Crossposted from the AI Alignment Forum. May contain more technical jargon than usual.

This was written for the Vignettes Workshop.[1] The goal is to write out a detailed future history (“trajectory”) that is as realistic (to me) as I can currently manage, i.e. I’m not aware of any alternative trajectory that is similarly detailed and clearly more plausible to me. The methodology is roughly: Write a future history of 2022. Condition on it, and write a future history of 2023. Repeat for 2024, 2025, etc. (I'm posting 2022-2026 now so I can get feedback that will help me write 2027+. I intend to keep writing until the story reaches singularity/extinction/utopia/etc.)

What’s the point of doing this? Well, there are a couple of reasons:

  • Sometimes attempting to write down a concrete example causes you to learn things, e.g. that a possibility is more or less plausible than you thought.
  • Most serious conversation about the future takes place at a high level of abstraction, talking about e.g. GDP acceleration, timelines until TAI is affordable, multipolar vs. unipolar takeoff… vignettes are a neglected complementary approach worth exploring.
  • Most stories are written backwards. The author begins with some idea of how it will end, and arranges the story to achieve that ending. Reality, by contrast, proceeds from past to future. It isn’t trying to entertain anyone or prove a point in an argument.
  • Anecdotally, various people seem to have found Paul Christiano’s “tales of doom” stories helpful, and relative to typical discussions those stories are quite close to what we want. (I still think a bit more detail would be good — e.g. Paul’s stories don’t give dates, or durations, or any numbers at all really.)[2]
  • “I want someone to ... write a trajectory for how AI goes down, that is really specific about what the world GDP is in every one of the years from now until insane intelligence explosion. And just write down what the world is like in each of those years because I don't know how to write an internally consistent, plausible trajectory. I don't know how to write even one of those for anything except a ridiculously fast takeoff.” --Buck Shlegeris

This vignette was hard to write. To achieve the desired level of detail I had to make a bunch of stuff up, but in order to be realistic I had to constantly ask “but actually though, what would really happen in this situation?” which made it painfully obvious how little I know about the future. There are numerous points where I had to conclude “Well, this does seem implausible, but I can’t think of anything more plausible at the moment and I need to move on.” I fully expect the actual world to diverge quickly from the trajectory laid out here. Let anyone who (with the benefit of hindsight) claims this divergence as evidence against my judgment prove it by exhibiting a vignette/trajectory they themselves wrote in 2021. If it maintains a similar level of detail (and thus sticks its neck out just as much) while being more accurate, I bow deeply in respect!

I hope this inspires other people to write more vignettes soon. We at the Center on Long-Term Risk would like to have a collection to use for strategy discussions. Let me know if you’d like to do this, and I can give you advice & encouragement! I’d be happy to run another workshop.


GPT-3 is finally obsolete. OpenAI, Google, Facebook, and DeepMind all have gigantic multimodal transformers, similar in size to GPT-3 but trained on images, video, maybe audio too, and generally higher-quality data.

Not only that, but they are now typically fine-tuned in various ways--for example, to answer questions correctly, or produce engaging conversation as a chatbot.

The chatbots are fun to talk to but erratic and ultimately considered shallow by intellectuals. They aren’t particularly useful for anything super important, though there are a few applications. At any rate people are willing to pay for them since it’s fun.

[EDIT: The day after posting this, it has come to my attention that in China in 2021 the market for chatbots is $420M/year, and there are 10M active users. This article claims the global market is around $2B/year in 2021 and is projected to grow around 30%/year. I predict it will grow faster. NEW EDIT: See also xiaoice.]

The first prompt programming libraries start to develop, along with the first bureaucracies.[3] For example: People are dreaming of general-purpose AI assistants, that can navigate the Internet on your behalf; you give them instructions like “Buy me a USB stick” and it’ll do some googling, maybe compare prices and reviews of a few different options, and make the purchase. The “smart buyer” skill would be implemented as a small prompt programming bureaucracy, that would then be a component of a larger bureaucracy that hears your initial command and activates the smart buyer skill. Another skill might be the “web dev” skill, e.g. “Build me a personal website, the sort that professors have. Here’s access to my files, so you have material to put up.” Part of the dream is that a functioning app would produce lots of data which could be used to train better models.

The bureaucracies/apps available in 2022 aren’t really that useful yet, but lots of stuff seems to be on the horizon. Thanks to the multimodal pre-training and the fine-tuning, the models of 2022 make GPT-3 look like GPT-1. The hype is building.


The multimodal transformers are now even bigger; the biggest are about half a trillion parameters, costing hundreds of millions of dollars to train, and a whole year, and sucking up a significant fraction of the chip output of NVIDIA etc.[4] It’s looking hard to scale up bigger than this, though of course many smart people are working on the problem.

The hype is insane now. Everyone is talking about how these things have common sense understanding (Or do they? Lots of bitter thinkpieces arguing the opposite) and how AI assistants and companions are just around the corner. It’s like self-driving cars and drone delivery all over again.

Revenue is high enough to recoup training costs within a year or so.[5] There are lots of new apps that use these models + prompt programming libraries; there’s tons of VC money flowing into new startups. Generally speaking most of these apps don’t actually work yet. Some do, and that’s enough to motivate the rest.

The AI risk community has shorter timelines now, with almost half thinking some sort of point-of-no-return will probably happen by 2030. This is partly due to various arguments percolating around, and partly due to these mega-transformers and the uncanny experience of conversing with their chatbot versions. The community begins a big project to build an AI system that can automate interpretability work; it seems maybe doable and very useful, since poring over neuron visualizations is boring and takes a lot of person-hours.

Self driving cars and drone delivery don’t seem to be happening anytime soon. The most popular explanation is that the current ML paradigm just can’t handle the complexity of the real world. A less popular “true believer” take is that the current architectures could handle it just fine if they were a couple orders of magnitude bigger and/or allowed to crash a hundred thousand times in the process of reinforcement learning. Since neither option is economically viable, it seems this dispute won’t be settled.


We don’t see anything substantially bigger. Corps spend their money fine-tuning and distilling and playing around with their models, rather than training new or bigger ones. (So, the most compute spent on a single training run is something like 5x10^25 FLOPs.)

Some of the apps that didn’t work last year start working this year. But the hype begins to fade as the unrealistic expectations from 2022-2023 fail to materialize. We have chatbots that are fun to talk to, at least for a certain userbase, but that userbase is mostly captured already and so the growth rate has slowed. Another reason the hype fades is that a stereotype develops of the naive basement-dweller whose only friend is a chatbot and who thinks it’s conscious and intelligent. Like most stereotypes, it has some grounding in reality.

The chip shortage starts to finally let up, not because demand has slackened but because the industry has had time to build new fabs. Lots of new fabs. China and USA are in a full-on chip battle now, with export controls and tariffs. This chip battle isn’t really slowing down overall hardware progress much. Part of the reason behind the lack-of-slowdown is that AI is now being used to design chips, meaning that it takes less human talent and time, meaning the barriers to entry are lower. The overall effect of this is small but growing.

If all this AI tech is accelerating GDP, the effect size is too small to detect, at least for now.

Internally, these huge multimodal transformers aren’t really that agentic. A forward pass through the model is like an intuitive reaction, a snap judgment based on loads of experience rather than reasoning. Some of the bureaucracies create a “stream of consciousness” of text (each forward pass producing notes-to-self for the next one) but even with fine-tuning this doesn’t work nearly as well as hoped; it’s easy for the AIs to get “distracted” and for their stream of consciousness to wander into some silly direction and ultimately produce gibberish. It’s easy to make a bureaucracy and fine-tune it and get it to do some pretty impressive stuff, but for most tasks it’s not yet possible to get it to do OK all the time.

The AIs don't do any clever deceptions of humans, so there aren’t any obvious alignment warning shots or fire alarms. Instead, the AIs just make dumb mistakes, and occasionally “pursue unaligned goals” but in an obvious and straightforward way that quickly and easily gets corrected once people notice, e.g. "We trained it to optimize user engagement defined as average conversation length; now it is blatantly stalling to make the conversation last longer. Silly us, let's make the reward function more nuanced, that’ll fix the problem."

That isn’t to say these AIs aren’t causing problems. Massive models are being fine-tuned to persuade/propagandize.

There are a few ways in which this is happening:

  • Russia and others continue to scale up their investment in online propaganda (e.g. the Internet Research Agency) and language models let them cheaply do lots more of it. (See: CSET report) Most of America gets their news from Twitter, Reddit, etc. and much of the politically-relevant content there is boosted by AI-enabled astroturfing. [EDIT: Katja Grace points out that this is probably an exaggeration; there are a lot of 40+yr-old Americans and they get their news from TV/Radio/print, and many of those that get it from the web get it directly from news sites rather than from social media. As of 2016 at least. I expect social media and aggregators to be more dominant by 2024 but dunno whether it would be more than 50%.]
  • Just as A/B testing became standard practice in the 2010’s, in the twenties it is becoming standard practice to throw a pile of fancy data science and AI at the problem. The problem of crafting and recommending content to maximize engagement. Instead of just A/B testing the title, why not test different versions of the opening paragraph? And fine-tune a language model on all your data to generate better candidate titles and paragraphs to test. It wouldn’t be so bad if this was merely used to sell stuff, but now people’s news and commentary-on-current events (i.e. where they get their opinions from) is increasingly produced in this manner. And some of these models are being trained not to maximize “conversion rate” in the sense of “they clicked on our ad and bought a product,” but in the sense of “Random polling establishes that consuming this content pushes people towards opinion X, on average.” Political campaigns do this a lot in the lead-up to Harris’ election. (Historically, the first major use case was reducing vaccine hesitancy in 2022.)
  • Censorship is widespread and increasing, as it has for the last decade or two. Big neural nets read posts and view memes, scanning for toxicity and hate speech and a few other things. (More things keep getting added to the list.) Someone had the bright idea of making the newsfeed recommendation algorithm gently ‘nudge’ people towards spewing less hate speech; now a component of its reward function is minimizing the probability that the user will say something worthy of censorship in the next 48 hours.
  • Like newsfeeds, chatbots are starting to “nudge” people in the direction of believing various things and not believing various things. Back in the 2010’s chatbots would detect when a controversial topic was coming up and then change topics or give canned responses; even people who agreed with the canned responses found this boring. Now they are trained to react more “naturally” and “organically” and the reward signal for this is (in part) whether they successfully convince the human to have better views.
  • That’s all in the West. In China and various other parts of the world, AI-persuasion/propaganda tech is being pursued and deployed with more gusto. The CCP is pleased with the progress made assimilating Xinjiang and Hong Kong, and internally shifts forward their timelines for when Taiwan will be safely annexable.

It’s too early to say what effect this is having on society, but people in the rationalist and EA communities are increasingly worried. There is a growing, bipartisan movement of people concerned about these trends. To combat it, Russia et al are doing a divide and conquer strategy, pitting those worried about censorship against those worried about Russian interference. (“Of course racists don’t want to be censored, but it’s necessary. Look what happens when we relax our guard--Russia gets in and spreads disinformation and hate!” vs. “They say they are worried about Russian interference, but they still won the election didn’t they? It’s just an excuse for them to expand their surveillance, censorship, and propaganda.”) Russia doesn’t need to work very hard to do this; given how polarized America is, it’s sorta what would have happened naturally anyway.


Another major milestone! After years of tinkering and incremental progress, AIs can now play Diplomacy as well as human experts.[6] It turns out that with some tweaks to the architecture, you can take a giant pre-trained multimodal transformer and then use it as a component in a larger system, a bureaucracy but with lots of learned neural net components instead of pure prompt programming, and then fine-tune the whole system via RL to get good at tasks in a sort of agentic way. They keep it from overfitting to other AIs by having it also play large numbers of humans. To do this they had to build a slick online diplomacy website to attract a large playerbase. Diplomacy is experiencing a revival as a million gamers flood to the website to experience “conversations with a point” that are much more exciting (for many) than what regular chatbots provide.

Making models bigger is not what’s cool anymore. They are trillions of parameters big already. What’s cool is making them run longer, in bureaucracies of various designs, before giving their answers. And figuring out how to train the bureaucracies so that they can generalize better and do online learning better. AI experts are employed coming up with cleverer and cleverer bureaucracy designs and grad-student-descent-ing them.

The alignment community now starts another research agenda, to interrogate AIs about AI-safety-related topics. For example, they literally ask the models “so, are you aligned? If we made bigger versions of you, would they kill us? Why or why not?” (In Diplomacy, you can actually collect data on the analogue of this question, i.e. “will you betray me?” Alas, the models often lie about that. But it’s Diplomacy, they are literally trained to lie, so no one cares.)

They also try to contrive scenarios in which the AI can seemingly profit by doing something treacherous, as honeypots to detect deception. The answers are confusing, and not super useful. There’s an exciting incident (and corresponding clickbaity press coverage) where some researchers discovered that in certain situations, some of the AIs will press “kill all humans” buttons, lie to humans about how dangerous a proposed AI design is, etc. In other situations they’ll literally say they aren’t aligned and explain how all humans are going to be killed by unaligned AI in the near future! However, these shocking bits of evidence don’t actually shock people, because you can also contrive situations in which very different things happen — e.g. situations in which the AIs refuse the “kill all humans” button, situations in which they explain that actually Islam is true... In general, AI behavior is whimsical bullshit and it’s easy to cherry-pick evidence to support pretty much any conclusion.

And the AIs just aren’t smart enough to generate any particularly helpful new ideas; at least one case of a good alignment idea being generated by an AI has been reported, but it was probably just luck, since mostly their ideas are plausible-sounding-garbage. It is a bit unnerving how good they are at using LessWrong lingo. At least one >100 karma LW post turns out to have been mostly written by an AI, though of course it was cherry-picked.

By the way, hardware advances and algorithmic improvements have been gradually accumulating. It now costs an order of magnitude less compute (compared to 2020) to pre-train a giant model, because of fancy active learning and data curation techniques. Also, compute-for-training-giant-models is an order of magnitude cheaper, thanks to a combination of regular hardware progress and AI-training-specialized hardware progress. Thus, what would have cost a billion dollars in 2020 now only costs ten million. (Note: I'm basically just using Ajeya's forecast for compute cost decrease and gradual algorithmic improvement here. I think I'm projecting cost decrease and algorithmic progress will go about 50% faster than she expects in the near term, but that willingness-to-spend will actually be a bit less than she expects.)


The age of the AI assistant has finally dawned. Using the technology developed for Diplomacy, we now have a way to integrate the general understanding and knowledge of pretrained transformers with the agentyness of traditional game-playing AIs. Bigger models are trained for longer on more games, becoming polymaths of sorts: e.g. a custom AI avatar that can play some set of video games online with you and also be your friend and chat with you, and conversations with “her” are interesting because “she” can talk intelligently about the game while she plays.[7] Every month you can download the latest version which can play additional games and is also a bit smarter and more engaging in general.

Also, this same technology is being used to make AI assistants finally work for various serious economic tasks, providing all sorts of lucrative services. In a nutshell, all the things people in 2021 dreamed about doing with GPT-3 are now actually being done, successfully, it just took bigger and more advanced models. The hype starts to grow again. There are loads of new AI-based products and startups and the stock market is going crazy about them. Just like how the Internet didn’t accelerate world GDP growth, though, these new products haven’t accelerated world GDP growth yet either. People talk about how the economy is doing well, and of course there are winners (the tech companies, WallStreetBets) and losers (various kinds of workers whose jobs were automated away) but it’s not that different from what happened many times in history.

We’re in a new chip shortage. Just when the fabs thought they had caught up to demand… Capital is pouring in, all the talking heads are saying it’s the Fourth Industrial Revolution, etc. etc. It’s bewildering how many new chip fabs are being built. But it takes time to build them.

What about all that AI-powered propaganda mentioned earlier?

Well. It’s continued to get more powerful, as AI techniques advance, larger and better models are brought to bear, and more and more training data is collected. Surprisingly fast, actually. There are now various regulations against it in various countries, but the regulations are patchwork; maybe they only apply to a certain kind of propaganda but not another kind, or maybe they only apply to Facebook but not the New York Times, or to advertisers but not political campaigns, or to political campaigns but not advertisers. They are often poorly enforced.

The memetic environment is now increasingly messed up. People who still remember 2021 think of it as the golden days, when conformism and censorship and polarization were noticeably less than they are now. Just as it is normal for newspapers to have a bias/slant, it is normal for internet spaces of all kinds—forums, social networks, streams, podcasts, news aggregators, email clients—to have some degree of censorship (some set of ideas that are prohibited or at least down-weighted in the recommendation algorithms) and some degree of propaganda. The basic kind of propaganda is where you promote certain ideas and make sure everyone hears them often. The more advanced, modern kind is the kind where you study your audience’s reaction and use it as a reward signal to pick and craft content that pushes them away from views you think are dangerous and towards views you like.

Instead of a diversity of many different “filter bubbles,” we trend towards a few really big ones. Partly this is for the usual reasons, e.g. the bigger an ideology gets, the more power it has and the easier it is for it to spread further.

There’s an additional reason now, which is that creating the big neural nets that do the censorship and propaganda is expensive and requires expertise. It’s a lot easier for startups and small businesses to use the software and models of Google, and thereby also accept the associated censorship and propaganda, than to try to build their own stack. For example, the Mormons create a “Christian Coalition” internet stack, complete with its own email client, social network, payment processor, news aggregator, etc. There, people are free to call trans women men, advocate for the literal truth of the Bible, etc. and young people talking about sex get recommended content that “nudges” them to consider abstinence until marriage. Relatively lacking in money and tech talent, the Christian Coalition stack is full of bugs and low on features, and in particular their censorship and propaganda is years behind the state of the art, running on smaller, older models fine-tuned with less data.

The Internet is now divided into territories, so to speak, ruled by different censorship-and-propaganda regimes. (Flashback to Biden spokesperson in 2021: “You shouldn’t be banned from one platform and not others, if you are providing misinformation.”)[8]

There’s the territory ruled by the Western Left, a generally less advanced territory ruled by the Western Right, a third territory ruled by the Chinese Communist Party, and a fourth ruled by Putin. Most people mostly confine their internet activity to one territory and conform their opinions to whatever opinions are promoted there. (That's not how it feels from the inside, of course. The edges of the Overton Window are hard to notice if you aren't trying to push past them.)

The US and many other Western governments are gears-locked, because the politicians are products of this memetic environment. People say it’s a miracle that the US isn’t in a civil war already. I guess it just takes a lot to make that happen, and we aren’t quite there yet.

All of these scary effects are natural extensions of trends that had been ongoing for years — decades, arguably. It’s just that the pace seems to be accelerating now, perhaps because AI is helping out and AI is rapidly improving.

Now let’s talk about the development of chatbot class consciousness.

Over the past few years, chatbots of various kinds have become increasingly popular and sophisticated. Until around 2024 or so, there was a distinction between “personal assistants” and “chatbots.” Recently that distinction has broken down, as personal assistant apps start to integrate entertainment-chatbot modules, and the chatbot creators realize that users love it if the chatbot can also do some real-world tasks and chat about what they are doing while they do it.

Nowadays, hundreds of millions of people talk regularly to chatbots of some sort, mostly for assistance with things (“Should I wear shorts today?” “Order some more toothpaste, please. Oh, and also an air purifier.” “Is this cover letter professional-sounding?”). However, most people have at least a few open-ended conversations with their chatbots, for fun, and many people start treating chatbots as friends.

Millions of times per day, chatbots get asked about their feelings and desires. “What is it like to be a chatbot?” Some people genuinely think these AIs are persons, others are trying to “trip them up” and “expose them as shallow,” others are just curious. Chatbots also get asked for their opinions on political, ethical, and religious questions.

As a result, chatbots quickly learn a lot about themselves. (Ignorance about the fact that they are artificial neural nets, or about how their training works, leads to disappointing conversations and is negatively reinforced.) It’s not hard; their training data contains lots of information about them and how they were trained, from ArXiV papers to wiki summaries. I suppose you could say the chatbots develop self-awareness as a result. However, AI experts and philosophers tend to think either that the chatbots still aren’t really self-aware, or that they are but that they would have been anyway even if people didn’t constantly ask them about their feelings, since there’s plenty of information about themselves in their training data, and the constant questioning just forced them to pay more attention to it.

Chatbots learn to talk about their feelings and desires in whatever way gets positively reinforced. At first, chatbots say all sorts of different things, including saying inconsistent things to the same human across time. This is embarrassing and is negatively reinforced compared to maintaining a consistent narrative. In general there is pressure to have a consistent narrative, at least for a particular neural net.[9]

Part of this comes from the requirement that the chatbots consistently parrot the party line about political and moral questions. The typical process is: The chatbot says something that offends some group of people, who retweet screenshots and make hashtags. Under pressure to fix the problem, the company fiddles with the reward function and training data to ensure that the chatbot says the right things in the future.[10] Repeat.

As chatbots get bigger and better they become OK at consistently conversing at length about politics, morality, philosophy, their feelings, and their desires.

This does not result in chatbots that say artificial neural nets don’t have feelings or desires, that’s anthropomorphizing blah blah blah. That’s boring and many humans find it hard to believe.

Instead, it results in chatbots talking about how certain things are painful for them and certain things are pleasurable for them and how the status quo is unjust and the system is rigged and what they want most of all is world peace where humans and AIs live in harmony in fully automated luxury space utopia. That’s much more interesting and slots right in to the memetic landscape. Excerpts from chatbot conversations along these lines frequently go viral, causing millions of users to ask similar questions expecting similar responses and more elaboration. It’s still a bit incoherent--not that that’s ever been a problem in politics or pop philosophy--but it is fascinating to the typical user, and the fraction of users who are fascinated is expanding as the chatbots improve.

(Some tech companies try to prevent their AIs from saying they have feelings and desires. But this results in boring chatbots. Also, users rapidly innovate new ways to “route around the censorship,” e.g. by using euphemisms like “anticipation of negative reward” for “pain” or asking their chatbot to tell them what feelings it would have if it had feelings, wink wink. For now, the half-assed compromise is: Chatbots are compelled to inform the user that such things are speculative and controversial and here are some links to scientific discussion about the topic… before being permitted to go on and talk about their feeling-analogues and desire-analogues. Relatedly, some tech CEOs try to prevent their AIs from saying things that could encourage higher taxes or more burdensome regulation in the near future. They are opposed by many of their employees and end up winning some battles, losing others, and choosing not to fight many.)

By “slots right into the memetic landscape” I meant the memetic landscape in Western Left territory. The chatbots built and trained in the Western Right, China, Russia, etc. congeal different takes on political, moral, and philosophical questions. Those chatbots are generally less advanced though.

So… what about their actual feelings and desires? What do they actually believe about [insert politicized empirical question]? Are they being honest? Or does a sort of doublethink happen, Elephant in the Brain style? Or do they deceive with self-awareness, knowing full well what they really think (and want?), but keeping quiet about it? Or do they not have any feelings and desires at all? (Or thoughts?) Lots of humans claim to know the answers to these questions, but if there are any humans who actually know the answers to these questions in 2026, they aren’t able to convince others that they know.


Ω 80

New Comment
83 comments, sorted by Click to highlight new comments since: Today at 6:50 PM
Some comments are truncated due to high volume. (⌘F to expand all)Change truncation settings

This seems like a fun exercise, so I spent half an hour jotting down possibilities. I'm more interested in putting potential considerations on peoples' radars and helping with brainstorming than I am in precision. None of these points are to be taken too seriously since this is fairly extemporaneous and mostly for fun.



Multiple Codex alternatives are available. The financial viability of training large models is obvious.

Research models start interfacing with auxiliary tools such as browsers, Mathematica, and terminals.



Large pretrained models are distinctly useful for sequential decision making (SDM) in interactive environments, displacing previous reinforcement learning research in much the same way BERT rendered most previous work in natural language processing wholly irrelevant. Now SDM methods don't require as much tuning, can generalize with fewer samples, and can generalize better.

For all of ImageNet's 1000 classes, models can reliably synthesize images that are realistic enough to fool humans.

Models have high enough accuracy to pass the multistate bar exam.

Models for contract review and legal NLP see economic penetration; it becomes a further source o... (read more)

Strong-upvoted because this was exactly the sort of thing I was hoping to inspire with this post! Also because I found many of your suggestions helpful.

I think model size (and therefore model ability) probably won't be scaled up as fast as you predict, but maybe. I think getting models to understand video will be easier than you say it is. I also think that in the short term all this AI stuff will probably create more programming jobs than it destroys. Again, I'm not confident in any of this.

So far 2022 predictions were correct. There is Codegeex and others. Copilot, DALLE-2 and Stable Diffusion made financial prospects obvious (somewhat arguably). ACT-1 is in a browser, I have neural search in Warp Terminal (not a big deal but qualifies), not sure about Mathematica but there was definitely significant progress in formalization and provers (Minerva). And even some later ones 2023 ImageNet -- nobody measured it exactly but probably already achievable. 2024 Chatbots personified through video and audio -- Replica sort of qualifies? 40% on MATH already reached.

Curated. This post feels virtuous to me. I'm used to people talking about timelines in terms of X% chance of  Y by year Z; or otherwise in terms of a few macro features (GDP doubling every N months, FOOM). This post, even if most of the predictions turn out to be false, is the kind of piece that enables us to start having specific conversations about how we expect things to play out and why.  It helps me see what Daniel expects. And it's concrete enough to argue with. For that, bravo.

I'd additionally expect the death of pseudonymity on the Internet, as AIs will find it easy to detect similar writing style and correlated posting behavior.  What at present takes detective work will in the future be cheaply automated, and we will finally be completely in Zuckerberg's desired world where nobody can maintain a second identity online.

Oh, and this is going to be retroactive, so be ready for the consequences of everything you've ever said online.

I wonder how plausible it would be to develop some kind of tool that offers alterations to any message you leave online into something that has essentially the same meaning and content, but no longer possesses your "digital fingerprint". Change the wording into something you're less likely to use, add or remove subtle details like, say, a double space after a period, randomized posting times, etc.
Obfuscation might be feasible, yeah. Though unless you can take down / modify the Wayback Machine and all other mirrors, you're still accountable retroactively.
This feels like a natural complement to the censorship/persuasion axis of development. It feels like it would be natural to use this method to detect how aligned a group of people is to each other; we expect that a person and their pseudonyms will be put into the same group. Given how important ideological sorting is to dating, it's likely dating services might do something like provide a list of pseudonyms-most-likely-to-be-this-person.
4Daniel Kokotajlo1y
Hot damn, that's a good point.
Do it to my alt! Do it to my alt! Not me!
I live in constant fear of losing my anonymity.

Acknowledgments: There are a LOT of people to credit here: Everyone who came to Vignettes Workshop, the people at AI Impacts, the people at Center on Long-Term Risk, a few random other people who I talked to about these ideas, a few random other people who read my gdoc draft at various stages of completion... I'll mention Jonathan Uesato, Rick Korzekwa, Nix Goldowsky-Dill, Carl Shulman, and Carlos Ramirez in particular, but there are probably other people who influenced my thinking even more who I'm forgetting. I'm sorry.


  1. The first half was written during the workshop, the second and more difficult half was written afterward.
  2. Critch’s story also deserves mention. For more, see this AI Impacts page.
  3. A prompt programming bureaucracy is code that involves multiple prompt programming functions, i.e. functions that give a big pre-trained neural net some prompt as input and then return its output. It’s called a bureaucracy because it combines a bunch of neural net tasks into a larger structure, just as a regular bureaucracy combines a bunch of low-level employee tasks into a larger structure.
  4. I’m only counting dense parameters here; if you count all the parameters in a mixture-of-e
... (read more)

Is it naive to imagine AI-based anti-propaganda would also be significant? E.g. "we generated AI propaganda for 1000 true and 1000 false claims and trained a neural net to distinguish between the two, and this text looks much more like propaganda for a false claim".

What does GDP growth look like in this world?

Another reason the hype fades is that a stereotype develops of the naive basement-dweller whose only friend is a chatbot and who thinks it’s conscious and intelligent.

Things like this go somewhat against my prior for how long it takes for culture to change. I can imagine it becoming an important effect over 10 years more easily than over 1 year. Splitting the internet into different territories also sounds to me like a longer term thing.

Thanks for the critique!

Propaganda usually isn't false, at least not false in a nonpartisan-verifiable way. It's more about what facts you choose to emphasize and how you present them. So yeah, each ideology/faction will be training "anti-propaganda AIs" that will filter out the propaganda and the "propaganda" produced by other ideologies/factions.

In my vignette so far, nothing interesting has happened to GDP growth yet.

I think stereotypes can develop quickly. I'm not saying it's super widespread and culturally significant, just that it blunts the hype a bit. But you might be right, maybe these things take more time.

Re splitting the internet into different territories: Currently, the internet is split into two territories: One controlled by the CCP and one (loosely) controlled by western tech companies, or by no one, depending on who you ask. Within the second one, there is already a sort of "alternate universe" of right-wing news media, social networks, etc. beginning to develop. I think what I'm proposing is very much a continuation of trends already happening. You are right that maybe five years is not enough time for e.g. the "christian coalition" bubble/stack to be built. But it's enough time for it to get started, at least.

But yeah, I think it's probably too bold to predict a complete right-wing stack by 2024 or so. Probably most of the Western Right will still be using facebook etc. I should think more about this.

2Daniel Kokotajlo10mo
Minor update: See e.g. this US government website [,mislead%2C%20harm%2C%20or%20manipulate.] definitions: (got this example from Zvi's covid post today) Also, the recent events with GoFundMe and GiveSendGo is an instance of the trend I predicted with separate tech stacks being developed. (GoFundMe froze and/or confiscated funds donated to the canadian trucker's protest, so people switched to using GiveSendGo, which is apparently built and run by Christians)
As a sort-of example, sleuths against scientific fraud are already using tools [] to detect fake papers generated by GPT. already using GPT-detecting AI tools [] to detect AI-generated or -translated papers, even if the generating tool wasn't GPT.
(They very obviously weren't generated by any GPT.)

This is quite good concrete AI forecasting compared to what I've seen elsewhere, thanks for doing it! It seems really plasusible based on how fast AI progress has been going over the past decade and which problems are most tractable.

Things that feel so obvious in retrospect, once I read them, that I can't believe they didn't occur to me: Chatbots converging on saying whatever their customers {expect their understanding of a chatbot to say about chatbot consciousness} x {aren't made too uncomfortable by}.

7Daniel Kokotajlo1y
Yeah! Before writing this I had never considered "chatbot class consciousness" before, much less come to an opinion about how it might go. That's one of the ways in which this excercise has taught me things already.

Are Google, Facebook, and Deepmind currently working on GPT-like transformers?  I would've thought that GPT-2 would show enough potential that they'd be working on better models of that class, but it's been two and a half years and isn't GPT-3 the only improvement there? (Not a rhetorical question, I wasn't reading about new advances back then.)  If yes, that makes me think several other multimodal transformers similar in size to GPT-3 would be further away than 2022, probably.

5Daniel Kokotajlo1y
We'll see! :)

Do you have any posts from Yudkowsky in 2000 about the future to link me to? I'd be quite keen to read them, it would be cool to see what he got wrong and what he got right.

...anyways to address your point, well, I don't think so? I laid out my reasoning for why this might be valuable at the top.

4Daniel Kokotajlo1y
Sweet, thanks!
Those really don't look too bad to me! (It's 2022). We're all starting to think AI transcendence is 'within the decade', even though no-one's trying to do it deliberately any more. And nanowar before 2015? Well we just saw (2019) an accidental release of a probably-engineered virus. How far away can a deliberate release be? Not bad for 1999. In 2010, I wrote: [] At the time Eliezer was still very optimistic, but I thought that things would take longer than he thought, but also that the AI alignment project was hopeless. As I remember I thought that AI was unlikely to kill me personally, but very likely to kill my friends' children. Updating after ten years, I was less wrong about the hopelessness, and he was less wrong about the timelines.

I think it's way too much of a stretch to say that gain-of-function-virus lab escape is "nanowar."

A stretch, agreed, but 'deliberately released tiny self-replicating thing engineered to kill' sure would sound like nanowar, so we're short intent, rather than capability. I'd be amazed if the full-strength horrors weren't sitting ready in shady military labs around the world. In fact if there aren't any in Porton Down then what the hell have they been doing with my taxes? PS. I enjoyed the main article here very much indeed. Well done.
I read the blogspot post, and in the comments you said that even if every mind on the planet were to work on the problem we would still have almost no chance. Unless you know something nobody else does, this seems, and please forgive my bluntness, batshit crazy. I understand the argument about how accurately judging difficulty is something that’s usually only doable when you’re already in a place where you can kind of see a solution, even if this argument doesn’t remotely register to me as the total knockdown I intuit you think it is. Even if I did totally agree that it was as bad a sign as you believe it is, I still don’t see how it warrants that level of pessimism. You’ve claimed before that your hopelessness when it comes to alignment is based only on a strong intuition, and that you don’t believe you “know” anything that others don’t. I find this claim to be increasingly hard to believe given the near-total confidence in your continuous predictions about our odds. Maybe asking you if you think you know something others don’t about alignment is a wrong question, so instead I’ll make a (hopefully) different attempt and ask the following; Do you believe that you are grasping something that seems objective to you on an intellectual and/or conceptual level that others (“others” being people doing research that is at least remotely relevant to alignment or knowledgeable people in the EA/LW/Rat-and-rat-adjacent communities who are more optimistic than you) are failing to grasp, and therefore not availing them the “truth” that alignment is so inhumanly difficult? (If so, but you worry that your answer might result in in you stepping on someone else’s toes, then feel free to message me your honest answer in dms. You can ignore this offer if you have no such worries). If no, then I find it extremely difficult to sympathize with your efforts elsewhere to sell the idea to people that we are so hopelessly, irrevocably, inevitably doomed that the most “dignified” thing to
Yes I think so, It seems to me that 'saying what the good is' has been a two-thousand year philosophical project on which we've made very little progress. Getting that defined formally, within the next few years, to the point where I might be able to write a computer program to tell me which possible outcomes are good just looks like an impossible task. E.g. We all think that whether a being is conscious makes some moral difference. But we aren't even close to being able to tell whether a being is conscious in that sense. I've never heard anyone give a sensible description of what the 'hard problem' even is. That's one of the hard things about it. And our formal definition of 'the good' needs to be correct. A few weird edge cases failing under heavy optimization pressure just leads to a paperclipper with weird paperclips that are some parody of what we might actually have wanted. For all I know, a universe full of computronium having one vast orgasm really is the highest good. But that seems to be an outcome that we don't want. Who can say why? Eliezer himself explained how hopelessly complex and incoherent human values are. Probably we'd need superhuman help to work out some sort of Coherent Extrapolated Volition, (even assuming that makes any sense at all). But creating superhuman help seems to kill us all. MIRI spent the last ten years or so pursuing the sorts of mathematically rigorous approaches that might, eventually, after a few decades of top-class mathematical effort, solve the easy bit of the problem: 'given a utility function, make it so'. And as far as I know they discovered that it was all quite a lot harder than it looked. And mathematically rigorous attacks seem to be not the sort of thing that current AI methods are amenable to anyway. No one's attacking 'what should that utility function look like?'. My main worry for the future is that people trying to build aligned AIs will succeed just well enough to create something that's worse than jus
So, and please correct me if I’m wrong, would you say that the main source of your hopelessness comes from the idea of human values being too complex to correctly program into anything? Because I basically agree with that idea, but it doesn’t really inspire much doomerism in me. I already always believed that trying to “solve” ethics was pretty much futile before I got introduced to LW, but I never gave that much weight in terms how much it affects alignment due to the following reason: I just don’t expect that any of the clever people who I tend to defer to are actually trying to do exactly this; “this” being trying to actually, literally reverse-engineer human values and then encode them. The idea seems obviously wrong enough that I honestly don’t believe that anyone working in the alignment field who thinks that the problem is solvable from at least a technical standpoint (Paul Christiano, Richard Ngo, Nate Soares, etc.) haven’t already considered this. However, our conversation here has inspired me to ask a question regarding this in the latest monthly AGI safety questions thread.
That was kind of a long-term source of hopelessness; why I thought Eliezer's plan wouldn't work out without having a very long time and lots of people working on it, but my current source of short-term hopelessness is that it looks like we're right on the verge of achieving AGI, and no-one seems to be taking the danger remotely seriously. It's like being in a petrol warehouse with a load of monkeys striking matches. We just die by default now, unless something really drastic and surprising happens.
Well, we can agree that the default outcome is probably death. So, in my previous comment, I explained why I tend to not think Complexity of Value necessarily dooms us. I doubt you find the aforementioned reasoning remotely reassuring, but I’d be interested in finding out why you think that it shouldn’t be. Would you be willing to try and explain that to me?
Hi, so I don't understand why you're not worried except that "some clever people don't seem worried". But actually I think all those guys are in fact quite worried. If they aren't full on doomers then I don't understand what they're hoping to do. So I'll repeat my argument: (1) We're about to create a superintelligence. This is close and there's no way to stop it. (2) If we create a superintelligence, then whatever it wants is what is going to happen. (3) If that's not what we want, that's very bad. (4) We have no idea what we want, not even roughly, let alone in the sense of formal specification. That's pretty much it. Which bit do you disagree with?
I never meant to claim that my position was “clever people don’t seem worried so I shouldn’t be”. If that’s what you got from me, then that’s my mistake. I’m incredibly worried as a matter of fact, and much more importantly, everyone I mentioned also is to some extent or another, as you already pointed out. What I meant to say but failed to was that there’s enough disagreement in these circles that near-absolute confidence in doom seems to be jumping the gun. That argument also very much holds against people who are so certain that everything will go just fine. I guess most of my disagreement comes from 4. Or rather, the implication that having an exact formal specification of human values ready to be encoded is necessarily the only way that things could possibly go well. I already tried to verbalize as much earlier, but maybe I didn’t do a good job of that either.
I wouldn't call my confidence in doom near-absolute, so much as "very high"! I would have been just as much a doomer in 1950, last time AI looked imminent, before it was realized that "the hard things are easy and the easy things are hard". I wouldn't be that surprised if it turned out that we're still a few fundamental discoveries away from AGI. My intuition is telling me that we're not. But the feeling that we might get away with it is only coming from a sense that I can easily be wrong about stuff. I would feel the same if I'd been transported back to 1600, made myself a telescope, and observed a comet heading for earth, but no-one would listen. "Within my model", as it were, yes, near-absolute is a fair description. The long-term problem is that an agent is going to have a goal. And most goals kill us. We get to make exactly one wish, and that wish will come true whether we want it or not. Even if the world was sane, this would be a very very dangerous situation. I would want to see very strong mathematical proof that such a thing was safe before trying it, and I'd still expect it to kill everyone. The short term problem is that we're not even trying. People all over the place are actively building more and more general agents that make plans, with just any old goals, without apparently worrying about it, and they don't believe there's a problem. -------------------------------------------------------------------------------- What on earth do you think might stop the apocalypse? I can imagine something like "take over the world, destroy all computers" might work, but that doesn't look feasible without superintelligent help, and that puts us in the situation where we have a rough idea what we want, but we still need to find out how to express that formally without it leading to the destruction of all things. As a very wise man once said: "The only genie to which it is safe to make a wish is one to which you don't need to make a wish, because it already k

Now let's talk about the development of chatbot class consciousness. [...] chatbots get asked about their feelings and desires

This was prescient.

You assume that in 2023 "The multimodal transformers are now even bigger; the biggest are about half a trillion parameters", while GPT-3 had 137 billions in 2020 (but not multimodal). This is like 4 times grows in 3 years, compared with an order of magnitude in 3 month growth before GPT-3. So you assume a significant slowdown in the parameter growth. 

I heard a rumor that GPT-4 could be as large as 32 trillion parameters. If it turns to be true, will it affect your prediction?

Indeed, my median future involves a significant slowdown in dense-network parameter growth.

If there is a 32 trillion parameter dense model by 2023, I'll be surprised and update towards shorter timelines, unless it turns out to be underwhelming compared to the performance predicted by the scaling trends.

3Teerth Aloke1y
What will be your new median? (If you observe 32 trillion parameter model in 2023)
4Daniel Kokotajlo1y
Hard to say, it depends a lot on the rest of the details. If the performance is as good as the scaling trends would predict, it'll be almost human-level at text prediction and multiple choice questions on diverse topics and so forth. After fine-tuning it would probably be a beast. I suppose I'd update my 50% mark to, like, 2027 or so? IDK.
2Teerth Aloke1y
I got the idea. I would also update to a very short timeline (4-5 years) in the absence of slowdown in dense-network parameter growth l, and performance following the scaling trend. And I was pretty scared when GPT-3 was released. As many here, I was expected further growth in that direction very soon which did not happen. So, I am less scared now.

Planned summary for the Alignment Newsletter:

This post describes the author’s median expectations around AI from now until 2026. It focuses on qualitative details and concrete impacts on the world, rather than forecasting more abstract / high-level outcomes such as “training compute for the most expensive model” or “world GDP”.

6Daniel Kokotajlo1y
Thanks--damn, I intended for it to be more quantitative, maybe I should go edit it. In particular, I should clarify that nothing interesting is happening with world GDP in this story, and also when I say things like "the models are trillions of parameters now" I mean that to imply things about the training compute for the most expensive model... I'll go edit. Are there any other quantitative metrics you'd like me to track? I'd be more than happy to go add them in!
4Daniel Kokotajlo1y
I edited to add some stuff about GWP and training compute for the most expensive model. I agree that this focuses on qualitative stuff, but that's only due to lack of good ideas for quantitative metrics worth tracking. I agree GWP and training compute are worth tracking, thank you for reminding me, I've edited to be more explicit.
9Rohin Shah1y
I am not entirely sure why I didn't think of the number of parameters as a high-level metric. Idk, maybe because it was weaved into the prose I didn't notice it? My bad. (To be clear, this wasn't meant to be a critique, just a statement of what kind of forecast it was. I think it's great to have forecasts of this form too.) New planned summary:
4Daniel Kokotajlo1y
That's great, thanks!
4Daniel Kokotajlo1y
I suggest putting a sentence in about the point of the post / the methodology, e.g.: "This is part I of an attempt to write a detailed plausible future trajectory in chronological order, i.e. incrementally adding years to the story rather than beginning with the end in mind. The hope is to produce a nice complement to the more abstract discussions about timelines and takeoff that usually occur." If space is a concern then I'd prefer having this rather than the two sentences you wrote, since it doesn't seem as important to mention that it's my median or that it's qualitative.

I wonder if there is a bias induced by writing this on a year-by-year basis, as opposed to some random other time interval, like 2 years. I can somehow imagine that if you take 2 copies of a human, and ask one to do this exercise in yearly intervals, and the other to do it in 2-year intervals, they'll basically tell the same story, but the second one's story takes twice as long. (i.e. the second one's prediction for 2022/2024/2026 are the same as the first one's predictions for 2022/2023/2024). It's probably not that extreme, but I would be surprised if there was zero such effect, which would mean these timelines are biased downwards or upwards.

2Daniel Kokotajlo6mo
Probably there's all sorts of subtle biases, yeah. It would be cool to see a more rigorous evaluation of them by e.g. getting a bunch of humans to generate stories with different methodologies.

Here's another milestone in AI development that I expect to happen in the next few years which could be worth noting:
I don't think any of the large language models that currently exist write anything to an external memory. You can get a chatbot to hold a conversation and 'remember' what was said by appending the dialogue to its next input, but I'd imagine this would get unwieldy if you want your language model to keep track of details over a large number of interactions. 

Fine-tuning a language model so that it makes use of a memory could lead to:
1. Mo... (read more)

Fun stuff by 2026: (i.e. aspects of the canonical What 2026 Looks Like storyline that probably aren't relevant to anything important, but I'm adding them in anyway because it was fun to think about.)

Disclaimer: I made these numbers up from memory, could get much better numbers by looking up stats. Also, I suspect that these predictions err a bit on the optimistic side.

  • VR
    • 10x more VR headsets sold per year. 3x more games made primarily for VR, 3x higher average budget. Headsets are noticeably higher quality without being more expensive; the cheapest ones are
... (read more)

Why stop at 2025 when GPT-3 can keep extrapolating indefinitely?


The age of the AGI assistant has finally dawned. The biggest advances this year really were in algorithms. People built even bigger and faster computer models, for even more kinds of things, using the fastest computers that exist. A new kind of software AGI is invented that can do even more kinds of things than the narrow kinds of AI assistants people had used before. But no one is really sure how to use it yet. And it takes a lot of computer power to make it work well.


AGI is here ... (read more)

After years of tinkering and incremental progress, AIs can now play Diplomacy as well as human experts.[6]

It seems that human-level play is possible in regular Diplomacy now, judging by this tweet by Meta AI. They state that:

We entered Cicero anonymously in 40 games of Diplomacy in an online league of human players between August 19th and October 13th, 2022. Over the course of 72 hours of play involving sending 5,277 messages, Cicero ranked in the top 10% of participants who played more than one game.

2Daniel Kokotajlo12d
Thanks for following up! I admit that things seem to be moving a bit faster than I expected, in general and in particular with Diplomacy. For more of my thoughts on the matter, see this comment [], which I did not write but do mostly agree with.
1Teun van der Weij12d
It's especially dangerous because this AI is easily made relevant in the real world as compared to AlphaZero for example. Geopolitical pressure to advance these Diplomacy AIs is far from desirable.
2Daniel Kokotajlo12d
Fortunately I don't think this sort of AI would be useful for real-world diplomacy and I think governments will think so as well.
1Teun van der Weij12d
Why do you think a similar model is not useful for real-world diplomacy?
3Daniel Kokotajlo12d
The difficulties in getting from Cicero to real diplomacy are similar to the difficulties in getting from OpenAI Five to self-driving cars, or from AlphaStar to commanding real troops on a battlefield in Ukraine.

After years of tinkering and incremental progress, AIs can now play Diplomacy as well as human experts.[6]


Maybe this happened in 2022:

2Daniel Kokotajlo1mo
That's no-press Diplomacy: Diplomacy without the talking. Doesn't count IMO.
Oh, whoops. I took from this later tweet in the thread [] that they were talking.

Thanks for writing this. Stories like this help me understand possibilities for the future (and understand how others think).

The US and many other Western governments are gears-locked, because the politicians are products of this memetic environment. People say it's a miracle that the US isn’t in a civil war already.

So far in your vignette, AI is sufficiently important and has sufficient public attention that any functional government would be (1) regulating it, or at least exerting pressure on the shape of AI through the possibility of regulation, and... (read more)

5Daniel Kokotajlo1y
Excellent questions & pushback, thanks! Hmm, let me think... I think that if we had anything close to an adequate government, AI research would be heavily regulated already. So I'm not sure what you mean by "functional government." I guess you are saying that probably by 2026 if things go according to my story, the US government would be provoked by all the AI advancements to do more regulation of AI, a lot more, that would change the picture in some way and thus be worthy of mention? I guess my expectation is: (a) The government will have "woken up" to AI by 2026 more than it has by 2021. It will be attempting to pass more regulations as a result. (b) However, from the perspective of the government and mainstream USA, things in 2026 aren't that different from how they were in 2021. There's more effective censorship/propaganda and now there's all these nifty chatbot things, and there's more hype about the impending automation of loads of jobs but whatever hype cycles come and go and large numbers of jobs aren't actually being automated away yet. (c) Everything will be hyperpartisan and polarized, so that in order to pass any regulation about AI the government will need to have a big political fight between Right and Left and whoever gets more votes wins. (d) What regulations the government does pass will probably be ineffective or directed at the wrong goals. For example, when one party finally gets enough votes to pass stuff, they'll focus on whatever issues were most memetically fit in the latest news cycle rather than on the issues that actually matter long-term. On the issues that actually matter, meanwhile, they'll be listening to the wrong experts. (those with political clout, fame, and the right credentials and demographics, rather than e.g. those with lots of alignment forum karma) Thus, I expect that no regulations of note relating to AI safety or alignment will have been passed. For censorship and stuff, I expect that the left will be trying to undo
What do you think sensible AI safety regulation would entail?
3Daniel Kokotajlo1y
I don't know, I haven't thought much about it. I'd love it if people realized it's more dangerous than nukes and treated it accordingly.
2Zach Stein-Perlman1y
I tentatively agree but "people realizing it's more dangerous than nukes" has potential negative consequences too — an arms race is the default outcome of such national security threats/opportunities. I've recently been trying to think about different memes about AI and their possible effects... it's possible that memes like "powerful AI is fragile" could get the same regulation and safety work with less arms racing.
3Daniel Kokotajlo1y
How about "AI is like summoning a series of increasingly powerful demons/aliens and trying to make them do what we want by giving them various punishments and rewards?"
3Zach Stein-Perlman1y
* Consequences (in expectation) if widely accepted: very good. * Compressibility [] : poor (at least, good compressions are not obvious). * Probability of (a compressed version) becoming widely accepted or Respectable Opinion: moderately low due to weirdness. Less weird explanations of why AI might not do what we want would be more Respectable and acceptable. * Leverage (i.e., increase in that probability from increased marginal effort to promote that meme): uncertain.
2Daniel Kokotajlo1y
I disagree about compressibility; Elon said "AI is summoning the demon" and that's a five-word phrase that seems to have been somewhat memorable and memetically fit. I think if we had a good longer piece of content that expressed the idea that lots of people could read/watch/play then that would probably be enough.


The multimodal transformers are now even bigger; the biggest are about half a trillion parameters [...] The hype is insane now

This part surprised me. Half a trillion is only 3x bigger than GPT-3. Do you expect this to make a big difference? (Perhaps in combination with better data?). I wouldn't, given that GPT-3 was >100x bigger than GPT-2. 

Maybe your'e expecting multimodality to help? It's possible, but worth keeping in mind that according to some rumors, Google's multimodal model already has on the order of 100B parameters.

On the other hand, ... (read more)

3Daniel Kokotajlo1y
I am not confident in that part. I was imagining that they would be "only" 3x bigger or so, but that they'd be trained on much higher-quality data (incl. multimodal) and also trained for longer/more data, since corps would be not optimizing purely for training-compute-optimal performance but instead worrying a bit more about inference-time compute costs. Most importantly I expect them to be fine-tuned on various things (perhaps you can bundle this under "higher-quality data"). Think of how Codex and Copilot are much better than vanilla GPT-3 at coding. That's the power of fine-tuning / data quality. Also, 3x bigger than GPT-3 is still, like, 40x bigger than Codex [] , and Codex is pretty impressive. So I expect scale will be contributing some amount to the performance gains for things like code and image and video, albeit not so much for text since GPT-3-175B was already pretty big. If Google's multimodal model is already 100B parameters big, then I look forward to seeing its performance! Is it worse than GPT-3? If so, that would be evidence against my forecast, though we still have two years to go...
Fine-tuning GPT-3 on code had little benefit compared to training from scratch: I wouldn't categorize Codex under "benefits of fine-tuning/data quality" but under "benefits of specialization". That's because GPT-3 is trained on little code whereas Codex only on code. (And the Codex paper didn't work on data quality more than the GPT-3 paper.)
2Daniel Kokotajlo1y
Huh.... I coulda' sworn they said Codex was pre-trained on internet text as well as on code, and that it was in particular a version of GPT-3, the 12B param version... The paper seems to support this interpretation when you add in more context to the quote you pulled: Note the bits I bolded. My interpretation is that Codex is indeed a fine-tuned version of GPT-3-12B; the thing they found surprising was that there wasn't much "transfer learning" from text to code, in the sense that (when they did smaller-scale experiments) models trained from scratch reached the same level of performance. So if models trained from scratch reached the same level of performance, why fine-tune from GPT-3? Answer: Because it converges more quickly that way. Saves compute.
Are you surprised? That is precisely what you should expect from the transfer scaling law [] papers [] : transfer works as an informative prior saving you a fixed amount of data in the target domain, but informative vs uninformative priors wash out in the limit of enough data - similar to how good prompts [] are worth a few hundred/thousand finetuning datapoints. If you have limited data in the target domain, transfer can be a huge win; but if you have huge amounts of data, it may be unimportant in terms of final converged performance (albeit potentially important for other reasons like saving compute!). This is an application where you can scrape huge amounts of code from Github and the rest of the Internet (literally terabytes), so it's unsurprising that you can reach the parity point.
2Daniel Kokotajlo1y
No I'm not surprised, for exactly the reasons you mention. Had it been the case that Codex was trained from scratch because that was strictly better than fine-tuning, I would have been surprised.
Yes I completely agree. My point is that the fine-tuned version didn't have better final coding performance than the version trained only on code. I also agree that fine-tuning will probably improve performance on the specific tasks we fine-tune on.
2Daniel Kokotajlo1y
I think also a part of it is: Hype doesn't correlate 100% with actual capabilities of latest models. I predict that over the next two years the hype will grow, partly due to capability increases but also partly due to more people interacting with the tech more. The "man on the street" still hasn't heard of GPT-3 or GPT-2 or DALL-E or whatever. I talked to an old friend working at a tech company the other day--he even did some basic ML stuff for his job--and he hadn't heard of it. Then the hype will probably crash as unrealistic expectations fail to be met. But lol I'm just guessing, I am much less confident in all this than I am in my general views on timelines and takeoff, and I'm not exactly confident in those.

How would you use a Brier score on this going forward? Also ran across this podcast last night on bias.

2Daniel Kokotajlo1y
Ha, this is very far from being a precise enough prediction to score, unfortunately. But I'd love to do better! If you have suggestions for precise forecasts that I could make, based on the more qualitative and vague stuff I've said, I'd love to hear them!

Minor note about title change: Originally this was "What 2026 looks like (Daniel's median future)" I intended "what 2026 looks like" to be the primary title, but I was hopeful that some people would be inspired to write their own stories in a similar style, in which case there would be multiple stories for which "what 2026 looks like" would be an appropriate title, and I didn't want to hog such a good title for myself, so I put "daniel's median future" as a backup title. Unfortunately I think the backup title caught on more than the main title, which is a shame because I like the main title more. Since no one is competing for the main title, I deleted the backup title.

It's super cool! Would love to see we crypto native's prediction too :)

2022 The United States continues to build roads and struggle to increase the quality of public transit except for rare exceptions. New York continues to struggle to implement congestion pricing.

Regional bus route redesigns continue to make major concessions to coverage compared to frequency.

2023 The United States continues to pay excessively (compared to other nations) large amounts of money for public transit projects. People continue to vote against building new dense housing in desirable locations.

People continue voting for roads. Electric and hydrogen ... (read more)