don’t forget the political environment:
- locally, there’s a meaningful “break up big tech” current which could make it politically difficult to simultaneously sell AI as a paradigm shift and monopolize it for yourself via the legal apparatus. cynically, firms might view regulation as a path to achieve similar ends but with fewer political repercussions, less blatant than if they leveraged patents.
- globally, the country which presently enjoys the lead in AI sees itself in an economic battle against a competitor unlikely to respect its intellectual property claims. to the degree which states view AI through any lens related of “national defense”, there will be some push to maintain competitiveness at least on the global stage.
ponkaloupe
Karma: 126
-
it’s odd to leap to things like housing markets and consumer debt without considering the demographics of startup employees. i believe your graphs are national averages, so are these employees expected to hold more or less debt relative to average? more or less likely to be homeowners v.s. renters? more or less likely to live in specific regions of the country?
-
the initial shock of covid 3.5 years ago was just massive. i get that it was in many ways transformative and not strictly destructive, but still hypotheticals like “a hundred billion decrease in VC funding” just seem so miniscule in comparison. simultaneously we see how the impacts of a sharp shock got dispersed pretty far across time with covid, and this VC bubble popping isn’t nearly as sharp a shock as we’ve known (call it 18mo, based on those burn rates, vs 2mo over which covid hit the whole country).
-
cascading failures are notoriously difficult to predict. seems to me the real worry is not the title of this post but that the systems which have arrested cascading failures may be eroding. good for bringing up national debt, actually, would be interesting to just embrace this fully and consider food/energy security & geopolitics — but that would make for a pretty different piece.
-
It seems mostly correct to accept the new calculations in the Improved COTI, which represent a −25% adjustment, and then include the 13% adjustment for taxes, resulting in about a −13% adjustment. This still represents an increase in the cost of thriving.
is COTI actually an inverted measure of the literal “cost of thriving”? i.e. the index goes up when the cost goes down? otherwise, this apparent inverted sign (a −13% change in COTI representing an “increase in the cost of thriving”) is throwing me for a loop.
In broad terms, families with children have seen large reductions in their federal income tax burden, largely due to the introduction and expansion of the child tax credit.
If we want to measure the COTI, as per its original justification, it seems correct to more or less accept the ‘improved COTI’ of −25% instead of −36%, reflecting the errors in health care premiums and college sticker versus effective prices, and various minor fixes. We then must take taxes into account, which should leave us with about a −13% change from 1985 to 2023.
here it happens again: tax costs have decreased since 1985, but somehow that manifests as an increase to the “cost of thriving” index (-25% —> −13%).
To learn gravity, you need additional evidence or context; to learn that the world is 3D, you need to see movement. To understand that movement, you have to understand how light moves, etc. etc.
for the 3d part: either the object of observation needs to move, or the observer needs to move: these are equivalent statements due to symmetry. consider two 2D images taken simultaneously from different points of observation: this provides the same information relevant here as were there to be but 2 images of a moving object from a stationary observer at slightly different moments in time.
in fact then, you don’t need to see movement in order to learn that the world is 3D. making movement a requirement to discover the dimensionality of a space mandates the additional dimension of time: how then could we discover the 4 dimensional space-time without access to some 5th dimensional analog of time? it’s an infinite regress.
similarly, you don’t need to understand the movement of light. certainly, we didn’t for a very long time. you just need to understand the projection from object to image. that’s where the bulk of these axiomatic properties of worldly knowledge reside (assumptions about physics being regular, or whatever else you need so that you can leverage things like induction in your learning).
rationally, automating more tasks in my life should make for an easier life that’s subject to fewer demands. rationally, when this isn’t the case — when individuals each working to automate more things causes them to instead be subjected to more demands (learn new skills, else end up on the street), you shouldn’t expect doubling down on this strategy to be long-term viable.
rationally, if you’re predicting the proportion of people able to stay afloat to be always decreasing up to the singularity — a point at which labor becomes valueless — you shouldn’t expect to still be afloat come that moment.
“rationally”, you’re doomed unless you can slide into a different economic system wherein you do observe the benefits of automation. idly watching your peers get rolled over by that bus is bad for your future as it further separates you from the potential exit ramps. the viable solutions to your problem require collective action. that doesn’t put it entirely out of league with rationality, but if it’s not clear from my tone (i apologize if it reads too strong) i believe you’re thinking of this way too narrowly. i think you’re leaning too far toward a Spock type of rationality for what is increasingly a social problem.
> The tradeoff for connecting with similar people is not connecting with people different from us.disagree. as you say, micro-communities are aligned very narrowly. which means that if you pair any two random individuals from the same micro-community, they’ll be extremely similar along only one particular metric, but randomly different across every metric not relevant to that community. the easiest example of this is nationality: to the degree LW is a micro-community, it connects people of many different nationalities. perhaps the disappointment is that although you’re connecting with different people, you aren’t connecting over your differences.
> Widespread camaraderie is becoming rarer and rarer. Most of us live in highly polarized societies where the vast majority of people won’t even date across political party lines. More than ever, we should be looking for ways to create camaraderie at scale and soften the dividing impact of micro-communities.
this may be a misleading use of “divide” in the context of “polarization”. there’s a pretty clear hierarchy here: a societal foundation upon which micro-communities are built. most micro-communities, like ours, aren’t antagonistic to other micro-communities nor can they be too antagonistic to the foundation unless they seek literal suicide. in the realm of micro-communities, cross-community disagreements are resolved via distance. but every micro-community is subject to the demands of that massive, volatile foundation. lesswrong.com dies if the nations to which we belong decide they don’t want us using their cables to communicate. if not for the power of the macro over the micro, we would not experience the level of division we do today. just look up: is it the micro-communities, or the macro society/state, which wields the power to legislate and enforce every thing which is today the topic of polarization?
to consider your example: i think it’s quite right to reject a date from anyone who participates in political action focused on destroying the micro-communities which allow you or i to be as we are. i don’t mind different *philosophies*: i’m perfectly happy to have a companion with whom to debate the existence of God—just so long as they aren’t fighting crusades about it. but coupled to every “political belief” is a vote, and behind that, an action in the real world.
during some period of the 18th century, we lived in a politics of construction. legislating things like the Bill of Rights. *creating* choice in religion, choice in speech, choice in which aspects of our life we share with whom (protections from unreasonable “search and seizure”). we created the very protections which allow for micro-identity and micro-communities.
today’s is a politics of destruction. it’s about restricting what a person can do to or put in their body; restricting what one can teach to the youth; restricting the very bits and information you and i are allowed to share with each other (not just the tainted label of “free speech”, but everything from DRM to intrusive surveillance). the mistake of the day is not that we refuse to date across political boundaries, but that we fail to recognize the very real violence beneath our political abstractions.
> 9/11 broke Americans’ sense of micro-identity. All of a sudden, which team you supported or which political party you agreed with seemed to not matter. Instead, people defaulted to a higher level of identity—being an American.
and this was pivotal to everything you discuss around community. the post-9/11 “camaraderie” led directly to this lasting atmosphere of “if you see something, say something” mutual distrust; to an increased normalization of aggression and dehumanization in our social hierarchies: being fondled by a TSA agent is just a *normal* component of domestic travel; and to the surveillance state which views every micro deviation—vital to micro-community—as cause for suspicion.
if this is what political camaraderie creates, then **i don’t want that**.
> When the fate of humanity is at risk, we all take on the identity of human.
and of course, EA seeks to promote this type of identity even without the risk. it’s also not a terrible place to find those “serendipitous, unscripted, and raw interaction[s]” you treasure. it’s especially interesting if you were find camaraderie in any of its spaces… as a collection of growing micro-communities, promoting an ideal of cooperation as it grows toward that dominant macro-community that’s today largely void of camaraderie: as EA grows how does the sense of camaraderie change, and if its ideals really did become the universal, would that bring about the type of “at scale” camaraderie you dream of?
> If camaraderie at scale makes ordinary days better, wouldn’t it be awesome to experience it more often? Or at least recreate it on a smaller scale within our own lives and communities. [...]
> What’s missing from the current stack is moments of serendipitous, unscripted, and raw interaction—moments where we find connection in places we least expect it.i just want to promote in-person conferences and conventions of any kind here. to build on my claim that everyone in this micro-community is meaningfully different in ways we don’t display in this medium, IRL gatherings are exactly that opportunity to explore those differences. and especially the nominally entertainment-focused conventions (anything from Comicon to EMFcamp): these specifically create an atmosphere where you can feel both comfortable enough and inspired enough to be creative and spontaneous with complete strangers. the most amazing ones are at their core a chaotic swirling of intensely human passion, some unidentifiable thing that just wants for everyone there to impart a little bit of themselves into whatever’s being created and to see eachother in the product. i can’t tell you the number of times i’ve walked past some conference hall at the end of the day, see 4 people painstakingly stacking 1000 chairs, walk inside to help and ten minutes later there’s 20 of us and the room’s spotless. that alone is satisfying, and it’s a short journey from there to far deeper acts of camaraderie.
that camaraderie might not be so immediately focused on “changing the world” or operating “at scale”, but it does exist in the moment, it’s strong enough to perpetuate itself, and it might provide insights for anyone ambitious enough to recreate it in new environments.
i’d love for anyone to present the argument against this. eq says it’s things like karaoke which make friendships great. the friends i know who are eager to do karaoke are the same ones who will start wild, speculative conversation when we’re idly sitting in the living room together. they’re the interesting people.
the people in my life whom, come the first lull in smalltalk after dinner get uncomfortable and declare “great meal, time to go” instead of opening themselves up for those late-night intimate conversations, are the same people who would turn down an invitation to karaoke.
interesting friends are fun friends. “boring” is the opposite of both “fun” and “interesting”. so if the latter two mean something different to the author than to me, perhaps we agree by saying “build non-boring friendships”?
OpenAI estimated that the energy consumption for training GPT-3 was about 3.14 x 10^17 Joules.
sanity checking this figure: 1 kWh is 1000 x 60 x 60 = 3.6 MJ. then GPT-3 consumed 8.7 x 10^10 kWh. at a very conservative $0.04/kWh, that’s $3.5B just in the power bill — disregarding all the non-power costs (i.e. the overheads of operating a datacenter).
i could believe this number’s within 3 orders of magnitude of truth, which is probably good enough for the point of this article, but i am a little surprised if you just took it 100% at face value.
Should there be an opt-out from A.I. systems? Which ones? When is an opt-out clause a genuine choice, and at what point does it become merely an invitation to recede from society altogether, like saying you can choose not to use the internet or vehicular transport or banking services if you so choose.
the examples given are all networks, with many of the nodes human. if “receding from society” means being less connected with the other humans, then there’s no debate: to opt out of these networks is necessarily to “recede from society”.
but LLMs don’t have this property. they aren’t a medium used to bridge connections between individuals: rather things like chatbots exist explicitly to replace human-human interactions with human-machine interactions, and presently they also serve as knowledge repositories: a single massive node with only one connection, to you, the user. to opt out of this form of human-machine interaction at present is not to recede from society, but rather the opposite.
will this change? surely. but i wouldn’t trust the author’s analogy to be at all useful in understanding how.
Instincts to punish people are how actual humans precommit.
i think you could equally frame this as “people precommit due to an expectation of reciprocity”. like, i don’t generally follow through on my commitments to plans with friends because i fear punishment for breaking them. it’s more that i expect whatever amount i invest into the friendship will be reciprocated (approximately).
you could frame the fallout of a commitment failure as “punishment”, but if the risk of punishment exceeded the benefit of cooperation that would discourage me from pre-committing; from interacting with the thing at all. if i thought my crush would beat me should i break things off with him, then i’d simply never ask him out to begin with and we’d probably both be worse off for that.
no love for it from me either, i’m sorry to say. the “society only exists when we overcome our base sexual desires” meme is tired. my university days were simultaneously my most promiscuous and my most productive (subjectively, measured by my extra-curricular contributions to technology). that’s a sample size of 1 (or dozens? depends how you measure it), but Huxley doesn’t even claim a single sample for the opposing view — much less an experiment, despite claiming this foundational assumption as “scientific”.
are complex systems like societies path-dependent? absolutely. the example of decentralized Swedish production arising after centralized English production is intriguing, in that this diversity appears to be predicated on the two societies having been only loosely connected prior to this — suggesting that this sort of divergence become more difficult as societies become more globalized (the opposing point of view being that globalization means those people with similar, but niche, divergent interests can more easily locate and collaborate with eachother). but that’s sort of the only interesting thing i could scrape from that intro, and it’s 80% my own extrapolation.
UBI will always have some power imbalance. if not due to how that income is provided, then by how that income is exchanged for the basic goods. if we want to universally provide for the basic needs, while avoiding that kind of power imbalance, it seems sensible to focus exactly on that: automate more and more of the housing/food production chain, and distribute the tools for that to decrease the power of whichever hierarchies might otherwise bar access to them.
so Universal Basic Income is the practical implementation for providing basic needs for as long as there’s actually a significant labor requirement in that loop: but further into the utopian future it will need to shift to Universal Basic Production, where individuals/households/communities are granted both the power and responsibility of operating whatever machinery actually does the providing.
consider a few scenarios around these two characters: a possibly-depressed Pierre and probably-sociopathic Eliza:
these characters chat IRL and Pierre ends his life.
these characters enact that same scenario on-stage at a theater.
these characters enact that same scene in a videogame via the player selecting dialogue options.
it’s scenario 1 which is horrific. in scenario 2, a Pierre-like viewer is far less likely to end his life after leaving the theater, ditto with scenario 3.
i think some of us already think of these chatbots as “acting out a role” — that’s what a bunch of prompt engineering is about. sometimes we’re explicit in telling that chatbot what “kind” of actor it’s chatting with. getting tsundere output from a chatbot is an example that requires role-playing for both actors. the weird part, then, is why do users end up relating to the experience as if it’s form (1) instead of (2) or (3)? is it possible (and good?) to explicitly shift the experiences into form (2) or (3)? instead of presenting the user a textbox that’s supposed to represent them, should we rather be presenting them a scene of two actors, and placing them in control of one of those actors?
the second-order effects of turning off the WiFi are surely comprised of both positive and negative effects, and i have no idea which valence it nets out to.
these days homes contain devices whose interconnectedness is used to more efficiently regulate power use. for example, the classic utility-controlled water heater, which reduces power draw when electricity is more expensive for the utility company (i.e. when peakers would need to come online). water heaters mostly don’t use WiFi but thermostats like Nest, programmable light bulbs, etc do: when you disrupt that connection, in which direction is power use more likely to change?
i have my phone programmed so that when i go to bed (put it in the “i’m sleeping” Do Not Disturb mode) it will automatically turn off all the outlets and devices — like my TV, game consoles, garage space heater — which i only use during the day. leaving any one of these on for just one night would cancel weeks of gains from disabling WiFi.
interoperability. we take it forgranted everywhere else in life: when you have to replace a fridge it’s easy because they all have the same electrical/water hookups. replace a door, same thing: standardized size, hinges, and knobs. going further, i’ve been upgrading the cabinets/drawers in my kitchen: they’re standard size so i can buy 3rd party silverware inserts, or even inserts made specifically to organize anything that’s k-cup shaped. i replaced the casters on my office chair with oversized carpet-friendly wheels: standardized attachments. so many things in the physical world are made to be interoperable because it facilitates mass production and allows for any company to innovate in any sliver they see. it’s cheaper for producers, and improves the consumer experience.
i assure you those causes and benefits aren’t restricted to the physical world. i read this post in my RSS client, even as my roommate was fiddling with the router because all my RSS feeds get saved for offline reading in the background, before i even decide to read them. simultaneously that RSS standard allows LessWrong to get more reach.
i confront the crux of your post differently: “how do i navigate adversarial relationships (with a business)”? increasingly my approach is to just not engage (or engage less). when it comes to mid-size group stuff, it’s usually pretty easy: LW is just better than Facebook, reddit, or anything that sees its users as a resource to extract from.
for smaller groups or1-to-1 things i choose SMS over Discord; for the people where that’s too low-bandwidth and IRL hangouts aren’t practical, treat any monopoly replacement (signal, telegram, etc) as explicitly ephemeral: as these services switch to value capture we hop ship without losing anything. the world is large enough that there are plenty of substitute activities even if you disengage from Facebook, say. but it’s easier to adopt a policy of “don’t engage” a priori, rather than integrate them into your life and then decide to cut back on them..
further down on that page:
We are also now offering dedicated instances for users who want deeper control over the specific model version and system performance. By default, requests are run on compute infrastructure shared with other users, who pay per request. Our API runs on Azure, and with dedicated instances, developers will pay by time period for an allocation of compute infrastructure that’s reserved for serving their requests.
Developers get full control over the instance’s load (higher load improves throughput but makes each request slower), the option to enable features such as longer context limits, and the ability to pin the model snapshot.
Dedicated instances can make economic sense for developers running beyond ~450M tokens per day.
that suggests one shared “instance” is capable of processing > 450M tokens per day, i.e. $900 of API fees at this new rate. i don’t know what exactly their infrastructure looks like, but the marginal costs of the compute here have got to be still an order of magnitude lower than what they’re charging (which is sensible: they do have fixed costs they have to recoup, and they are seeking to profit).
commenting on the body, separate from the incident that prompted this. when i was in school:
the occasional professor would invite the class for drinks after a test.
a subset of students i TA’d would invite me for tea.
a subset of students would bake food and share it after exams.
no mention of relationships yet. but all these activities are exactly those avenues by which people learn about each other and by which they form bonds. the professors i bonded with were exactly those professors whose office hours i attended most. and vice versa for the students i bonded with attending more of my office hours.
student/teacher bonding means the student is more comfortable asking the teacher for help, means the teacher better understands how to frame things in a way the student will get. if you ran the study, you would surely find correlation between this and course scores. does that mean this style of bonding is unethical?
if you ran the study and found that informal socializing didn’t decrease any student’s learning outcome, but it did increase some outcomes non-uniformly, would that be unethical?
it seems to me that the vast majority of times where relationships cause power problems is when one of the members is in competition with another. the argument that the criteria for competition among employees in the workplace shouldn’t involve sex is largely an argument that people shouldn’t be coerced into participating in a competition they don’t want to be a part of. aiming for consensus w.r.t. which criteria you should apply such that all workers prefer to be in such competition is a somewhat limited prospect. maybe there’s some progress that can be made there, but i expect the bulk of progress will actually be in finding ways to not force everyone into the same competition. in the post-scarcity world where you don’t have to compete in the workplace to stay alive, this consent issue would largely disappear. until then, the best we can do is fragment the competitive pools: less hierarchical workplaces such that the effect of any one competition is radically reduced, and more employment choices so that those who want work and life to be separate can avoid entering into direct competition with those who want work and life to overlap.
More importantly, if we have some one value, that values are to be valued, so much as to enact for, not only to want them—then we have a value which has no opposite in utilitarianism.
sounds a little like Preference Utilitarianism.
this observation means, if we align to mere values of humanity: AI can simply modify the humans, so to alter their values and call it a win; AI aligns you to AI. In general, for fulfillment of any human value, to make the human value it, seems absolutely the easiest, for any case.
here “autonomy”, “responsibility”, “self-determination” are all related values (or maybe closer to drives?) that counter this approach. put simply, “people don’t like being told what to do”. if an effective AI achieves alignment via this approach, i would expect it to take a low-impedance path where there’s no “forceful” value modification, coercion is done by subtler reshaping of the costs/benefits any time humans make value tradeoffs.
e.g. if a clever AI wanted humans to “value” pacifism, it might think to give a high cost to large-scale violence, which it could do by leaking the technology for a global communications network, then for an on-demand translation systems between all human languages, then for highly efficient wind power/sail design, and before you know it both the social and economic costs to large-scale violence is enormous and people “decide” that they “value” peaceful coexistence.
i’m not saying today’s global trade system is a result of AI… but there are so many points of leverage here that if it (or some future system like it) were, would we know?
if we wanted to avoid this type of value modification, we would need to commit to a value system that never changes. write these down on clay tablets that could be preserved in museums in their original form, keep the language of these historic texts alive via rituals and tradition, and encourage people to have faith in the ideas proposed by these ancients. you could make a religion out of this. and its strongest meta-value would necessarily be one of extreme conservatism, a resistance to change.
i’m naive to the details of GPT specifically, but it’s easy to accidentally make any reduction non-deterministic when working with floating point numbers — even before hardware variations.
for example, you want to compute the sum over a 1-billion entry vector where each entry is the number 1. in 32-bit IEEE-754, you should get different results by accumulating linearly (1+(1+(1+…))) vs tree-wise (…((1+1) + (1+1))…).
in practice most implementations do some combination of these. i’ve seen someone do this by batching groups of 100,000 numbers to sum linearly, with each batch dispatched to a different compute unit and the 10,000 results then being summed in a first-come/first-serve manner (e.g. a queue, or even a shared accumulator). then you get slightly different results based on how each run is scheduled (well, the all-1’s case is repeatable with this method but it wouldn’t be with real data).
and then yes, bring in different hardware, and the scope broadens. the optimal batching size (which might be exposed as a default somewhere) changes such that even had you avoided that scheduling-dependent pitfall, you would now see different results than on the earlier hardware. however, you can sometimes tell these possibilities apart! if it’s non-deterministic scheduling, the number of different outputs for the same input is likely higher order than if it’s variation strictly due to differing hardware models. if you can generate 10,000 different outputs from the same input, that’s surely greater than the number of HW models, so it would be better explained by non-deterministic scheduling.
I’m relatively OOTL on AI since GPT-3. My friend is terrified that we need to halt it urgently: I couldn’t understand his point of view; he mentioned this book to me. I see a number of pre-readers saying the version they read is well-suited exactly for convincing people like me. At which point: if you believe the threat is imminent, why delay the book four months? I’ll read a digital copy today if you point me to it.