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 -
Culture War appropriate? Okay, I'm hearing everything about how AI is coming for our jerbs/gonna make us so productive, eventually only six humans will be working in the entire economy and the rest of us will be livin' large on that sweet, sweet UBI from all the yuuuuge economic gains.
But what does that mean in actuality if I'm not a software engineer type?
For example, a Substack comedian (literally, that was his day job) has a post up about AI and how this is all hysteria, nobody is going to lose their jobs, it's merely the usual sort of dip in the economy and the sectors most affected are:
Okay. I fall into the "secretaries and administrative assistants" bucket and I know Sweet Fanny Adams about AI. My exposure to it in the workplace is with the free Copilot Microsoft has bundled in with Microsoft 365 and, apart from annoying me with "Do you want me to write that email?" (no thanks, I think I can figure how to say "I got that invoice, thanks" all by my little ownsome), I see no use for it.
But! AI is going to be the wave of the workplace future! So, for all you who know and use the thing and are up on the different models, here's an example of a task I routinely need to do in my job. Can AI (Copilot or whatever) do this, or most of it, for me?
A request from our auditors:
Where this information is located:
How do I use AI to take all this drudgery off my hands? Can I ask/tell it "here's the details of how to log in, now go ahead lil' Copilot and pull out all that info and make a nice, tidy spreadsheet out of it all"? Or do I have to hand-hold it every step of the way, in which case I am just as well off to do it all myself?
Well since you aren't on the unemployment/UBI queue yet, it follows that AI can't do your job. AI is pretty bad at using interfaces designed for humans.
But the hard part has already been done. I can prompt out a shader-evolution program, it basically evolves, based on human ratings of who propagates, generations of graphical animations, patterns of energy that would make pretty screensavers, save the ones I like to a folder.
That's easy, a couple of prompts.
Or I can have it write out fun skits, like the Iran/Hormuz situation mashed up with the "Ship Front fell off" skit: https://youtube.com/watch?v=3m5qxZm_JqM
The raw intelligence is there, there are just certain fiddly parts and a level of precision that's missing. A skit doesn't necessarily need to be precise, if code doesn't work right you can just tell it what went wrong and have it improve. Whereas with sending payment documents to people, precision is quite important.
I use it for work too, mildly superhuman proofreading and some other tedious document processing tasks, some of which need a certain level of judgement and wisdom. Precision is pretty good. But for these I have to use my custom scripts for that specific hours-long task, scripts that I've tried and tested over some time. And they really need to be more 'document input, document output' tasks rather than 'messing around with portals and niche software'.
But that's now, they're working on making agents reliable and cost effective. Eventually they'll get there. I can't understand a scenario where it's possible to create fairly complicated pieces of custom code with AI but logging into outlook, sending emails, logging into whatever govt portal, getting a 2FA code... is some domain reserved for humans alone. In the long term either AI has to be dumb as a rock/incapable of anything but specialized tasks like chess/go or it can be general purpose, doing anything a human can do.
More options
Context Copy link
The true robot apocalypse where we all lose our jobs will only come when Clippy returns at the head of his AI army. It's a millenarian cult now and I'm going to treat it like one.
I'm sorry but the constant 'but AI is gaining in capabilities' ignores one thing- yes, yes it is, but those capabilities don't actually correspond to anything. Yes, it can reduce secretarial workload, and that will over time drastically reduce the number of secretaries, but mostly through not replacing retirees. Good thing the population's shrinking! In the meantime, I get a raise because AI hardware raises the wages for people that work on cooling systems(and electricians, too). The lower working class, that could theoretically be replaced by trained monkeys(but training the monkeys is not cost-competitive with $15/hr), yeah, a computer that knows how to flip burgers was possible in, like, 1982- still hasn't happened, because hands, you know, Claude doesn't have them. McDonald's uses lots of robots in their kitchen. None of them have hands, they're not a replacement for local teenagers.
More options
Context Copy link
You could instruct an AI system like Claude Cowork (https://youtube.com/watch?v=ZeWfksNXlbU) to actually go do the work for you in your browser, and it would be reasonably accurate but slow (probably slower than doing it yourself - but you can go do other work instead while it's running). My guess is describing the task roughly the way you did in this post, and then going back and forth with it a few times (handholding it the first few times) and asking it to come up with a better prompt for itself afterwards - will get you to a pretty good state of accuracy.
Generally the long-term vision for AI transformation in enterprises is that:
My nightmare scenario. Government signs a contract with, say, Microsoft and now we have to store everything where Copilot can get at it and I loathe Copilot, Edge, etc.
More options
Context Copy link
More options
Context Copy link
AI can definitely do this if you give it the signin endpoint or your auth token. Since you don't have access to the code base for these websites, the main blocker is reversing the API structure. This is pretty easy as far as reversing goes, since your client should use the API with unobfuscated JavaScript. Yes, this is probably beyond non-software engineer technical competence to do, even with an AI assistant, but it's still possible to use AI to automate tasks like this, especially if you use the same handful of websites for all tasks like this.
Why not just use browser automation tooling? The frontier models seem to be reasonably good at reading screenshots, and that way you're completely indistinguishable from a regular user.
More options
Context Copy link
More options
Context Copy link
Reminder to myself to reply.
I work on something very very close.
We have people in our beta running exactly this query.
Yep
Yep
Data connection is hard....esp if there is no way to export data from a closed source.
But as long as you have something - an API/MCP (software connection to data) or CSV export (manual export), it should be doable. If you are willing to really burn money, you can give the computer permission to run your computer for you, in which case it will open the browser page and manually navigate to the specific information it is looking for. (even closed sources are viewable from a website). Although, the latter is very expensive, so no company is really working on it. But, it not a capability gap....just something companies think is not monetarily feasible YET.
The former 2 are definitely doable right now. They are closer to how Google search worked in 2000. Good but not perfect. But just like Google search in 2000, the steps to become better are very obvious to the experts internally. It is pretty much all hands on deck going straight towards it. But it will take between a few months to a couple of years to get there.
More options
Context Copy link
I had my first "AI at work" experience the other day when I sat through a luncheon meeting presented by a rep for one of the big legal research companies. It was billed as a continuing education event but was really just a sales pitch for their AI products. The guy was able to cite two uses for AI in the legal field:
That's all well and dandy, but I don't do either of those things very often. This wasn't presented as "the technology is quickly changing and you'll be able to do more in the future" as much as "this is all you can do within the bounds of ethics and without exposing yourself to a malpractice suit". The idea that law firms will consist of a few partners handling a suspiciously large number of cases by prompting AI to generate outputs is pure fantasy. The people who think that AI will take over everything do so on the assumption that all work boils down to a set of deliverables that simply need to be generated, when that's not the case. If I'm looking to generate deliverables, I can already have a paralegal do all the drafting and research and just put my name on it, because there's nothing that says you need a law license to do legal research or draft documents.
What the client is paying for is for someone to take responsibility for the case, and it would be irresponsible of me to "handle" a case about which I knew nothing. Most of my time is spent reviewing and analyzing facts. Sure, an AI may theoretically be possible that can determine what's relevant and formulate a strategy better than I can, but the AI is not going to be responsible for its output. I'm never going to trust AI with tasks I wouldn't trust to support staff, no matter how much I trust my support staff (and they're great, btw), because the client doesn't want to hear about how it's the paralegal's fault. If I allow AI to do all my work for me, and I go into negotiations missing something, that's a pretty big matzo ball hanging out there. It's not that I'm perfect, or even necessarily better than AI theoretically could be, but the client is ultimately trusting me to make the relevant decisions, and I can't make them without a thorough knowledge of the case. It's the same problem with autonomous vehicles. I said a decade ago that they would never catch on, not because of any technical limitation, but because auto manufacturers aren't going to take responsibility for them. We've already seen this with Tesla being very aggresive in their defense of lawsuits stemming from autopilot. I don't necessarily disagree with Tesla's stance on this as things stand now, but if a vehicle is truly autonomous then an accident isn't caused by negligence on the part of the driver but on products liability on the part of the seller and manufacturer. As long as auto makers take the stance that the owners of vehicles are ultimately responsible for them, true AVs will never exist.
The other big issue is data security. You can tell me all day long about how great Claude, or ChatGTP, or Gemini are, but in the legal world using any of these is a complete nonstarter. Any lawsuit is going to deal with confidential data, and some suits are going to deal with little but confidential data. At the very minimum, we need to use settlement histories to evaluate potential settlement value of a case. Google literally built its business around data harvesting, and the tech sector as a whole doesn't have a stellar reputation for protecting client data. Regardless of whatever "opt out" provisions are allegedly in place, no law firm in their right mind would take the risk of feeding reams of data into a chatbot if there's any risk whatsoever that that information will show up later in a chatbot response. And no, this isn't the same as companies feeding their proprietary code bases into chatbots; the code's confidential status is subject to the discretion of management. An attorney does not have the discretion to reveal confidential information, especially if that information will be harmful to the client in the wrong hands.
This is before you even get to the fact that the current technology is underwhelming even for legal research. It looks good in demos but as soon as you try to use it for anything it proves its inadequacy. For document summarizing, 5,000 pages sounds generous, but it's rare that I'm concerned about finding information in a single document. I said this in a comment last week, but the utility would be more like "search all the depositions we have on file and pull all the ones where a witness testified about X". Well, we have tens of thousands of depositions on file, most in PDF but some in a special format used for court transcriptions. Conservatively assuming 100 pages per depo and 140 words per page, that comes out to something like half a billion tokens of context required, before we even consider that PDFs take more tokens than plain text, and a lot more if they haven't been OCR'd (which most of these haven't). Even the document functions described by the sales rep weren't that good; the example he gave was that if you were searching medical records for mentions of cancer it could broaden the search to include mentions of specific cancers.
This is well-within the technological capability of 2010 AI systems (using techniques that look nothing like chatgpt and where concepts like "context windows" don't apply). So the lack of a tool that does this for you has nothing to do with technical limits on AI, but only on someone's desire to build it/market it to you.
I'm sure rov wants to search by concept or something not keyword.
I understand. Modern LLMs have not meaningfully impacted search at all. Our ability to semantically search documents in a particular technical domain has not advanced tremendously in the last decade. word2vec (2012) and BERT (2017) were both meaningful step changes (but only a small step). Nothing in the generative/LLM era has meaningfully impacted search.
I want to know how this is going to impact things when one side is using it to write and distribute memo's and other corporate information and the other side is using it to summarize and automate replies to said information and 'no' side in reality is paying attention to what the other is doing. The rift and chasm I imagine opening up is going to bring chaos to organizations and cause more harm than it solves.
I read a story the other day about how new candidates on the job market are using them to build their resumes and in response HR departments are using them in tandem with the applicant tracking systems (ATS) to filter through candidates and reject the ones that are suspected to be influenced by an AI / LLM. It's a lose / lose proposition for both parties in the long run.
More options
Context Copy link
I would argue that there's a tremendous gulf in capability between the post-bert world of embeddings plus cross encoder rankings, and the ancient days of 2010.
I agree that the genai era does shit for search though, we barely addressed the needle in the haystack problem, let alone asking more complex queries. And LLMs are horrendously incapable of returning all matches for a query, even a super simple one.
Good God, that's a frustration that I totally forgot about in my various AI criticisms. I have, on multiple occasions specifically asked it to "Provide a comprehensive list of X", which list would include dozens of items, and it instead provides 5 examples, complete with shitty summaries I didn't ask for, which examples are the most obvious and well-known examples (obviously pulled from the Wikipedia page) that I wouldn't need a trillion-dollar technology to find.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
Keyword search doesn't even require AI.
More options
Context Copy link
LLMs aren't needed for that:
https://en.wikipedia.org/wiki/Concept_search
They don't work well and have been obsoleted by newer llm based techniques, mostly bert based (not to be confused with gpt/genai)
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
And that sort of 2010 era AI only needs a decent GPU to run (if even), not massive datacenters dedicated to it.
2010 was before AI on GPU was really a thing.
Well ackchually AI on GPU goes at least back to 2006.
More options
Context Copy link
More options
Context Copy link
You would have needed a decent full rack of machines in a data center to do a moderately large scale search task back then. No GPUs though. These days, that same task can be done on a single machine that cost <<$5k. Hardware efficiency gains have been massive.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
Thanks to everyone who replied, I have a better grasp on what AI (at present) can and can't do for me.
More options
Context Copy link
One thing I've taken to doing to help me use AI in my job (which I do a lot these days, and it's growing) is to use as good an AI as I have access to (Claude Opus right now) how to use an LLM to solve some problem I have, and using that as a guide to use other, cheaper LLMs to accomplish whatever task I have. I couldn't help but notice that your comment is essentially a prompt that i might enter into an AI, just dorected at a bunch of NIs. If I were in your situation, I'd try copying and pasting your entire comment into Claude and then go from there.
More options
Context Copy link
AI could probably do all these things if you handhold it, but these things are all general computer use which is something that AI is currently quite bad at so I doubt current models could do these types of things consistently. Bad performance at general computer use is a colossal bottleneck for all sorts of things where progress has been pretty slow.
The idea we're not going to have jobs in the future due to AI is just classic Lump of Labor fallacy. Something like 80% of the world's population were simple peasants in 1500, while after the Industrial Revolution that number dropped to ~1% for most advanced countries. If national economies could go from the vast majority of people simply scratching out enough food for themselves and their immediate families to fully industrialized societies without widespread unemployment, then the same can happen for AI.
Similarly, even though they're no longer needed as draft animals, horses are doing just fine.
In 1900 there were about 21 million of them in the US, whereas now that mechanization has made them much cheaper to feed there are, let me see... oh, no. Oh, no no no.
For anyone wondering, about 6-7 million horses in the US today.
Mules, being sterile, have fared worse.
More options
Context Copy link
Human population decline is already priced in thanks to demographic transition. No AI required.
I'm pretty sure that, if you could ask them, a horse would rather be a random horse today than a random horse 130 years ago.
But would they rather be a random horse 130 years ago, or not exist at all? Because that's what happened to the majority of horses that would have been here today; they were simply never born.
But given fertility rates, that's the case for people too- essentially voluntarily.
Calling the fertility crisis "essentially voluntarily" is disingenuous.
From Kevin Dolan's 2023 Natal Conference speech:
More options
Context Copy link
More options
Context Copy link
I can believe that the median horse life 130 years ago was net negative. Seeing a horse being beaten literally broke Nietzsche's brain.
Even if it was net positive, unless you'd like to bite the bullet of the repugnant conclusion, I don't find it to be too compelling.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
Horses are gigantic temperamental finnicky assholes who die if you look at them funny. Especially racehorses.
Counterpoint, Hell is other people.
More options
Context Copy link
My neighborhood has feral horses. They're fine, follow the rules, look both ways when crossing a street, and look after themselves.
Afaik feral horses and work horses are a lot better than racehorses but my personal experience is 99.5% racehorses who are bred for explosiveness via chronic inbreeding and thus are the worst.
More options
Context Copy link
More options
Context Copy link
I must be looking at them wrong. I've had too much experience with them and keep a beady eye on every one I encounter, but I've yet to kill one with a funny look.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
Yes, it will probably take several years to change the system, so that the users are just uploading the documents to the AI assisted system directly, and you are guiding them through how to find/upload them.
My job is offering pleasant experiences to children. It is already not strictly necessary. I suppose in the future an AI could order the supplies and curriculum. Actually I would prefer it to to organize a more distributed system for providing experiences to children, so they aren't stuck all condensed with up to a thousand children in a single school, and some of them freak out about that.
More options
Context Copy link
2 and 3: I expect the Claude Chrome add-on could handle those, depending on who each website is set-up. Two issues though: first, Claude is actually pretty slow going through Chrome, and no guarantee it wouldn't be faster to just do it yourself. Second, you're opening yourself up for all kinds of security risks at the moment.
1: As a below post says, agents can now use computers directly, but it's not great yet. Not knowing much about the situation, I expect an MCP would be enough, but it heavily depends on the software.
Tbh, the easiest solution would be to see if any of these have some kind of "Export to csv" option, then just chuck those files at Claude/ChatGPT/Whatever and tell it what you need
Computer use itself is mostly probably a stopgap. The long term solution is that those data warehouses that they need to extract information from will. have some way for the AI to auth into on your behalf and grab the files it needs. Computer use in the way people in this thread are talking is like building only human shaped robots.
More options
Context Copy link
Most of them would have, but if I go that far by myself, it's probably easier just to sort and filter manually and forget asking AI. This is the kind of routine work which I think is not considered when the talk about "administrative work will be automated away" is going on, be that alarm about job losses or boosting about 'it'll make everyone so productive'. Yeah, if all I did was write emails and reports, sure it'd be easier to tell Copilot 'just email Jenny about I got your email' but it's not. It very much is not.
AI agents are still in that awkward spot: "I spent 3 hours automating a task that would have taken me 5 minutes manually". Except previously only those skilled with computers could try this on a small range of tasks, and now anyone can do it for pretty much anything.
You absolutely could spend a week setting up Claude or Codex or whatever, putting in all the connectors and skills you need, and running a bunch of trial runs to fully automate a big chunk of your job. You'll still end up bottlenecked by stuff. Depends how much you actually enjoy working I suppose
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
Computer use is still pretty ass and I absolutely would not trust AI to navigate websites for you that aren't on the clearnet. I.e. if it requires a login tied to you, no good. If it's freely available via Google, go for it.
For your tasks, you can absolutely leverage AI to do the data processing steps required after you download the files.
I'm still in the process of figuring out how to use it for effective spreadsheet work. It's hit or miss, but the hits are fantastic. I find I need much more precise prompts for spreadsheet work, whereas for code/general tasks you can usually get away with saying "do X" whereas with spreadsheets I find I need to be a bit more precise "use X to accomplish Y, you'll probably want to Z first" .
It does sound like what you're doing involves some reconciliation/bridging. AI LOVES doing this. I've been using it to sift through shitty client accounting (my clients seem to deliberately hire the worst accountants) and it's fantastic. Just be warned, if the answer is there, it will find it eventually, but if the answer isn't there, it will come up with the most insane plugs to get there. So you have to understand the scope/problem, but it's saving me a LOT of headaches along the way. I haven't had to manually do a bridge between documents in a few months, thank god.
But I still have to do the downloading myself, and by that stage it's quicker for me to go "I need to sort out the 9 employees for centre 2 and sum up their gross wages in that period" with Excel rather than hand-hold an AI through all the steps about "no, don't include that staff member but do include this staff member" and so forth.
It may very well be
If this is a regular task and the various tacit rules wrt employees are stable, it's potentially worth codifying this so you can streamline the one step.
But that is a nonzero amount of work (frequently a lot of work) to set up a new workflow and it's your life!
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
Anything you need to sign into you'd probably need to get yourself. Not like something like that would be impossible for ai to do or get by itself long term, but for now it's both not secure enough to trust with your password, and not everything here is behind some sort of AI accessible API. Access mechanisms will need to be built out if you want AI to have access. Right now, it's fundamentally able to understand plain text and images. Anything beyond that, it needs to have tooling to allow it to get that info in such a format, and that tooling needs to basically be built out of plain text-operable mechanisms, think like command-line operable commands, http requests, etc.
More options
Context Copy link
More options
Context Copy link