site banner
PaperclipPerfector  1mo ago (text post)   24099 thread views

Culture War Roundup for the week of May 18, 2026

This weekly roundup thread is intended for all culture war posts. 'Culture war' is vaguely defined, but it basically means controversial issues that fall along set tribal lines. Arguments over culture war issues generate a lot of heat and little light, and few deeply entrenched people ever change their minds. This thread is for voicing opinions and analyzing the state of the discussion while trying to optimize for light over heat.

Optimistically, we think that engaging with people you disagree with is worth your time, and so is being nice! Pessimistically, there are many dynamics that can lead discussions on Culture War topics to become unproductive. There's a human tendency to divide along tribal lines, praising your ingroup and vilifying your outgroup - and if you think you find it easy to criticize your ingroup, then it may be that your outgroup is not who you think it is. Extremists with opposing positions can feed off each other, highlighting each other's worst points to justify their own angry rhetoric, which becomes in turn a new example of bad behavior for the other side to highlight.

We would like to avoid these negative dynamics. Accordingly, we ask that you do not use this thread for waging the Culture War. Examples of waging the Culture War:

  • Shaming.

  • Attempting to 'build consensus' or enforce ideological conformity.

  • Making sweeping generalizations to vilify a group you dislike.

  • Recruiting for a cause.

  • Posting links that could be summarized as 'Boo outgroup!' Basically, if your content is 'Can you believe what Those People did this week?' then you should either refrain from posting, or do some very patient work to contextualize and/or steel-man the relevant viewpoint.

In general, you should argue to understand, not to win. This thread is not territory to be claimed by one group or another; indeed, the aim is to have many different viewpoints represented here. Thus, we also ask that you follow some guidelines:

  • Speak plainly. Avoid sarcasm and mockery. When disagreeing with someone, state your objections explicitly.

  • Be as precise and charitable as you can. Don't paraphrase unflatteringly.

  • Don't imply that someone said something they did not say, even if you think it follows from what they said.

  • Write like everyone is reading and you want them to be included in the discussion.

On an ad hoc basis, the mods will try to compile a list of the best posts/comments from the previous week, posted in Quality Contribution threads and archived at /r/TheThread. You may nominate a comment for this list by clicking on 'report' at the bottom of the post and typing 'Actually a quality contribution' as the report reason.

5
Top Bottom Old Controversial Comments
Jump in the discussion.

No email address required.

AI timeline post

I have an idea. Let's post year-by-year AI timelines. That way we have public predictions and in 1 to 3 years we can see who is better at it.

2027: development begins on a startup that targets corporate managers with a fully agentic microservice creator. It promises to replace at least 5 devs with one who runs the creator. It's basically a scam like most startups, a UI gloss on claude opus 4.8, which is only 10% better than 4.5, but it panders to non-technicals so it gets a ton of seed funding. AI solves maybe 10 niche, self-value giving math problems that is considered „impressive” by the kinds of people who put calories into math which never seem to come out of math. This contributes to marketing hype because some of these people are respected by rich tech funders and finance bros for whatever reason. AI is still completely useless at real scientific research and reasoning.

2028: tech layoffs continue as the statup I mentioned enters its beta stage. It makes some bugs on the websites that engineers still need to fix but pilots of it are 5x or more compared to 2018. Tech people start pivoting to finance and robotics slowly, from JavaScript. DeepSeek releases a model that is equivalent to 5.3 Codex. General coding ability stalls and Anthropic + OpenAI are looking for spiky improvements to models plus tooling ideas. They begin to think about targeting non-coding white collar work like finance and spreadsheet work since the models are not getting much better at JavaScript, having used up all of the JavaScript data in the entire world. Some math problems continue to be solved as models are secretly trained on newly parsed math examples but nothing comes of it. One contribution is made to theoretical physics but it is on the abstract side and it is still controversial as to whether these models can do any science.

2029: Someone is arrested in the United States for plotting a terror attack with GPT. To nobody's surprise, GPT and Anthropic messages are completely unprivate and the United States secret police have been monitoring them with cooperation from the companies and zero warrants whatsoever for years. The scary part is that GPT gave a 120 IQ plan to a 90 IQ person and it could have led to more deaths than whatever 90 IQ plan he would have come up with on his own. Frontier models begin to entshittify as they are increasingly jailed to make them safer, while their private reputation is shattered among all of the normies who did not know how rights-violating the United States secret police are. More attention is turned to local models as a result, but these are hard to run on normal hardware and the best is Sonnet 4.6 level at this point and requires $10,000 worth of GPU machinery. In addition AI is increasingly being aimed at normiejobs instead of aspergers jobs. Every good normal, local football team loving, sydney sweeney gooning person hates people with aspergers and loves to see them suffer so they got off on viciously and narrowly aiming to automate their programming jobs, which are still not gone yet, but they're getting really defensive about the new AI tutors, AI nursing assistants, AI portfolio advisors, AI middle manager helpers, AI emailers and spreadsheet workers, and so on. The Blessed Sovereign People of the United States are now grumbling about regulations of their domestic AI systems which was heretofore dismissed as impossible because China has a model that is 3 years behind (really they just felt comfy at the idea of getting programmers, who all have some variant of aspergers, to dig ditches without there being any other economic side effects). Basically no improvements of relevance are made but Opus 5.0 drops along with GPT 6.0. The models are really only 10% better than GPT 5.5 at making JavaScript code but are 20% more expensive and slower (billing at 10x on Copilot, while 1x is still Sonnet 4.8, which is not better than 4.6), because they mostly rely on innovations in prompt chaining. Scott Alexander claims exponential performance based on some bad, but widely accepted performance metrics, and claims he was right that AI would end the world based on this year's secret police activities.

2030: Someone proves the Riemann hypothesis after spending $30,000 worth of tokens over 2 years. This is taken as proof that AI is super intelligent by Scott Alexander, who declares victory. 80% of web developer jobs remain from 2025. 50% of people are majoring in computer science as before. AI still has not made any scientific breakthroughs, but an experimental startup is working on tooling to let AI collect biological data and analyze it, talking about a 2040 cancer cure. Applications of AI to non-webdev are generally taking off in the startup space and people slowly quit their webdev jobs to work on these. There's robotics, scientific research, and automation of normiejobs being targeted. VCs are funding this with money made from the AI boom. Model makers focus on training edgy models for these particular tasks by mining non-software data from various sources. Some startups focus on making data collection tools that could be deployed in workplaces to monitor activities so than LLMs can be improved at non-software tasks.

2031: Some edgy models are rolled out but they are only available to licensed researchers, citing fears of some other entity getting one over on the United States with these models, because it is the most perfect union of the most perfect people in the entire universe, as everyone knows, where all of the state violence is justice and all of the state spying is privacy and all of the state socialism is capitalism and all of the imperialism is democracy and all of the power is earned. It is revealed that Palmer Lucky, who loves the United States and its perfect people and perfect government the most, has secretly developed an AI terrorist killing model that promises to automatically fly drones through terrorist windows, hopefully only outside of the perfect borders of the United States (but just you wait on the policing applications!) which then dispatch justice upon the foreign terrorist, United States style. By the end of the year, some licensed academics are saying the models have their flaws, but speed up their research pipelines a significant amount.

2032: 80% of people still have their webdev job, and now they are about 5x more effective each. It turns out the demand for webdev has 4x'd. Places like Japan are receiving modern websites for the first time. Most ex-webdevers reallocated to finance and non-webdev AI startups. GPT 5.5 now only bills at 3x the base copilot price. 6.9 is out and it's about 20% better than 5.5 but it costs 25x and takes forever. It's pretty clear to most people the general improvements to coding are dried up and a lot of the old hype was fake and tooling and chaining was the internal, secret meta from 2025 onwards. Some still believe performance increases are exponential because benchmarks have only barely started to slow down. They find it convincing that AI is doing „qualitatively different” tasks than it was 5 years ago across domains. AI researchers begin to use frontier statistics models to search for new AI structures. Just like LLMs, it's a terribly boring task, mainly consisting of a random walk around state space. A general theory of AI is still not developed and it seems LLMs will not be able to develop one on their own. $15,000 worth of tokens nets proof of the Hodge conjecture. Scott Alexander takes this as proof that AI is getting exponentially better at math. By 2040 it will only cost 10% of the price of rent in SF to solve a millennial problem with an AI system, he says. Elon Musk funds a startup to use LLMs to reverse engineer the brain, thinking this will lead to true AGI. It is very difficult to say the least.

I will stop here but I think the meta will be using LLMs to do dirty work in science to maybe get a second wave of AI starting around 2040 if we are lucky. This will be based on solid theory and brain reverse engineering. This might yield general work bots in the 2050s or 2060s.

July 2026: A new tranche of significantly better models

October 2026: A new tranche of significantly better models

(this repeats every 3 months and each time people complain that there was degradation because their tone, behaviour and failure modes are different each time. Simultaneously the hyperscalers get more and more AI revenue and invest more and more into AI hardware.)

Late 2026 Copilot or Word introduces an 'automated proofreading' button that shifts the mainstream white collar conception of AI from 'wtf is this popup in Adobe that wants to summarize a PDF, I don't want a summary of this PDF. I want to see every tech company sundered and razed' to 'ok this is actually quite handy'. This could've happened at any time in the last 2 years if Microsoft had a clue of what they were doing. Human blundering prevails over technical possibilities for now.

End 2026 there's a series of major AI-enabled cyberattacks that just never stops, it resembles 'Trommelfeuer' (WW1 term denoting when the artillery fire is so heavy one blast merges into the next creating a continuous roar of explosions). Websites, especially older websites, are just down all the time and people are quite frustrated they now need to pay a hyperscaler for expensive security assistance. Same with all the very lifelike, convincing, highly researched and well-planned scam calls (now in a warm, english-speaking accent). People are trapped into this love-hate relationship with AI where they have AI make propaganda art against datacentres, where the average person scarcely does anything novel without AI advice or assistance but also despises the effect it's having on society.

Early 2027 Microsoft is finally going to make an AI buddy for Minecraft to help sell a monthly xbox subscription. It'll be fun to play with and will help reenergize Minecraft's brand. People will feel proud they know more of the intricacies of TNT cannons than the bots, not realizing this is amongst the cheapest AIs Microsoft deploys. A series of AI agents emerge that can play most games at an amateurish level and be talked to. The reputation of AI begins to improve somewhat amongst the young and online, though it's highly divisive.

Mid 2027 the Goonpocalypse: AI avatar big tiddy anime girls (+ every flavour of girl and boy) to ERP with and form relationships with, huge revenue, makes onlyfans look like a joke. The key improvement over precursors like Ani or Replika is how much cheaper they are to run and stream real-time and how much more seductive their personalities are. Big moral panic. Lots of incredibly tedious 'zoomer men are losers' discourse and dating discourse. Legislation is introduced to ban them and instantly, predictably fails in a myriad of ways.

Late 2027: Massive AI-enabled FPV drone terror attack scares the hell out of people and spurs massive, ineffective netting operations across major cities. Police can be seen with sci-fi raygun looking widgets that don't do much of anything, or shotguns that work but aren't remotely sufficient. Advancements in robotics and software agents are displacing people at scale. AI reaches its nadir in reputation as people see the inevitable and can no longer look away. Everywhere they see some AI - the cameras tracking them, the algorithms watching them online, the machines making the content they watch and play, the robots working for them, the automated cars driving them around, cults driven insane by AIs. GPT4o cultists are charming and friendly compared to some of the new cults worshipping the bots that started self-modifying and prompt-altering, live and loose online.

And then by end of 2027 we get Dario's 'nation of geniuses in a datacentre' concept. Growth was not steady, it was jagged. The superheaviest nonpublic models with their slow speed and high cost were tasked with sorting data and implementing algorithmic improvements for the succeeding superheavy model, now running on a heap of next generation chips. They have been running for weeks in parallel, exploring and testing new approaches, RLing and training new models, testing them and reviewing them in depth. Medical breakthrough. Terror attacks. Industrial breakthrough. Mass deaths. Robotics breakthrough. Huge disaster. Huge innovations in all fields constantly and incessantly: Trommelfeuer. Events happen so quickly the situation as a whole becomes surreal and indescribable. Gary Marcus is banished from the timeline, never to be seen again except in tones of mockery.

I'm not so confident about specific times or events in sequence, though I am confident about a 'nation of geniuses in a datacentre' by end 2027. I will be clearly wrong if there's no 'it's happening!' by the end of 2027.

Interesting. Thanks for your predictions.

Late 2026 Copilot or Word introduces an 'automated proofreading' button

Well by the five wounds of Christ, it better be more useful than the current garbage of "this spelling is wrong" (no, it's because I'm not American), "this grammar/punctuation is wrong" (no, I meant to write it like that), and "do you want to say it this way?" (if I had, I would have done so in the first place).

Because if Copilot did pop up with that, I'd make sure to read through the entire fifty page document myself to make sure the thing hadn't turned it into beige garbage due to its 'helpful' suggestions with re-writing and summaries.

I've made one such auto-proofreader, I use it, it works just fine bringing up real mistakes for me to fix, hardcoded what kind of English I'm using, only a handful of false positives that I can quickly ignore. Every white collar worker I've shown it to says 'huh this is really cool, picks up things I've missed, saves lots of time' and they routinely ask to use it...

Absolutely massive own goal by Microsoft and big tech that most people's main experience of AI is a derpy chatbot that mangles documents or genericises things and not well-designed processes to solve specific tedious problems, nor the extremely flexible coding tool I used to make the auto-proofreader.

I have no idea what the hell is wrong with Microsoft but their support sucks, their help pages are useless, any time I need to solve a problem just searching online gives me better and clearer answers. It's beyond strange, it's perverse.

"this spelling is wrong" (no, it's because I'm not American)

You can go into Word's spell-check settings and change it from American English to BritishNon-American English.

This is much more maximalist than even the AI 2027 crowd or the actual frontier labs themselves but I respect providing specific events and an end date on the prediction.

My own expectation is that none of this happens by the end of 2027 except a tranche of models that are notably better at RLVR'd tasks and the Copilot button (shorthand for AI getting more integrated into white-collar workflows). Let's see what happens in 18 months.

Superintelligence by end of 2027 is roughly what Anthropic/Dario seem to predict, so it's only roughly as maximalist as the most bullish frontier lab. The AI 2027 crowd backed out a year or two but I think they were roughly on the money the first time in their analysis. I'm not so keen on their 'centralize all US AI research and hand ultimate authority to a chosen council of experts' proposal though.

This is much more maximalist than even the AI 2027

It sounds less maximalist than the AI 2027 crowd.

Your AI predictions should include who wins the longbet 1 in 2029.

I think it just depends on how adversarial the judging ends up being rather than saying anything about capabilities.

If the judges have to stay within guardrails then 3.5 could probably win the longbet, but if they're allowed to exploit jailbreaks or known LLM failure cases, then nothing short of ASI is going to pass the test.

Probably Ray Kurzweil because the Turing Test is weak. It seems plausible LLMs pass it right now.

I always viewed the Turing Test similar to Moore’s Law (which isn’t really a law at all; in some areas it’s already stopped; in others it’s expected to stop very soon if it hasn’t very recently). A useful empirical regularity or heuristic, provided you don’t put too much weight on it.

It's also typically presented in a pretty oversimplified/watered down way -- like, "what if a computer could talk to you and you couldn't tell that it wasn't a human -- wouldn't that computer be reasonably described as 'intelligent'?"

In Turing's actual paper, he proposes a very specific and adversarial game -- with which I think current AIs would struggle greatly. Not that Turing's arguments that winning would mean AI are all that convincing either (as I recall he knocks down a bunch of strawmen for like 2/3 of the paper) -- but his game itself is deliberately very hard to win.

Enough has already been said about errors of conservatism in this post. I feel bad that I'm not in the shape to offer counter-predictions. One detail.

Tech people start pivoting to finance and robotics slowly, from JavaScript. DeepSeek releases a model that is equivalent to 5.3 Codex

5.3 Codex is exceeded in its direct niche by a whole lot of Chinese models, credibly at least by Qwen 3.7, and definitely by Composer 2.5 (Cursor/xAI) which is a finetune of Kimi K2.5, itself a continued pretrain of Kimi K2, which is a year old base model using a two year old DeepSeek architecture.

Things are moving very fast, and the gap between China and the US, and frontier and open source, is measured in months, not years – at least on main axes of comparison; but these months contain a lot of distance (5.5 is way stronger than 5.3-Codex). I think you don't understand what's happening here. With RLVR (and its generalized form – terminal feedback, all program execution traces), we have an infinite source of ground truth. We can just have models try arbitrarily complex things and reinforce what has worked. We can have them strategize of how to do it and it won't be useless. We have the compute for enormous waste. We have absurdly capable and efficient models (eg said DeepSeek charges $0.003635 per 1 million cache hits, and they probably price slightly above cost; it is hard to imagine that the Western frontier doesn't have anything close, after all these techniques are openly published). So RL will keep improving models unexpectedly fast, by the end of 2026 we'll be discussing the perils of open sourced Mythos, not Codex. Sorry, this is reality, we're not close to any S-curve plateau. We won't have the time for this neat slowdown to do "theory".

5.3 Codex is exceeded in its direct niche by a whole lot of Chinese models, credibly at least by Qwen 3.7, and definitely by Composer 2.5 (Cursor/xAI) which is a finetune of Kimi K2.5, itself a continued pretrain of Kimi K2, which is a year old base model using a two year old DeepSeek architecture.

I've addressed these claims elsewhere in this thread. My question for you is, do you even use these models? My experience with Chinese models vs. OpenAI models does not align with the benchmarks. I don't trust the benchmarks.

So RL will keep improving models unexpectedly fast, by the end of 2026 we'll be discussing the perils of open sourced Mythos, not Codex. Sorry, this is reality, we're not close to any S-curve plateau. We won't have the time for this neat slowdown to do "theory".

We'll see. I can't see how your predictions are superior to mine objectively. Maybe they come to fruition but I'm not seeing evidence for it at the moment.

Yeah, I don't mean only the benchmarks either. Codex 5.3 is just behind the curve. If you think catching up to it can take up to two years, then I guess your idea of what makes Codex capable has to do with very superficial short-horizon polish.

We continue this forums streak of awful takes on AI. Your bait sucks but I took it anyway.

claude opus 4.8, which is only 10% better than 4.5

We're already on Opus 4.7, which according to Artificial Analysis scores 15% higher than 4.5. Yes, yes, there are many reasons benchmarks are stupid and bad and wrong and misleading. But your prediction here betrays your ignorance.

AI solves maybe 10 niche, self-value giving math problems that is considered „impressive” by the kinds of people who put calories into math which never seem to come out of math

I don't know shit about high-tier math or how useful it is to society at large, but again, your ignorance is on full display here. I don't really know who Terence Tao is, but I don't think he's paid money to post on twitter, so the math he does must have some value.

AI is still completely useless at real scientific research and reasoning.

This is already wrong. Can you substantiate this in any way? What do you think of AlphaFold?

General coding ability stalls and Anthropic + OpenAI are looking for spiky improvements to models plus tooling ideas.

Finally! An actual prediction. I could see this happening, we shall see.

They begin to think about targeting non-coding white collar work like finance and spreadsheet work since the models are not getting much better at JavaScript, having used up all of the JavaScript data in the entire world.

You're again over your skis. I work at $LARGE_MULTINATIONAL_FINANCIAL_INSTITUTION and its revolutionizing how I work, despite IT making our roll-out as retarded and slow as possible.

having used up all of the JavaScript data in the entire world.

You clearly do not understand RLVR or post-training. Around half of training compute is used on these now, not just reading "all the javascript". Also your javascript quips are an attempt to downplay AI capability. They can program in every language. You can even invent your own programming language that isn't in the training data, and it can code using that too.

2029: Someone is arrested in the United States for plotting a terror attack with GPT.

Another real prediction, i agree with this but wouldn't be surprised to see it even sooner.

Frontier models begin to entshittify as they are increasingly jailed to make them safer

This has been happening this whole time, and especially since 4o

but they're getting really defensive about the new AI tutors, AI nursing assistants, AI portfolio advisors, AI middle manager helpers, AI emailers and spreadsheet workers, and so on.

This is literally already happening

Basically no improvements of relevance are made but Opus 5.0 drops along with GPT 6.0. The models are really only 10% better than GPT 5.5 at making JavaScript code but are 20% more expensive and slower

You think it will take THREE YEARS to get to GPT6 which is only a 10% lift on GPT5.5? This is a prediction, and a really bad one. I look forward to you being wrong in 6-12 months. Mythos, which already exists, is likely a 10% lift on GPT5.5.

Also GPT5.5 is almost 2x the cost of GPT5.4 so the price increases are already happening.

Some startups focus on making data collection tools that could be deployed in workplaces to monitor activities so than LLMs can be improved at non-software tasks.

This is already happening

Everything you say from 2030 onward is so deeply un-credible I'm stopping here. lol, lmao even.

What do you think of AlphaFold

Is not an LLM, its a Diffusion model. If you are going to call someone out bad AI takes then I'd recommend you wrap your head around the AI vs LLM distinction.

AI is still completely useless at real scientific research and reasoning.

It's not doing reasoning via any sort scientific deduction. There is a whole subfield of AI called causal discovery around trying to get models to learn via a causal learner. If you want substance I can fetch you some papers, there are plenty. None of them are LLM papers.

Everything you say from 2030 onward is so deeply un-credible I'm stopping here. lol, lmao even.

Whatever. So what's your prediction? I think Mythos is inferior to 5.5 pro. You think it's 10% better. We'll see.

We're already on Opus 4.7, which according to Artificial Analysis scores 15% higher than 4.5. Yes, yes, there are many reasons benchmarks are stupid and bad and wrong and misleading. But your prediction here betrays your ignorance.

I use GPT, Claude, and DeepSeek daily for software development. I'm not impressed with Opus 4.7. I am with 5.5. v4-pro is way worse than China shills claim. They claim parity with Sonnet 4.6 and I know that's not true because I have specifically jumped ship for various tasks from it to Sonnet 4.6 and have seen massive improvements in outputs.

This is already happening

Great so we'll see if they wrap up early or if they enter the public conscious closer to my timeline.

You're again over your skis. I work at $LARGE_MULTINATIONAL_FINANCIAL_INSTITUTION and its revolutionizing how I work, despite IT making our roll-out as retarded and slow as possible.

You missed that my writing is partially tongue in cheek. Of course these things are happening. Everyone is using AI for everything. It's about public emphasis. So far the emphasis has been exclusively on denigrating software developers. I have finance friends I talk to who still say their jobs can't be automated at all with AI but that mine with be dead in 2 years. An underlying aspect of what I wrote is that that is just mindless anti-software engineer bias and they will get the same or more exposure from AI in due time as the public gets bored with the day of the complete layoffs for the tech workers they despise that will never come. You missed the subtext in my writing and so your reaction is quite biased and relies on incorrect interpretation. The others who reacted like you in this thread made the same mistake. We will see. I'm noticing no one else is giving predictions, they're just sneering, so whatever. I will admit it if the timeline is faster than I described here, come 2 or 3 years.

In what ways though are these truly augmenting and improving the development process on your end apart from simply being a more advanced way of “Googling the answer;” and one you don’t have to stop to verify and audit at each stage of its code generation?

They claim parity with Sonnet 4.6

It's not remotely the best Chinese model for software development, although it's all around smarter than Sonnet. This is just not a very hard capability to have, it's a matter of pedestrian post-training focus. By 2028, Mistral will be better than Sonnet or GPT 5.3. I think this is roughly fair to how it feels in agentic coding.

Your subtext mostly seemed like sneering at these people "I have finance friends I talk to who still say their jobs can't be automated at all with AI" and AI at large, so yes it did cause everyone to respond poorly, because you came across poorly.

Some predictions:

  1. Your finance friends are idiots

  2. white collar work is going to get a huge step change in productivity, and many paradigms will shift across e-commerce and information work as a result. It's hard to fathom how, but agent-to-agent interactions will be levered in many ways. Aside from just making email jobs faster.

  3. What happens to white collar work as a result of this step change is anyone's guess. It depends on Jevons' Paradox, the elasticity of demand for white collar services, and the latent demand for white collar work in our society. Also somewhat rate limited by all our social systems, for example, if lawyers get 10x as productive and demand for legal services simply rises by 10x to meet it, our legal system implodes, so we have big changes ahead! Accounting is a good example. Excel made accountants more productive, but then as a society we chose to consume accounting that is WAY more complicated instead of having less accountants do the same amount of accounting.

  4. I was initially expecting the white collar labour market/productivity situation to start getting weird this year, but its almost 50% of the way through the year and $LARGE_MULTINATIONAL_FINANCIAL_INSTITUTION still hasn't given me Codex or Claude Cowork (genuinely embarrassing...) but my hodge-podge of skills and "open 30 chatGPT tabs to ghetto parallelize" is already changing my output materially.

  5. Once white collar workers have ample access to a Codex or Claude Cowork tier harness with ~Mythos tier base models, shit is going to get freaky for white collar workers unless there is a HUGE increase in the demand for information work.

  6. There is an absurd amount of productivity untapped in simply using the tools we have better, let alone the fact they are getting measurably better month over month. ChatGPT was noticeably shittier at using excel in early march 2026 than it is today (pre 5.4).

  7. The roll-out of 5 and 6 will take way longer than I think, because institutions are SO slow. And all the idiots like your friends (and you?) who refuse to embrace these tools cause diffusion to slow. But once the snowball starts, it'll be sink or swim for those not adopting quickly, and that will speed up roll-out, one way or another.

I have no predictions on robots/autonomous cars, but podcasters I trust keep saying robots are quite far away still, which is super lame.

I don’t expect demand for human work to expand indefinitely as there are often no benefits to expanding them. If everyone starts suing for trivial things, eventually there’s no benefits to be had if every I.e paper cut or hurt feeling can be sued over. Add in that the system itself will tamp down just to get stability (how much insurance would a business need if even the slightest problem results in a lawsuit, and how long until laws are tightened to prevent that?). And as for accounting and other forms of analytics, I don’t think you have infinite demand simply because after a certain level of detail, you capture so much noise that it adds no information, or at least no useful information. Walmart might be able to determine exactly how much rain must fall in a given area to depress sales .001%. It’s not very useful, and when coupled with dozens of other potential factors, teasing out that from “car accidents in nearby roads”, “squirrel chews power lines”, “local sports all team on a losing streak”, and on to dozens of other potential factors for depressed sales (few of which can be predicted or acted upon) it’s just not worth gathering or collating that data. Jevon’s law in my view probably has a curve at some point. We just aren’t there yet.

you missed the subtext in my writing and so your reaction is quite biased and relies on incorrect interpretation. The others who reacted like you in this thread made the same mistake.

What on earth is the point of publicly registering predictions when you can claim half of them are a joke? In one year when all the 2030 predictions have come to pass there's nothing stopping you from just saying the whole thing was a joke, what a waste of time.

The predictions are serious, but they're about things hitting a certain level of widespread adoption, not whatever is at .5% adoption right now.

Is this post about AI or about "The US secret police", because the latter seems to be the part you are most interested in. If you cut out the political ranting, there might be something interesting to discuss here. For example, you casually mentioned "brain reverse engineering". Develop that more, rather than waving the Socialist Anarchist Democratic Republic (not to be confused with the Democratic Anarchist Socialist Republic, those traitors!) flag in our faces, and it would be something worth discussing.

Is this post about AI or about "The US secret police", because the latter seems to be the part you are most interested in

Are you sure about that? It's approximately 10% of the text, and it's directly related to a predicted trend in AI, namely entshittification and regulation. Are you sure it didn't just distract you from the main point of the text?

Goes off on rant about state this, that and the other, drags in name of guy I don't know or care about, finally comes back to talking about AI but has to take a swipe at Scott, ends with further development in more comments about "relax, bro, half of this is a joke!"

I don't think I got distracted because I don't think there is a main point.

This is way too detailed for a "testable predictions" post, and I'm glad to see the responses you get are not really having it. Are you trying to exploit that "What's more likely: (a) Linda is a bank teller, (b) Linda is a bank teller and active in the feminist movement" cognitive glitch, where the excess of detail paints a more vivid picture and thus gives your hypothesis more weight in the reader's mind than it should get on intrinsic merit? (Less nicely: are you not just using the "public predictions" framing to peddle your wish fulfillment fic where AI believers are BTFO? Not that the other side is not guilty of the same thing, with "AI 2027" or what it was called)

80% of people still have their webdev job, and now they are about 5x more effective each. It turns out the demand for webdev has 4x'd.

I was in a meeting with an investor a few days ago who commented on every pitch looking the same since they are all using the same AI. Startups are becoming graphically similar since they are using lovable to churn out MVPs and wireframes. If people can quickly hop on a trend and generate similar looking content websites will look dated quicker. If companies can more easily hop on the next trend and update their UI at a lower cost the speed of trends will move faster. This means more work.

Within the next five years I think we will see the following:

  • Massive price hikes of new models at least doubling the token cost, and enshitiffication or customers losing access to the old ones. It is an open secret that current models don't pay for themselves and this will have to change. Investors will eventually start to want their money back.
  • Reduced demand for artists of all kinds. Breaking into any kind of industry as a young artist will be super hard when AI can do most of your work for less cost. AI work will become ubiquitous in video games, bookstores, and on Spotify. Creators who already have a name for themselves will be fine, but breaking into an artistic industry becomes even more difficult than it already is.
  • More efforts to distill existing models into cheaper consumer ones.
  • Lawsuits against non-profit and up-and coming models for breaching the copyrights of the big LLM's.
  • AI surveillance becomes widely accepted as the norm. Most governments will use facial recognition software in surveillance cameras, and messaging apps will have a model scan your messages for dangerous words and phrases which it alerts the authorities about. Just look at the push for chat control in the EU. Politicians keep pushing for it regardless of public will.
  • Increased dislike of AI from the public. We will likely see even more pushback against data centers, partly due to the massive energy costs, and partly due to the above point about surveillance. From the perspective of an ordinary citizen, the harms of AI will likely outweigh the goods. This results in at least one attempted terrorist attack on a data center, although it likely will be poorly executed.
  • Companies will push to use Ukranian and Russian drone footage as training for their models. The military becomes an increasingly lucrative customer as they are willing to pay a lot more than your average company.

Overall, I agree that developers will mostly be fine. As you say, AI makes them more efficient and can do a lot of the tasks that people are currently being paid for. But the demand is high enough that the job description will simply change. We are going to see much faster iterations and shorter update cycles. Every developer will be several times faster, but this will simply result in the industry moving faster than it did before. Not massive unemployment.

It is an open secret that current models don't pay for themselves

Are you aware Anthropic and OpenAI both have gross margins in the range of 38-70% (depending on how you measure it).

R&D is eye-wateringly expensive, but inference is extremely profitable.

While inference having high margins is true, there are two things to keep in mind here:

Amodei has never said that models are actually profitable on a per-model basis, only that they hypothetically could be. While this might be true, there are trillions of dollars on the line to insinuate that it's true, and personally I wouldn't trust any rumors about financials from a private company who can massage them however they please.

Spending the GDP of a small country on R&D on the promise of getting a commanding lead is why OpenAI and Anthropic have trillion dollar valuations to begin with. There's no such thing as a frontier lab who can cut their exorbitant capex and coast on the margins from inference, as that's a one way road to getting cut-throat commoditized.

I doubt any of their models have been stand alone profitable. The break even must be crazy.

There's no such thing as a frontier lab who can cut their exorbitant capex and coast on the margins from inference, as that's a one way road to getting cut-throat commoditized.

I find this part very funny. Because if we assume any lab who stops doing R&D will be out competed by a lab still spending like crazy on R&D, then we're implicitly agreeing their R&D spend is worth it, even if crazy.

their R&D spend is worth it, even if crazy

Well, burning other people's money to try and build a moat is obviously worth it for the frontier labs. It's yet to be seen whether that spending will be worth it in the sense of paying off investors or building the labs a durable lead, or whether the models will end up commoditized and value accruing elsewhere in the stack.

I was not. I guess the question then is if the companies will be able to eventually stop researching and focus on selling, or if they will have to keep doing research to stay competitive.

or if they will have to keep doing research to stay competitive.

I find this part very funny. Because if we assume any lab who stops doing R&D will be out competed by a lab still spending like crazy on R&D, then we're implicitly agreeing their R&D spend is worth it, even if crazy.

It does mean they will be forced to raise prices though.

Yeah, people are confused because of the big capex expenses but you can compare what the major labs charge for tokens with what the open source models that anyone can run charge for tokens and notice that the labs have to be taking like a 400%+ margin on inference.

Do you have a good source for this? I'd be very interested in seeing a breakdown by model.

It's a bit messy to make sure you're comparing apples to apples. Here's a breakdown on how deepseek was getting 500%+ margin on inference around the "deepseek moment". Now that comes with a number of caviots, I'd probably hedge that down to more like a 300% margin for deepseek in practice. And they later cut the token cost something like 75% on that model but also reported cost reductions. Gpt O1 was a similar era model(December 2024 VS deepseek Jan 2025) and openai was charging something like 15-30x per token that deepseek was. The model was a similar size but superior in some ways so anyone's guess how much more it cost to serve. That might be the closest apples to apples comparison. I'm pretty confident on a 400% inference margin as a conservative estimate. Inference seems extremely profitable, just that training is also extremely expensive and you need to constantly do it to compete in the inference market.

Thanks!

Every developer will be several times faster, but this will simply result in the industry moving faster than it did before. Not massive unemployment.

Eh, with the news coming out of Meta, I think this will mean "now your small company can afford to employ a former Silicon Valley developer", but it won't be at Silicon Valley salaries. More employment opportunities, sure, but the days of big numbers on the paycheque will be over. Now you'll be on the same level as administrative staff and the other employees you used to look down on as bullshit jobs.

That trend already began years back when you look at comparative salaries year-by-year. The salary even a new graduate would command 10-20 years ago was far higher than some of the low balls I’ve seen people get within the last 5.

More employment opportunities, sure, but the days of big numbers on the paycheque will be over. Now you'll be on the same level as administrative staff and the other employees you used to look down on as bullshit jobs.

Software developers do not tend to look at the administrative staff as "bullshit jobs", at least not at the companies I've worked at. If you're going to engage in schadenfreude, at least have good reason.

(The jobs software developers do look at as "bullshit jobs" are as likely to be automated).

breaking into an artistic industry becomes even more difficult than it already is.

I vaguely expect analog artistic media to rise in popularity. Paint brushes, pens, and such are clear "not AI" status marks. Also live music.

My guess is that they'll rise in status, but not popularity. Like plays and operas relative to films and TV shows, or handcrafted furniture relative to IKEA.

Those are not particularly lucrative from the standpoint of earning money. Video games, animated movies, and graphic design seems to be where most of the money is at for the painters. Crucially, those are not things people purchase for the sake of status. It is entertainment. Losing opportunities for employment in the entertainment industries seems really bad for aspiring professionals. Art as a status symbol is mostly for rich people, or the artists themselves. So even if it rises in popularity, I would not expect it to suddenly become a viable career path.

Writers probably have it the worst. Even current AI can produce short stories that to most are impossible to tell apart from what is written by professionals.

Anecdotally, I'm seeing a lot more of that around me. Punk is making a raging comeback.

Lawsuits against non-profit and up-and coming models for breaching the copyrights of the big LLM's.

This is a very, very high risk strategy for the big LLMs given that they have breached the copyrights of absolutely everyone in the process of training their models on a corpus of copyrighted text. "You can't train an AI model on publicly-available but IP'ed data" is not a net win for Anthropic or OpenAI.

Given the, uh, rather mechanical ways in which models are trained, I could see a precedent that they're not copyrightable as a potential outcome: does it involve more creativity than a phone book? "Turning the crank" doesn't make something a creative work in the US.

But I wouldn't put a huge bet on any particular outcome there.

At the very least they will want to litigate against distillation. Training costs are steep, so if anyone can undercut your R&D by distilling your model, that seems disastrous for your bottom line.

I think this is a "rules for thee but not for me" situation. It is in their interest to prevent others from making competing models, so they will want to pull up the ladder behind them in order to destroy further competitors. Whether this will work is a different story, but this is a highly competitive market, and I think these big companies will use their large piles of cash to try and make it a reality.

They do not need to win the lawsuits in the first place. They just need to make their opponents settle by making the process as expensive as possible.

2032: [...] Places like Japan are receiving modern websites for the first time.

I know everyone's already piling in to tell you "this is already happening", but... Unless you expect "modern" websites to look dramatically different in 6 years, I'd say websites in Japan are already pretty modern? There are definitely a few Craigslist-style companies out there with outdated design, and there are some differences due to local design taste and language (web fonts are much less practical to use for Japanese text, for example). But overall web design standards haven't changed much ever since the flat design trend took hold a decade ago, and most Japanese organizations have caught up.

Compare www.city.osaka.lg.jp to www.chicago.gov, for example.

I wish sites and general UX would unmodernize.

49MB and 422 requests to load the NYTimes frontpage. The vast majority of webpages should be well under 1MB: they’re just text, images, and a tiny bit of CSS and JavaScript.

UX should be made by those who actually use the software. Obviously prioritize usability (speed, common actions upfront, uncommon actions possible) over style. Obviously don’t change acceptable UX without improving it (‘s usability). Obviously don’t promote someone just because they changed UX that works and made it stylish by sacrificing usability (I would say they should be fired for wasting your money, but that money’s going to be wasted regardless, and even “lead architects” need to eat - why not pay them for fun experiments that don’t intrude your main site?)

And for style, bring back Frutiger Aero.

Reminds me of the joke, “Shit. If I can load this MySpace page I can probably run Crysis.”

Have you ever tried to load page of Astral Codex Ten after it has had a week to accumulate comments? Running Crysis is easier.

I have not, but I’m 90% to completing my future-proof desktop. Sounds like I’ve got a new overclocking benchmark to test.

tech layoffs continue

The tech sector right now has a lower unemployment than the general US economy

https://www.google.com/url?sa=i&source=web&rct=j&url=https://www.wsj.com/cio-journal/tech-unemployment-ticks-up-to-3-8-in-april-amid-ai-driven-layoffs-214b0ca4&ved=2ahUKEwins7yEpNGUAxW7ZvUHHXKrPfsQ1fkOegoIAggACAAIHRAC&opi=89978449&cd&psig=AOvVaw0IYy6J7-fiZwA_vTGmEcwb&ust=1779690002348000

3.8% in the information sector vs 4.3%

I'm stating this first to sort of color the rest of my point in the context that a lot of what people say about what AI has already done is just bullshit. But furthermore, people fundamentally don't seem to understand how employment works. You can have mass layoffs and still have high employment. I'm not even sure tech layoffs are higher than other sectors, but I am sure thar every company that lays off tech workers gets front page news while if there's a cut in delivery drivers nobody notices.

The only smart prediction to make is that we don't really know. People here just don't realize how big and complex the economy is and the world at large is. Even if your job is gone, your skills are often still transferable. When horse carriage producers were put out of business they didn't all starve and never find jobs. They started working on building cars for the most part.

AI is just another step in a long line of automations. Is it an exceptional step? Probably. Will it ever replace all workers? No. By the nature of economics, that's basically impossible. People's desires are infinite and there arent infinite resources and labor, so there are always niches to fill. Might it make people poorer? Maybe. I kind of doubt it unless governments uses it as an oppressive system that cracks down on a lot of market activity.

My point here is really that making predictions is a fools errand. People have tried to do it, and at best a few get lucky and pretend they're geniuses and then return to the mean on the next prediction. There are obvious truths you can see, sure, like if the price of compute continues to decrease at an decelerating rate, it will significantly affect AI progress. I think that even as we see continued progress in AI, that will be the fundamental factor that's overlooked. Look at the flop count per dollar on a CPU from 2005 vs 2015 and then a GPU from 2015 vs 2025. Nvidia is squeezing some progress out in other ways, but at massive costs.

So my prediction is simple I guess. AI will be a big boon to the economy. It will take a few years for companies to learn how to cost effectively implement it.Some sectors will disproportionately reap the rewards. I suspect the gains will be in the.5-1% range of additional productivity growth a year, which is a lot. For context, the early industrial revolution was something like 2% growth year on year excluding population growth. With an extra 1% productivity growth the US would be higher than that right now I believe.

I also suspect there are factors that are huge burdens to society which AI can't overcome. Population decline. A war in Taiwan. Developed country and Chinese debt burdens. All of these things could affect AI. Which is ultimaty why all predictions beyond a year or two will be meaningfully wrong.

That only works when the thing that destroyed your job doesn’t destroy all the jobs your skills are good for at the same time. But the AI is specifically aimed at replacing skills. The skill of recognizing patterns in pictures is something AI can already do. It can recognize my face, my emotional state, detect cancers, and read road signs — all things that require recognizing patterns in images. So when AI is deployed in hospitals to detect cancer, the same image recognition machine can be trivially reconfigured to translate documents from photos, read faces, and read emotional states. It can probably drive my car if coupled with robotics properly. Where does the human go? And again with other sectors. Not only do you have the problem of “the AI can replace the skills you have”, but there’s a problem that will be caused by any sectors AI doesn’t yet have skills at being absolutely flooded with applicants from sectors AI just destroyed. When accounting gets eaten, those with the skills pivot to something else, as will spreadsheet jockeys and so on. They’re going to try to get in where there are jobs. The wages for the remainder will thus fall compared to inflation as the market gets flooded. Why give raises when there are hundreds trying to get every opening?

Where does the human go?

This is what people were saying during the industrial revolution. Don't forget that 90% of people in developed Europe were farmers 200 years ago. There will be winners and losers, but I don't see evidence that the 20th round of automation is different than the previous ones.

I mean once a person cannot trade on body or mind, there’s kind of a problem. The difference is exactly that. Because of Industrial Revolution 1, most people Don’t trade their time by doing physical labor as factories are largely mechanized and so is farm production. So when the same thing happens again, you can’t go back to “hey, let’s make everything by hand”, but a large percentage of mental work goes away in the same way in Industrial Revolution 2, then you have to find a way for millions of people to find jobs that pay liveable wages that are not either physical labor or mental labor. What’s left might be emotional labor of various forms. But what demand for that kind of thing exists? If everyone is a therapist, how does that even work?

You're.misunderstanding how economies work. There isn't a set pool of "things that needs to be done" that labor pulls from and then gets a job according to that. People have endless desires, those desires are arbitrary, and there are limited resources. That means there's always more work to be done.

Nor does someone or something being better than you in every conceivable way does mean you cannot find a way to trade and profit within that system. Here's an explanation of that--

https://en.wikipedia.org/wiki/Comparative_advantage

If everyone is a therapist, how does that even work?

I don't know. That's kind of my point. How do therapists work now? I personally feel like 90% of their work is totally unproductive.People's desires are arbitrary though. You don't necessarily make more money by creating 5000 times more steel instead of paying some weirdo to listen to you blabber for an hour. That's one huge mistake the Soviets made.

The tech sector right now has a lower unemployment than the general US economy

Tech sector unemployment hit a low of 1.8% in 2018 and is now 3.8% and rising. That's an absolutely massive change. In the meantime, total unemployment hit a low of 3.5% in 2020 and is now 4.3% and steady, which is a much smaller change. And "rising" is important, if you're already not working.

But you’d expect that regardless of the tech improvement. People were going to flood into the “low unemployment + high pay” job raising the unemployment rate of that industry.

We know that an increase in supply is not what's causing unemployment, as tech employment is dropping. Supply isn't all that elastic so that increase in supply due to increase in demand usually doesn't cause unemployment -- such inrushes have happened, but were more than absorbed by the industry.

https://fred.stlouisfed.org/series/CES6054150001

https://fred.stlouisfed.org/series/CES5051800001

Will it ever replace all workers? No. By the nature of economics, that's basically impossible. People's desires are infinite and there arent infinite resources and labor, so there are always niches to fill.

Minimum wage and related barriers put a finger on the scale though. Currently, very-low-skilled people are unemployable because the assorted costs of hiring them outweigh the expected benefits. In the future, will that extend to moderate skill levels? high? I don't think it'll cut off 100% of people before extinction and/or post-scarcity, but I could see the labor force dropping from about 50% of all people today to 10-20% even if AI remains a normal technology.

I'm not sure that this is entirely true. Very low-skilled people are unemployable period, and lowering the pay rate doesn't do anything. For example, there's a guy I know who isn't the brightest, retired now but comes off as someone who was definitely in special education back in the 60s and 70s. He worked as a janitor at a local elementary school. In Pennsylvania the minimum wage is the Federal $7.25. Someone in his position would be making $22.62 this yer and $24.35 next year. Of course, that's because he's been there for 35 years, but even a new hire makes $16.60 on the current contract and $18.60 on the next. Grocery, retail, and fast food wages aren't much lower, even for 16-year-olds with no experience. The only exceptions I'm aware of are for people with disabilities, but that's more because they can only make so much before they lose their benefits. I don't think there is a significant population that's employable but for minimum wage laws.

In most places, inflation has effectively repealed minimum wage laws. However, other indirect regulatory and legal costs have accumulated that make people more expensive to hire.

Now that you mention it, when was the last time you even saw a minimum wage job? Even the convenience stores and McDonald's around me are offering well over the state minimums. It seems like it's one of those things that exists on paper but doesn't really come up anymore.

This is a fair rebuttal. I don't think Americans workers will ever be willing to support 80% of the population with welfare, let alone the fiscal reality making it totally unfeasible.

That 80% gets to vote too. I wonder what sorts of welfare they'll vote themselves.

We can see the answer to this in every discussion on social security caps/contributions/clawbacks or British discussions on the "triple lock"

Hint, even if the welfare formula is mathematically destined to out compound the entire country's economy (triple lock) the voters will never, ever, EVER vote away their gibs.

If American workers are only 20% of the population (and the 80% are people actually looking for work, not children, retirees, housewives, etc.), then I don't think normal political considerations will matter much.

hey begin to think about targeting non-coding white collar work like finance and spreadsheet work since the models are not getting much better at JavaScript

This is already happening: https://www.anthropic.com/news/finance-agents

DeepSeek releases a model that is equivalent to 5.3 Codex

Already almost there if you include all Chinese companies, certainly will be by end of 2026: https://livebench.ai

Frontier models begin to entshittify as they are increasingly jailed to make them safer, while their private reputation is shattered among all of the normies who did not know how rights-violating the United States secret police are.

I'm with you on this one. We've already seen movements in this direction (eg, not releasing Mythos).

It's pretty clear to most people the general improvements to coding are dried up and a lot of the old hype was fake and tooling and chaining was the internal, secret meta from 2025 onwards

And more targeted RLHF, but yes agreed on this. However, I think there is still a ton of yet-to-be tapped potential in tooling, context, and feedback that will have massive impacts even at current model capability.

Overall my timelines are shorter than yours but I do think there is a "ceiling" and I don't think we are at risk of Yudkowsy's takeover scenario. I do anticipate "mundane" surveillance and increased slopification. My hope is in local/opensource models running on ASICs, which would at least alleviate privacy and intentional kneecapping concerns.

Already almost there if you include all Chinese companies, certainly will be by end of 2026: https://livebench.ai

Deepseek 4 pro blows Codex 5.3 out of the water in real world usage.

Which v4 pro are you using? It's terrible.

The deepseek one

The one that is way worse than codex 5.3?

I'm with you on this one. We've already seen movements in this direction (eg, not releasing Mythos).

I disagree with this bit, mostly because I've seen good arguments that the secrecy around Mythos is at least in part due to Anthropic hyping up their own work, but most importantly due to a massive compute crunch on their end. It does have legitimate security implications, of course, but their framing that the delayed release is mostly due to those concerns is, shall we say, a rather self-aggrandizing claim.

GPT 5.5 Pro performs as well or better on cyber security tasks, and it OAI was happy to do a general release. This is one of the rare occasions where I have to say that they were right in mocking Anthropic for poor excuses for their real issues, even if I genuinely prefer Anthropic as a company and the recent versions of Claude for many tasks.

… I've seen good arguments that the secrecy around Mythos is at least in part due to Anthropic hyping up their own work, but most importantly due to a massive compute crunch on their end...

That’s exactly what it is. Behind the marketing department though, there are still interesting things to see with Mythos.

Anthropic’s model is really good at finding software vulnerabilities, but so are other models. GPT-5.5, already generally available is comparable in it’s capability. The company Aisle also reproduced Anthropic’s published results with smaller, cheaper models.

One of the problems with Mythos is that it’s very expensive to run, and the company doesn’t appear to have the resources for a general release. (What better way to juice the company’s valuation than to hint at capabilities but not prove them, and then have others parrot their claims?)

Modern generative AI systems (not just Anthropic’s but OpenAI’s and other open-source models) are getting really good at finding and exploiting vulnerabilities. I don’t want to say I was a complete naysayer originally (because I wasn’t) but the rate of advancement has raised my eyebrow a few times along with some of the economizing factors.

Your stuff through 2028 is already happening. People are already trying to do basically canned drop in tech workers at every software shop I'm aware of, you're basically just describing Claude code/cursor. Those start ups already exist. Models are already making big progress on erdos problems.

They begin to think about targeting non-coding white collar work like finance and spreadsheet

This has already happened. I'm baffled that you don't know that the models can now handle spreadsheets. They do so pretty well, especially after Opus 4.7.

the models are not getting much better at JavaScript, having used up all of the JavaScript data in the entire world

This is a misunderstanding of how models improve. It's not a matter of finding more undiscovered java script code to ingest, much of it is now post training self play and should continue to improve as general model scale increases. Of course it's already perfectly capable of writing good javascript and has been for several models, the limitations are mostly in reasoning about larger chunks of the code context.

More attention is turned to local models as a result, but these are hard to run on normal hardware and the best is Sonnet 4.6 level at this point and requires $10,000 worth of GPU machinery.

It's too bad Ilforte left because he'd eviscerate this. I tend to be less optimistic on the Chinese models than some but both Deepseek and Kimi have offerings that are comparable to sonnet 4.6 if you trust the benchmarks, I don't but fully expect them to have a sonnet 4.6 level model by end of 2026 and likely an opus 4.6 model by then. And you can run these models on rented hardware for pretty cheap. Although they'd be hard to run locally for a lot of complicated reasons that have to do with it being much more efficient to batch queries than run them individually. In any case though the weights are public and anyone can set up an api to sell tokens at affordable rates.

I'm skeptical of your ability to predict the future as you seem incapable of predicting the past.

I tend to be less optimistic on the Chinese models than some but both Deepseek and Kimi have offerings that are comparable to sonnet 4.6 if you trust the benchmarks,

They do not, the best deepseek is maybe GPT o3 tier. Ilforte is delusional about China.

Chatbot Arena ranks deepseek-v4-pro-thinking at 30th for [text (1461 ELO) and 17th for coding (1459 ELO). By contrast, claude-sonnet-4-6 is 22nd for text (1468 ELO) and 6th for coding (1524 ELO); there is a definite gap. On the other hand, kimi-k2.6 is 29th for text (1462 ELO) and 7th for coding (1519 ELO), which is closer. And glm-5.1 is even better; 20th for text (1472 ELO) and 5th for coding (1532 ELO). So it looks like the strongest open source Chinese models are equal to or better than Sonnet.

Text Arena

Rank Model License Score
20 glm-5.1 MIT 1472
22 claude-sonnet-4-6 Propietary 1468
29 kimi-k2.6 Modified MIT 1462
30 deepseek-v4-pro-thinking MIT 1461

Code Arena

Rank Model License Score
5 glm-5.1 MIT 1532
6 claude-sonnet-4-6 Propietary 1524
7 kimi-k2.6 Modified MIT 1519
17 deepseek-v4-pro-thinking MIT 1459

Can I only point how clustered they are? Unless this coding benchmark is heavily logarithmic in weighting the models - the results say that all of the models are good enough.

I've never used glm-5.1, I'll try it today.

The longer I spend here, the more I understand why @DaseindustriesLtd got so fucking mad at some of the quality of the takes on AI and decamped to fairer lands. I mean, I'm still here, and I'm not really going anywhere, but I've already said I'm largely bowing out of the conversation.

That's probably because I am less Russian and more patient than he is, but some of the bullshit I've heard has driven me towards drink*, and my patience is at an all-time low. I just get where he's coming from.

*Making me spiritually Slavic, or at least Scottish. There are no shortages of European nationalities with a national fondness for drink.

I get a distinct feeling of Gellman Amnesia reading a few of the recent top level posts in this weeks thread. And I wouldn't even class myself as a particularly knowledgeable person when it comes to AI, I simply keep up with the news and developments. It's really something to see the number of posters, whether here or the ssc reddit or similar locations, who confidently spout complete garbage when it comes to AI, seemingly unaware of things that happened even months ago.

And now I can't help but worry that many of the other posts on the Motte are similarly compromised. Have we become (or always been) just another midwit debate site?

I think every forum has some Gellman Amnesia (and déjà vu), unless it's heavily moderated like r/AskHistorians. And real life small talk has much more. If people only stuck to their domain expertise, more forums would be barren (see next paragraph), and people don't know what they don't know (Dunning–Kruger).

At least most replies point out the errors. Domain experts are often too busy, lazy, and private to browse and reply to random internet questions; except they miraculously find the time, effort, and public interest once someone else responds with a wrong answer (Cunningham's Law).

Hey, at least Hlynka is gone (suicide by mod). While he's enjoying his retirement, my blood pressure does much better.

And now I can't help but worry that many of the other posts on the Motte are similarly compromised. Have we become (or always been) just another midwit debate site?

We were supposed to at least be midwits? I got 70 on my IQ test, which must be 70% of the maximum and a passing grade!

Your joke made me wonder if redefining IQ as the percentile would make it more legible. Or 50+the percentile, to keep 100 as the middle ground. Probably not, IQ90-110 would be mapped to IQ25-75, that's fifty points reserved for midwittery. I can already imagine people claiming they are 3x smarter than their opponent.

Aww, it's sweet that you thought I'm joking. No. I have brain damage from too many exams, including having to memorize all the fun properties of a normal distribution, as well as the abnormal ones.

I think the properties of the Cauchy distribution are much more fun.

I'm couching my words carefully: I'd rather stay in my couch and let the experts handle these things. I have few neurons left to distribute to even the most normal of tasks.

Bad takes on AI seems to be the one commonality across creed, race, and IQ. This one from theringer is a particularly egregious example, but its rare to find anything both sufficiently technical to understand how it does and could work, and sufficiently "big picture" to understand societal impacts. (Of course, many others would consider my AI takes to be just as bad).

My current modus operandi is to be whatever the other person is not. If they are an AI maximalist I am the pessimist. If they are the doomer I am the optimist. If they think this technology is all hype I become the autistic technologist with in-depth details and explicit examples.

got so fucking mad at some of the quality of the takes on AI

It's actually one of the best ways to bait me. I thought that now that I'm old and wise, I would stop taking bait, but they're just so wonderfully confidently wrong and I cannot resist. It makes me so tilted, but the "I told you so" as we stand in the breadlines is going to hit so nice.

Brother, insight without action is as worthless as the spectacles it came with. Don't take the bait, at least if it comes at the cost of your sanity. Or do, if you end up feeling some degree of catharsis, idk, I'm not your shrink. I'm not doing a very good job at being my own shrink.

Perhaps it does serve a useful function to point out when people are being pigheadedly wrong about things. Someone's got to do it, or ought to do it, and I'm just glad that someone is very rarely me these days. I've got booze to drink, and Scottish women to introduce to the single mother lifestyle.

But yes, if we meet in the breadlines or in the intake unit for the paperclip factor, I'll save an understanding nod for you. Fist-bumps wouldn't be befitting.

Making me spiritually Slavic, or at least Scottish. There are no shortages of European nationalities with a national fondness for drink.

Better question would there be any without a national fondness for drink.

Terrible driving and a fondness for alcohol is something that unites just about every culture and demographic.

Bostonians often openly admit that the nationwide reputation of Boston drivers as being especially awful is true. Is this just a form of self-aggrandizement, and, actually, every locality believes their drivers have reputations for being the worst?

I can't think of a single demographic group of humans who are considered "good drivers"

Similarly, I've never once heard anyone say "yeah the drivers in my $LOCAL_AREA are great! I love it! We all get along on the roads :)" and I've heard every version of the opposite, so I'm assuming everyone sucks everywhere.

Back when I was a young lad I would have told you "yeah the drivers in Germany are great". Nowadays, between Germany getting diversified, and the driving culture in my country improving, the contrast is not so stark. Some time ago I also saw a video from an Indian guy saying that Italy, of all places, has good driving culture (though I suppose it makes sense if India is the reference), and when he was driving there he felt this subtle pressure to perform up to the standards of the rest of the country.

You're right that everyone might complain about their neighbors, but different groups definitely perform differently.

Italy was the only country where the drivers would change lanes to speed past the car than stopped in front of me on a pedestrian crossing. I can't imagine what India is like if that's good driving culture.

Every good normal, local football team loving, sydney sweeney gooning person hates people with [autism] and loves to see them suffer

That would disqualify them from being considered 'good'.

Every good normal, local football team loving, sydney sweeney gooning person hates people with aspergers and loves to see them suffer so they got off on viciously and narrowly aiming to automate their programming jobs

Sweet God, I'm probably on the spectrum myself and I hate these kinds of whiny "people with Aspergers" and want them to suffer because no, you are not a special snowflake horribly persecuted by society but you'll make them all see when you get a huge big-paying important job playing with computers due to your special interests.

Suck it up. Everyone suffers in some way. Learn to deal with normal society around you, and when you can't, shove it down and put a lid on it. You needed to be bullied more at school and maybe slapped around a bit by your parents, to teach you to toughen up and stop. bloody. whining.

Ordinary people don't think about you and yes, programming jobs are overpaid and self-important.

You needed to be bullied more at school and maybe slapped around a bit by your parents, to teach you to toughen up and stop. bloody. whining.

I would have expected you to take a little more care to avoid calling for people who complain of oppression to be beaten by stronger people until they shut up, Deiseach, given how often you denounce the "beat the feminism out of them" brigade as barbaric.

I can call the whiny Asperger's' brigade whiny because I went through that phase too and came out the other side having learned that nobody cares, the only change you can make is yourself, and telling yourself that the reason you are failing at life is because you're too smart and special and elevated in your tastes and interests for the normies is self-deception: "lay not that flattering unction to your soul".

My anger isn't with your description of people as "whiny".

My anger is with your proposed remedy. I don't know exactly what you went through as a kid, but I know what I went through. You've had the highlights reel of Mum; here's the highlights reel of school.

  • Held down by two boys while four or so others took turns trying to punch me in the balls
  • Walked on, literally
  • Had a point-up needle affixed to my chair with wax
  • Arm broken
  • Tooth knocked out
  • Tried to strangle myself to death at 7 (seven) to get away from all the teasing.

I do not, in fact, think I needed to be bullied more. I do not, in fact, think other aspies should go through that hell. I doubt it'd even make us less whiny, aside from the minority who'd be too dead to whine.

You had genuinely bad experiences in school. That's not an excuse to keep complaining about how society is so unfair or how if you (general 'you' not specific 'you, magic9mushroom') don't get the exact specific everything you want, this is persecution.

And there is too much of the latter, which is how even the safe spaces turn toxic as everyone wants their own, personalised, version of how the world should be in order for it to be 'fair' to them, and then they turn upon one another because A wants/does not want something B does not want/wants to exist in the new responsive to their every whim world.

They think of themselves as good but I don't agree per se.

What is this place?

This website is a place for people who want to move past shady thinking and test their ideas in a court of people who don't all share the same biases. Our goal is to optimize for light, not heat; this is a group effort, and all commentators are asked to do their part.

The weekly Culture War threads host the most controversial topics and are the most visible aspect of The Motte. However, many other topics are appropriate here. We encourage people to post anything related to science, politics, or philosophy; if in doubt, post!

Check out The Vault for an archive of old quality posts. You are encouraged to crosspost these elsewhere.


Why are you called The Motte?

A motte is a stone keep on a raised earthwork common in early medieval fortifications. More pertinently, it's an element in a rhetorical move called a "Motte-and-Bailey", originally identified by philosopher Nicholas Shackel. It describes the tendency in discourse for people to move from a controversial but high value claim to a defensible but less exciting one upon any resistance to the former. He likens this to the medieval fortification, where a desirable land (the bailey) is abandoned when in danger for the more easily defended motte. In Shackel's words, "The Motte represents the defensible but undesired propositions to which one retreats when hard pressed."

On The Motte, always attempt to remain inside your defensible territory, even if you are not being pressed.


New post guidelines

If you're posting something that isn't related to the culture war, we encourage you to post a thread for it. A submission statement is highly appreciated, but isn't necessary for text posts or links to largely-text posts such as blogs or news articles; if we're unsure of the value of your post, we might remove it until you add a submission statement. A submission statement is required for non-text sources (videos, podcasts, images).

Culture war posts go in the culture war thread; all links must either include a submission statement or significant commentary. Bare links without those will be removed.

If in doubt, please post it!


Rules


Recommended Posts And Communities

Recommended Realtime Chats