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 -
Of the job categories listed in the 1960 US Census, supposedly only one has disappeared completely: elevator operator.
Nevertheless, I think the complexity argument might fail for accountants. Complexity actually favors AI. They are good at it.
When you talk to rich people, everyone is looking for an accountant. They all think that their accountant is bad and want a better one. I've only had bad ones personally. Now I just do things myself because I don't want to deal with errors. AI has been a godsend. I think it's likely that AI will do my taxes in April 2028.
But there is a silver lining for accountants: knowing which rules to bend or break. If it treats taxes like it treats code, AI will happily fill out a 500 page form so you can pay $0.04 to the state of Arkansas for some K1 you got from an oil services company. The accountant knows to throw it in the trash. They can tell you to ignore that letter from IRS because it doesn't get serious until the nth notice. They know how to ignore the rules that contradict the other rules. They know that certain tax dodges are okay because all their clients do them and never get audited.
I think it's somewhat similar to driving. The system only works because people know when to break the rules. Will AI break the rules? If so, will they be liable for that? Humans might stay in the loop just so they can break/bend the rules.
"Computer" is also no longer a job category thanks primarily to the machine named after said profession.
More options
Context Copy link
Hilariously I've been in elevators with an operator.
I'm pretty sure another silver lining for accountants is that "my freakin' AI did it wrong" is probably not an excuse that the IRS will accept. They will accept "my accountant did it wrong."
Was this in a heavily Jewish area, perchance?
New York City, so yes.
More options
Context Copy link
More options
Context Copy link
Definitely. It’s the same problem that self driving has where every Tesla that crashes with FSD is nationwide news even though it’s 10x (or more?) safer than average drivers.
Those things only apply on average random samples. As soon as they hit a consistent failure mode, you end up with Problems.
More options
Context Copy link
I think it's worse. In the case I am familiar with, (sample size of two) as far as I can tell, when a guy did something on his own, he did okay. When he tried to use AI to do about the same thing, SNAP!
So either the AI screwed up big time 50% of the time, or 100% of the time. Either way, not great, and seems worse "per task" than self-driving cars.
Weird. This is so far from my experience that I wonder what you guys are even doing? Like, you only tried AI twice?
If you’re a senior dev, and AI is not banned by your org, and you’re not personally using AI, that seems very inefficient.
Oh no, I've tried AI lots more than twice. Definitely witnessed more than one hallucination.
I'm not in software. The mistake I'm talking about wasn't made by me, so I don't have any insight into how it was made or how often the user was using it, or what model.
To give you the sort of idea of what happened without doxxing myself, imagine someone turned in their PhD dissertation and it (or parts of it) turned out to be AI written, and the AI had made stuff up.
Was he using a "cutting edge" model? No idea. I've had Opus goof stuff up for me before, so I don't necessarily trust even the higher-end models not to make mistakes from time to time.
Fair enough. I trust your lived experience. But listen to software devs when they say that AI works now, it hallucinates a lot less, and it has gotten remarkably better in the last 12 months, even the last 6 months. I use it every day. It is does not hallucinate much.
Harnesses such as Claude Code and Codex have changed the game.
I think that AI has been gravitating towards software development (perhaps to the detriment of other skills), in part because software development helps develop AI! But software development is fairly validation oriented – code runs or it doesn't. Someone who is lazy enough to have an AI write text with their name on it may also be too lazy to validate that text, or even too lazy to have another AI fact-check it.
You can say "yeah they are a poor user" and fair enough, but people will say things like "AIs have PhD level intelligence" or "AIs can act as a smart research intern" and then (in my experience) from time to time they go and make stuff up. Certainly lazy PhDs and research interns do this, all the time, but I don't think that demonstrates intelligence.
I'm not an AI-denier, I use it relatively consistently and I like it for research – but a lot of that is because Google is bad now (and their AI overview is atrocious, I've seen it present conflicting information), and you can use AI to get you links that you can check with your own eyeballs, and you can give it specific prompts that help mitigate other bad tendencies. But the fact that you have to do these things means that you cannot necessarily just expect AI cruising around on its own recognizance (or in the hands of someone who is not intelligent and/or used to using AI) to operate at that high level.
Sweet summer child. One of the undocumentd laws of programming is that if you have even number of errors you will get program that works mostly ok.
I have had way too many debug sessions that were - here is the error, how the fuck this thing worked at all, ahhh - here is the other error.
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
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
AI fails for accounting not because of complexity but because of care. Accountants have to get the numbers right, and occasionally hallucinating shit isn't going to cut it. Maybe it would work for mob accounting, but I suspect the mob wants their made-up shit to be less random.
I doubt it - the IRS don't start with your fingernails when you make an error.
More options
Context Copy link
If you have three independent AIs do the numbers and they match, they’re likely right.
More options
Context Copy link
The hallucination argument is losing power. Harnesses will fix hallucination much like they have for coding.
I'm not going to straight up ask Astra to do my taxes, but I'm fully confident that Astra+harness is capable. So now we just wait for the harness. No doubt lots of smart people are working on it.
Accountants don't get the numbers right anyway which is why I do my own taxes now. It's a boring, frustrating job and attracts people who like money more than having interesting work with predictable effects.
Harnesses fixing hallucinations is news to me. I just spent an hour today putting a Junior's head on straight after codex sent them down the garden path for a day and a half.
It's not really about "harnesses" as some dedicated tooling, it's about meta-instructions and models generally becoming good at following instructions over long horizons. Like, I just wrote in my AGENTS.md some stuff to the effect of "periodically spawn subagents for independent blind audits, create teams with different scopes and domain-customized personas, check with web access, sometimes step back, iterate until you get clean results", and even the latest DeepSeek will diligently execute all that and massively improve its reliability. Astra Ultra won't even need such handholding. They'll still make mistakes, rabbithole into irrelevant minutiae, perform suboptimally on cost, but for labs this is all easy to improve with More RL, and humans are flawed too. I am sure that a junior's level can be automated today.
And yet.
ChatGPT Sol absolutely.
insists on
writing.
Like this.
In a shitty.
broken up.
style.
If I ask it to demonstrate a writing concept or somesuch and there are people in the scene. It will happily tell what the problem is and how to supposedly prompt it away but will then keep on writing like that. So much for "good at following instructions".
ChatGPT Astra is meaningfully better than Sol. I of course agree it's still disgusting. It's also nice news for me because literally Fields Medal winners are getting automation anxiety before I have to worry about competition in bespoke shitposts.
The issue with writing style is that they're not built to write for humans. These are agents. The process by which they output almost every given token over their training is getting rewarded conditional on its contribution to the final outcome of a programming (or programming-shaped) task.
Be thankful you don't have to deal with raw CoTs (unreadable now).
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
No, it most definitely is not.
Do you use these tools on a regular basis? My experience with hallucinations is that they are extremely reduced.
Same. I can't recall seeing one with Codex and 5.6-Sol.
More options
Context Copy link
It is still a problem with whatever models people are using. Just had to deal with a pretty bad incident at work a couple of weeks ago (thankfully not one due to any failures on our part).
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
More options
Context Copy link