LLMs are very helpful for transcribing handwritten historical documents, but sometimes those documents contain language/ideas that a perfectly aligned LLM will refuse to output. Sometimes as a hard refusal, sometimes (even worse) by subtly cleaning up the language.
In my experience the latest batch of models are a lot better at transcribing the text verbatim without moralizing about it (i.e. at "understanding" that they're fulfilling a neutral role as a transcriber), but it was a really big issue in the GPT-3/4 era.
I have a project where I'm using LLMs to parse data from PDFs with a very complicated tabular layout. I've been using the latest Gemini models (flash and pro) for their strong visual reasoning, and they've generally been doing a really good job at it.
My prompt states that their job is to extract the text exactly as it appears in the PDF. One data point to be extracted is the race of each person listed. In one case, someone's race was "Indian". Gemini decided to extract it as "Native American". So ridiculous.
I was attempting to help someone who runs a small shop selling restored clothing set up a gemini pipeline that would restage images she took of clothing items with bad lighting, backgrounds, etc.
Basically anything that showed any “skin” on a mannequin it would refuse to interact with. Even just a top, unless she put pants on the mannequin.
Photoshop's AI tools will fail constantly if you try to say, remove an extraneous wire or a tree branch etc in a photo that has women showing any bare arms or legs etc. Works fine with men with no shirts on.
It pops up some moralizing text and refuses to continue.
It doesn't surprise me at all that this is going on. There are lots of social media fan pages that are run by real people who post real content 99% of the time but are willing to post promo material for a fee. Usually that fee is pretty high, easily $100-500 depending on the account's follower count, with different price points for how long it stays up (pay more for a permanent post, pay less and it gets deleted after X number of hours). It's really effective because those accounts already have a well-established presence and function as tastemakers.
It's simple (both in terms of gameplay and graphics) and it's the fastest Civ game to complete a full playthrough. Later releases made the game slower and more complex.
This has been the message at the F100 that one of my relatives works at. The CEO's increasingly aggressive message to their hundreds of thousands of employees is that they should figure out how to get 10x faster with AI or their job is on the line. The average non-technical white collar employee doesn't know the details of how LLMs work or any of the day-to-day changes in tooling that we see in the tech industry. All they see is elites pouring all their resources into a machine that will result in Great Depression 2 if it succeeds. Millions of people whose lives depend on their $50k office job in Middle America are hoping and praying that it fails.
I live in an area that's not a tech hub and lots of people get confrontational when they find out I work in tech. First they want to know if I'm working on AI, and once they're satisfied that the answer is no, they start interrogating me about it. Which companies are behind it, who their CEOs are, who's funding them, etc. All easily Googleable, but I'm seen as the AI expert because I work in tech.
My career is built on people not knowing how to Google lmao (IT)
To most people, AI is chatGTP. Maybe Gemini.
Claude? No idea.
VS Code, Cursor, Antigraivity, Claude Code? Blank stares.
Same as when the computer came, some will fall behind. Excel monkeys copy pasting numbers will go, copywriters, written word jobs = already gone. Art for simple images = AI now all done by one person.
Unless you want a Soviet system where jobs are kept to keep people busy.
> I don't care that when I click "delete", the item may not disappear from the screen immediately.
The disconnect here between tech people and non-tech people is that most users do care about stuff like this.
I run a popular website as a solo project so all the feedback/complaints are routed to me, and one thing I've learned is that users really don't want websites to "feel old". Sure, they want it to be fast, but they also want all the bells and whistles like loading indicators and animations.
If you show Hacker News to someone who's not a developer, especially if they're under 30-35, their reaction to the layout and functionality will be visceral disgust. I really can't stress enough how much modern users hate the traditional plain HTML look. If you're trying to convince users to use your site and it looks or functions anything like HN, they'll get angry and close the tab within seconds to look for an alternative. Even if you've made a SPA with plenty of bells and whistles, users will still get upset if anything feels "clunky", which is often user-speak for "this component needs animations and a transition state". They don't know or care that all the fancy stuff increases the complexity of the codebase.
Every software project hits a point where the super clean abstractions the developers came up with start to clash with the messy way it's used in the real world. This is the frontend version of that. We have no choice but to give users the UX they want.
It's interesting to observe that fame (and the money that usually comes with it) seems to follow something like a log scale. People usually don't become gradually more famous in a linear way. They're more likely to spend a few years with 50k listeners and then get a big hit and get 1 million listeners overnight, then the next big jump is 20 million, and so on.
It's possible to be semi-famous and still able to go to the grocery store and pump your own gas without getting recognized. The local sports radio guys don't need an entourage, even if they do get recognized. But as a rising artist, you hit a point where you can no longer go out in public at all. It's really shocking when it happens because it's so abrupt. My dad's famous friend was a regular at a local restaurant and wasn't bothered for a long time, even when his name/face started showing up in the media. Then one day another customer shouted his name and he got mobbed by fans, and he realized he couldn't go out to eat like a normal person anymore. I think Charli crossed that line with the success of her album Brat last year. It's the point where you start to ask yourself if it's really worth it, and maybe consider going full recluse like Thomas Pynchon. (That's not even getting into the online stan culture stuff that Charli talks about in the article.)
> I think Charli crossed that line with the success of her album Brat last year.
In Hollywood, that line gets crossed at a surprisingly low level. I am friends with Josh Sussman, who played Jacob Ben Israel on Glee. I occasionally visit him in LA, and we can’t go anywhere in public without getting constantly stopped by people wanting photos. It’s exhausting.
I didn't watch it myself, but Glee was a very popular show. Since Josh Susman was a recurring character, it's unsurprising that he'd have a large fanbase (especially in LA).
In the words of Adam Ant: it took us 3 years to be famous overnight.
I also heard about Matt Lucas, of Little Britain fame. He was slowly plugging away at it, and was about to give up. At around 30 years old, he teamed up with David Walliams, describing it as the last roll of the die. Their popularity exploded.
Morgan Freeman didn't become famous until he was in his 50's. Someone asked him if he was upset that it took so long. His response was: "No, because it didn't have to happen at all."
It's fascinating to me that her new album's name is "Wuthering Heights", the name of Kate Bush's debut and number 1 single from 1978. Kate Bush is well known (in the circles of people who know about this sort of thing) and as fiercely independent and self-controlled artist. I hope Charlii manages her career and fame as well as Kate has over the decades.
As I understand, Charli’s album is the soundtrack to a movie called Wuthering Heights. Which is loosely based on the 19th century novel of the same name. And that novel was also the inspiration for the Kate Bush song.
This is a big issue for young people, too. Every white collar career path is very on-rails now - you're expected to get a degree in XYZ and then get a fresh grad job as a Level 1 XYZer and so on.
So the stakes are drastically higher for 18 year olds picking their college majors. It's effectively a life commitment for a specific career path, and there's a lot of anxiety among students because they don't know if the career path they're betting the farm on will still be be viable by the time they graduate. There's also a sense that if you can't manage to find work in the field you majored in within a year or two of graduating, you've fallen off the track and are condemned to DoorDashing forever.
I'm always amazed at how many older people I know (especially 60+) spent their twenties directionless and then started a decent paying career in their 30s, often by simply learning how to do something and getting a job doing it. I'm not sure what policy platform would make that possible again, but accomplishing it would alleviate a ton of the anxiety that young people have today.
It's been years now, but I used to be involved in the trading market for Team Fortress 2. There were people who did TF2 trading as a full-time job, exploiting arbitrage between markets and holding items that were expected to increase in value (and sometimes using bots to farm items).
The Mann vs. Machine update in 2012 added a new game mode that would give players loot as a reward for completing missions. Players who didn't care about the trading market (i.e. the vast majority) would look up trading sites after a gaming session to offload their stuff fast without caring much about the value. People who described themselves as "quickbuyers" would aim for the people who wanted quick and simple transactions and then sell the item elsewhere for a markup. I did this for a while and averaged $5-10 of profit per day, usually 20-30 cents per transaction. Someone treating it as a full-time job could probably have made a lot more, like $20+ per day.
That was pretty good money if you a) were a kid with no living expenses, or b) lived in a developing country where the money went farther. I was in group A. Any time I wanted to buy a Steam game, I'd put up my quickbuyer listing on the trading sites and save up the money. But I suspected at the time that most people in the scene were in group B. If I were Valve I'd struggle to pull the trigger on a major change to the in-game economy knowing that it would affect the livelihood of a nontrivial number of people in countries like Venezuela, and maybe that's the reason they seem to have hesitated for so long.
That's wild! I thought that would have disappeared with Doordash, etc.
Can you pay a doordasher when the food arrives? I assume that's all through CCs.
I'm sure DoorDash doesn't allow it. But a lot of older people call for pizza the way they've always done for decades, so it's common enough that the pizza places (at least in my low-crime suburban area) have decided to keep allowing it.
They usually have some sort of system where your address is connected with your phone number after your first order, so they must be able to see that you've called X times and paid reliably in the past.
I think remote work gets increasingly hard to manage the larger a company gets.
My parents both worked for the same Fortune 500 company when COVID hit and the thousands of employees in their branch had to abruptly transition to WFH. Something like 10% of employees just disappeared, never to be heard from again. Lots of people who had been perfectly fine employees in the office ended up getting fired because with WFH they couldn't manage to stay at their desk and get their work done. That division of the company was seriously crippled for about six months.
My own job is with a small business that has been remote-only since before COVID and it's all been great. They've never even needed to "prune" anyone who abused remote work. I guess they're good at determining how reliable someone will be during interviews. We're all adults and there's a high level of trust that we're all doing our jobs, but the team is small enough that it would take a maximum of a single day to notice if someone is slacking.
But when the company gets really large, they sometimes have to manage to the lowest common denominator, and "we're all adults" becomes an increasingly shaky assumption. So I kind of understand where the anti-WFH CEOs are coming from if they were at the helm of a massive company and saw all kinds of chaos during COVID. But I also think small, geographically distributed teams can massively outperform if you hire the right people.
In my experience the latest batch of models are a lot better at transcribing the text verbatim without moralizing about it (i.e. at "understanding" that they're fulfilling a neutral role as a transcriber), but it was a really big issue in the GPT-3/4 era.
reply