China has made several efforts to preserve their chip access, including smuggling, buying chips that are just under the legal limit of performance, and investing in their domestic chip industry.[1]
This post centres around an email I sent to the Center for AI Safety (CAIS) expressing concern about their 2023-08-15 newsletter's coverage of US-China competition in the AI space[2], but the overall point is broader. There are some ways of discussing the topic of international relations regarding AI which strike me as un-nuanced in a counterproductive and dangerous way, by hiding certain truths or emphasising others, and supporting a conflict-oriented mindset.
In writing about this, I'm also gesturing at something about the more general topic of 'how to think and write about politically-charged topics'.
Jump to the summary if you are in a hurry.
This conversation really is important, which is why I think it's worth a public message discussing particular statements, but this should be understood as constructive criticism and part of a broader conversation which society, and especially the community of those focused on AI safety, needs to have. The CAIS newsletters are worth a (not unquestioning!) read, including the edition in question.
The particulars in this message serve as good exemplars of the problems and questions I have, and I'd be interested in responses from CAIS but even more so in remarks more broadly on the topic from anyone interested. The public conversation about this appears from my perspective to sometimes be broken. If that is so, I would like it to be rectified, and if not, I would like to be put right myself, the better to prioritise in my own work!
Specifics, CAIS case study
Here, quoted[3], is my response to the CAIS letter, which serves as an initial dialogue opener and the core of this post:
Hello,
I've been a supportive reader for some time and am myself an AI safety researcher. I'm generally very impressed and encouraged by your newsletters! - but I was disappointed and concerned by the phrasing (and mindset it can encourage) regarding US-China competition in the letter dated 16th August ('US-China Competition on AI Chips, ...').
In general I'm very wary of messaging which could inflame a them-vs-us mindset, in short mainly because I think it a) destroys humans' ability to think sensibly and b) tends to foreclose win-win outcomes. I expect these brief points to be clear and to have rich referents in your mental pictures of the world, but please correct me if not!
I think your letter skirted close to dangerously simplified presentation in this way. I would not normally spend my time or yours on a criticism of one section of a newsletter, but in this case I consider it worthwhile because your letter is close to the Pareto frontier on nuance, correctness, helpfulness, and reach (and it pays to try to nudge such things in good directions), but this kind of message needs to be delivered with care to avoid misunderstanding and harm.
Hopefully my pointing this out, accompanied by a few select quotes, is enough to encourage you to carefully take this criticism into account, but I'd happily expand more if you like!
Without further ado, a few quotes and my response:
The US and China have been competing for access to these chips for years.
Kind of true, but really US-based and China-based international corporations (as well as other orgs) have sought access to this scarce resource. Competition for this particular resource is mostly zero-sum across all of these entities, importantly including intra-US and intra-China.
Where market-share is near zero-sum (i.e. for direct competitors) the market-share outcomes may be greater-than-linearly zero-sum in this resource (your minor/temporary lack of chips could be my major gain of market-share), which might better warrant the term 'competing', but this effect is actually much stronger intra-country/bloc rather than inter-, due to respective markets! i.e. Google and Microsoft really care about each other's chip access in a way that they only do to a weaker degree about Alibaba's.
Interjection: To emphasise, 'the US and China have been competing' doesn't literally preclude belief in intra-bloc competition. But there's a strong implicature that intra-bloc is (relatively) unimportant, while in fact the mechanism I mentioned here increases intra-bloc competition (which I think is borne out by observation to date).
When governments have paid attention, they have indeed made moves which adjust share (but also supply, as you've noted later, making it nonzero-sum, in chips at least). It's unclear (to me) exactly what incentives have motivated each move, but certainly they're not the actions of monolithic or coherent entities 'The US' and 'China'. And it's certainly not the case that where such activity changes chip share, it's collected by the acting entity. Non-governments have also made moves adjusting chip share, for example the case you cite of Nvidia (a 'US' company) deliberately rules-lawyering the US gov in order to supply more chips to various China-based companies!
Typically when nations are used as agentic subject nouns it refers to the government and/or military of said country. I don't think there's a reading of these statements in those terms which is true, and I'm not aware of any other plausible reading which is true.
China has made several efforts to preserve their chip access, including smuggling, buying chips that are just under the legal limit of performance, and investing in their domestic chip industry.
I dislike this sentence and think it is false! Who is this 'China'? Did said unified entity carry out all of these activities? Was it coherently pursuing all of these 'several efforts' to some particular end?
Interjection: I feel that I was unkind in my tone here. The kind of claim exemplified in the letter and in other places has proved impossible for me to map to something sincerely resembling reality without caveating so much as to be essentially starting from scratch i.e. it looks like a kind of non-proposition or emotive filler. I would be very interested to hear from people who have a better parse on this to help me understand! My crux discussion below the rest of the email contains my current best attempts.
Meanwhile, the United States has struggled to build American chip manufacturing capacity, and has taken further steps to prevent Americans from investing in Chinese technology.
This one is poor for the same reasons, though not quite as bad (perhaps because the authors are American/anglophone and have a closer perspective on the nuance).
The discussion after this point in the letter is relatively good and nuanced! It names (some of) the individual orgs and companies, and makes clearer the multiplicity of others. All further references to nations as agentic subject nouns appear to be consistent with a conventional reading referring to the respective governments.
I'm interested to know how the rather good detail got paired with a rather harmful introduction and I urge you to consider the processes and thinking which gave rise to this section of the otherwise good letter.
Thanks,
Oly
This was a brief email intended to convey something I expected to be quickly-graspable with a few pointers. CAIS content suggests that they have a broader familiarity with associated facts, but it seems to be digested/compressed here in a way which is needlessly and harmfully lossy.
In particular, the statements abstract very neatly over pre-drawn boundaries (national i.e. 'US' and 'China') and furthermore assign a greater sense of coherence and agency to those abstractions than is warranted. At least some possible such statements must be true-ish (or we would not have those abstractions), but this convenient compression happens in too many conversations to be a coincidence! Said pre-existing abstraction boundaries are already salient in the information ecosystem, and laden with emotional and political baggage. This same phenomenon (it sometimes seems like a pre-written bottom line but in implicature?) appears in other publications by other orgs and in verbal conversations I've witnessed or been part of.
The fact that I, a relative governance rookie (I'm focused mainly on technical matters), struggle to rectify or understand this makes me wonder: am I missing something? Is there a relevant factor I'm unaware of? More concerningly, is there some terrible equilibrium which prevents more involved people from speaking more clearly here? I think more likely the abstractions ('US' and 'China') have a background potency which distorts perceptions and shapes how people communicate.
Possible cruxes and areas of high uncertainty
Implicit in my own discussion is the background assumption that the main concern is about possible direct or near-direct impacts[4] of AI deployments by government entities (or military). This seems to me the obvious reading when people use countries as agentic subject nouns. ('China [the governmental entity] has made several efforts'[5].) In this framing, I can't rectify some of the things people say with reality. But some alternative concerns might fit the bill, and if so, this is evidence that people are talking past each other, and we should aim to frame concerns more clearly!
I've never focused intently on this area, but have had a handful of conversations about this over the years, and among relevant cruxes seems to be a family of questions along the lines of
How quickly/totally/coherently could US gov/CCP capture AI talent/artefacts/compute within its jurisdiction and redirect them toward excludable destructive ends? Under what circumstances would they want/be able to do that?
People's intuitions here appear to differ a lot, and data might be hard to come by!
It seems plain that nations are not currently meaningful players in AI development and deployment, absent conspiracy-level secrecy. So to support the apparent take that they are, we may need to imagine that they could ably/rapidly become meaningful players in AI development and deployment, hence the above cruxes.
Depending on the answers to these questions, one might perceive various goings-on which happen to occur under one or other jurisdiction to have greater import on the international stage and perhaps to warrant treating national or multi-national blocs as more coherent entities than they really are at present, for the purposes of AI discussion. ('China [the government] has [allowed/encouraged made] several efforts [because eventually they will probably seize the gains/means]'[7].)
Other possible cruxes, more guesswork:
Perhaps the concern is about indirect (e.g. economic) impacts of non-government entities' AI activities leading to some (risky) change in balance of power (between existing governments/blocs)
Then, abstracting references to lots of individuals and groups via their home country might be a move which is writer-intuitive, even if nonstandard and reader-confusing. ('China [the impersonal collective economic entity] has made several efforts'[8].)
The additional step in this type of theory, namely that indirect effects cause a risky change in balance of power, should really be spelled out if it is loadbearing
The use of countries as agentic subject nouns is difficult to justify under this reading
Perhaps the concern is indeed about direct impacts, but wielded by non-government entities (who remain the major players in development and deployment of AI)
If so, the conversation should be about general/global resource/capabilities rather than inter-bloc 'competition'...
...unless we also posit that inter-bloc non-government conflict is liable to be much worse than intra-bloc[9]
Similarly, if these are loadbearing assumptions, they really ought to be spelled out clearly
The use of countries as agentic subject nouns could be justified at a stretch here, but only by first spelling out the reasoning
Perhaps the concern is that, regardless of the actual impact of AI resources, apparent competition could lead to inflammation of traditional conflict, or weaken defenses against such inflammation
Then, reporting on the apparent competition via a mention and with explicit caveat, would make sense! ('China [gov/military] has [been perceived as having] made several efforts... [but the reality is more nuanced]'[10].)
Alternatively, reporting on actual conflict could be used as evidence for the claim (that apparent competition inflames conflict), but only by also pointing to the stated or implied reasons for the conflict. ('China has made several efforts... [in each case citing US provocation as justification]'[11].)
In either case, there are additional claims being made that can not be left implicit, and require supporting argument
Speculation
As it is, for me, the evidence seems to suggest that an AI race, if it is happening at all, is being run by (mainly US- and UK-based) companies with little or no oversight from governments or militaries. Rather, governments are in a position to collectively act to diffuse the race! And they appear as likely to do this as to exacerbate it, from my limited viewpoint.
Separately, a lack of reliable alignment techniques and performance guarantees makes AI-powered belligerent national interest plays look more like bioweapons than like nukes - i.e. minimally-excludable - and perhaps mutually-knowably so! This presently damps the incentive to go after them. But proliferation of naively-aligned AI ('figure out what I want and make it happen') might make harm plays more excludable, exacerbating lose-lose or race game dynamics ('go and steal/destroy their stuff but don't let that happen to my stuff'). This concern in part motivates consideration of multi-principal-multi-agent delegation and the cooperative AI agenda.
Summary and takeaways
Un-nuanced coverage and discussion has the potential to inflame harmful confusion and us-vs-them mentality, which diminish the chance of safe outcomes.
China has made several efforts to preserve their chip access, including smuggling, buying chips that are just under the legal limit of performance, and investing in their domestic chip industry.
China [the governmental entity] has made several efforts...[5:1]
China [the government] has [allowed/encouraged made] several efforts [because eventually they will probably seize the gains/means]...[7:1]
China [the impersonal collective economic entity] has made several efforts...[8:1]
[People and organisations in] China [have has] made several efforts...[8:2]
China [gov/military] has [been perceived as having] made several efforts... [but the reality is more nuanced] [10:1]
China has made several efforts... [in each case citing US provocation as justification] [11:1]
...?
When compressing discussion of political topics, be extra wary of compression which coincidentally abstracts over already-charged us-them divides (and be careful when phrasing comes too easily, lest you write bottom lines first)! You're more likely to be wrong (because your information ecosystem biases toward thinking in these terms, and because you might be mildly-to-severely mind-killed on the matter), and being wrong is more likely to be harmful (by reinforcing those dynamics in others).[12] The same vigilance applies to reading and listening.
My (not very informed) take is that governments are at this point as likely to want to defuse as to exacerbate an AI race, and those of us with any privileged insight or influence should avoid one-sided discussion of the matter (if anything preferring to focus on constructive, collaborative possibilities, the better to raise them to salience and generate common knowledge).
Most of my remarks here are somewhat weakly held (if forcefully stated) and it seems important to gather perspectives on this. Inform me! All responses will be gratefully received.
I'm supportive of some of CAIS' work, and the content of their newsletters (they have impressive breadth), and theirs is far from the only outfit which appears to produce confused or confusing messaging on the topic of US-China competition. In fact they seem to be better than many! ↩︎
My response is at quote level 1. Excerpts from the CAIS letter are within, at quote level 2. I interject a little for the purposes of this post, without quotation. ↩︎
e.g. deployment for weapons control or for offensive R&D (bio, materials, ...) ↩︎
'China [the governmental entity] has made several efforts' is a fairly standard use of language; governments are at least somewhat coherent and also do things with consequences and subsequent plans 'in mind'. This sentence has the disadvantage of being baldly false, though[6]. ↩︎↩︎
(unless we posit a near-hivemind coherence to the people of China, which is absurd and obscene) ↩︎
'China [the government] has [allowed/encouraged made] several efforts [because eventually they will probably seize the gains/means]' is a big stretch of the language, but at least somewhat consistent. To support this reading, though, there's a substantial additional claim that needs to be justified. ↩︎↩︎
'China [the impersonal collective economic entity] has made several efforts' would be a rather nonstandard use of language; economies of billion+ people do not make 'efforts' with consequences or subsequent plans. Leaving aside the implicature of agency, though, this sentence is a closer fit to reality. '[People and organisations in] China [have has] made several efforts' would be even better. ↩︎↩︎↩︎
Should conflict between non-territorial entities be worse for inter-bloc than intra-bloc? I think the point I made previously about zero-sum market-share competition suggests the opposite. But humans' destructive jingoistic/xenophobic tendencies are real, and a point in favour. ↩︎
'China [gov/military] has [been perceived as having] made several efforts... [but the reality is more nuanced]' is in large part one of the messages of this post! I don't think the original letter in question can have meant this, but I do maintain it as a hypothesis for the more general case of compressed discussion of political things. People are often imagining third-party reactions when discussing political topics, and sometimes use-mention distinctions fail to come across. ↩︎↩︎
'China has made several efforts... [in each case citing US provocation as justification]' is something that could legitimately be said, and has a clear meaning, even if in this particular case it is false if we understand 'China' to be the CCP or military. ↩︎↩︎
I'd tentatively go further and suggest you ought to train yourself to be appalled when you catch yourself doing this without justification because only then do you stand a chance of thinking clearly about politics. ↩︎
Sounds about right?
This post centres around an email I sent to the Center for AI Safety (CAIS) expressing concern about their 2023-08-15 newsletter's coverage of US-China competition in the AI space[2], but the overall point is broader. There are some ways of discussing the topic of international relations regarding AI which strike me as un-nuanced in a counterproductive and dangerous way, by hiding certain truths or emphasising others, and supporting a conflict-oriented mindset.
In writing about this, I'm also gesturing at something about the more general topic of 'how to think and write about politically-charged topics'.
Jump to the summary if you are in a hurry.
This conversation really is important, which is why I think it's worth a public message discussing particular statements, but this should be understood as constructive criticism and part of a broader conversation which society, and especially the community of those focused on AI safety, needs to have. The CAIS newsletters are worth a (not unquestioning!) read, including the edition in question.
The particulars in this message serve as good exemplars of the problems and questions I have, and I'd be interested in responses from CAIS but even more so in remarks more broadly on the topic from anyone interested. The public conversation about this appears from my perspective to sometimes be broken. If that is so, I would like it to be rectified, and if not, I would like to be put right myself, the better to prioritise in my own work!
Specifics, CAIS case study
Here, quoted[3], is my response to the CAIS letter, which serves as an initial dialogue opener and the core of this post:
Interjection: I'm referring to the vicinity of mind-killing politics
Interjection: To emphasise, 'the US and China have been competing' doesn't literally preclude belief in intra-bloc competition. But there's a strong implicature that intra-bloc is (relatively) unimportant, while in fact the mechanism I mentioned here increases intra-bloc competition (which I think is borne out by observation to date).
Interjection: I feel that I was unkind in my tone here. The kind of claim exemplified in the letter and in other places has proved impossible for me to map to something sincerely resembling reality without caveating so much as to be essentially starting from scratch i.e. it looks like a kind of non-proposition or emotive filler. I would be very interested to hear from people who have a better parse on this to help me understand! My crux discussion below the rest of the email contains my current best attempts.
This was a brief email intended to convey something I expected to be quickly-graspable with a few pointers. CAIS content suggests that they have a broader familiarity with associated facts, but it seems to be digested/compressed here in a way which is needlessly and harmfully lossy.
In particular, the statements abstract very neatly over pre-drawn boundaries (national i.e. 'US' and 'China') and furthermore assign a greater sense of coherence and agency to those abstractions than is warranted. At least some possible such statements must be true-ish (or we would not have those abstractions), but this convenient compression happens in too many conversations to be a coincidence! Said pre-existing abstraction boundaries are already salient in the information ecosystem, and laden with emotional and political baggage. This same phenomenon (it sometimes seems like a pre-written bottom line but in implicature?) appears in other publications by other orgs and in verbal conversations I've witnessed or been part of.
The fact that I, a relative governance rookie (I'm focused mainly on technical matters), struggle to rectify or understand this makes me wonder: am I missing something? Is there a relevant factor I'm unaware of? More concerningly, is there some terrible equilibrium which prevents more involved people from speaking more clearly here? I think more likely the abstractions ('US' and 'China') have a background potency which distorts perceptions and shapes how people communicate.
Possible cruxes and areas of high uncertainty
Implicit in my own discussion is the background assumption that the main concern is about possible direct or near-direct impacts[4] of AI deployments by government entities (or military). This seems to me the obvious reading when people use countries as agentic subject nouns. ('China [the governmental entity] has made several efforts'[5].) In this framing, I can't rectify some of the things people say with reality. But some alternative concerns might fit the bill, and if so, this is evidence that people are talking past each other, and we should aim to frame concerns more clearly!
I've never focused intently on this area, but have had a handful of conversations about this over the years, and among relevant cruxes seems to be a family of questions along the lines of
People's intuitions here appear to differ a lot, and data might be hard to come by!
It seems plain that nations are not currently meaningful players in AI development and deployment, absent conspiracy-level secrecy. So to support the apparent take that they are, we may need to imagine that they could ably/rapidly become meaningful players in AI development and deployment, hence the above cruxes.
Depending on the answers to these questions, one might perceive various goings-on which happen to occur under one or other jurisdiction to have greater import on the international stage and perhaps to warrant treating national or multi-national blocs as more coherent entities than they really are at present, for the purposes of AI discussion. ('China [the government] has [allowed/encouraged
made] several efforts [because eventually they will probably seize the gains/means]'[7].)Other possible cruxes, more guesswork:
Speculation
As it is, for me, the evidence seems to suggest that an AI race, if it is happening at all, is being run by (mainly US- and UK-based) companies with little or no oversight from governments or militaries. Rather, governments are in a position to collectively act to diffuse the race! And they appear as likely to do this as to exacerbate it, from my limited viewpoint.
Separately, a lack of reliable alignment techniques and performance guarantees makes AI-powered belligerent national interest plays look more like bioweapons than like nukes - i.e. minimally-excludable - and perhaps mutually-knowably so! This presently damps the incentive to go after them. But proliferation of naively-aligned AI ('figure out what I want and make it happen') might make harm plays more excludable, exacerbating lose-lose or race game dynamics ('go and steal/destroy their stuff but don't let that happen to my stuff'). This concern in part motivates consideration of multi-principal-multi-agent delegation and the cooperative AI agenda.
Summary and takeaways
Un-nuanced coverage and discussion has the potential to inflame harmful confusion and us-vs-them mentality, which diminish the chance of safe outcomes.
When compressing discussion of political topics, be extra wary of compression which coincidentally abstracts over already-charged us-them divides (and be careful when phrasing comes too easily, lest you write bottom lines first)! You're more likely to be wrong (because your information ecosystem biases toward thinking in these terms, and because you might be mildly-to-severely mind-killed on the matter), and being wrong is more likely to be harmful (by reinforcing those dynamics in others).[12] The same vigilance applies to reading and listening.
My (not very informed) take is that governments are at this point as likely to want to defuse as to exacerbate an AI race, and those of us with any privileged insight or influence should avoid one-sided discussion of the matter (if anything preferring to focus on constructive, collaborative possibilities, the better to raise them to salience and generate common knowledge).
Most of my remarks here are somewhat weakly held (if forcefully stated) and it seems important to gather perspectives on this. Inform me! All responses will be gratefully received.
Center for AI Safety, AI Safety Newsletter #19, 2023-08-15 ↩︎
I'm supportive of some of CAIS' work, and the content of their newsletters (they have impressive breadth), and theirs is far from the only outfit which appears to produce confused or confusing messaging on the topic of US-China competition. In fact they seem to be better than many! ↩︎
My response is at quote level 1. Excerpts from the CAIS letter are within, at quote level 2. I interject a little for the purposes of this post, without quotation. ↩︎
e.g. deployment for weapons control or for offensive R&D (bio, materials, ...) ↩︎
'China [the governmental entity] has made several efforts' is a fairly standard use of language; governments are at least somewhat coherent and also do things with consequences and subsequent plans 'in mind'. This sentence has the disadvantage of being baldly false, though[6]. ↩︎ ↩︎
(unless we posit a near-hivemind coherence to the people of China, which is absurd and obscene) ↩︎
'China [the government] has [allowed/encouraged
made] several efforts [because eventually they will probably seize the gains/means]' is a big stretch of the language, but at least somewhat consistent. To support this reading, though, there's a substantial additional claim that needs to be justified. ↩︎ ↩︎'China [the impersonal collective economic entity] has made several efforts' would be a rather nonstandard use of language; economies of billion+ people do not make 'efforts' with consequences or subsequent plans. Leaving aside the implicature of agency, though, this sentence is a closer fit to reality. '[People and organisations in] China [have
has] made several efforts' would be even better. ↩︎ ↩︎ ↩︎Should conflict between non-territorial entities be worse for inter-bloc than intra-bloc? I think the point I made previously about zero-sum market-share competition suggests the opposite. But humans' destructive jingoistic/xenophobic tendencies are real, and a point in favour. ↩︎
'China [gov/military] has [been perceived as having] made several efforts... [but the reality is more nuanced]' is in large part one of the messages of this post! I don't think the original letter in question can have meant this, but I do maintain it as a hypothesis for the more general case of compressed discussion of political things. People are often imagining third-party reactions when discussing political topics, and sometimes use-mention distinctions fail to come across. ↩︎ ↩︎
'China has made several efforts... [in each case citing US provocation as justification]' is something that could legitimately be said, and has a clear meaning, even if in this particular case it is false if we understand 'China' to be the CCP or military. ↩︎ ↩︎
I'd tentatively go further and suggest you ought to train yourself to be appalled when you catch yourself doing this without justification because only then do you stand a chance of thinking clearly about politics. ↩︎