ska invita
back on the other side
I doubt it , I think its literally scrabbling around for a techno fix to the economy
AI - Starmers 50 point plan to be released (13/1/25)
- Thread starter Artaxerxes
- Start date
tommers
It was... Rebekah Vardy's account
I think there's a basic naming issue. When people say AI they can mean anything from a chatbot to something identifying symptoms and diagnosing people,or "improving"pictures on your phone or whatever that pothole thing is that they're talking about.
Basically people use it for anything "clever", because there's loads of funding in it cos loads of people like starmer don't know what it means and they're all desperate for something to come along and fix everything and make the line go up again.
And, once again, they've confused "good for the economy" with "good for businesses". If 2 million people lose their jobs then that isn't good for the economy.
Basically people use it for anything "clever", because there's loads of funding in it cos loads of people like starmer don't know what it means and they're all desperate for something to come along and fix everything and make the line go up again.
And, once again, they've confused "good for the economy" with "good for businesses". If 2 million people lose their jobs then that isn't good for the economy.
Artaxerxes
Look out, he's got a gnu!
It really is just a series of post it note project plan ideas
Rob Ray
Weight is meaningless
Yep, and it's been deliberate obfuscation on the part of Big Tech, because saying "large language models which are really good at certain things but systemically unreliable in others and very definitely not even approaching human-like levels of reasoning or flexible problem solving " is not great for marketing to gullible politicians.
I mean fuck can you imagine how these things are going to struggle with implementing the constantly shifting, contradictory, self-defeating and cynical machinations of Westminster policy at ground level? Fudged data, 20-year-old computers running key departmental systems on Windows 95, the fact legislation can't possibly account for every edge case you come across etc etc? The mind boggles.
Last edited:
Crispy
The following psytrance is baṉned: All
Computers are useful, I think we can all agree.
Neural networks are useful. Image recognition, live translation etc.
Large Language Models can be useful when there is a mountain range of good data to train on. I've heard enough programmers say with a straight face that it's revolutionised their work to believe them. Like having a second programmer to bounce ideas off or spit out boring bits of code, a constant buddy. Sounds great, and it's all thanks to the wealth of existing code and discussion of that code.
But the way it works is so horrifyingly inefficient it's like a bad joke. Constantly feeding the same data through the network to get the next token, taking kilowatts to do what a small part of the human brain does with microwatts. This power usage, and paying back the even more tremendous power cost of training, makes them fundamentally uneconomic for the mass market. They've run out of general purpose data to train them on so the capabilities can't improve, the costs aren't going down because the kW/FLOPS is limited by physics, and the general public just aren't willing to actually pay for what they can do.
Meanwhile the tech giants have sunk hundreds of billions into hardware and training, with nothing world-changing to show for it. It's a bubble, it will pop, and it will be ugly for those giants when it does. This is a foolish move by the govt.
Neural networks are useful. Image recognition, live translation etc.
Large Language Models can be useful when there is a mountain range of good data to train on. I've heard enough programmers say with a straight face that it's revolutionised their work to believe them. Like having a second programmer to bounce ideas off or spit out boring bits of code, a constant buddy. Sounds great, and it's all thanks to the wealth of existing code and discussion of that code.
But the way it works is so horrifyingly inefficient it's like a bad joke. Constantly feeding the same data through the network to get the next token, taking kilowatts to do what a small part of the human brain does with microwatts. This power usage, and paying back the even more tremendous power cost of training, makes them fundamentally uneconomic for the mass market. They've run out of general purpose data to train them on so the capabilities can't improve, the costs aren't going down because the kW/FLOPS is limited by physics, and the general public just aren't willing to actually pay for what they can do.
Meanwhile the tech giants have sunk hundreds of billions into hardware and training, with nothing world-changing to show for it. It's a bubble, it will pop, and it will be ugly for those giants when it does. This is a foolish move by the govt.
teqniq
DisMembered
This x100 essentially. When I was at uni back in the early 90's one of our lecturers had a thing for expert systems. To my my what is currently being touted as AI it pretty much a jumped-up version of that with lots more processing power available. Imo we are a long way from genuine AI and by that I mean a true general-purpose AI.
Rob Ray
Weight is meaningless
A really good example of the sheer level of inefficiency is if you watch ai try to recreate Minecraft. And I do mean recreate, over and over and over again. Because every time you move in the game it's generating the next frame from scratch, guessing based on a sample. Every time.
The player's comment about "no object permanence" here is spot on, and absolutely destroys its use value in any number of fields.
The player's comment about "no object permanence" here is spot on, and absolutely destroys its use value in any number of fields.
UnderOpenSky
baseline neural therapy
No, because it has a use. It's rather over hyped at the moment, I suspect in a couple of years we will stop using the term (as it's not actually AI) but it will just be built in to everything we use. But you won't think this is AI. It will be a smarter search, a better way to analyse large amounts of data etc
SpookyFrank
A cheap source of teeth for aquarium gravel
I expect the people who make AI tools face the problem of having their work ripped off by others without credit or payment. But then that's what AI does to everyone else so tough shit for them I guess.
That's why Open AI are charging $200 PCM for their Pro service.
Claw something back before the penny drops.
Rob Ray
Weight is meaningless
There's a quite useful breakdown on Adam Conover's show across two episodes which goes into depth about both the market situation driving Big Tech's AI push and the broader state of things for those with a bit of podcast time on their hands. Zitron is a bit over-bearish about predicting burst bubbles but the base analysis seems pretty sound.
(That'll be it from me on YouTube links)
(That'll be it from me on YouTube links)
krink
I'll do it this afternoon
I'm a lowly data admin/processor for a local council benefits section.
Crispy
The following psytrance is baṉned: All
Nah LLMs are nothing like Expert Systems. That was a pre neural network approach where scientists thought that all of human knowledge could be encoded in a systematic way with logical rules, categories, relationships etc. (CAT is a subset of MAMMAL, CAT chases MOUSE, CAT is DOMESTICATED and so on). Big efforts were made and there are still some huge ES databases out there.
LLMs find their own relationships between words and word fragments inside a very high-dimensional space. Places and directions in this space have "meaning" but it all emerges by random association during training, represented by a sea of numbers. It's completely opaque to an outside observer and critically cannot self-modify. If you want to truly encode new information in an LLM, you have to add that information to the training data in sufficient volume for it to be heard over the noise, and re-train.
I think this is the best short explainer for how they work:
(with tons more detail in his other videos if you're curious)
Some other approach is needed for a truly useful learning AI. Something that doesn't have separate training data and input data, so that it can genuinely learn and adapt.
But LLMs are only any good at information that resembles the information they've been trained on. Get outside the big hump in the middle of the normal distribution and they get much more loosey-goosey, turning into bullshitters. So if your job is just moving boring, non-novel information around and making powerpoints about it, a) a LLM can probably do it and b) it's probably a Bullshit Job (in the Graeber sense).
tommers
It was... Rebekah Vardy's account
Yeah that was brilliant.
UnderOpenSky
baseline neural therapy
What I'm curious to see is how good it is for querying a SIEM which is plugged into thousands of endpoints, servers and networking gear. So maybe not novel. But potentially quite useful. We'll see anyway. It's not installed yet.
UnderOpenSky
baseline neural therapy
Or if the dam thing could look at everyone's calendars and tell me when I can get all the people I need in the same room, that would be quite handy.
Rob Ray
Weight is meaningless
You Are All Available On Friday At 4.45pm. I Have Scheduled Your Meeting And Made It Mandatory.
"I'm ill"
Based On Recent Workplace Studies This Is Not Accurate. Your Role Has Been Terminated.
krink
I'll do it this afternoon
Reading this thread has eased my stress a little so thanks to the u75 boffins.
I also read something about how the huge amounts of power needed for AI means some areas will have to build better infrastructure that can generate and carry that power. Apparently some musk ai computer site is threatening to cause blackouts in memphis. The UK has a more robust power network compared to the creaky USA ones but do we have the capacity for the extra energy?
I also read something about how the huge amounts of power needed for AI means some areas will have to build better infrastructure that can generate and carry that power. Apparently some musk ai computer site is threatening to cause blackouts in memphis. The UK has a more robust power network compared to the creaky USA ones but do we have the capacity for the extra energy?
Artaxerxes
Look out, he's got a gnu!
This is the Outlook Scheduling Assistant, and a cattle prod for those who miss a meeting.
tommers
It was... Rebekah Vardy's account
Well,that's why they're talking about building "mini nuclear reactors".
tommers
It was... Rebekah Vardy's account
I don't trust it. So I scroll through all the ads to check the actual results anyway.
Serge Forward
Just enjoyin' my coffee.
This looks interesting but both together it's two and a half hours long. Anyone fancy fleshing out Rob Ray's summary so I don't have to watch it?
Artaxerxes
Look out, he's got a gnu!
Could ask an AI to do it
Rob Ray
Weight is meaningless
From the show notes:
1.
ChatGPT burns obscene amounts of cash daily with little return, Google's AI dispenses useless and sometimes dangerous advice, and a recent study showed that tech companies will soon run out of new training data to improve their AI models. If AI is really so costly, unreliable, and limited, what happens to the industry that has bet so big on it? This week, Adam talks with journalist and influential tech critic Ed Zitron of wheresyoured.at to discuss the impending burst of the AI bubble, the hubris of Silicon Valley, and how we suffer under big tech's "Rot Economy."
2.
Adam sits with Arvind Narayanan and Sayash Kapoor, computer scientists at Princeton and co-authors of "AI Snake Oil: What Artificial Intelligence Can Do, What It Can’t, and How to Tell the Difference." Together, they break down everything from tech that's labeled as "AI" but really isn’t, to surprising cases where so-called "AI" is actually just low-paid human labor in disguise. Find Arvind and Sayash's book at factuallypod.com/books
Crispy
The following psytrance is baṉned: All
For the Zitron interview, the first half is all about Enshittification in tech (chasing Number Go Up while actively making the actual products worse).
The second half is about the AI bubble.
You could read Zitron instead of listening to him.
1st half: Never Forgive Them
2nd half: The Subprime AI Crisis
He doesn't half write a lot of words though so get a comfy chair!
magneze
🎧
As well as Zitron, this blog is worth a look: Pivot to AI
tommers
It was... Rebekah Vardy's account
"AI agents mean Salesforce won’t hire software engineers in 2025, apparently"
Strong opening
Similar threads
