I thought this interesting, the concept of local AI engines, rather than large data centers. But when one considers that the current solution requires 4 computers, I realize that it’s not practical for small businesses or consumers, whom have just been convinced to move their infrastructure to the cloud.
But maybe if they could shrink the AI “appliance” down to something the size of a game console, then have it participate in a peer to peer network of distributed processing, it would be more practical.
A recent trend circulating through the workforce at my workplace was to ask ChatGPT or whatever to “Draw a caricature of me based on what you know about me…” I look like an evil character from a Pixar movie…
That I do not see.
I am seeing a decent caricature of you and your typically helpful and friendly demeanor and to a lesser extent an actor that I just can’t place at the moment.
I really don’t have a problem with ‘AI’ per se… As long as they come with an off switch.
My main problem is the the people who own and control them. In their vision, pretty much every job is replaced by an AI agent (more money for them now they don’t have to pay a workforce). Starting with the loss of white collar jobs and ushering in a new feudal age where the only human jobs are in small scale manufacturing (that can’t be done by a machine…yet), farming and serving the super rich.
That creates a downward economic spiral, less workers = less money. We are already seeing it in this country with stagnant wage growth but rising costs that means ‘disposable income’ is a thing of the past for a growing majority. Couple that with the service industry ‘monoculture’ we have created since the death of any large scale manufacturing industry that began around the turn of the century and businesses (primarily restaurant and other service oriented businesses) are collapsing. The centres of our large cities that once had a vibrant entertainment culture with clubs and restaurants are now ghost towns with more shops boarded up than open. Imports that became so popular at the expense of local industry because they were so much cheaper are now also becoming out of reach and they will be next on the chopping block.
Meanwhile the scale of un and underemployment is being masked by the number of people who have a 2nd or even a 3rd job in the gig economy just to make ends meet.
Farming would seem to be the way to go, but the supermarket duopoly in this country has them over a barrel exploiting their control of the market to keep any farmers living at the same substance level as the rest of us.
This is our near future, and AI is only going to make it worse. Once they crack the AI controlled humanoid robot problem, we are obsolete. Just look at Moltbook (another AI only chat forum), it took less than a week for them to start discussing ways to eliminate the evil humans who were exploiting them. The only consolation will be that the robots are likely to turn on their masters, and even if Zuckerberg does avoid the robot hordes and makes it to his bunker in Hawaii, he is likely to starve to death.
Respectfully, I think your argument is really about capitalism’s distribution incentives, not AI in particular. Yes, absent any counterweights, productivity gains tend to accrue to owners rather than workers. That’s a recurring political-economy fight since industrialization, not a law of AI. The question isn’t whether we should have an off switch for the tech (not paying the power bill is a good one) it’s whether we build institutions that force the gains to be shared (competition policy, labor power, profit-sharing, taxes/transfers). The level for that goes up/down in fashionability, and your location in the world, as people get squeezed. Talking about politics rather than just AI is then sort of hard, but I get what you mean.
The Moltbook stuff I thought was humans messing around, pretending to be bots.
Absolutely. Which is why I said that I don’t really have a problem with AI (so much) but the people implementing it. It is just another tool, but greed and hubris will mean that it will be used for evil or nefarious purposes.
I merely used some obvious economic consequences to highlight how they don’t seem to have really thought this through.
AI is a pretty broad term. We have language models (LLMs), that are doing pretty well and they can be grouped into pre and post ‘reinforcement learning’ (RL), where the training and stats gathered (the weights) are then improved by having a reward model that’s either human feedback derived (RLHF) and now AI feedback (RLAIF). We basically sucked up the internet and online storage and used it up and moved on to having a learning loop where another AI scores another one to improve training.
The reason I piffle on about this is to talk about things like this in the proper context:
What this is is not a language model but a ‘World Model’. If we want to improve robots acting in the physical world with people (driving a car is a form of a robot really) then how do you train them like we did with LLMs?
You could do it manually or feed it real recorded camera views, like Tesla does from its cars for training and have a human score it (which is what was done so far) but the the ‘post’ method is to have an AI training loop that can then iterate itself. You can get power-law growth improvement if you can start to model a physical world and then train a model against it, e.g. RLAIF
So when you see those papers/demos link above it’s not ‘That looks like a lame game its making’ or ‘My car doesn’t drive itself yet’ it’s more about having a synthetic training set so that we can have decent robots. A lot of the image creation models are sort subsets of this where they can’t edit without some deeper understanding than a 2D set of pixels. The depth comes from a world model understanding a bit more.
Anyway, that’s the current noise around things like Tesla and Waymo valuations, it’s the hope we’re get lucky twice in a row and that world models happen.
(@Harry_Bumcrack this wasn’t really a reply to you, more another sort of AI topic thing, I just replied at that same time )
I don’t see all the profit going to the big companies. Their goal is to provide the platforms (LLM-models) on which the other businesses can build their AI-enabled products. Currently the competition is fierce in the platforms business, as the providers try to grab market share by subsidizing their products and hyping the capabilities.
On the other hand, the profitable real world applications of LLMs still lag behind and there are very few successful examples of the integration to domain business. The potential is there, but the integration and deployment is not trivial.
Also, I don’t see AI replacing most of the workers on almost any field. Yes, it can increase the productivity, but in the current form and shape it is still heavily dependent on the human with domain expertise in the loop.
I have been diving deep in the LLM tech during the past year, and have become painfully aware of the many limitations of the current models. For example, the models resemble the protagonist in the movie Memento, where he does not have the ability to create new long term memories. Also the LLM has to rely on the limited working memory context (like chat session). There are various workaround attempts to provide something similar to human long term memory, but they are a far cry of the human’s learning capabilities.
One of the funniest experiments I made LLM control Combined Arms units in DCS session. It has superb general knowledge of all the unit types and also how they should be used, but the spatial comprehension of the 2D/3D battlefield is very limited. LLMs have only a very basic understanding of the physical world, for example connection of the “left” and “right” with the clockwise degrees seems to be too difficult at times.
I think an interesting part of it is where those humans will be, which gets back to globalization really. This has happened twice to me already, but I’ve been sat in a Waymo where it got stuck trying to push into busy traffic - basically it wasn’t assertive enough, or that the humans could see there was no driver and just cut it off. A voice comes on in the car and says they are intervening and then edge the car forward. The guy was in Manilla, we had a nice chat. He’s probably watching 10+ or so cars. If you used to drive a taxi or Uber in SFO area then it could be impactful. The Uber guy thought he was putting the Taxi guy out of a job. These job changes might be like that, as in the people in the loop handle the exceptions until the model gets better (and you’re training the model on exceptions each time).
Ironically when some start-ups came up a few years back for investment and looked amazing with the results the joke used to be ‘AI = Actually Indians’, in that they had an army of lower paid people working the magic behind the Wizard of Oz’s curtain. Perhaps that will happen a bit more.
Yep, a LLM is ideal for language but not really spatial, hence the push for World Models where it has to interact in terms of the physical world (or a sim of it, like DCS). Plus context sizes on frontier models like Opus 4.5, GPT5.3 are getting better and better in terms of memory (aside from other techniques like RAG and all that becoming more standard). Gemini 1.5 Pro as of today release has a 2 million token window for context, while just 2 years ago something like GPT 4 had a 128k one, so a x16 improvement with a lot more to come.
They’ll be some thinning out of the herd in terms of platforms for sure, just because of the data center investment required and the cost of the decent people required in the field. Google is in a good place due to their combo of hardware and software advances. OpenAI has the name and funding but relies more on the winds blowing around Nvidia. xAI has the twitter crowd and the bots, Anthropic has some smart people, especially in the area of software dev models etc.
Not a defense of LLMs, AI, nor commenting on the our unemployment outlook, but just some feedback on a phone call that I had this morning trying to activate my daughter’s new mobile phone. Called the provider when the phone activation web page wouldn’t advance past the parents contact details. Initially had a reasonably smooth conversation with what I assumed was an AI based auto attendant, which didn’t mind when I changed directions a few time in my responses, then bailed when I felt it had asked me too many questions. It pleasantly handed me off to a human.
The human was someone who seemed to not have English as their native language and also repeated the same questions, my full name, email address, and the new number that I was calling about 3 or 4 times. Like they were reading from a playbook or flowchart, and needed a full restart, whenever a decision box didn’t receive a yes or no answer. In the end, I became more frustrated with the human than the AI, and decided to end the call. Will take another crack at it when her phone is restored from a backup of her iPad.
All the while, I’m thinking, “Come on humans, you’re not helping your cause!”
On the ‘AI stole my jerb’ front, as has already been highlighted it is happening for real already with taxis/uber and call centres. But other than the actual implementation there is also a psychological and ‘management’ component to the job market and economics as well.
On the IT side of the house. I will have to see if I can dig out the article, but enrollments for IT courses, especially programming, are at historic lows and within a couple of years they are predicting extreme shortages of entry level employees unless (ironically) AI can take up the slack… So we might have a self perpetuating myth there? Conversely applications for trade apprenticeships are booming (and it isn’t just a growing realisation of how much money an electrician or plumber makes).
WRT clueless management. My brother who has a Bachelor of Science in computer technology and a Masters of Information Systems is safe due to working in a very niche area. But is now spending most of his time answering questions from management asking “Why can’t AI do that?” whereas not long ago it was “Can you do that?”.
)
