Bro. You’re talking about replacing people at programming jobs, sure, maybe. But I’m literally watching people program shit with LLMs. The proof is in the pudding. A guy I know was annoyed with the app for a given task the creators had loaded with DRM and refused to update, and now working with nothing but LLMs, intelligence, and patience, he has an app with features and functions the humans were too lazy to implement.
Meanwhile right off the bat you link to a tweet where you accuse people of not paying attention to your points. Well. I guess I’ll have to wait for part four of your magnum opus to see you address the one in my pinned post, to the extent you will bother. Prediction: You will elide the most important elements and focus on what you think are the weak points. Nobody will be convinced of anything.
We have our biases and we’re all wearing them on our sleeves. Don’t front like that’s just us.
–
EDIT to add: The strawberrry thing. You can trick humans into embarrassing themselves too. Doesn’t mean the human is useless.
EDIT to add: I am not sufficiently educated to understand 99% of what you wrote, so take with grains, but your summation boiled down to the “it’s just collage” argument, which, again, is contradicted by the evidence of them producing original constructions.
Yes, they’re made out of information obtained by training, but THAT IS EXACTLY WHAT HUMANS DO. They’re just doing it very differently. Arguably worse, and you may have made a successful argument to that effect, if I could understand it without a graduate degree. But it doesn’t make them useless anymore than brain damaged humans are useless.
Ableist against my robot siblings, tsk tsk tsk.
EDIT to add: Not all humans learn by letters. Some dyslexic people learn their phonetic languages symbolically, recognizing the shape of the word rather than individual letters within it. Cool cognitive feat. Some AIs have some version of this, whether it’s up to speed or not yet. I have no idea how it works, but you can just try it out and watch ’em go. oh-oh it’s magic.
EDIT to add: I can see the labor argument on one hand. On the other, at my job, the material we work with is complex enough that literally nobody in the organization is right 100% of the time on it. When we mess up, it can cause people financial damage, up to and including losing their homes, because of hard technical limits on processing time for certain operations.
As soon as the LLMs reach better than human success rate, it would be immoral to let me keep my job, even while it puts my home in jeopardy. More people are at risk from our human failures than from our unemployment.
The question is whether our employer is going to have the wisdom to wait until the LLMs are demonstrably superior to the median employee in the organization. Magic Eightball says “Not Fucking Likely.” HJ Eightball says “It will never be superior to humans” and I’m like, are you and I looking at the same human species?
EDIT to add: You’re literally asking us to ignore the evidence of our senses and experiences, like a priest.
–
that’s it, i’m done for now! looking forward to loneliness part three, genuinely.
At least you stick to LLMs, not mix LLMs and LLM’s.
Also, you allow comments.
Literally everyone in the programming space has been talking about AI coding for the past year. Not everyone likes it. Everyone knows it is capable, and becoming more so. Like you say, the proof is in the pudding.
The notable thing about the strawberry test, is that most search results are at least a year old. People stopped talking about it maybe because models now answer correctly. A problem that demonstrates something interesting about LLMs, but otherwise of dubious importance, nonetheless solved.
The hallucination problem is more evergreen. If you speak to programmers, it’s one that materially affects their work, and is viscerally frustrating. It doesn’t make the models incapable of programming though.
The bro is full of it. Mental masturbation is how my teachers used to describe this kind of stuff, where you just say a bunch of smart sounding things, while everyone around you is actually doing stuff.
Anyone actually working in any field that involves programming understands just how incredible these models are and uses them A LOT. And if you’re in “team tabs” and stuff isn’t working for you because if it – you are welcome to stop using LLMs and continue writing your bad code for another what – year or two max?
Addendum – the reason by the way, why LLMs produce bad code when you use tabs is not because LLMs suck at tabs. It’s because people who use tabs are on average quite a bit worse at programming and the LLM is getting down to your level.
You can get similar results of LLMs performing better or worse depending on how you talk to them. Talk like an
idiot* and it’ll level with you.*edited by beeb per my possibly outdated and overapplied ableist language policy
I’m a bit of an outsider here, but I do both understand what Hornbeck is saying as well as what Bebe is saying and my opinion is; they are talking about different things huddling under the same LLM umbrella.
I’ve seen some fantastic results from LLMs from precisely-tailored training sets. I don’t believe the publicly available, companion AI, style LLMs perform to the same level. Or rather, they are trained on sets with both good and poor data, and from sources where the data is muddy and can be applied in multiple ways.
My impression of the current state of LLMs is that they can be useful tools, but the quality of the results has a lot to do with both the quality of the tool and the skill of the user. A skilled wood carver can only do so much with a dull chisel. The craftsman knows that to get the best results the tool has to be sharpened. For an LLM this means honing the training set to the task. An unskilled worker using a well-honed tool will get more variable results, in some cases it will be astounding in other cases it will not reflect anything that the worker intended. One of the problem with LLMs is that in a general sense the unskilled user may not recognize that the LLM generated slop, while an expert user will immediately be able to identify it.
This isn’t just a facet of LLMs, it’s true of most tools. All tools have limitations, as do all tool users. Does everyone remember the multi-tool craze a couple decades ago? A lot of people carried a single tool on their belt which was a screwdriver/pliers/wire cutter/leather punch/awl/can opener/etc. (I think for many of them only the corkscrew was ever used). A couple decades earlier there were Swiss Army knives with the same idea. One tool to rule them all, one tool to bind them. However, while these tools are fine in a pinch, none of them are as good at any specific task as a tool designed specifically for that task. You can use a Phillips screwdriver on a Reed & Prince screw, but it won’t work as well.
As far as I can see, this applies to LLMs as well. There really is a huge market for LLMs for specific tasks, it exists and is growing. But these are not generic LLMs which are marketed as companions or used for making unrealistic porn. These are built with specific training sets, and a lot are probably being built for specific companies for specific purposes. Even if another corporation wants to perform the same task, they will probably want to use their own, proprietary, training sets.
Bebe’s example above of the fellow building code using an LLM is an example of someone who has a very specific task in mind, and knows enough about how it all works to be able to properly generate prompts and prune areas where the LLM sprouted nonsense. I don’t know what LLM this person is using, but I would expect that the more general the training set the more specific their prompts have to be. They also need patience and skill, a complex problem won’t be solved by an LLM in a single pass, of course a human couldn’t write a complex app in a single pass either.
Hornbeck, on the other hand, is looking at how the LLM works and see that tokens are messy, have a lot of overlap, can quickly follow an unexpected trail, and can easily lead the non-experts (often called managers) to conclude they are doing a better job at the task then they really are. The managers look at an LLM and think it could be like a CNC Milling machine. A CNC Milling machine can make thousands of identical parts, an LLM can spit out thousands of working code blocks. Which would be fine if all code blocks are identical. But to get variation in a CNC Milling machine a skilled operator is needed to change the parameters (or put in a new paper tape in the old days), to get controlled variation in an LLM you need a skilled LLM operator.
Automation reduced the number of workers needed in manufacturing, but replaced a lot of those workers with people with skills to manage the automation. There will undoubtably be some false starts, but I expect LLMs (as they exist today) will still need skilled operators to generate the best results.
love ya flex, so reasonable
One thought I had a few weeks ago was that to some degree the LLMs are getting closer to what we wanted computers to do forty years ago. We all remember Star Trek: Save the Whales where Scotty picks up the mouse and speaks into it expecting the Mac to respond with a svelte voice. (And we desperately try to forget the next 30 seconds where Scotty apparently can not only learn an obsolete CAD program instantly, but also masters a QWERTY keyboard in milliseconds. Ahh, movie magic.)
What we were envisioning at the time, once we started linking all the computers together, was for them to be able to search through their databases and answer ANY question we might ask. Answer it to the depth of our understanding, with the ability to teach us more if we wanted. We wanted to be able to ask questions, get accurate answers, from something with infinite patience, willing to
dumb (with a “b”) it down* simplify it if we didn’t understand, or respond to follow-up questions at our speed, on our schedule.You know what? We have that. And we find it’s as full of human flaws and miss-information as other humans are. But it does have patience, it can
dumb things down* simplify things, or follow non-sequiturs, and is willing wait without interaction for days.It’s halfway there. Halfway to where we would like it to be. Not a god, not a friend, but a repository of knowledge which can be accessed at any time. Want to know which berries in the wood are edible? If you have internet access, and can learn to parse out the bullshit, you can. Want to know why Picasso and Braque invented cubism, you can learn it, even get translations of their own reasons (and you can judge for yourself if they succeeded in their goals). Have a specific task, like taking the gross output of logs harvested in Armenia and learn how much of that became toilet paper in Brazil? Well, it can’t do that yet, but give it a few years and maybe.
The problems with LLMs now are mainly human ones. Because humans can’t even agree on what is real or not, the LLMs can’t either. Frankly, I don’t want the LLMs to offer their own opinion on what constitutes reality. I’m not certain it will match any of our own ideas.
*censored by beeb per my possibly outdated and overapplied policy on ableist language
I’m struck by the way my posts attract nuanced and middle-ground comments sometimes, how even the pro-AI people tend to automatically concede some amount of ground to the anti-side, and how the anti-side has none of it. at least not on FtB. by not acceding to the utility of this technology in any way, while people are finding it useful, the anti-AI argument is making itself unreasonable, making that position seem absurd, quixotic.
i am always copping to the fact this will make our lives worse in various ways. in my post above, i mention i could lose my home and fuck up the lives of everyone who depends on me if my boss decides they’re good enough. the best argument in HJ’s post was the one he quoted, because it was grounded in easily observed realities we’re all facing. i’d contend the real problem is still capitalism and its good buddy fascism, not the tech, and i’d quibble on the word choice some, but otherwise agree with it.
i’m not seeing a flexibility to mirror my own on the opposite side. as much of a partisan as i am on this issue, that’s saying something. c’mon. try to see another point of view for a second, AI haters.
–
Cory Doctorow has been on a jag about that.
[evidence] (crossed out by blog hostess for reasons, see below)
Pluralistic: Criticizing the everything machine (06 Jun 2026)https://pluralistic.net/2026/06/06/applied-counterescatology/
Pluralistic: Refining humanity (05 Jun 2026)https://pluralistic.net/2026/06/05/defining-humanity/
Pluralistic: Delusion as a service (04 Jun 2026)https://pluralistic.net/2026/06/03/mission-space/
Pluralistic: The tedious power of storytelling (02 Jun 2026)https://pluralistic.net/2026/06/02/must-we-pretend/
Pluralistic: Molly Crabapple’s “Here Where We Live Is Our Country” (01 Jun 2026)https://pluralistic.net/2026/06/01/doikayt/
i don’t need more of the anti-AI position hosted in my comments, man. i wasn’t sure if you meant he was saying leftists risk looking unreasonable or if you meant he was acting like hj, so i looked at his blog before you posted that, and everything i was able to see was completely un-nuanced AI hate – as i’d have expected from the coiner of enshittification.
Ah, right. Sorry.
Point I intended, he was a thing.
Coined ‘enshittification’.
Wrote some damn good stuff.
Was clever, was cluey.
Obs, not endorsing it.
Saying that it’s one of those topics.
Sorts people out, sorta. Ahem.
(FWIW, I just saw Mano’s new post)
—
You know if I pollute your blog you can delete my stuff.
Not a problem, and I know you would indicate it.
You got my full permission.
I know I forgot your rule in that other thread, used bad words.
Not intentional. Was more or less stream of consciousness.
(Like this apology)
thanks for the pology, brudder. see u around.
I’ve been thinking about the fears people have of AI taking away all the jobs, and while I think that in an ideal society we wouldn’t need to have jobs to live, we can find parallels in the recent past of other disruptive technology where people at the time had similar fears.
Much like today’s scare about AI, when computers first hit the desktop the claim was that they would eliminate all secretarial jobs, all the printing, scheduling, collating, stapling, note-taking, dictation, etc., could be done by computers, making some of the only jobs which were majority filled by women (and low paying) obsolete. So let’s look at the job numbers over the last 60 years.
From the Bureau of Labor Statistics we find the following, looking only a non-farm jobs in January.
1974 – 78.1 million jobs
1984 – 92.7 million jobs
1994 – 112.6 million jobs
2004 – 130.7 million jobs
2014 – 137.6 million jobs
Let me continue, as I accidently hit some button combination which posted.
2024 – 151.0 million jobs.
Even with a large disruption in the types of jobs as computers enters the office, jobs continued to grow.
Obviously, this begs the question of, “Why, then, is it so hard to find a job?’
That is probably more related to the number of people who want/need to work. There are a large number of reasons why employee wages have not kept up with inflation over the last fifty years. One of them is the growth of the available workforce. A large part of that was the entry of woman into the workforce. Now, please don’t miss-understand me. I am 100% behind woman working. There are far more positive results of women entering the workforce than any negative impact on the negotiating power of workers. Women who have money, even a little bit of money from a job which pays unequally, have a lot more power over their own lives that a woman who is dependent on another to live. And it’s not just women who benefit from women entering the workforce, all of society does in many ways. But it is also true that an expanded workforce can dilute workers rights and protections. There is far less job security if an employer can hire a replacement for you tomorrow than if they would struggle for weeks.
There are other reasons for women to enter the workforce too, when workers do not get part of the improvements in productivity they generate, but prices continue to rise, the household with two (or more) full time wage earners becomes more common. So even if people are not really interested in joining the workforce, they have to do so to survive.
Again, a better form of society might allow a household to comfortably subsist on a single income, but that is not what our current US society requires.
I know bringing up the entry of women into the workforce is a very loaded topic, and mentioning that some negatives may have resulted because of that is inviting angry retorts. Very few things are 100% positive, when a change occurs there will hopefully be more positive effects than negative ones. But to ignoring negative impacts is a form of blindness which we should avoid.
That was a long digression, I know. But any discussion about jobs and wages which is not jingoism will necessarily have to explain a couple not-obvious trends.
My point, at the end of the day, is that even through the adoption of another disruptive technology jobs still existed, and the number of them even grew. I don’t expect the LLMs will be more disruptive than the introduction of the desktop computer. Certainly the types of jobs shifted somewhat, and there are still people who are upset about how the introduction of the desktop computer eliminated their positions. Yet, the introduction of the desktop computer also created jobs.
I would certainly prefer a society where we don’t have to spend 45+ hours a week earning a paycheck. But I’m not worried about my job being taken over by LLMs anytime soon. Not because my job couldn’t be performed by an advanced LLM (although I don’t think they could at the current time), but because there are a lot of jobs out there and churn happens
I would probably not post this, as I re-read it, it is not nearly as clear as I would like. But as I accidently posted the first part of this thought. I guess I’ll post the second part and see how much opprobrium I generate. I hope the reader understands.
this is where the argument HJ quoted at the end of his article stings the most. the ai boom is being so grossly boomy specifically because the creeps foresee firing everyone. very shortsighted, because then it’s institute UBI or the whole economy tanks, and you’re paying your bodyguards in bullets and water from your skull mountain base, until they decide to use your skin for a new flag. but then, that’s the essence of crapitalism.
many jobs will most certainly be replaced by AI before AI is ready to actually do those jobs, and shit will be a mess. fingers cross’d i can pay off my mortgage before it happens to me.