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.
Jump in the discussion.
No email address required.
Notes -
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.
We continue this forums streak of awful takes on AI. Your bait sucks but I took it anyway.
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 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.
This is already wrong. Can you substantiate this in any way? What do you think of AlphaFold?
Finally! An actual prediction. I could see this happening, we shall see.
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 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.
Another real prediction, i agree with this but wouldn't be surprised to see it even sooner.
This has been happening this whole time, and especially since 4o
This is literally already happening
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.
This is already happening
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.
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.
Great so we'll see if they wrap up early or if they enter the public conscious closer to my timeline.
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.
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:
Your finance friends are idiots
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.
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.
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.
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.
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).
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.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link