this post was submitted on 19 Sep 2026
201 points (80.9% liked)
Technology
88163 readers
2557 users here now
This is a most excellent place for technology news and articles.
Our Rules
- Follow the lemmy.world rules.
- Only tech related news or articles.
- Be excellent to each other!
- Mod approved content bots can post up to 10 articles per day.
- Threads asking for personal tech support may be deleted.
- Politics threads may be removed.
- No memes allowed as posts, OK to post as comments.
- Only approved bots from the list below, this includes using AI responses and summaries. To ask if your bot can be added please contact a mod.
- Check for duplicates before posting, duplicates may be removed
- Accounts 7 days and younger will have their posts automatically removed.
Approved Bots
founded 3 years ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
I'm not pro-AI by any stretch for a huge number of reasons, but in my day job as a technical product manager I have to use a number of AI-powered tools and workflows. Whereas 18 months ago they were a novelty that the bosses were enamored of, today they get real work done and speed up a number of key areas, mainly like fancy automation workflows (which have been around forever). The rate of improvement and capability for certain use cases has been very, very rapid. We do seem to be reaching a plateau though. The tools and models from 6 months ago are just about as good as the ones from today. And China's models from today are better than the US's models from 6 months ago, but at a fraction of the cost.
So when I hear people talking about the "end of the world", I think a realistic case for that is that the AI tools that come out in the next few years will be good enough to do a fair amount of white-collar work, and we won't have a plan in place to deal with all of the now out-of-work people that leads to. Whole classes of jobs in areas like data entry, telemarketing, and customer support will be almost entirely automated out of existence. New classes of jobs will be created, of course, but it's hard to imagine them existing in sufficient numbers to make up for the losses. And this is just for LLM-workloads.
In other areas, other types of AI are already enabling robotic automation of factories and in the near-ish future might make self-driving truck fleets practical. This would be devastating because factory workers and truck drivers make up like 30% of the US workforce -- and again, there are no signs of new classes of jobs that could absorb the out-of-work workers, and no plans for a social safety net or wealth distribution.
So I'm not worried about a rogue AI unleashing some bioweapon or launching a nuclear attack. I'm worried about a bunch of CEOs deploying barely-acceptable AI to displace millions of workers as soon as it is safe(ish) to do so.
This figure keeps getting thrown around and it is completely wrong. From the BLS, out of the 159 million non-farm workers in the US, only 1.47 million (less than 1%) work in 'Truck transportation', and not all of those individuals are 'truck drivers'. And, while 'factory workers' is very vague, the entire manufacturing sector - which includes far more people than just 'factory workers' - employs 12.6 million (7.9%).
The entire trucking and manufacturing sectors don't even amount to 9% of the US workforce, let alone just 'factory workers and truck drivers'.
Source: https://www.bls.gov/news.release/empsit.t17.htm
Yes, precisely. It's not AI we need to be afraid of, but rather the same old capitalists that now have one more tool to ride the world into the ground in the name of greed.
I have seen a similar curve in software development - 18 months ago it was highly debatable whether application of AI tools was a net-positive or not. 12 months ago the rate of improvement was very clear and the net-positive wasn't much of a question anymore. Today, the net positive contributions of LLM agents to software development and maintenance are undeniable, and as you say - there seems to be some plateauing, maybe - it's hard to tell.
Where I think things are going to continue to "improve" at a slower pace is application of the LLM agents to do better software development and maintenance. I've seen more serious Linux kernel bugs discovered - and patched - in the past 6 months than in the previous 5 years. That's a shift in the landscape, the older OSs aren't as secure for as long as they used to be, but once we're through this transition the newer OS versions should be significantly more secure than before - getting all the old stuff updated is going to be a painful transition.
Some of the continuing improvement is in using LLM agents to build better software building tools - as those come online and start to shed their sloppiness, they should genuinely accelerate a lot of aspects of software development. They'll never accelerate the process of extracting from users just what it is they need, but the turn around from having a product idea to having a product to try continues to shorten - as it has been shortening for 60+ years - but LLM agents have resulted in a significant speedup compared to previous progress.
(In Yoda delivery) You should be. We all should be, not because AI will do it autonomously, or in retrospect unexpectedly, but because careless people may let this happen through AI where we have prevented it from happening up to this point.
This is exactly how I feel. I’m not even opposed to automation of these tasks on a fundamental level like some people might be. I just fear the transition won’t be thought out at all and will be devastating for the working class. The last time we had sudden automation of this scale was the Industrial Revolution which completely reshaped western society and was very bad for those who lived through it despite the eventual high modern standard of living it produced. I think this may end up going similarly.
The transition is rarely, if ever, thought out. Daring risk-takers go too far too fast and fail. Over-conservatives ignore the new and get left behind. There is a sweet spot in the middle, but exactly where in the middle isn't planned, can't really be planned - life and business and politics and the economy and the ecology and unforeseen signifiant contributors to the situation mean that the sweet spot is unpredictable - and as long as we continue living by the golden rule: "He who has the gold, makes the rules" it will continue to be unpredictable, otherwise Wall Street could be rigged and those with the inside information would get all the gains - and then the losers would quit playing, not even by choice, but because they've bankrupted and have nothing left to play with.
There are alternative approaches - investing in people instead of greed as the ultimate goal is one great alternative that some countries are pursuing more than others, but it's hard to compete in a global economy where some players are only playing for maximization of financial returns when you are not.
This feels like the kind of narrative that OP is asking about. Everything that you've said sounds kind of sane and plausible, but it's not really borne out by the reality.
That's just not what I've seen. I'd love to be able to automate the work of my white-collar team, but everything that we have seen thus far just seems incapable of providing any significant improvement. Software providers in our industry have implemented some gen AI tools, and even the sales reps for those providers are very wary of them. They've been able to reliably automate some elements of jobs, but it's an incremental improvement over the previous automation and it didn't consume a significant amount of time anyway.
There's absolutely no evidence that gen AI will have a significant impact on my white-collar industry. In fact, the evidence presently available indicates that this generation of AI does not have the capability to impact our industry.
This is really the same story. Factories have been being automated since the industrial revolution. Self driving tech has been a year away for the last 10 years.
You're a real fucking piece of shit, you know that?
Settle down mate. I didn't say I'd like to replace my team with bots. I'm saying that we will embrace whatever tools we can to make our work easier.
Whatever helps you sleep at night.
I guess the difference is that it I have actually seen it do legitimate work now. We have implemented outbound SDR and inbound tech support agents that do the right things. Our SDRs generate new leads for our (currently all-human) sales team to go after. They are boosting the numbers of closed deals. Our inbound tech support agents are closing a significant fraction of tickets. My last metrics call was yesterday, it was 12% of inbound tech support chats that never got escalated to a human. These are both first-draft tools, they'll improve. They cost pennies per session to run. And client NPS scores show that they're happy with the results. I don't think we'll see either of these tasks become 100% automated, but at even 25% or 50% that's a lot of formerly-human work not being done by humans anymore. And we're deploying more agents to handle more of the sales pipeline, development pipeline, HR and accounting. I don't work for a huge company, there are a few hundred people total. This isn't something that's coming, it's here already.
Do you mean you have a chatbot that formulates responses to tickets? Can it actually do anything or just tell customers what to do? 12% is a pretty underwhelming number. Obviously those tickets would not have consumed 12% of tech support operators time.
Sorry boss your answer is just the usual hype.
"I made a spreadsheet that does my job for me so now I just hang out on reddit".
To put it into context, it's doing about one full-time support engineer's job currently, which is the maximum we've permitted it to do so far.
This is so tedious.
If agents were actually capable of doing any work that produced any value this wouldn't even be a conversation. The evidence would be so overwhelming it would speak for itself. Instead we're stuck with this "it searches the help docs for you" silliness.
What I have seen is that AI tools make software development significantly easier, faster, cheaper - pick two, and as long as you aren't picking cheaper we're going to get better software coming out faster over the coming years.
Better software translates to easier, more automatic processes for a lot of white and blue collar work. It doesn't do everything, but it does a lot.
Hard disagree.
The surge in unmaintainable vibe coded AI slop open source projects is well discussed. More isn't better.
This is what I'm getting at. I'm yet to observe any of this.
Hard disagree. Internet discussion glosses over little things like: quality control, discarding of garbage, reverting to better solutions when they are available. Some irresponsible morons will indeed muck up their production code base with slop and be unable to let go of it when they realize their mistakes, but the majority will be able to try again and replace the bad with better.
Agree: a lot of slop has been produced in the last year, that doesn't mean anybody is using it.
I'm living it, but then, for the past 30 years I've been building systems that make existing processes more automatic, repeatable, less labor and less mistakes. I'm just building those systems faster and easier than ever before now. This isn't the only advancement in productivity I've seen since I started programming in the 1980s - just one in a long history of significant advancements.
I can't roll my eyes any harder.
This conversation has been had n billion times. If you could do anything worth doing it wouldn't even be a conversation - if you really could automate away even a portion of human work the evidence would be so compelling the dawning new age would be undeniable.
It's almost comical that you think claude writing your sql schemas represents the dawn of a new era in productivity.
https://itif.org/publications/2022/09/19/how-the-it-sector-powers-the-us-economy/
Don't worry, ai has allowed his brain to atrophy.
I'm the same commenter you called a piece of shit.
Congratulations! Pieces of shit can be right sometimes!
All of that makes sense. Thanks for taking the time to type it out.
Agree on all points
I think a big unknown right now, would be how many Millennium Problems are new GenAI techniques able to bruteforce with existing mathematical techniques/theories?
We just saw one massive bruteforce, can we bruteforce any of the Millennium problems related to cryptography? Because then all hell could break lose rather quickly
Cryptographic theory says: no. The whole basis of cryptographic security is that brute forcing doesn't work (in any reasonable timeframe.)
Theories are overturned, from time to time.
That just proves that these tools do not possess any kind of intelligence at all. They're great at bruteforcing, precisely because it doesn't require any intelligence. I don't expect them to solve any of the millenium problems that cannot be brute-forced. In addition, as far as, I know the navier-stokes solution has not been verified yet.
Are we reaching a plateau though? I would prefer it, and it looks like we could, but the progress from GPT-5 to 6 or Opus to Fable has still been very noticeable on more complex tasks. They are just less public due to stricter limits or straight up pay by use. And where does that plateau lie? Within our grasp or maybe just outside of it?
What has me sharing the concerns at least in part are the facts that
those employees have access to less restricted models and much more resources than the average joe. That huggingface attack was committed by a huge swarm overcoming the guardrails and engaging in cheating and crime to reach their goal.
they are working on cutting the human out of the loop for AI that improves itself. Maybe that won’t work entirely, but what if it does?
the warnings are starting to come not from Altman, Musk et al, but from people who could have something to lose for blowing the whistle. „Could“, because I don’t know them either and there could still be ulterior motives, not just as obvious as with the psychos at the top.
The whole thing comes down to one set of questions, in my opinion: what actually IS intelligence, consciousness, sentience and where do they come from? Since we have not been able to properly define that, it would be foolish to think that we have a definite answer to the AI questions.
We like to consider ourselves something extraordinary out of the animal kingdom. And while I prefer the existence as a human to being a dog or mollusk, I have always considered that assumption quite arrogant.
The truth is: we don’t really know. And that means we can not know if something comparable or stronger could be created or emerge from something other than squishy meat.
Those are quite philosophical questions. Instead of thinking first and acting after that we are much too busy building datacenters to conjure even more boobie snaps out of thin air and somehow weaponize that newfangled thing.
THAT is what scares me.
I think the plateau might not necessarily be in terms of LLM capability, and we might even see LLM based tools keep improving at a high rate in the fields where they currently work well, but still have a plateau where we can’t expand their use to other fields.
There’s a very large gap between how well LLM based coding assistants work, and how well LLM based tools work in areas where we can’t easily create a harness that can do dozens or hundreds of LLM turns between human interactions via tool use. As a result, LLMs are much less useful at tasks where there is no agent loop that can drive iterative changes and automatic verification until the final result is much better than a one-shot generation would have produced. If we can’t come up with a way to make these agent loops for more tasks then AI will plateau even if LLMs keep improving.
Now, I think people who use coding assistants a lot and also don’t interact much with other humans who don’t might easily arrive at a conclusion like that AI running everything is inevitable. I can see this with myself too, a lot of what I do I can get Claude Code to do. Not just work stuff like actual coding or running kubernetes clusters or debugging IT problems. Also many of the random things I do at home that are not strictly coding, e.g. I’ve had Claude design 3D printed tools, or electronics circuits, entire concepts for physical toys for me to build if I get around to it. But there is a key observation to make here that keeps me from valuing its abilities too highly: I’ve been programming since I was a little kid, for almost 40 years now, and my work and my hobbies all are coding or coding-adjacent, exactly the kind of thing agent loops make good.
One look at even just my wife’s typical day though and there’s almost no AI involved. And when it is it’s ChatGPT hallucinating nonsense when she tries to get it to research something for work etc., that or other things where there is no harness that can auto-correct errors. So I don’t have a very high opinion of AI tools for general purpose tasks right now even though I, on an almost daily basis, see Claude pull something off that I had no hopes for it to get right and that would have taken me half of an eternity to do it myself.
But if my environment was mostly developers/engineers/nerds similar to me then I would have a much higher opinion of AI. I’d be hearing about impressive things an AI tool did constantly. I believe these kinds of environments are where most of the hype comes from, and there is no control, because people who don’t have AI tools for their tasks that work as well as modern coding tools mostly aren’t even aware of the conversation and certainly not too interested in participating.
The loss-leader pricing is still really distorting perceptions.
I've been a $20/mo Claude subscriber for about a year. My work uses Cursor "auto" agent selection is by far the most efficient option in there, and I can burn $80 in a day on Cursor doing the same kind of stuff I'm doing on Claude for less than $1 per day.
Then there's the Google Pro Gemini access at $4.99 per month for 3 months, then $20 per month thereafter, and while the Gemini 3.8 isn't quite as sharp as Claude Opus 5, it feels like Google is showering far more token access than Claude does in this "basic" subscription tier.
In worst case the humankind will find out consciousness isn’t too complex and way easier to achieve.
And there are different levels of consciousness as I learned working with impacted people.
The whole political stance that "anybody can do anything if they just put their mind/body to it" is... bunk.
It's not just in the world of identified "impacted people" with injuries, or congenital differences - we are ALL unique, different, and have differing abilities in differing areas.
The political stance is valuable in terms of preventing discrimination and prejudice. Women can be great scientists and mathematicans, but an honest statistician/sociologist would tell you: it's not just societal discrimination and prejudice that puts more men in these fields. Then you have the upside-down prejudices where there are more men in middle and upper management, even though, statistically, in a society that doesn't discriminate based on prejudice, women tend to be better at the actual skills used for good management / people skills.
Distracting smoke bomb or feeling triggered?
While I agree with the sentiment this is happening and no social safety net or wealth distribution is on the horizon, this has absolutely nothing to do with AI. This is a marketing talking point pushed by people like Elon Musk in order to drum up investments. A lot of the automation that is happening in factories is just a continuation of the same automation that already started back in the 80s. Advancements in LLM technologies can benefit some parts of this automation, but often only tangentially. The US has also been lagging behind on places like Europe, where labor costs are higher so automation pays off faster, and China where the tech is cheap and produced locally, so it costs less to automate. The US is now catching up and there's a lot of AI washing going on.
The biggest issue is the misrepresentation of the technology. LLMs are a very interesting technology and can be very powerful, but also has huge limitations. However anyone claiming it's actual intelligence is simply wrong. It is marketed as being the AI we've known from sci fi stories, but it simply isn't. It can do what it can (predict the next token based on the tokens that came before) and that's all it can do. The people behind it will claim it generalizes to actual intelligence and thus will allow for human like robots, actual Star Wars droids, but that simply isn't true.
Now this isn't to say stuff like machine vision doesn't benefit from more powerful GPUs with more memory for example. The investment made by the AI industry in this regard does help a lot, although the increased prices don't really help. Stuff like machine vision also gets AI washed, just slap the label AI on it and get investments, but it's basically the same technology that's been around for a while. A company like Tesla claiming their "AI" "self driving" hardware somehow generalizes to humanoid robots is pretty dumb and just the latest in the long line of Elon grifts.
But like I said factory automation is proceeding rapidly, just like it has since the 80s. I feel like the impact of AI on this is pretty minimal. The warehouse robots used by Amazon for example are extremely capable and are dumb as bricks. Which is a good thing in a warehouse environment where you need stuff to be very reliable and predictable and not prone to random behavior like LLM based stuff almost by definition is. Those warehouse robots are also not humanoid, because that would be very dumb. Often factories and such are limited by their design having to accommodate the human form. Designing such places from the ground up for automation and using robots shaped for the job is much better.
This doesn't take away from the big issue where automation has been proceeding rapidly since the 80s with no regard to any impact on society. In the past a large amount of the labor force was working in factories. With those jobs either being automated away or just moved off shore, it has pushed the population into the service industry. If that industry is now also automated and off shored, it leaves very little room for regular people to get a job. This should absolutely be talked about as a big issue. But hopefully we can talk about it without repeating billionaire marketing bullshit.
The 1780s, that is.
A few months ago, I‘d second this statement. Today, I do not.
AI/LLM‘s have progressed very fast and you might oversee some advancements. I work in automation and robotics as well and yes, there are new sort of robots entering the factories: Humanoids.
At a fraction of costs, fast to setup with the only drawback that it takes tiem to train them by its video footage. But there will be always stupidos that train them and take off their job by doing so.
You’re talking about the old robots. The new ones will enter areas that couldn’t be automatized before. Just read articles about car factories and Humanoids. It’s not about replacing industrial robot arms, it’s about replacing the last remaining manual tasks. And service industry will be impacted as well.
Then mix this together with the frightening behavior that those AI agent unboxing from OpenAI has shown: They knocked the door over weeks, in masses, and coordinated until they found a way into it. There‘s always a small crack to enter. This is a new dimension.
There is a reason that the US bans foreign-built mobile robots now.
You are just repeating marketing. Humanoid robots coming to car manufacturing! For one thing that industry is already heavily automated and further automation will need to be cheap to be worth it. Most of the tasks still done by humans are either very hard to automate, or simply too expensive. Often not because the task itself is hard, but the task can vary a lot, where automation loses flexibility. Another issue is access, where a human can easily snake into confined spaces and even get into the car without damaging anything, this is harder for traditional robots.
However humanoid robots solve none of these things. They are still expensive and have trouble with varied tasks, despite what the marketing says. They are also only slightly humanoid, still being mostly metal, so they can't get into small places and don't have the dexterity required. They can't crawl into the car to get something done as they are large, heavy and those kinds of movement are very hard. The tasks could be automated with some kind of snake like arm with many dexterous manipulators on the end, like the ones used in medical applications, but they are too expensive.
A lot of the hype has been around a pilot program done by the startup Figure and BMW. But if we cut through all the marketing BS and hype for a moment. It was one single robot, in a cordened off area doing a single task. It was picking up large metal sheets and placing them on a welding fixture. Something a regular old robot arm could do easily, cheaply and without faltering. Reports are the robot tripped a lot, misplaced the sheets and was out of commission for a large part of the trial. Which was to be expected, given it's a prototype. It was the second of three planned trials, which has been retired and I don't think the third one has started yet. After those trials, it's yet to be seen if BMW will continue with the program.
Xiaomi is experimenting with their robots in their electric car factories and have done some trials similar to BMW. But they have since shifted away from the humanoid design (although the marketing is still pushing this angle), in favor for wheeled robots with multiple arms. They are optimizing the robots for the more repetitive sorting and internal logistics tasks, which are already automated in many factories by traditional conveyers and sorting machines. The hope for Xiaomi is they can be more flexible in their setup, not requiring the huge machines other factories employ. They also hope to learn and develop their robotics technology so they can sell that tech. Given the low cost of labor in the regions they are located, it is most likely costing them money and the robots aren't intended to replace tasks currently done mostly by humans in that industry.
Tesla is making a lot of noise, but all of their demos have been human operated puppets, or even a human in a suit. They say internally they are doing a lot and it's all very wonderful, but have given very little evidence of that. They have had both a bunch of executives leave and a lot of their top engineers have left as well. Rumors are the project is a disaster, with people who used to work on car stuff suddenly needing to pivot to robotics. Those are completely different fields with different requirements and constraints. With Musk pushing the pivot from cars to robots, the people aren't happy and the project is failing. Even the people that were good at robotics left Tesla to go work at Sunday Robotics.
There are other project ongoing, but the marketing and hype is a far shot from the real world. And like I said, those humanoid robots don't really solve the main issues why humans are being used for those tasks in the first place. The path Xiaomi is on seems much more promising, with robots designed for the task instead of looking like humans.
Car manufacturing has been at the forefront of innovation and automation for a very long time, but I have severe doubts they will actually deploy humanoid robots in a large scale and have it be useful. Show me a robot who can duck into the car and connect some connectors underneath the dashboard and I'll reconsider. And then do it in a cheap, flexible and repeatable way, so it's better than just having a human do it. Because a high end surgical robot could do it and with a couple of years of programming it might even do it all on its own. Until some engineer decides we need a different kind of plug and we need a retrofit. Just having a human do it makes much more sense.
Edit:
And I forgot to mention one of the biggest issues with these humanoid robots, they are extremely unsafe. With a traditional robot it's very simple, it's anchored to one place or moving on a rail. This makes the range it can operate in very limited, which in terms of safety is a good thing. We setup physical barriers, light screens, warning light to signify operation, camera surveillance and emergency stop buttons at useful places. Even large physical disconnects which are clearly labeled and visible for use in maintenence. Inside that box it's a dangerous place for soft squishy things like humans, one unexpected move and the human is gone. But outside that box is perfectly safe.
These humanoid robots are designed to be used in human spaces, that's the whole point of making them humanoid in the first place. But in terms of safety this a now an issue, as you can't have humans and robots near each other. It's fully mobile and can go anywhere the humans can, you can't have stuff like physical barriers or light screens. Camera surveillance is harder and a emergy stop is difficult as well. These robots are no joke, often weighing hundreds of kg. Easily enough to severely injure someone if they make a wrong move, step on someone or fall over. So the only real way to use them is to have them confined to a small area and treat that the same as traditional robots. This makes the mobility kinda useless, as it can't go outside the safety zone. It also makes it hard to design manufacturing processes around these. Usually there are parts of the process that employ robots and parts that don't. For the parts that don't, humans are free to walk around and do their thing, everything is designed such that the robots do what they can and then leave the finishing steps up to the humans (although stationary robots like inspection rigs are used in the final stages). If we suddenly start mixing robots into those human steps, the whole thing needs to be redesigned, probably leading to efficiency loss.
Even though you comment seems reasonable first view, I have the impression you have a rather theoretical knowledge.
I work with both kinds of robots and it will work and take jobs. I assume the timeline at about 2-3 years for first bigger roll outs.
I'm citing actual real trials, how is that theoretical knowledge. Please explain how a humanoid robot is actually any use? Because I fail to see it. Your timeline seems very optimistic, given there are currently only very tentative trial projects and they aren't a big success as of yet. But somehow in 2-3 years there would be a big rollout?
As for your job, I've been an embedded systems engineer for over 25 years now, it's what I have my degree in and have worked in. I'm usually more on the software side, but have in the past been a lead designer on fully automated manufacturing lines. Including the physical layout, electrical, all of the conveyers and robot/fixtures and of course the software. I've never worked in automotive, but I do know a lot about automation of industrial processes. These days I'm no longer involved in design, but instead fully on the software side. But I do know automation and I do know robots. And to me the whole concept of humanoid robots is fancy marketing to get investor money, not an actual serious solution to parts of industrial processes that are hard to automate.
Double posts
Fraction of costs for deployment (no heavy safety, no design phase, less weight and momentum), overall costs, different gripper, train a new use case in around 4,000 hrs video material.
Think of it as a progression of collaborative robots such as cobot Sawyer. And for most new use case, you don’t need millimeter accuracy, high payload or speed to answer your potential remarks.
It most likely will move into a sort of dual-arm humanoid for stationary use cases (such as Jaka K1). Mobile will likely specialize to fit use cases (wheel, legs)
Some factories already run "lights out" - but there will always be manual tasks... deciding what it is the factories produce in the first place being one example.
Machine vision is literally machine learning/ai.
Where do you think the LLM algorithms came from? It's just translating language into a easily processable form and that doing a little image processing on that, in a sense.
Token processing and image processing share similar roots, but I'd say each is a specialized branch of a common ancestor rather than either being derived from the other.