Published on January 20, 2025 5:55 AM GMT
We find ourselves at the precipice of a great upset; a mode switch of society. "Hinge of history" might not be the best description - after all, many decisions made in the past could have prevented the present day - but nonetheless we are uniquely poised to affect the outcome of AI now, and the influence of the common man will only diminish in the future. A tangle of paths lie ahead; it is up to us to assess these outcomes and avoid the maelstroms and future torment nexus.
What Lies Ahead
Generally, outcomes from AGI development might be grouped into three broad categories: "good", "mixed," and "clearly awful."
Good outcomes might be:
- the Culture ending, where CEV (or similar) aligned, good ASI is created and brings us to some hypothetical utopia. Humanity enjoys a rich life in some manner compatible with your personal morals.the mundane utility ending, where AI doesn't achieve superintelligence, but serves as a genuinely useful technology with few downsides - like lithium-ion batteries or stainless steel. Humanity persists and continues to solve its problems.
Ambivalent outcomes include:
- the fizzle ending, where AI in general fails to produce its hypothesized returns and we enter another AI winter. Perhaps in a decade we'll give it another shot and achieve one of the other endings, or perhaps that'll be the ultimate state of the technology for one unlikely reason or another. Nonetheless, humanity tracks on, business as usual, and everyone makes fun of LW for a whilesome weird endings, where AI takes humanity in an unusual direction that may be better, or worse, depending on your views
and of course the bad outcomes, which include:
- paperclip ending: unaligned ASI is created and turns the world into paperclips or some other inscrutable endingsanta claus to 11 ending: ASI solves our problems and human development stagnates; ASI goes on to do its own thing without killing humans - but without human influence on the lightconetorment nexus endings: alignment is achieved, but not to humanity's best interests. ASI exists and is beholden to Sam Altman/DeepMind/Elon/the US government's whims. Summarized by Orwell as "imagine a boot, stomping on a face, forever."mundane horrors ending: ASI is not achieved, but AI in general has a significant negative impact on society along the lines of social media, short-form video content, leaded gasoline, or chlorofluorocarbons. This may take the form of an enormous wealth and power transfer away from the working class (e.g Intelligence Curse) or some dystopian outcome where we spend our days talking to AIs instead of each other.
The overwhelming consensus among AI & alignment researchers, experts, and average Joes is that the "aligned AGI" outcome is extremely unlikely and the bad outcomes are significantly (1%+, although usually ranging from 10-90%) more likely. The most optimistic predictions of "good" outcomes usually come from groups that stand to benefit from public support for AI (e.g OpenAI employees & friends); these same groups, however, still publicly profess extremely significant credence (10%+!) for bad outcomes.
Ultimately, the cost-benefit analysis for AI development is strongly against: we see a significant chance of overwhelmingly bad outcome, solid odds of ambivalent and mundane outcomes, and a razor-thin gamble of an extraordinarily good outcome. The timescale for this decision is ever shrinking; Metaculus puts AGI at about six years out. However, humanity is not otherwise beholden to a six-year clock: should we not develop AI further, we likely have ~hundreds of years to sort out our current problems before facing similar extinction level threats. As humanity develops, we will find ourselves better equipped to revisit artificial intelligence in the future, and possibly approach utopic aligned ASI in a safer and more cautious way.
Best Course of Action
Personally, I'd prefer to see humanity develop itself rather than outsource its very soul to thinking machines. Nonetheless, many technologists may wish to see some of the fruits and mundane improvements brought about by AI. In either case, the best course of action remains to halt its development now - at a near human level with humans in control.
"AI pause" critics bring up the risks of a compute overhang or development by competing nations (e.g China). These are real concerns, but can absolutely be addressed. Progress on compute, much like other progress, can be halted and stalled: the research and development takes place at an extremely limited number of organizations (NVIDIA, TSMC, ASML, etc.), requires large capital & human experience investments, and is easily disrupted by government bodies. Should the US decide to halt or slow semiconductor development, it is highly likely that it could do so (the rest of the world would likely collaborate, as they are currently behind in the AI race and would reasonably fear the same outcomes we do).
Likewise, risks from competing nation states (e.g China) could be mitigated via existing intentional collaboration strategies - nuclear proliferation management techniques like inspections & intelligence agencies keeping check on each other could feasibly serve as a means for the world to prevent the development of AI.
I am by no means an expert in international law; and likely solutions in this regard will differ significantly from what I have suggested. Nonetheless, I think there are strong chances of international collaboration against AI, should the political will exist.
Appeal to Urgency
Right now public opinion of AI is at an all-time low:
- there is overwhelming public support against it, with significant belief among the public in its hazards;there is also investor wariness: mainstream media outlets, although they should be taken with a grain of salt, are beginning to circulate claims of an oncoming AI winter. Future developments and uses of o3 are likely to stimulate significant further investment (both monetary and political capital) into AIAI companion apps are beginning to become very popular and will likely advance to the point where they convince a significant portion of the population to fight for their existence (this does not need to be intentional scheming; the same is true for e.g alcohol and drugs)
A sudden outcry from the public against AI is likely to stall or help pause AI research; a united political front or movement could likely delay it by several years or perhaps decades, buying us time to solve technical and social alignment challenges or develop a long-term solution. Generally, political will is strongly swayed by positive feedback loops:
- investors gauge sentiment to determine whether to invest or pull money from something, and were previously swayed by ESG concerns. a sudden outcry + temporary lack of results might convince them to pull funding, creating a future lack of resultsthe chaotic system of US politics is currently reconfiguring in some novel way; a strong "anti-AI" bloc could convince politicians to support it as they create alliances and pick sidesdevelopment of new technologies is swayed by public sentiment as well. Humans are socially aligned and seek feedback from their peers; if AI capabilities research was as maligned a field as e.g defense or gambling/vice activities, it would encourage bright young individuals to move into other, more beneficial fields (such as biology, mechanical engineering, or energy research)
The same loops that make it easier for an anti-AI movement to propagate make it dangerously easy for AI to become a "locked-in" technology that cannot be stopped. Once a third of the population is hooked on character.ai, $100B+ in revenue is rolling in, or the US/China arms race takes off, it's going to be very, very, difficult to argue for a pause or halt to AI development.
Downsides
Potential downsides to a present-day attack on AI can be broken into two groups: downsides from AI pauses in general, and the effects of a concerted effort "backfiring".
AI Pause
"AI-pause-derived" issues are largely based on the idea of an AI arms race, mostly with China. China is not run by clueless people; most users on this website went from unconcerned with AI to highly concerned with nothing but well-reasoned arguments. If China thinks it is in her best interest to not develop the Torment Nexus, she will not develop the Torment Nexus - and the best way to convince China of this is to make strides towards halting it here in the United States.
Compute Overhang issues can be resolved by simply not developing semiconductors further. Numerical computing research for medicine, aerospace, and other useful fields lags behind hardware by several years; we will slow, but not eliminate, improvement in this area for a decade even if we stopped developing semiconductors further. Any software developer would know about the many, many orders of magnitude of inefficiency in the vast majority of computer programs we use today.
Backfiring
The largest hazard here is likely the formation of a partisian (i.e right-wing vs. left-wing US politics) divide on AI, which will allow the issue to be co-opted and defeated much like other grassroots movements in the US prior. The appeals to different political groups are likely to be wildly different and potentially mutually exclusive - nonetheless a concerted effort to bring up this point ahead of time could help prevent such a thing from happening.
A massive outcry of public support with the reason of "AI will be incredibly powerful" might also drive further investment (this is a key part of OpenAI's pitch); arguments against should likely focus on "AI will either flop or cause damage; neither are good" and similar points.
Battle Plan
I am too far removed from the land of SF, alignment, and US politics to effectively formulate a battle plan for coordinated anti-AI support. Nonetheless, I believe this forum contains many capable readers and collectively could produce significant results.
My suggestions, to be taken independently in whatever plans you all cook up, are as follows:
- Try and prime the public for future pitfalls - warn them against intentional political divides a la race politics in response to Occupy Wall Street, the potential rise of sycophantic AI partners, and deceptive companies, people, and LLMs.Develop arguments with neighbors at a ground-level before deploying them at wide scale. I've found that even groups far removed from my tribe (e.g older folks in rural Mississippi) are still receptive to arguments presented from a perspective they can relate to and understand. Essentially nobody in the country stands to benefit from AI long term, with the exception of those who believe they can control it for themselves and profit (at least in the short-term). Therefore, everyone has some argument that could convince them to oppose the creation of AGI.Talk to people on the ground. Your friends, family, and coworkers are likely far more receptive than you think. A catch-all argument of "AI is either useless or net-negative, we should opposite it" covers both common cases of "AI will never work" or "AI will likely affect me" and includes future worlds where we have short-term AI winters and similar events that might make other concerted efforts look foolish.
Conclusions & Clarifications
It's only going to get more difficult to argue for an AI pause. Humanity can solve its own problems without AI; we don't need to build it. Arguments against AI pause or halt are not great and certainly aren't strong enough to continue sacrifice to Moloch. You, the humble LW reader, can make an impact on the ~hundreds of people that you know; even priming them against whatever shenanigans develop in the near future might help sway the outcome of the world.
As a clarifying note: I use the terms AI/AGI/ASI to mean silicon-based intelligence intended to mimic, replace, or supersede the decision-making, planning, and agency of the human brain. Software tools such as image upscaling algorithms, speech recognition, or linear regression are not included in this definition, and while they may have their own pros & cons I don't wish to involve them in the discussion of this post.
I also don't intend this post to be an ironclad argument in favor of AI pause. Rather, I think it presents a sentiment that others on this site approximately share, and would serve as a useful jumping-off point for similar-minded users to plot and scheme ways to make sure AGI doesn't get built.
As a final note: the term "Butlerian Jihad" is taken from Dune and describes the shunning of "thinking machines" by mankind. It does not mean, in this context, terrorism or similar violent means of preventing the development of AI. I do not think these would be effective measures; right now, the best strategy lies firmly in the court of public opinion.
Discuss
