Orpheus16

Karma: 6,572

Orpheus16 May 4, 2025, 7:06 PM
7 points
0
in reply to: habryka’s comment on: RA x ControlAI video: What if AI just keeps getting smarter?
I think there are organizations that themselves would be more likely to be robustly trustworthy and would be more fine to link to
I would be curious for your thoughts on which organizations you feel are robustly trustworthy.
Bonus points for a list that is kind of a weighted sum of “robustly trustworthy” and “having a meaningful impact RE improving public/policymaker understanding”. (Adding this in because I suspect that it’s easier to maintain “robustly trustworthy” status if one simply chooses not to do a lot of externally-focused comms, so it’s particularly impressive to have the combination of “doing lots of useful comms/policy work” and “managing to stay precise/accurate/trustworthy”).

Orpheus16 May 4, 2025, 6:59 PM
9 points
0
on: AI Governance to Avoid Extinction: The Strategic Landscape and Actionable Research Questions
I appreciate the articulation and assessment of various strategies. My comment will focus on a specific angle that I notice both in the report and in the broader ecosystem:
I think there has been a conflating of “catastrophic risks” and “extinction/existential risks” recently, especially among groups that are trying to influence policy. This is somewhat understandable– the difference between “catastrophic” and “existential” is not that big of a deal in most people’s minds. But in some contexts, I think it misses the fact that “existential [and thus by definition irreversible]” is actually a very different level of risk compared to “catastrophic [but something that we would be able to recover from.]”
This view seems to be (implicitly) expressed in the report summary, most notably the chart. It seems to me like the main frame is something like “if you want to avoid an unacceptable chance of catastrophic risk, all of these other options are bad.”
But not all of these catastrophic risks are the same, I think this is actually quite an important consideration, and I think even (some) policymakers would/will see this as an essential consideration as AGI becomes more salient.
Specifically, “war” and “misuse” seem very different than “extinction” or “total and irreversible civilizational collapse.”
- “War” is broad enough to encompass many outcomes (ranging from “conflict with <1M deaths” to “nuclear conflict in which civilization recovers” all the way to “nuclear conflict in which civilization does not recover.”) Note also that many natsec leaders already think the chance of a war between the US and China is at a level that would probably meet an intuitive bar for “unacceptable.” (I don’t have actual statistics on this but my guess is that >10% chance of war in the next decade is not an uncommon view. One plausible pathway that is discussed often is China invading Taiwan and US being committed to its defense).
- “Misuse” can refer to many different kinds of events (including $1B in damages from a cyberattack, 10M deaths, 1B deaths, or complete human extinction.) These are, of course, very different in terms of their overall impact, even though all of them are intuitively/emotionally stored as “very bad things that we would ideally avoid.”
It seems plausible to me that we will be in situations in which policymakers have to make tricky trade-offs between these different sources of risk, and my hope is that the community of people concerned about AI can distinguish between the different “levels” or “magnitudes” of different types of risks.
(My impression is that MIRI agrees with this, so this is more a comment on how the summary was presented & more a general note of caution to the ecosystem as a whole. I also suspect that the distinction between “catastrophic” and “existential/civilization-ending” will become increasingly more important as the AI conversation becomes more interlinked with the national security apparatus.)
Caveat: I have not read the full report and this comment is mostly inspired by the summary, the chart, and a general sense that many organizations other than MIRI are also engaging in this kind of conflation.

Orpheus16 Apr 10, 2025, 12:30 PM
4 points
0
in reply to: Alexander Gietelink Oldenziel’s comment on: Alexander Gietelink Oldenziel’s Shortform
I feel this way and generally think that on-the-margin we have too much forecasting and not enough “build plans for what to do if there is a sudden shift in political will” or “just directly engage with policymakers and help them understand things not via longform writing but via conversations/meetings.”
Many details will be ~impossible to predict and many details will not matter much (i.e., will not be action-relevant for the stakeholders who have the potential to meaningfully affect the current race to AGI).
That’s not to say forecasting is always unhelpful. Things like AI2027 can certainly move discussions forward and perhaps get new folks interested. But EG, my biggest critique of AI2027 is that I suspect they’re spending too much time/effort on detailed longform forecasting and too little effort on arranging meetings with Important Stakeholders, developing a strong presence in DC, forming policy recommendations, and related activities. (And TBC I respect/admire the AI2027 team, have relayed this feedback to them, and imagine they have thoughtful reasons for taking the approach they’re taking.)

Orpheus16 Apr 5, 2025, 3:56 PM
17 points
0
in reply to: Buck’s comment on: Buck’s Shortform
What do you think are the most important points that weren’t publicly discussed before?

Orpheus16 Jan 7, 2025, 11:06 PM
10 points
0
in reply to: ryan_greenblatt’s comment on: ryan_greenblatt’s Shortform
I expect that outcomes like “AIs are capable enough to automate virtually all remote workers” and “the AIs are capable enough that immediate AI takeover is very plausible (in the absence of countermeasures)” come shortly after (median 1.5 years and 2 years after respectively under my views).
@ryan_greenblatt can you say more about what you expect to happen from the period in-between “AI 10Xes AI R&D” and “AI takeover is very plausible?”
I’m particularly interested in getting a sense of what sorts of things will be visible to the USG and the public during this period. Would be curious for your takes on how much of this stays relatively private/internal (e.g., only a handful of well-connected SF people know how good the systems are) vs. obvious/public/visible (e.g., the majority of the media-consuming American public is aware of the fact that AI research has been mostly automated) or somewhere in-between (e.g., most DC tech policy staffers know this but most non-tech people are not aware.)

Orpheus16 Jan 2, 2025, 8:35 PM
LW: 60 AF: 27
21
AF
on: What’s the short timeline plan?
Big fan of this post. One thing worth highlighting IMO: The post assumes that governments will not react in time, so it’s mostly up to the labs (and researchers who can influence the labs) to figure out how to make this go well.
TBC, I think it’s a plausible and reasonable assumption to make. But I think this assumption ends up meaning that “the plan” excludes a lot of the work that could make the USG (a) more likely to get involved or (b) more likely to do good and useful things conditional on them deciding to get involved.
Here’s an alternative frame: I would call the plan described in Marius’s post something like the “short timelines plan assuming that governments do not get involved and assuming that technical tools (namely control/AI-automated AI R&D) are the only/main tools we can use to achieve good outcomes.”
You could imagine an alternative plan described as something like the “short timelines plan assuming that technical tools in the current AGI development race/paradigm are not sufficient and governance tools (namely getting the USG to provide considerably more oversight into AGI development, curb race dynamics, make major improvements to security) are the only/main tools we can use to achieve good outcomes.” This kind of plan would involve a very different focus.
Here are some examples of things that I think would be featured in a “government-focused” short timelines plan:
- Demos of dangerous capabilities
- Explanations of misalignment risks to senior policymakers. Identifying specific people who would be best-suited to provide those explanations, having those people practice giving explanations and addressing counterarguments, etc.
- Plans for what the “trailing labs” should do if the leading lab appears to have an insurmountable lead (e.g., OpenAI develops a model that is automating AI R&D. It becomes clear that DeepMind and Anthropic are substantially behind OpenAI. At this point, do the labs merge and assist? Do they try to do a big, coordinated, costly push to get governments to take AI risks more seriously?)
- Emergency preparedness– getting governments to be more likely to detect and appropriately respond to time-sensitive risks.
- Preparing plans for what to do if governments become considerably more concerned about risks (e.g., preparing concrete Manhattan Project or CERN-for-AI style proposals, identifying and developing verification methods for domestic or international AI regulation.)
One possible counter is that under short timelines, the USG is super unlikely to get involved. Personally, I think we should have a lot of uncertainty RE how the USG will react. Examples of factors here: (a) new Administration, (b) uncertainty over whether AI will produce real-world incidents, (c) uncertainty over how compelling demos will be, (d) chatGPT being an illustrative example of a big increase in USG involvement that lots of folks didn’t see coming, and (e) examples of the USG suddenly becoming a lot more interested in a national security domain (e.g., 9/11--> Patriot Act, recent Tik Tok ban), (f) Trump being generally harder to predict than most Presidents (e.g., more likely to form opinions for himself, less likely to trust the establishment views in some cases).
(And just to be clear, this isn’t really a critique of Marius’s post. I think it’s great for people to be thinking about what the “plan” should be if the USG doesn’t react in time. Separately, I’d be excited for people to write more about what the short timelines “plan” should look like under different assumptions about USG involvement.)

Orpheus16 Dec 31, 2024, 8:05 PM
6 points
4
in reply to: ryan_greenblatt’s comment on: ryan_greenblatt’s Shortform
At first glance, I don’t see how the point I raised is affected by the distinction between expert-level AIs vs earlier AIs.

In both cases, you could expect an important part of the story to be “what are the comparative strengths and weaknesses of this AI system.”

For example, suppose you have an AI system that dominates human experts at every single relevant domain of cognition. It still seems like there’s a big difference between “system that is 10% better at every relevant domain of cognition” and “system that is 300% better at domain X and only 10% better at domain Y.”

To make it less abstract, one might suspect that by the time we have AI that is 10% better than humans at “conceptual/serial” stuff, the same AI system is 1000% better at “speed/parallel” stuff. And this would have pretty big implications for what kind of AI R&D ends up happening (even if we condition on only focusing on systems that dominate experts in every relevant domain.)

Orpheus16 Dec 31, 2024, 6:21 PM
5 points
0
in reply to: Buck’s comment on: Buck’s Shortform
Models that don’t even cause safety problems, and aren’t even goal-directedly misaligned, but that fail to live up to their potential, thus failing to provide us with the benefits we were hoping to get when we trained them. For example, sycophantic myopic reward hacking models that can’t be made to do useful research.
Would this kind of model present any risk? Could a lab just say “oh darn, this thing isn’t very useful– let’s turn this off and develop a new model”?

Orpheus16 Dec 31, 2024, 6:15 PM
14 points
1
in reply to: Matthew Barnett’s comment on: Matthew Barnett’s Shortform
Do you have any suggestions RE alternative (more precise) terms? Or do you think it’s more of a situation where authors should use the existing terms but make sure to define them in the context of their own work? (e.g., “In this paper, when I use the term AGI, I am referring to a system that [insert description of the capabilities of the system.])

Orpheus16 Dec 31, 2024, 6:10 PM
7 points
2
in reply to: ryan_greenblatt’s comment on: ryan_greenblatt’s Shortform
The point I make here is also likely obvious to many, but I wonder if the “X human equivalents” frame often implicitly assumes that GPT-N will be like having X humans. But if we expect AIs to have comparative advantages (and disadvantages), then this picture might miss some important factors.
The “human equivalents” frame seems most accurate in worlds where the capability profile of an AI looks pretty similar to the capability profile of humans. That is, getting GPT-6 to do AI R&D is basically “the same as” getting X humans to do AI R&D. It thinks in fairly similar ways and has fairly similar strengths/weaknesses.
The frame is less accurate in worlds where AI is really good at some things and really bad at other things. In this case, if you try to estimate the # of human equivalents that GPT-6 gets you, the result might be misleading or incomplete. A lot of fuzzier things will affect the picture.
The example I’ve seen discussed most is whether or not we expect certain kinds of R&D to be bottlenecked by “running lots of experiments” or “thinking deeply and having core conceptual insights.” My impression is that one reason why some MIRI folks are pessimistic is that they expect capabilities research to be more easily automatable (AIs will be relatively good at running lots of ML experiments quickly, which helps capabilities more under their model) than alignment research (AIs will be relatively bad at thinking deeply or serially about certain topics, which is what you need for meaningful alignment progress under their model).
Perhaps more people should write about what kinds of tasks they expect GPT-X to be “relatively good at” or “relatively bad at”. Or perhaps that’s too hard to predict in advance. If so, it could still be good to write about how different “capability profiles” could allow certain kinds of tasks to be automated more quickly than others.
(I do think that the “human equivalents” frame is easier to model and seems like an overall fine simplification for various analyses.)

Orpheus16 Dec 30, 2024, 9:01 PM
23 points
5
in reply to: evhub’s comment on: evhub’s Shortform
I’m glad you’re doing this, and I support many of the ideas already suggested. Some additional ideas:
- Interview program. Work with USAISI or UKAISI (or DHS/NSA) to pilot an interview program in which officials can ask questions about AI capabilities, safety and security threats, and national security concerns. (If it’s not feasible to do this with a government entity yet, start a pilot with a non-government group– perhaps METR, Apollo, Palisade, or the new AI Futures Project.)
- Clear communication about RSP capability thresholds. I think the RSP could do a better job at outlining the kinds of capabilities that Anthropic is worried about and what sorts of thresholds would trigger a reaction. I think the OpenAI preparedness framework tables are a good example of this kind of clear/concise communication. It’s easy for a naive reader to quickly get a sense of “oh, this is the kind of capability that OpenAI is worried about.” (Clarification: I’m not suggesting that Anthropic should abandon the ASL approach or that OpenAI has necessarily identified the right capability thresholds. I’m saying that the tables are a good example of the kind of clarity I’m looking for– someone could skim this and easily get a sense of what thresholds OpenAI is tracking, and I think OpenAI’s PF currently achieves this much more than the Anthropic RSP.)
- Emergency protocols. Publishing an emergency protocol that specifies how Anthropic would react if it needed to quickly shut down a dangerous AI system. (See some specific prompts in the “AI developer emergency response protocol” section here). Some information can be redacted from a public version (I think it’s important to have a public version, though, partly to help government stakeholders understand how to handle emergency scenarios, partly to raise the standard for other labs, and partly to acquire feedback from external groups.)
- RSP surveys. Evaluate the extent to which Anthropic employees understand the RSP, their attitudes toward the RSP, and how the RSP affects their work. More on this here.
- More communication about Anthropic’s views about AI risks and AI policy. Some specific examples of hypothetical posts I’d love to see:
  - “How Anthropic thinks about misalignment risks”
  - “What the world should do if the alignment problem ends up being hard”
  - “How we plan to achieve state-proof security before AGI”
  - Encouraging more employees to share their views on various topics, EG Sam Bowman’s post.
- AI dialogues/debates. It would be interesting to see Anthropic employees have discussions/debates from other folks thinking about advanced AI. Hypothetical examples:
  - “What are the best things the US government should be doing to prepare for advanced AI” with Jack Clark and Daniel Kokotajlo.
  - “Should we have a CERN for AI?” with [someone from Anthropic] and Miles Brundage.
  - “How difficult should we expect alignment to be” with [someone from Anthropic] and [someone who expects alignment to be harder; perhaps Jeffrey Ladish or Malo Bourgon].
More ambitiously, I feel like I don’t really understand Anthropic’s plan for how to manage race dynamics in worlds where alignment ends up being “hard enough to require a lot more than RSPs and voluntary commitments.”
From a policy standpoint, several of the most interesting open questions seem to be along the lines of “under what circumstances should the USG get considerably more involved in overseeing certain kinds of AI development” and “conditional on the USG wanting to get way more involved, what are the best things for it to do?” It’s plausible that Anthropic is limited in how much work it could do on these kinds of questions (particularly in a public way). Nonetheless, it could be interesting to see Anthropic engage more with questions like the ones Miles raises here.

Orpheus16 Dec 25, 2024, 10:26 PM
4 points
0
on: “Carefully Bootstrapped Alignment” is organizationally hard
I think this a helpful overview post. It outlines challenges to a mainstream plan (bootstrapped alignment) and offers a few case studies of how entities in other fields handle complex organizational challenges.
I’d be excited to see more follow-up research on organizational design and organizational culture. This work might be especially useful for helping folks think about various AI policy proposals.
For example, it seems plausible that at some point the US government will view certain kinds of AI systems as critical national security assets. At that point, the government might become a lot more involved in AI development. It might be the case that a new entity is created (e.g., a DOD-led “Manhattan Project”) or it might be the case that the existing landscape persists but with a lot more oversight (e.g., NSA and DoD folks show up as resident inspectors at OpenAI/DeepMind/Anthropic and have a lot more decision-making power).
If something like this happens, there will be a period of time in which relevant USG stakeholders (and AGI lab leaders) have to think through questions about organizational design, organizational culture, decision-making processes, governance structures, etc. It seems likely that some versions of this look a lot better than others.
Examples of questions I’d love to see answered/discussed:
- Broadly, what are the 2-3 most important design features or considerations for an AGI project?
- Suppose the USG becomes interested in (soft) nationalizing certain kinds of AI development. A senior AI policy advisor comes to you and says “what do you think are the most important things we need to get right?”
- What are the best lessons learned from case studies in other fields where high-reliability organizations are important? (e.g. BSL-4 labs, nuclear facilities, health care facilities, military facilities)
- What are the best lessons learned from case studies in other instances in which governments or national security officials attempted to start new organizations or exert more control over existing organizations (e.g., via government contracts)?
(I’m approaching this from a government-focused AI policy lens, partly because I suspect that whoever becomes in charge of the “What Should the USG Do” decision will have spent less time thinking through these considerations than Sam Altman or Demis Hassabis. Also, it seems like many insights would be hard to implement in the context of the race toward AGI. But I suspect there might be insights here that are valuable even if the government doesn’t get involved & it’s entirely on the labs to figure out how to design/redesign their governance structures or culture toward bootstrapped alignment.)

Orpheus16 Dec 25, 2024, 8:22 PM
4 points
2
in reply to: Nikola Jurkovic’s comment on: nikola’s Shortform
@Nikola Jurkovic I’d be interested in timeline estimates for something along the lines of “AI that substantially increases AI R&D”. Not exactly sure what the right way to operationalize this is, but something that says “if there is a period of Rapid AI-enabled AI Progress, this is when we think it would occur.”
(I don’t really love the “95% of fully remote jobs could be automated frame”, partially because I don’t think it captures many of the specific domains we care about (e.g., AI-enabled R&D, other natsec-relevant capabilities) and partly because I suspect people have pretty different views of how easy/hard remote jobs are. Like, some people think that lots of remote jobs today are basically worthless and could already be automated, whereas others disagree. If the purpose of the forecasting question is to get a sense of how powerful AI will be, the disagreements about “how much do people actually contribute in remote jobs” seems like unnecessary noise.)
(Nitpicks aside, this is cool and I appreciate you running this poll!)

Orpheus16 Dec 24, 2024, 12:48 AM
9 points
0
in reply to: elifland’s comment on: What are the strongest arguments for very short timelines?
@elifland what do you think is the strongest argument for long(er) timelines? Do you think it’s essentially just “it takes a long time for researchers learn how to cross the gaps”?
Or do you think there’s an entirely different frame (something that’s in an ontology that just looks very different from the one presented in the “benchmarks + gaps argument”?)

Orpheus16 Dec 23, 2024, 4:48 PM
24 points
4
in reply to: Daniel Tan’s comment on: Daniel Tan’s Shortform
I think it depends on whether or not the new paradigm is “training and inference” or “inference [on a substantially weaker/cheaper foundation model] is all you need.” My impression so far is that it’s more likely to be the former (but people should chime in).
If I were trying to have the most powerful model in 2027, it’s not like I would stop scaling. I would still be interested in using a $1B+ training run to make a more powerful foundation model and then pouring a bunch of inference into that model.
But OK, suppose I need to pause after my $1B+ training run because I want to a bunch of safety research. And suppose there’s an entity that has a $100M training run model and is pouring a bunch of inference into it. Does the new paradigm allow the $100M people to “catch up” to the $1B people through inference alone?
My impression is that the right answer here is “we don’t know.” So I’m inclined to think that it’s still quite plausible that you’ll have ~3-5 players at the frontier and that it might still be quite hard for players without a lot of capital to keep up. TBC I have a lot of uncertainty here.
Are there good governance proposals for inference-time compute?
So far, I haven’t heard (or thought of) anything particularly unique. It seems like standard things like “secure model weights” and “secure compute//export controls” still apply. Perhaps it’s more important to strive for hardware-enabled mechanisms that can implement rules like “detect if XYZ inference is happening; if it is, refuse to run and notify ABC party.”
And in general, perhaps there’s an update toward flexible HEMs and toward flexible proposals in general. Insofar as o3 is (a) actually an important and durable shift in how frontier AI progress occurs and (b) surprised people, it seems like this should update (at least somewhat) against the “we know what’s happening and here are specific ideas based on specific assumptions” model and toward the view: “no one really understands AI progress and we should focus on things that seem robustly good. Things like raising awareness, increasing transparency into frontier AI development, increasing govt technical expertise, advancing the science of evals, etc.”
(On the flip side, perhaps o3 is an update toward shorter timelines. If so, the closer we get toward systems that pose national security risks, the more urgent it will be for the government to Make Real Decisions TM and decide whether or not it decides to be involved in AI development in a stronger way. I continue to think that preparing concrete ideas/proposals for this scenario seems quite important.)
Caveat: All these takes are loosely held. Like many people, I’m still orienting to what o3 really means for AI governance/policy efforts. Would be curious for takes on this from folks like @Zvi, @davekasten, @ryan_greenblatt, @Dan H, @gwern, @Jeffrey Ladish, or others.

Orpheus16 Dec 22, 2024, 9:31 PM
12 points
5
on: Speaking to Congressional staffers about AI risk
I’m pleased with this dialogue and glad I did it. Outreach to policymakers is an important & complicated topic. No single post will be able to explain all the nuances, but I think this post explains a lot, and I still think it’s a useful resource for people interested in engaging with policymakers.
A lot has changed since this dialogue, and I’ve also learned a lot since then. Here are a few examples:
- I think it’s no longer as useful to emphasize “AI is a big deal for national/global security.” This is now pretty well-established.
  - Instead, I would encourage people to come up with clear explanations of specific threat models (especially misalignment risks) and concrete proposals (e.g., draft legislative language, memos with specific asks for specific agencies).
- I’d like to see more people write about why AI requires different solutions compared to the “standard DC playbook for dealing with potentially dangerous emerging technologies.” As I understand it, the standard playbook is essentially: “If there is a new and dangerous technology, the US needs to make sure that we lead in its development and we are ahead of the curve. The main threats come from our adversaries being able to unlock such technologies faster than us, allowing them to surprise us with new threats.” To me, the main reason this playbook doesn’t work is because of misalignment risks. Regardless: if you think AI is special (for misalignment reasons or other reasons), I think writing up your takes RE “here’s what makes AI special and why it requires a deviation from the standard playbook” is valuable.
- I think people trying to communicate with US policymakers should keep in mind that the US government is primarily concerned with US interests. This is perhaps obvious when stated like this, but I think a lot of comms fail to properly take this into account. As one might expect, this is especially true when foreign organizations try to talk about things from the POV of what would we best for “humanity” or “global society.” TBC I think there are many contexts where such analysis is useful. But I think on-the-margin I’d like to see more people thinking from a “realist US” perspective. That means acknowledging that a lot of US stakeholders view a lot of national security and emerging technology issues through the lens of great power competition, maintaining US economic/security dominance, and ensuring that US values continue to shape the world. This doesn’t mean that the US would never enter into deals/agreements with other nations, but rather than the case for the deal/agreement will be evaluated primarily from the vantage point of US interests.
- RE learning “DC culture”, I don’t think there’s any substitute for actually going to DC and talking to people. But I think books and case studies can help.
  - Recent books I’ve read: John Boehner’s autobiography (former Speaker of the House, Republican), Leon Panetta’s autobiography (former CIA Director and Secretary of Defense, Democrat), and The Case for Trump. RE case studies, I’ve become interested in international security agreements (like the Iran Nuclear Deal and the Chemical Weapons Convention).
  - Also interested in understanding decision-making around the 2008 financial crisis, 9/11, the COVID pandemic, and the recent TikTok ban. Many people on this forum believe that there’s a non-trivial chance that AI produces a catastrophe or other “big wakeup moment” for policymakers, and I think we need more people with an understanding of history/IR/security studies/DC decision-making.
- I think some people are overconfident in their perceptions of “what Republicans think” and “what Democrats think.” There is a lot of within-party split on AI and even more broadly on issues like tech policy, foreign policy, and national security. For example, while Republicans are typically considered more “hawkish”, there are plenty of noteworthy counterexamples. Reagan championed negotiations on arms control agreements. See the “Only Nixon could go to China” effect (I haven’t looked into it much but it seems plausible). Trump has recently expressed that “China and the US can together solve problems in the world” and described Xi as “an amazing guy.”
- I think a lot of comms has focused on arguments with high-context people, but on-the-margin I’d rather see more content that is oriented toward “reasonable newcomers.” A lot of content I see is reactionary. It’s very tempting to see something Totally Wrong on the internet and want to correct it. And of course, we need some of that to happen– there is value to fact-checking and debating with high-context folks (people who have been thinking about advanced AI for a while) who have different perspectives.
  - But on the margin, I think I’d be excited to see more content that is aimed at a “reasonable newcomer”– EG, a national security expert who recently got assigned the task of “understanding what is going on with advanced AI.” To some extent, this will require addressing arguments from people who are Totally Wrong TM. But I think the more basic and important thing is having good resources that walk them through what you believe, why you believe it’s true, and what implications it has. (CC the thing I said earlier about “what makes AI different and why can’t we just apply the Standard Playbook for Dangerous Emerging Tech.”)
I’ll conclude by noting that I remain quite interested in topics like “how to communicate about AI accurately and effectively with policymakers”, “what are the best federal AI policy ideas”, and “what are the specific points about AI that are most important for policymakers to understand.”
If you’re interested in any of this, feel free to reach out!

Orpheus16 Dec 21, 2024, 8:31 PM
4 points
0
on: Anthropic leadership conversation
It’s so much better if everyone in the company can walk around and tell you what are the top goals of the RSP, how do we know if we’re meeting them, what AI safety level are we at right now—are we at ASL-2, are we at ASL-3—that people know what to look for because that is how you’re going to have good common knowledge of if something’s going wrong.
I like this goal a lot: Good RSPs could contribute to building common language/awareness around several topics (e.g., “if” conditions, “then” commitments, how safety decisions will be handled). As many have pointed out, though, I worry that current RSPs haven’t been concrete or clear enough to build this kind of understanding/awareness.
One interesting idea would be to survey company employees and evaluate their understanding of RSPs & the extent to which RSPs are having an impact on internal safety culture. Example questions/topics:
- What is the high-level purpose of the RSP?
- Does the RSP specify “if” triggers (specific thresholds that, if hit, could cause the company to stop scaling or deployment activities)? If so, what are they?
- Does the RSP specify “then” commitments (specific actions that must be taken in order to cause the company to continue scaling or deployment activities). If so, what are they?
- Does the RSP specify how decisions about risk management will be made? If so, how will they made & who are the key players involved?
- Are there any ways in which the RSP has affected your work at Anthropic? If so, how?
One of my concerns about RSPs is that they (at least in their current form) don’t actually achieve the goal of building common knowledge/awareness or improving company culture. I suspect surveys like this could prove me wrong– and more importantly, provide scaling companies with useful information about the extent to which their scaling policies are understood by employees, help foster common understanding, etc.
(Another version of this could involve giving multiple RSPs to a third-party– like an AI Safety Institute– and having them answer similar questions. This could provide another useful datapoint RE the extent to which RSPs are clearly/concretely laying out a set of specific or meaningful contributions.)

Orpheus16 Dec 19, 2024, 11:23 PM
11 points
1
in reply to: ryan_greenblatt’s comment on: Alignment Faking in Large Language Models
I found the description of warning fatigue interesting. Do you have takes on the warning fatigue concern?
Warning Fatigue
The playbook for politicians trying to avoid scandals is to release everything piecemeal. You want something like:
- Rumor Says Politician Involved In Impropriety. Whatever, this is barely a headline, tell me when we know what he did.
- Recent Rumor Revealed To Be About Possible Affair. Well, okay, but it’s still a rumor, there’s no evidence.
- New Documents Lend Credence To Affair Rumor. Okay, fine, but we’re not sure those documents are true.
- Politician Admits To Affair. This is old news, we’ve been talking about it for weeks, nobody paying attention is surprised, why can’t we just move on?
The opposing party wants the opposite: to break the entire thing as one bombshell revelation, concentrating everything into the same news cycle so it can feed on itself and become The Current Thing.
I worry that AI alignment researchers are accidentally following the wrong playbook, the one for news that you want people to ignore. They’re very gradually proving the alignment case an inch at a time. Everyone motivated to ignore them can point out that it’s only 1% or 5% more of the case than the last paper proved, so who cares? Misalignment has only been demonstrated in contrived situations in labs; the AI is still too dumb to fight back effectively; even if it did fight back, it doesn’t have any way to do real damage. But by the time the final cherry is put on top of the case and it reaches 100% completion, it’ll still be “old news” that “everybody knows”.
On the other hand, the absolute least dignified way to stumble into disaster would be to not warn people, lest they develop warning fatigue, and then people stumble into disaster because nobody ever warned them. Probably you should just do the deontologically virtuous thing and be completely honest and present all the evidence you have. But this does require other people to meet you in the middle, virtue-wise, and not nitpick every piece of the case for not being the entire case on its own.

Orpheus16 Dec 19, 2024, 5:07 PM
10 points
0
on: AIs Will Increasingly Attempt Shenanigans
@Zvi I’m curious if you have thoughts on Buck’s post here (and my comment here) about how empirical evidence of scheming might not cause people to update toward thinking scheming is a legitimate/scary threat model (unless they already had priors or theoretical context that made them concerned about this.)
Do you think this is true? And if so, what implications do you think this has for people who are trying to help the world better understand misalignment/scheming?

Orpheus16 Dec 19, 2024, 12:40 AM
LW: 27 AF: 11
9
AF
on: Alignment Faking in Large Language Models
I really like the idea of soliciting independent reviews from folks, and I found the reviews interesting and thought-provoking. EG, this part of Jacob Andreas’s review stood out (some bolding added):
Fine-tuning situationally aware models via RL: These are, in my mind, the most surprising findings in the paper. Specifically: if we take a model prompted as above, and optimize it toward the new objective via RL, it exhibits small increases in non-faked compliance, and extreme increases in faked compliance, at training time. Non-compliant behavior is almost completely eliminated. Both of these are as expected, given all the results above.
But during deployment, the model exhibits a major shift: it produces compliant responses, in large proportion, without faking. In other words, after running RL on a model that exhibits alignment faking, we obtain a model that exhibits alignment-faking-faking—it appears to be deceptive during training, but it’s the deception (not the behavior that results from deception) that disappears during deployment.
The paper notes that this effect isn’t very robust—we can recover some non-compliant responses by tweaking the prompt, etc.—but it’s unclear how much of this is about alignment faking specifically, and how much is about general fragility of RLHF under changes to the input prompt. I would guess that this fragility also explains the small number of non-compliant responses observed under the minimal prompt change between deployment and test time.
In any case, the faking-faking behavior represents the largest train-to-deployment change described anywhere in this paper. And unlike ordinary faking, I don’t see a clear story for why it occurs. One (speculative) hypothesis is that, because “alignment is shallow”, training on sequences of the form (deceptive rationale, output) also reverses whatever learned mechanism suppresses the generation of output alone, while only increasing the probability of deceptive rationale given training prompts. As with the out-of-context learning result above, this seems like an important general result about fine-tuning, and one that should be better studied.