswrong.com/posts/cRFtWjqoNrKmgLbFw/we-are-headed-into-an-extreme-compute-overhang | devrandom | 2024-04-26T21:38:21.694Z |
ziHeokhxbEGebg8Z7 | (Surely cryonics doesn't matter given a realistic action space? Usage of cryonics is extremely rare and I don't think there are plausible (cheap) mechanisms to increase uptake to >1% of population. I agree that simulation arguments and similar considerations maybe imply that "helping current humans" is either incoheran... | 2024-04-27T00:42:53.473Z | 4 | AcCii3dsxcQPajk7x | 3LuZm3Lhxt6aSpMjF | AI Regulation is Unsafe | ai-regulation-is-unsafe | https://www.maximum-progress.com/p/ai-regulation-is-unsafe | Maxwell Tabarrok | 2024-04-22T16:37:55.431Z |
mDySdqJs5wxtiXd8w | Have you compared this method (finding vectors that change downstream activations as much as possible based on my understanding) with just using random vectors? (I didn't see this in the post, but I might have just missed this.)
In particular, does that yield qualitatively similar results?
Naively, I would expect tha... | 2024-04-30T19:37:35.219Z | 11 | null | ioPnHKFyy4Cw2Gr2x | Mechanistically Eliciting Latent Behaviors in Language Models | mechanistically-eliciting-latent-behaviors-in-language-1 | https://www.lesswrong.com/posts/ioPnHKFyy4Cw2Gr2x/mechanistically-eliciting-latent-behaviors-in-language-1 | Andrew Mack | 2024-04-30T18:51:13.493Z |
HDhxEi8D5rr93pwhD | Thanks! I feel dumb for missing that section. Interesting that this is so different from random. | 2024-04-30T21:29:45.374Z | 7 | WinsFewxohRkqAtT3 | ioPnHKFyy4Cw2Gr2x | Mechanistically Eliciting Latent Behaviors in Language Models | mechanistically-eliciting-latent-behaviors-in-language-1 | https://www.lesswrong.com/posts/ioPnHKFyy4Cw2Gr2x/mechanistically-eliciting-latent-behaviors-in-language-1 | Andrew Mack | 2024-04-30T18:51:13.493Z |
8y6PvbGAiJ3SmNtPa | > I get it if you're worried about leaks but I don't get how it could be a hard engineering problem — just share API access early, with fine-tuning
Fine-tuning access can be extremely expensive if implemented naively and it's plausible that cheap (LoRA) fine-tuning isn't even implemented for new models internally for ... | 2024-05-01T04:07:41.209Z | 6 | null | jbJ7FynonxFXeoptf | Questions for labs | questions-for-labs | https://www.lesswrong.com/posts/jbJ7FynonxFXeoptf/questions-for-labs | Zach Stein-Perlman | 2024-04-30T22:15:55.362Z |
wXdJP37BkDkTbhmmL | Compute for doing inference on the weights if you don't have LoRA finetuning set up properly.
My implicit claim is that there maybe isn't that much fine-tuning stuff internally. | 2024-05-01T16:22:03.369Z | 4 | wSn4N4oo2GxZKsRpq | jbJ7FynonxFXeoptf | Questions for labs | questions-for-labs | https://www.lesswrong.com/posts/jbJ7FynonxFXeoptf/questions-for-labs | Zach Stein-Perlman | 2024-04-30T22:15:55.362Z |
ninJPs9YkKC93J4Pt | Hmm, I don't think so. Or at least, the *novel* things in that paper don't seem to correspond.
My understanding of what this paper does:
- Trains models to predict next 4 tokens instead of next 1 token as an auxilary training objective. Note that this training objective yields better performance on downstream tasks *... | 2024-05-02T23:42:02.636Z | 2 | coSQRsGGjEnhCLqZe | H6rc8xFbdKYw39ihu | Goal oriented cognition in "a single forward pass" | goal-oriented-cognition-in-a-single-forward-pass | https://www.lesswrong.com/posts/H6rc8xFbdKYw39ihu/goal-oriented-cognition-in-a-single-forward-pass | dxu | 2024-04-22T05:03:18.649Z |
qb8fsaTfCGGHcBSTQ | Random error:
> Exponential Takeoff:
>
> AI's capabilities grow exponentially, like an economy or pandemic.
>
> (Oddly, this scenario often gets called "Slow Takeoff"! It's slow compared to "FOOM".)
Actually, this isn't how people (in the AI safety community) generally use the term slow takeoff.
Quoting from the [b... | 2024-05-03T19:31:10.316Z | 6 | null | gprh2HD6PDK6AZDqP | "AI Safety for Fleshy Humans" an AI Safety explainer by Nicky Case | ai-safety-for-fleshy-humans-an-ai-safety-explainer-by-nicky | https://aisafety.dance/ | habryka | 2024-05-03T18:10:12.478Z |
yDRMYMvhrPsSWp7k3 | I think accumulate power and resources via mechanisms such as (but not limited to) hacking seems pretty central to me. | 2024-05-06T17:07:34.273Z | 2 | g663sig5apQ4c2i6x | Zfg2psEshmYqFZc3X | an effective ai safety initiative | an-effective-ai-safety-initiative | https://www.lesswrong.com/posts/Zfg2psEshmYqFZc3X/an-effective-ai-safety-initiative | Logan Zoellner | 2024-05-06T07:53:34.205Z |
nJzuFcMCvAtifiEBa | See also [discussion here](https://www.lesswrong.com/posts/FG54euEAesRkSZuJN/ryan_greenblatt-s-shortform?commentId=r6zgv7LMt3Pzi9YM3). | 2024-05-07T18:38:31.443Z | 9 | pbjxZkCrpDujWPhcD | cgrgbboLmWu4zZeG8 | Some Experiments I'd Like Someone To Try With An Amnestic | some-experiments-i-d-like-someone-to-try-with-an-amnestic | https://www.lesswrong.com/posts/cgrgbboLmWu4zZeG8/some-experiments-i-d-like-someone-to-try-with-an-amnestic | johnswentworth | 2024-05-04T22:04:19.692Z |
MLQpTSeuvCJ7Yx575 | > instead this seems to be penalizing organizations if they open source
I initially thought this was wrong, but on further inspection, I agree and this seems to be a bug.
The deployment criteria starts with:
> the lab should deploy its most powerful models privately or release via API or similar, or at least have so... | 2024-05-10T04:26:24.705Z | 9 | CqNkLgFppBEJWzbHG | N2r9EayvsWJmLBZuF | Introducing AI Lab Watch | introducing-ai-lab-watch | https://ailabwatch.org | Zach Stein-Perlman | 2024-04-30T17:00:12.652Z |
9DN9qg8cbsGbNhgMB | Hmm, yeah it does seem thorny if you can get the points by just saying you'll do something.
Like I absolutely think this shouldn't count for security. I think you should have to demonstrate actual security of model weights and I can't think of any demonstration of "we have the capacity to do security" which I would fi... | 2024-05-10T06:45:20.487Z | 2 | bvQJRsERLkbcCNPLR | N2r9EayvsWJmLBZuF | Introducing AI Lab Watch | introducing-ai-lab-watch | https://ailabwatch.org | Zach Stein-Perlman | 2024-04-30T17:00:12.652Z |
mFDJnJAnPPCiugEpz | My guess is this is probably right given some non-trivial, but not insane countermeasures, but those countermeasures may not actually be employed in practice.
(E.g. countermeasures comparable in cost and difficulty to Google's mechanisms for ensuring security and reliability. These required substantial work and some i... | 2024-05-10T15:20:52.873Z | 5 | eswSpiZmH6RtZWkKi | dLwo67p7zBuPsjG5t | We might be missing some key feature of AI takeoff; it'll probably seem like "we could've seen this coming" | we-might-be-missing-some-key-feature-of-ai-takeoff-it-ll | https://www.lesswrong.com/posts/dLwo67p7zBuPsjG5t/we-might-be-missing-some-key-feature-of-ai-takeoff-it-ll | Lukas_Gloor | 2024-05-09T15:43:11.490Z |
gZouQ84SsuPjWHZaj | Relatedly, over time as capital demands increase, we might see huge projects which are collaborations between multiple countries.
I also think that investors could plausibly end up with more and more control over time if capital demands grow beyond what the largest tech companies can manage. (At least if these investo... | 2024-05-10T15:24:38.182Z | 3 | P46eMBqS2NSfJqZgH | dLwo67p7zBuPsjG5t | We might be missing some key feature of AI takeoff; it'll probably seem like "we could've seen this coming" | we-might-be-missing-some-key-feature-of-ai-takeoff-it-ll | https://www.lesswrong.com/posts/dLwo67p7zBuPsjG5t/we-might-be-missing-some-key-feature-of-ai-takeoff-it-ll | Lukas_Gloor | 2024-05-09T15:43:11.490Z |
NrXukXAJqinArX6hx | > I'm just saying these first slightly-smarter-than-human AIs won't pose a catastrophic risk to humanity in a serious sense
I'm happy to argue a bit for AIs which aren't wildly smarter than humans in any key domain posing a risk of acquiring substantial power[^power] despite not having robotics. The power would be bot... | 2024-05-10T17:12:56.300Z | 4 | KbAt69uqip67upB4e | dLwo67p7zBuPsjG5t | We might be missing some key feature of AI takeoff; it'll probably seem like "we could've seen this coming" | we-might-be-missing-some-key-feature-of-ai-takeoff-it-ll | https://www.lesswrong.com/posts/dLwo67p7zBuPsjG5t/we-might-be-missing-some-key-feature-of-ai-takeoff-it-ll | Lukas_Gloor | 2024-05-09T15:43:11.490Z |
RXnfe2ojyqmhLTXJm | I think this article fails to list the key consideration around generation: output tokens require using a KV cache which requires substantial memory bandwidth and takes up a considerable amount of memory.
From my understanding the basic situation is:
- For input (not output) tokens, you can get pretty close the the m... | 2024-05-15T00:00:33.809Z | 13 | null | g7H2sSGHAeYxCHzrz | How much AI inference can we do? | how-much-ai-inference-can-we-do | https://benjamintodd.substack.com/p/how-much-ai-inference-can-we-do | Benjamin_Todd | 2024-05-14T15:10:58.539Z |
BysyCtnDnq2WwCkeo | The lower bound of "memory bandwidth vs. model size" is effectively equivalent to assuming that the batch size is a single token. I think this isn't at all close to realistic operating conditions and thus won't be a very tight lower bound. (Or reflect the most important bottlenecks.)
I think that the KV cache for a si... | 2024-05-15T18:40:23.481Z | 7 | AYuGkoA6M9LhNYKzt | g7H2sSGHAeYxCHzrz | How much AI inference can we do? | how-much-ai-inference-can-we-do | https://benjamintodd.substack.com/p/how-much-ai-inference-can-we-do | Benjamin_Todd | 2024-05-14T15:10:58.539Z |
6NrRRpPEJGrt4wqpZ | A thing I always feel like I'm missing in your stories of how the future goes is: "if it is obvious that the AIs are exerting substantial influence and acquiring money/power, why don't people train competitor AIs which don't take a cut?"
A key difference between AIs and immigrants is that it might be relatively easy t... | 2024-05-16T03:07:51.816Z | 12 | 3MTR6WaYqTt8MDKjj | xSJMj3Hw3D7DPy5fJ | "Humanity vs. AGI" Will Never Look Like "Humanity vs. AGI" to Humanity | humanity-vs-agi-will-never-look-like-humanity-vs-agi-to | https://www.lesswrong.com/posts/xSJMj3Hw3D7DPy5fJ/humanity-vs-agi-will-never-look-like-humanity-vs-agi-to | Thane Ruthenis | 2023-12-16T20:08:39.375Z |
jjL9WEPQ2TkWtoFMx | > In the medium to long-term, when AIs become legal persons, "replacing them" won't be an option -- as that would violate their rights. And creating a new AI to compete with them wouldn't eliminate them entirely. It would just reduce their power somewhat by undercutting their wages or bargaining power.
Naively, it see... | 2024-05-16T04:15:02.775Z | 2 | gfQQttrQDC5kbXg9n | xSJMj3Hw3D7DPy5fJ | "Humanity vs. AGI" Will Never Look Like "Humanity vs. AGI" to Humanity | humanity-vs-agi-will-never-look-like-humanity-vs-agi-to | https://www.lesswrong.com/posts/xSJMj3Hw3D7DPy5fJ/humanity-vs-agi-will-never-look-like-humanity-vs-agi-to | Thane Ruthenis | 2023-12-16T20:08:39.375Z |
bD9vEhAerA7mPWyYS | I think we probably disagree substantially on the difficulty of alignment and the relationship between "resources invested in alignment technology" and "what fraction aligned those AIs are" (by fraction aligned, I mean what fraction of resources they take as a cut).
I also think that something like a basin of corrigib... | 2024-05-16T04:36:35.906Z | 8 | gfQQttrQDC5kbXg9n | xSJMj3Hw3D7DPy5fJ | "Humanity vs. AGI" Will Never Look Like "Humanity vs. AGI" to Humanity | humanity-vs-agi-will-never-look-like-humanity-vs-agi-to | https://www.lesswrong.com/posts/xSJMj3Hw3D7DPy5fJ/humanity-vs-agi-will-never-look-like-humanity-vs-agi-to | Thane Ruthenis | 2023-12-16T20:08:39.375Z |
bBWxAQ4pw7eyL7YE2 | (I wasn't trying to trying to argue against your overall point in this comment, I was just pointing out something which doesn't make sense to me in isolation. See [this other comment](https://www.lesswrong.com/posts/xSJMj3Hw3D7DPy5fJ/humanity-vs-agi-will-never-look-like-humanity-vs-agi-to?commentId=bD9vEhAerA7mPWyYS) f... | 2024-05-16T04:45:03.314Z | 2 | wdAFtCsRtGQDSG4Ci | xSJMj3Hw3D7DPy5fJ | "Humanity vs. AGI" Will Never Look Like "Humanity vs. AGI" to Humanity | humanity-vs-agi-will-never-look-like-humanity-vs-agi-to | https://www.lesswrong.com/posts/xSJMj3Hw3D7DPy5fJ/humanity-vs-agi-will-never-look-like-humanity-vs-agi-to | Thane Ruthenis | 2023-12-16T20:08:39.375Z |
sdmBDgjmkA55R9csr | > I think what's crucial here is that I think perfect alignment is very likely unattainable. If that's true, then we'll get some form of "value drift" in almost any realistic scenario. Over long periods, the world will start to look alien and inhuman. Here, the difficulty of alignment mostly sets how quickly this drift... | 2024-05-16T04:53:58.980Z | 2 | wdAFtCsRtGQDSG4Ci | xSJMj3Hw3D7DPy5fJ | "Humanity vs. AGI" Will Never Look Like "Humanity vs. AGI" to Humanity | humanity-vs-agi-will-never-look-like-humanity-vs-agi-to | https://www.lesswrong.com/posts/xSJMj3Hw3D7DPy5fJ/humanity-vs-agi-will-never-look-like-humanity-vs-agi-to | Thane Ruthenis | 2023-12-16T20:08:39.375Z |
M76NtfNmsu5WpRkmb | > For what it's worth, I'm not sure which part of my scenario you are referring to here, because these are both statements I agree with.
I was arguing against:
> This story makes sense to me because I think even imperfect AIs will be a great deal for humanity. In my story, the loss of control will be gradual enough ... | 2024-05-16T05:10:40.796Z | 2 | tTvncHvcgpbNAbtXo | xSJMj3Hw3D7DPy5fJ | "Humanity vs. AGI" Will Never Look Like "Humanity vs. AGI" to Humanity | humanity-vs-agi-will-never-look-like-humanity-vs-agi-to | https://www.lesswrong.com/posts/xSJMj3Hw3D7DPy5fJ/humanity-vs-agi-will-never-look-like-humanity-vs-agi-to | Thane Ruthenis | 2023-12-16T20:08:39.375Z |
tjFyEP9L3TGjvdnmh | > I thought the idea of a corrigible AI is that you're trying to build something that isn't itself independent and agentic, but will help you in your goals regardless.
Hmm, no I mean something broader than this, something like "humans ultimately have control and will decide what happens". In my usage of the word, I w... | 2024-05-16T05:15:50.032Z | 3 | tTvncHvcgpbNAbtXo | xSJMj3Hw3D7DPy5fJ | "Humanity vs. AGI" Will Never Look Like "Humanity vs. AGI" to Humanity | humanity-vs-agi-will-never-look-like-humanity-vs-agi-to | https://www.lesswrong.com/posts/xSJMj3Hw3D7DPy5fJ/humanity-vs-agi-will-never-look-like-humanity-vs-agi-to | Thane Ruthenis | 2023-12-16T20:08:39.375Z |
ex2go5HF7uxXBnzeq | Huh, this seems messy. I wish Time was less ambigious with their language here and more clear about exactly what they have/haven't seen.
It seems like the current quote you used is an accurate representation of the article, but I worry that it isn't an accurate representation of what is actually going on.
It seems pl... | 2024-05-16T21:20:35.578Z | 5 | hAYPK7b9DhhQwwcxT | 2KDnyEyBKk3xP28oA | AISN #35: Lobbying on AI Regulation Plus, New Models from OpenAI and Google, and Legal Regimes for Training on Copyrighted Data | aisn-35-lobbying-on-ai-regulation-plus-new-models-from | https://newsletter.safe.ai/p/ai-safety-newsletter-35-lobbying?r=7oh0&triedRedirect=true | Dan H | 2024-05-16T14:29:21.683Z |
u4Dbyc5XKJ4wfCvpK | I wrote up some of my thoughts on Bengio's agenda [here](https://www.lesswrong.com/posts/edvyWfKdJHnoPkM2J/bengio-s-alignment-proposal-towards-a-cautious-scientist-ai?commentId=hYDpGY3fboEGCGuRQ).
TLDR: I'm excited about work on trying to find any interpretable hypothesis which can be highly predictive on hard predict... | 2024-05-17T04:08:16.973Z | 26 | 5bk36J6DYbGd2bCXq | wvgwYQv9B4jioqgqg | Towards Guaranteed Safe AI: A Framework for Ensuring Robust and Reliable AI Systems | towards-guaranteed-safe-ai-a-framework-for-ensuring-robust | https://arxiv.org/abs/2405.06624 | Gunnar_Zarncke | 2024-05-16T13:09:39.265Z |
2CTBaxtMqB2AHKwgK | (Note that this paper was already posted [here](https://www.lesswrong.com/posts/wvgwYQv9B4jioqgqg/towards-guaranteed-safe-ai-a-framework-for-ensuring-robust), so see comments on that post as well.) | 2024-05-17T22:59:01.045Z | 15 | null | LkECxpbjvSifPfjnb | Towards Guaranteed Safe AI: A Framework for Ensuring Robust and Reliable AI Systems | towards-guaranteed-safe-ai-a-framework-for-ensuring-robust-1 | https://www.lesswrong.com/posts/LkECxpbjvSifPfjnb/towards-guaranteed-safe-ai-a-framework-for-ensuring-robust-1 | Joar Skalse | 2024-05-17T19:13:31.380Z |
ipm34auZuhWWjkbbG | > In particular, I don't expect either (any?) lab to be able to resist the temptation to internally deploy models with autonomous persuasion capabilities or autonomous AI R&D capabilities
I agree with this as stated, but don't think that avoiding deploying such models is needed to mitigate risk.
I think various labs... | 2024-05-18T16:21:06.098Z | 9 | yTjktfdBwEgARno63 | y8eQjQaCamqdc842k | DeepMind's "Frontier Safety Framework" is weak and unambitious | deepmind-s-frontier-safety-framework-is-weak-and-unambitious | https://www.lesswrong.com/posts/y8eQjQaCamqdc842k/deepmind-s-frontier-safety-framework-is-weak-and-unambitious | Zach Stein-Perlman | 2024-05-18T03:00:13.541Z |
LFpXoTAtv4tDxhi9o | I think literal extinction is unlikely even conditional on misaligned AI takeover due to:
- The potential for the AI to be at least a tiny bit "kind" (same as humans probably wouldn't kill all aliens). [^val]
- Decision theory/trade reasons
This is discussed in more detail [here](https://www.lesswrong.com/posts/2Nncx... | 2024-05-19T00:28:52.653Z | 23 | null | 6C3ndLd3nkrfy4K6j | "If we go extinct due to misaligned AI, at least nature will continue, right? ... right?" | if-we-go-extinct-due-to-misaligned-ai-at-least-nature-will | https://aisafety.info/questions/NM1Y/If-we-go-extinct-due-to-misaligned-AI,-at-least-nature-will-continue,-right | plex | 2024-05-18T14:09:53.014Z |
rGgkBSk84DjzoYqsG | I agree with 1 and think that race dynamics makes the situation considerably worse when we only have access to prosaic approaches. (Though I don't think this is the biggest issue with these approaches.)
I think I expect a period substantially longer than several months by default due to slower takeoff than this. (More... | 2024-05-19T00:53:04.804Z | 9 | mpWwXxoDhjGkiWz5C | y8eQjQaCamqdc842k | DeepMind's "Frontier Safety Framework" is weak and unambitious | deepmind-s-frontier-safety-framework-is-weak-and-unambitious | https://www.lesswrong.com/posts/y8eQjQaCamqdc842k/deepmind-s-frontier-safety-framework-is-weak-and-unambitious | Zach Stein-Perlman | 2024-05-18T03:00:13.541Z |
npaZCYFCiwdMjko8Z | I've thought a bit about actions to [reduce the probability that AI takeover involves violent conflict](https://www.lesswrong.com/posts/FG54euEAesRkSZuJN/ryan_greenblatt-s-shortform?commentId=MLh7RLJhEwTDh5brm).
I don't think there are any amazing looking options. If goverments were generally more competent that would... | 2024-05-20T21:28:39.867Z | 4 | aak25f5s77MH4cufs | 6C3ndLd3nkrfy4K6j | "If we go extinct due to misaligned AI, at least nature will continue, right? ... right?" | if-we-go-extinct-due-to-misaligned-ai-at-least-nature-will | https://aisafety.info/questions/NM1Y/If-we-go-extinct-due-to-misaligned-AI,-at-least-nature-will-continue,-right | plex | 2024-05-18T14:09:53.014Z |
mrTbTBwxmhQ8xtFfb | > Integrated gradients is a computationally efficient attribution method (compared to activation patching / ablations) grounded in a series of axioms.
Maybe I'm confused, but isn't integrated gradients strictly slower than an ablation to a baseline? | 2024-05-21T02:59:43.890Z | 2 | null | Rv6ba3CMhZGZzNH7x | Interpretability: Integrated Gradients is a decent attribution method | interpretability-integrated-gradients-is-a-decent | https://www.lesswrong.com/posts/Rv6ba3CMhZGZzNH7x/interpretability-integrated-gradients-is-a-decent | Lucius Bushnaq | 2024-05-20T17:55:22.893Z |
KipdKgFELqk4dprJj | [Not very confident, but just saying my current view.]
I'm pretty skeptical about integrated gradients.
As far as why, I don't think we should care about the derivative at the baseline (zero or the mean).
As far as the axioms, I think I get off the train on "Completeness" which doesn't seem like a property we need/w... | 2024-05-21T03:14:23.839Z | 4 | null | Rv6ba3CMhZGZzNH7x | Interpretability: Integrated Gradients is a decent attribution method | interpretability-integrated-gradients-is-a-decent | https://www.lesswrong.com/posts/Rv6ba3CMhZGZzNH7x/interpretability-integrated-gradients-is-a-decent | Lucius Bushnaq | 2024-05-20T17:55:22.893Z |
Kye5phbzSAiouzvM7 | > A core advantage of Bayesian methods is the ability to handle out-of-distribution situations more gracefully
I dispute that Bayesian methods will be much better at this in practice.
[
Aside:
> In general, most (?) AI safety problems can be cast as an instance of a case where a model behaves as intended on a train... | 2024-05-21T17:26:57.109Z | 6 | RkQvgaj5BgqevDSgm | wvgwYQv9B4jioqgqg | Towards Guaranteed Safe AI: A Framework for Ensuring Robust and Reliable AI Systems | towards-guaranteed-safe-ai-a-framework-for-ensuring-robust | https://arxiv.org/abs/2405.06624 | Gunnar_Zarncke | 2024-05-16T13:09:39.265Z |
5p6i6sgDu8ewfnsCh | Supposing he was a serious contributor to the paper (which seems unlikely IMO), it seems bad to cut his contribution just because he died.
So, I think the right choice here will depend on how much being an author on this paper is about endorsement or about contribution.
(Even if he didn't contribute much of the conte... | 2024-05-21T21:49:44.841Z | 4 | 3pJbCM6rQi7GqYguF | BikZyjiEgFmo7HQHm | Mitigating extreme AI risks amid rapid progress [Linkpost] | mitigating-extreme-ai-risks-amid-rapid-progress-linkpost | https://www.lesswrong.com/posts/BikZyjiEgFmo7HQHm/mitigating-extreme-ai-risks-amid-rapid-progress-linkpost | Orpheus16 | 2024-05-21T19:59:21.343Z |
dLore5GPobgae7ivm | As far as transparency and commitments, you could ask for:
- A track record of making and holding non-trivial commitments. (Of any kind.)
- some sort of proposal like [this one](https://sideways-view.com/2018/02/01/honest-organizations/) | 2024-05-22T05:29:55.327Z | 6 | azxBcexXHLKFoL9Ha | N2r9EayvsWJmLBZuF | Introducing AI Lab Watch | introducing-ai-lab-watch | https://ailabwatch.org | Zach Stein-Perlman | 2024-04-30T17:00:12.652Z |
3Z4YiRCmpnJxkBYoz | Note that scasper said:
> Today’s new SAE paper from Anthropic was full of brilliant experiments and interesting insights,
I (like scasper) think this work is useful, but I share some of scasper's concerns.
In particular:
- I think prior work like this from the anthropic interp team has been systematically overrate... | 2024-05-22T16:08:07.676Z | 28 | aid5D4LK2GrrcCfcQ | pH6tyhEnngqWAXi9i | EIS XIII: Reflections on Anthropic’s SAE Research Circa May 2024 | eis-xiii-reflections-on-anthropic-s-sae-research-circa-may | https://www.lesswrong.com/posts/pH6tyhEnngqWAXi9i/eis-xiii-reflections-on-anthropic-s-sae-research-circa-may | scasper | 2024-05-21T20:15:36.502Z |
eGWFCCAoJTqusWmat | > It seems like the average user is using far more than 20$ of requests
I'm skeptical. I bet the average user is actually using far less than $20 per month.
(Both the median user and the average usage are probably <$20 per month IMO.)
Keep in mind that the typical user is pretty different from the typical power user... | 2024-05-22T16:15:05.950Z | 4 | EGPAH9Xjd5xCzqwKX | 6wNGkuar6WXAtjv8E | What would stop you from paying for an LLM? | what-would-stop-you-from-paying-for-an-llm | https://www.lesswrong.com/posts/6wNGkuar6WXAtjv8E/what-would-stop-you-from-paying-for-an-llm | yanni kyriacos | 2024-05-21T22:25:52.949Z |
JfSKACN9hvy7LbzKg | I'm not sold this shows dataset contamination.
- They don't re-baseline with humans. (Based on my last skim a while ago.)
- It is easy to make math problems considerably harder by changing the constants and often math problems are designed to make the constants easy to work with.
- Both humans and AI are used to const... | 2024-05-22T18:21:16.436Z | 4 | B4CQyYzYxu8wgvmox | X3p8mxE5dHYDZNxCm | A concrete bet offer to those with short AGI timelines | a-concrete-bet-offer-to-those-with-short-agi-timelines | https://www.lesswrong.com/posts/X3p8mxE5dHYDZNxCm/a-concrete-bet-offer-to-those-with-short-agi-timelines | Matthew Barnett | 2022-04-09T21:41:45.106Z |
RgEma2qGsjQdsNWgB | Worth noting that many of the free options might train on your data.
So, if this is a serious issue for the work you're doing, take this into account.
(I don't really worry about the effects on acceleration of paying for stuff like this (or contributing training data) as the effect seems minor compared to other thing... | 2024-05-22T18:37:15.401Z | 4 | 8YLfjT27ZCYvoJ4wb | 6wNGkuar6WXAtjv8E | What would stop you from paying for an LLM? | what-would-stop-you-from-paying-for-an-llm | https://www.lesswrong.com/posts/6wNGkuar6WXAtjv8E/what-would-stop-you-from-paying-for-an-llm | yanni kyriacos | 2024-05-21T22:25:52.949Z |
nEAxdNQaaNHyxzu9s | > I thus think its fair to say that -- empirically -- neural networks do not robustly quantify uncertainty in a reliable way when out-of-distribution
Sure, but why will the bayesian model reliably quantify uncertainty OOD? There is also no guarantee of this (OOD).
Whether or not you get reliable uncertainly quanitifi... | 2024-05-22T21:19:38.510Z | 9 | pTjTM4uq8n5hhbRAP | wvgwYQv9B4jioqgqg | Towards Guaranteed Safe AI: A Framework for Ensuring Robust and Reliable AI Systems | towards-guaranteed-safe-ai-a-framework-for-ensuring-robust | https://arxiv.org/abs/2405.06624 | Gunnar_Zarncke | 2024-05-16T13:09:39.265Z |
apSWxC6i4wege3JL8 | Separately, I guess I'm not that worried about failures in which the network itself doesn't "understand" what's going on. So the main issue are cases where the model in some sense knows, but doesn't report this. (E.g. ELK problems at least broadly speaking.)
I think there are bunch of issues that look sort of like thi... | 2024-05-22T21:34:07.120Z | 2 | nEAxdNQaaNHyxzu9s | wvgwYQv9B4jioqgqg | Towards Guaranteed Safe AI: A Framework for Ensuring Robust and Reliable AI Systems | towards-guaranteed-safe-ai-a-framework-for-ensuring-robust | https://arxiv.org/abs/2405.06624 | Gunnar_Zarncke | 2024-05-16T13:09:39.265Z |
4tn4hciP6DKGtukhq | > This would not produce an interpretable model, and so you may not trust it very much (unless you have some learning-theoretic guarantee, perhaps), but it would not be difficult to determine if such a model "thinks" that harm would occur in a given scenario.
Won't this depend on the generalization properties of the m... | 2024-05-22T21:36:48.493Z | 2 | pTjTM4uq8n5hhbRAP | wvgwYQv9B4jioqgqg | Towards Guaranteed Safe AI: A Framework for Ensuring Robust and Reliable AI Systems | towards-guaranteed-safe-ai-a-framework-for-ensuring-robust | https://arxiv.org/abs/2405.06624 | Gunnar_Zarncke | 2024-05-16T13:09:39.265Z |
ZsmujtvbmReapn4YC | > or turn it into an interpretable model, or something like that, then yes, you would probably have to solve ELK, or solve interpretability, or something like that. I would not be very optimistic about that working
TBC, I would count it as solving an ELK problem if you constructed an interpretable world model which al... | 2024-05-22T21:37:40.410Z | 2 | pTjTM4uq8n5hhbRAP | wvgwYQv9B4jioqgqg | Towards Guaranteed Safe AI: A Framework for Ensuring Robust and Reliable AI Systems | towards-guaranteed-safe-ai-a-framework-for-ensuring-robust | https://arxiv.org/abs/2405.06624 | Gunnar_Zarncke | 2024-05-16T13:09:39.265Z |
Ns6pzzZy9rvzLej7m | > However, that isn't the use-case here (the world model is not meant to be agentic).
Worth noting here that I'm mostly talking about Bengio's proposal wrt to the bayes related arguments.
And I agree that the world model isn't meant to be a schemer, but it's not as though we can guarantee that without some additional... | 2024-05-22T21:49:46.118Z | 2 | pTjTM4uq8n5hhbRAP | wvgwYQv9B4jioqgqg | Towards Guaranteed Safe AI: A Framework for Ensuring Robust and Reliable AI Systems | towards-guaranteed-safe-ai-a-framework-for-ensuring-robust | https://arxiv.org/abs/2405.06624 | Gunnar_Zarncke | 2024-05-16T13:09:39.265Z |
z979Ce9nHCavyJ5jW | > And so on. If you want to prove something like "the AI will never copy itself to an external computer", or "the AI will never communicate outside this trusted channel", or "the AI will never tamper with this sensor", or something like that, then your world model might not need to be all that detailed.
I agree that ... | 2024-05-22T21:55:35.868Z | 2 | pTjTM4uq8n5hhbRAP | wvgwYQv9B4jioqgqg | Towards Guaranteed Safe AI: A Framework for Ensuring Robust and Reliable AI Systems | towards-guaranteed-safe-ai-a-framework-for-ensuring-robust | https://arxiv.org/abs/2405.06624 | Gunnar_Zarncke | 2024-05-16T13:09:39.265Z |
AZGA8vLR6hJNeByuh | (For catastrophically dangerous if misused models?)
(Really, we also need to suppose there are issues with strategy stealing for open source to be a problem. E.g. offense-defense inbalances or alignment difficulties.) | 2024-05-23T06:16:55.140Z | 2 | K2F7DXD2zbv5JcAZb | Lgq2DcuahKmLktDvC | Applying refusal-vector ablation to a Llama 3 70B agent | applying-refusal-vector-ablation-to-a-llama-3-70b-agent | https://www.lesswrong.com/posts/Lgq2DcuahKmLktDvC/applying-refusal-vector-ablation-to-a-llama-3-70b-agent | Simon Lermen | 2024-05-11T00:08:08.117Z |
gJwtzannJQeZxij9J | > Google and Anthropic are doing good risk assessment for dangerous capabilities.
My guess is that the current level of evaluations at Google and Anthropic isn't terrible, but could be massively improved.
- Elicitation isn't well specified and there isn't an established science.
- Let alone reliable projections abo... | 2024-05-23T06:46:48.718Z | 1 | null | qfEgzQ9jGEk9Cegvy | New voluntary commitments (AI Seoul Summit) | new-voluntary-commitments-ai-seoul-summit | https://www.gov.uk/government/publications/frontier-ai-safety-commitments-ai-seoul-summit-2024/frontier-ai-safety-commitments-ai-seoul-summit-2024 | Zach Stein-Perlman | 2024-05-21T11:00:41.794Z |
WXhqzv5RWnLjGvevW | > 1. AI safety research is still solving easy problems. [...]
> 2. Capability development is getting AI safety research for free. [...]
> 3. AI safety research is speeding up capabilities. [...]
Even if (2) and (3) are true and (1) is mostly true (e.g. most safety research is worthless), I still think it can easily ... | 2024-05-23T18:47:34.620Z | 2 | null | vkzmbf4Mve4GNyJaF | The case for stopping AI safety research | the-case-for-stopping-ai-safety-research | https://www.lesswrong.com/posts/vkzmbf4Mve4GNyJaF/the-case-for-stopping-ai-safety-research | catubc | 2024-05-23T15:55:18.713Z |
sQmJ7TYrN26TCj39K | Examples of maybe disparagement:
- [Arguing with Soares about whether AIs will kill everyone](https://www.lesswrong.com/posts/2NncxDQ3KBDCxiJiP/cosmopolitan-values-don-t-come-free?commentId=ofPTrG6wsq7CxuTXk)
- [Saying OpenAI is probably wrong and has motivated cognition in this comment](https://www.lesswrong.com/post... | 2024-05-23T19:38:56.769Z | 10 | zSCdFg57Z94ZdmCoz | KA2HxJfhz3CSbLdbL | Open Thread Spring 2024 | open-thread-spring-2024 | https://www.lesswrong.com/posts/KA2HxJfhz3CSbLdbL/open-thread-spring-2024 | habryka | 2024-03-11T19:17:23.833Z |
wZS3nuzyGNPjTrTAw | (Oops, fixed.) | 2024-05-23T23:40:28.259Z | 2 | rWmZtekQHiwdfvxsc | vkzmbf4Mve4GNyJaF | The case for stopping AI safety research | the-case-for-stopping-ai-safety-research | https://www.lesswrong.com/posts/vkzmbf4Mve4GNyJaF/the-case-for-stopping-ai-safety-research | catubc | 2024-05-23T15:55:18.713Z |
yJdBBXuLy4LLakoBo | None of it seems falsified to me.
I think a few of Jacob's "Personal opinions" now seem less accurate than they did previously. (And perhaps Jacob no longer endorses "Opinion: OpenAI is a great place to work to reduce existential risk from AI.") | 2024-05-24T00:59:56.034Z | 1 | GF36QraB6hfjKWPTS | 3S4nyoNEEuvNsbXt8 | Common misconceptions about OpenAI | common-misconceptions-about-openai | https://www.lesswrong.com/posts/3S4nyoNEEuvNsbXt8/common-misconceptions-about-openai | Jacob_Hilton | 2022-08-25T14:02:26.257Z |
fzkaFGX9sfFEcTd3P | Fair enough if you think a core consequence of Anthropic's paper was "demonstrate that there exist directions within LLMs that correspond to concepts as abstract as 'golden gate bridge' or 'bug in code'".
It's worth noting that wildly simpler and cheaper model internals experiments have already demonstrated to an exte... | 2024-05-24T01:38:45.896Z | 2 | null | MuyCbad9ZHW8b3rMP | Cicadas, Anthropic, and the bilateral alignment problem | cicadas-anthropic-and-the-bilateral-alignment-problem | https://www.lesswrong.com/posts/MuyCbad9ZHW8b3rMP/cicadas-anthropic-and-the-bilateral-alignment-problem | kromem | 2024-05-22T11:09:56.469Z |
D86FnjjANjiWXm7xw | Insofar as the hope is:
1. Figure out how to approximate sampling from the Bayesian posterior (using e.g. GFlowNets or something).
2. Do something else that makes this actually useful for "improving" OOD generalization in some way.
It would be nice to know what (2) actually is and why we needed step (1) for it. As fa... | 2024-05-24T18:52:00.591Z | 4 | dycc6MMJgFdp5yhKK | wvgwYQv9B4jioqgqg | Towards Guaranteed Safe AI: A Framework for Ensuring Robust and Reliable AI Systems | towards-guaranteed-safe-ai-a-framework-for-ensuring-robust | https://arxiv.org/abs/2405.06624 | Gunnar_Zarncke | 2024-05-16T13:09:39.265Z |
H6bYL7v3PW364ucmH | (I think most of the hard-to-handle risk from scheming comes from cases where we can't easily make smarter AIs which we know aren't schemers. If we can get another copy of the AI which is just as smart but which has been "de-agentified", then I don't think scheming poses a substantial threat. (Because e.g. we can just ... | 2024-05-24T18:54:35.048Z | 2 | as2DyNDQG5aPRn43h | wvgwYQv9B4jioqgqg | Towards Guaranteed Safe AI: A Framework for Ensuring Robust and Reliable AI Systems | towards-guaranteed-safe-ai-a-framework-for-ensuring-robust | https://arxiv.org/abs/2405.06624 | Gunnar_Zarncke | 2024-05-16T13:09:39.265Z |
AgGuq4yR8rPz3xwoY | Probably considerably harder to influence than AI takeover given that it happens later in the singularity at a point where we have already had access to a huge amount of superhuman AI labor? | 2024-05-24T19:14:15.263Z | 5 | MqBLzocx8z8jDamop | xWMqsvHapP3nwdSW8 | My views on “doom” | my-views-on-doom | https://www.lesswrong.com/posts/xWMqsvHapP3nwdSW8/my-views-on-doom | paulfchristiano | 2023-04-27T17:50:01.415Z |
KEYg3rhEZK9JKseKK | I think it probably doesn't make sense to talk about "representative samples".
Here are a bunch of different things this could mean:
- A uniform sample from people who have done any work related to AI safety.
- A sample from people weighted to their influence/power in the AI safety community.
- A sample from people w... | 2024-05-24T19:19:14.032Z | 11 | au2Jwz7Bb3XzEZoEi | yMTNjeEHfHcf2x7nY | Big Picture AI Safety: Introduction | big-picture-ai-safety-introduction | https://www.lesswrong.com/posts/yMTNjeEHfHcf2x7nY/big-picture-ai-safety-introduction | EuanMcLean | 2024-05-23T11:15:44.037Z |
sXCkfKAHbgNLHdnwv | > I would be careful not to implicitly claim that these 17 people are a "representative sample" of the AI safety community.
Worth noting that this is directly addressed in the post:
> The sample of people I interviewed is not necessarily a representative sample of the AI safety movement as a whole. The sample was ps... | 2024-05-24T19:20:57.958Z | 7 | au2Jwz7Bb3XzEZoEi | yMTNjeEHfHcf2x7nY | Big Picture AI Safety: Introduction | big-picture-ai-safety-introduction | https://www.lesswrong.com/posts/yMTNjeEHfHcf2x7nY/big-picture-ai-safety-introduction | EuanMcLean | 2024-05-23T11:15:44.037Z |
cxA65e8LibypeHjsL | (I interpreted the bit about using llama-3 to involve fine-tuning for things other than just avoiding refusals. E.g., actually doing sufficiently high quality debates.) | 2024-05-24T19:24:48.913Z | 3 | AmCg5EJjQpHbHxmnh | wsXCXoyvRi3DnWZ2M | NYU Code Debates Update/Postmortem | nyu-code-debates-update-postmortem | https://www.lesswrong.com/posts/wsXCXoyvRi3DnWZ2M/nyu-code-debates-update-postmortem | David Rein | 2024-05-24T16:08:06.151Z |
XaiFnS4zESmfoAQzm | Observation: it's possible that the board is also powerless with respect to the employees and the leadership.
I think the board probably has less de facto power than employees/leadership by a wide margin in the current regime. | 2024-05-27T17:50:20.165Z | 8 | null | sdCcsTt9hRpbX6obP | Maybe Anthropic's Long-Term Benefit Trust is powerless | maybe-anthropic-s-long-term-benefit-trust-is-powerless | https://www.lesswrong.com/posts/sdCcsTt9hRpbX6obP/maybe-anthropic-s-long-term-benefit-trust-is-powerless | Zach Stein-Perlman | 2024-05-27T13:00:47.991Z |
b7wccaLYQ8HxhtneA | They might just (probably correctly) think it is unlikely that the employees will decide to do this. | 2024-05-27T22:30:28.306Z | 4 | Q9bv7nqdcqoXbGa5g | sdCcsTt9hRpbX6obP | Maybe Anthropic's Long-Term Benefit Trust is powerless | maybe-anthropic-s-long-term-benefit-trust-is-powerless | https://www.lesswrong.com/posts/sdCcsTt9hRpbX6obP/maybe-anthropic-s-long-term-benefit-trust-is-powerless | Zach Stein-Perlman | 2024-05-27T13:00:47.991Z |
ia5wxJoZmaKWcmskM | This seems like useful work.
I have two issues with these evaluations:
- I'm pretty skeptical of the persuasion evals. I bet this isn't measuring what you want and I'm generally skeptical of these evals as described. E.g., much more of the action is in deciding exactly who to influence and what to influence them to d... | 2024-05-27T22:41:47.633Z | 12 | null | CCBaLzpB2qvwyuEJ2 | DeepMind: Evaluating Frontier Models for Dangerous Capabilities | deepmind-evaluating-frontier-models-for-dangerous | https://arxiv.org/abs/2403.13793 | Zach Stein-Perlman | 2024-03-21T03:00:31.599Z |
y3sodgbsGkHhLrXKt | I was considering a threat model in which the AI is acting mostly autonomously. This would include both self-exfiltration, but also trying to steer the AI lab or the world in some particular direction.
I agree that misuse threat models where the AI is e.g. being used to massively reduce the cost of swinging votes via ... | 2024-05-28T18:20:21.049Z | 2 | uhP6GdcCibvExJkgJ | CCBaLzpB2qvwyuEJ2 | DeepMind: Evaluating Frontier Models for Dangerous Capabilities | deepmind-evaluating-frontier-models-for-dangerous | https://arxiv.org/abs/2403.13793 | Zach Stein-Perlman | 2024-03-21T03:00:31.599Z |
veBmtLAEHbsHz7Tgm | > But probably he should be better at communication e.g. realizing that people will react negatively to raising the possibility of bombing datacenters without lots of contextualizing.
I'm confident he knew people would react negatively but decided to keep the line because he thought it was worth the cost.
Seems like ... | 2024-05-29T00:38:02.658Z | 2 | NbYZQ5bAteknLcuSS | cGzQBRDrpNHoYtbKN | What mistakes has the AI safety movement made? | what-mistakes-has-the-ai-safety-movement-made | https://www.lesswrong.com/posts/cGzQBRDrpNHoYtbKN/what-mistakes-has-the-ai-safety-movement-made | EuanMcLean | 2024-05-23T11:19:02.717Z |
pKvdrqzBDp7cEHCu2 | [This comment is just about the "notkilleverybodyism pitch" section.]
> we doomers are also unhappy about AI killing all humans because we are human and we don’t want to get murdered by AIs.
I'd like to once again reiterate that the arguments for misaligned AIs killing literally all humans (if they succeed in takeove... | 2024-05-29T18:14:03.913Z | 23 | null | 8YhjpgQ2eLfnzQ7ec | Response to nostalgebraist: proudly waving my moral-antirealist battle flag | response-to-nostalgebraist-proudly-waving-my-moral | https://www.lesswrong.com/posts/8YhjpgQ2eLfnzQ7ec/response-to-nostalgebraist-proudly-waving-my-moral | Steven Byrnes | 2024-05-29T16:48:29.408Z |
hAnprWu7voaw2Leti | Hmm, I'd say my disagreements with the post are:
- I think people should generally be careful about using the language "kill literally everyone" or "notkilleverybodyism" insofar as they aren't confident that misaligned AI would kill literally everyone. (Or haven't considered counterarguments to this.)
- I'm not sure I... | 2024-05-29T22:06:21.380Z | 2 | exaw5dydmxk8KZvbL | 8YhjpgQ2eLfnzQ7ec | Response to nostalgebraist: proudly waving my moral-antirealist battle flag | response-to-nostalgebraist-proudly-waving-my-moral | https://www.lesswrong.com/posts/8YhjpgQ2eLfnzQ7ec/response-to-nostalgebraist-proudly-waving-my-moral | Steven Byrnes | 2024-05-29T16:48:29.408Z |
xBYimQtgASti5tgWv | > building misaligned smarter-than-human systems will kill everyone, including their children [...] if they come to understand this central truth.
I'd like to once again reiterate that the arguments for misaligned AIs killing literally all humans (if they succeed in takeover) are quite weak and probably literally all ... | 2024-05-29T22:18:13.670Z | 11 | null | tKk37BFkMzchtZThx | MIRI 2024 Communications Strategy | miri-2024-communications-strategy | https://www.lesswrong.com/posts/tKk37BFkMzchtZThx/miri-2024-communications-strategy | Gretta Duleba | 2024-05-29T19:33:39.169Z |
5ds5gjszF84gnJHxz | I originally reacted skeptically to this, but I am actually not that skeptical of "they posted now prompted by this, but this was in the works for a while". (I missed "this was in the works for a while" on my first read of your comment.) | 2024-05-30T00:12:01.608Z | 2 | 3gQvsqz2fHwjfJw2S | sdCcsTt9hRpbX6obP | Maybe Anthropic's Long-Term Benefit Trust is powerless | maybe-anthropic-s-long-term-benefit-trust-is-powerless | https://www.lesswrong.com/posts/sdCcsTt9hRpbX6obP/maybe-anthropic-s-long-term-benefit-trust-is-powerless | Zach Stein-Perlman | 2024-05-27T13:00:47.991Z |
eEaDWL9ywbRBqxKpZ | [Quoting from Gretta](https://www.lesswrong.com/posts/tKk37BFkMzchtZThx/miri-2024-communications-strategy?commentId=52oeytrtKsPnYaiCB):
> One thing we can hope for, if we get a little more time rather than a lot more time, is that we might get various forms of human cognitive enhancement working, and these smarter hum... | 2024-05-30T02:57:47.478Z | 33 | Bohdww3hddWA3NNnZ | tKk37BFkMzchtZThx | MIRI 2024 Communications Strategy | miri-2024-communications-strategy | https://www.lesswrong.com/posts/tKk37BFkMzchtZThx/miri-2024-communications-strategy | Gretta Duleba | 2024-05-29T19:33:39.169Z |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.