The kind of regulation we need is not new: a licensing regime that requires a minimum safety standard before a system can be built and released. This is how we handle nuclear power, airplanes, buildings, elevators, hairdressers and sandwich makers. Is it too much to ask of trillion-dollar AI corporations, who claim to be building the most dangerous technology in history?
You need this to apply to all models, globally, forever. Which ain’t gonna happen, is it?
not sure including hairdressers in that list makes the case!
When I listened to the CATO Daily Podcast they regularly had stories about States where you needed a licence to practice such professions as hairdressing, coffin building and hair braiding.
Way back, florists. Don’t know if that’s still going on, but seems likely.
I’m pretty sure that what some people think of as AI is already regulated. I asked ChatGPT how was it possible for someone as stupid as David Lammy to qualify as a lawyer only to be told that doing badly on a TV quiz show was no indication of intelligence.So there!
The danger of all these LLMs is that they’re trained on datasets that have already been completely colonised/warped by the left (e.g. Reddit and Wikipedia). Musk seems aware of this, but I’m not sure he can do much about it.
I vaguely remember a minor shytestorm amongts the bien-pensants because apparently Musk or the Trainers had tagged anything from Wokeypedia as “potentially biased” and “low/moderate fidelity source”.
Dunno about Reddit, but no-one in his right mind would treat anything on there as Truth without checking up on some real sources..
It’s not even just about leftism. It’s about amateurs. They get trained on Wikipedia, on Reddit content, maybe on daily news stuff. Because that’s cheap. It spits out a reasonably coherent response.
The whole trick of what’s going on with these LLMs is everyone looks at them and thinks “wow, this is going to destroy so many jobs”, but it’s amateur level knowledge, common belief. Feeds in stuff from twats writing on Wikipedia or in the Times.
And amateur level stuff can be fine. Go on your holiday to Verona, Google Translate is good enough. Maybe it makes mistakes, but the cafe owner and you will figure it out together. It’s a whole different thing when you’re translating instruction manuals, contracts, movie subtitles. You’ll get embarassing or costly problems.
You dont have to be intelligent to be a lawyer. There is no creative thinking involved. I mean even Sadiq Khan passed and the CPS is full of lawyers. No really, it is !
Perhaps I’ll say I’m a lesbian transgender, get on Question Time and ask Lammy to explain the law of tort…
The CPS is full of lawyers, but damn few competent ones, given they can make £££ in private practice. See also: MPs.
Every one of the serious licensing systems mentioned have grown incrementally out of bitter experience. I expect AI to follow this pattern. Unless we can get an AI to map it out for us…
Regulating fast moving inventions is totally impractical. At the cutting edge of development the people who really know their stuff are too busy doing their stuff and getting excited over the new shiny thing.
Any regulator will be several steps behind the curve and cannot hope to truly understand the issue so inevitably becomes just a risk-averse clipboard full of tick-boxes*
*with a lanyard too, for ease of hanging, just to set Tim’s mind at rest.
Yep. There’s no complete mathematical understanding of how models work and they have emergent properties. It will just be bans and safety theatre.
Any regulator will be several steps behind the curve and cannot hope to truly understand the issue so inevitably becomes just a risk-averse clipboard full of tick-boxes
Which is why they have the precautionary principle and no innovation gets done in Europe.
“In early June, the company posted an article describing early signs of recursive self-improvement (RSI), a process in which an AI system devises ways to increase its own intelligence, leading to a greater ability to improve itself, and so on.”
Why do people think Anthropic are bothering with spending time and money writing press releases about their product? Where’s the money in that? The answer is that Anthropic’s primary goal is going public and making the current shareholders rich. It’s about convincing rubes that it’s going to be as powerful as Skynet, the next Microsoft or Google. So, every bit of this activity is a form of marketing. It’s about finding what sounds good, in the same way that car ads are shot on sunny days on empty roads in the highlands.
“Then on 12 June, the White House issued an export control directive banning access to Anthropic’s new frontier models, Fable 5 and Mythos 5, for all foreign nationals – including many of its own key researchers. Anthropic responded by shutting the models down altogether.”
It would be in Anthropic’s interest to pay Donny to do this. If you think this is regular business, you might think it means they can’t sell more product, but at this point, the “too dangerous” thing is great marketing. It will boost the share price when it floats.
“These two June events are closely related. A few months ago, Anthropic’s Claude Code became good enough that its leading researchers no longer write any code at all; they just describe ideas and experiments to Claude and it does all the work.”
Which I’m sure that they 100% factually are. It just isn’t quite as amazing as it sounds. What it means is, they feed in a highly detailed specification that takes as much time as writing code, let the model generate code, and then do more spec tweaks until it gets the code. All so they can say, hand on heart “we don’t write any code”. It’s all about getting that amazing sounding press release, even at the cost of some productivity. Anthropic are not in the business of making money from selling AI. They’re in the business of selling shares in the company at a high price.
If Anthropic could do this, productively, the whole programming industry would be dead right now.
“The recent changes in White House policy suggest we might not need a Chornobyl to spur real regulation, but perhaps only a Three Mile Island. The kind of regulation we need is not new: a licensing regime that requires a minimum safety standard before a system can be built and released. This is how we handle nuclear power, airplanes, buildings, elevators, hairdressers and sandwich makers. Is it too much to ask of trillion-dollar AI corporations, who claim to be building the most dangerous technology in history?”
But code can’t kill you. If you plug code into aircraft, it can, sure, but in a box, spitting out characters? Nah.
This fucking academic who has never written a useful line of code in his life doesn’t understand what we do with software. At the point where physical harm can happen, we get regulation and oversight. The likes of the FAA or the FDA will review what you’ve done. And no-one is going to let some AI anywhere near that stuff because it’s somewhat unreliable. Like all this “full self-driving” requires you to be awake and sober.
its leading researchers no longer write any code at all; they just describe ideas and experiments to Claude and it does all the work
Just as when I got to the point that my assembler was creating good enough machine code, I no longer had to create the machine code myself, I used the assembler to create the updated assembler code.
An entire history of programming languages could be written around this theme. The quest for natural language programming or, at least, a programming language that non-programmers can understand, always ends up with a new programming language at a higher lever that only programmers understand.
This feeds into my frustration at politicians and opinionators demanding “the lumen proletariat must be taught programming!!1!!!1!”
No. That’s like saying “electricity exists, everybody must be taught electrical engineering”. NO. 99.99999999999999999999% of people need to know how to use a light switch, *ALLOW* the rest with the interest and aptitude to *CHOSE* to learn to do the underlying technical engineering stuff.
“These two June events are closely related. A few months ago, Anthropic’s Claude Code became good enough that its leading researchers no longer write any code at all; they just describe ideas and experiments to Claude and it does all the work.”
If it’s still writing code in human interpretable programming languages it’s a cludge. No different to code monkies.
When you use these tools you give it a specification and using best guesses, it writes code. You can write in quite normal language. It’s good. But as I said, it makes best guesses based on that.
Version 1 of my wine scanner was “create a mobile site to take a photo of a wine label and look up the details, then merchants selling it and their prices”. It wasn’t a lot more than that. And it was not bad. But then, oh, it is looking for a $ sign for price. And then, it didn’t handle original and sale price properly. As you get down into the details, you have to very clearly express what you want, and that’s where you might as well code.
I like working this way. It does a lot of the boring basic stuff and I add the stuff that’s more interesting. Saves time and tedium, but you’re not going to do good stuff without knowing how to do the modifications to it.
Try thinking outside the box.
When I first started messing with logic circuits it was soldering individual.
components on boards. Made some things did useful stuff. For a parlour trick. you could build yourself a calculator would do math. Binary input/binary output. For those who couldn’t read raw binary you could even add digital input/output stages. But the logic gates were using binary
There is no difference, in principal, between that & a processor chip. Give or take a couple billion transistors. Programming languages & the rest of it were developed because us ‘umans have trouble handling bulk binary. But I find it very hard to believe somebody’s not going to make a machine can do it. Since machines run on binary. And once you’ve got that you can throw away all the cludge. Why would you want a machine composing human readable programming code? Layers upon layers of cludge.
“Why would you want a machine composing human readable programming code? Layers upon layers of cludge.”
Because there isn’t a way to make it go directly, reliably from human written specification to binary code. “find me the price on the page” means the software makes the best guess of what you mean by that. Now, you could explicitly write it out but at that point, you might as well just code it.
And that’s before we get into all the weird issues, like how the code approached a problem. Does it mean it did a load of inefficient database processing, network requests and so forth? You need to see the code it produced because it could be a disaster.
Sorry, WB, but I don’t see what you’re talking about as being the future of AI. It’s trivial. The potential’s a lot more than a personal assistant.
Does it mean it did a load of inefficient database processing, network requests and so forth?
No, that’s the shit you produce.
Incidentally I’ve just been wrestling with GoogleTranslate’s AI. It doesn’t matter how I phrase something in English it isn’t using the words I want in Spanish. And I do want those specific words or I’m not saying what I want to say. So fcuk it, the recipient’s just have to put up with my poor Spanish grammar. That’s a fair bit of time’s been wasted.
GoogleTranslate is now an utter PoS, thanks to AI.
Ah ha. So GLOBAL GOVERNMENT!!! is the answer.
As always.
No matter if the problem is Imminent Climate Armageddon, or chinese bioweapons on the loose, or “AI is tekkin or jerbs”, or whatever, the solution presented is always global communism.
It’s almost as if the communism is the whole point, just the excuse varies.
It’s almost as if the communism is the whole point, just the excuse varies.
Pretty much, although I think the even more fundamental point is that the people who veer towards that are damaged and ill. They want power, and the must be restrained for all our sakes.
Most of what’s on the internet is rot.
AI trains itself on what it finds on the internet.
We should not be surprised by the results.
[Did you mean Chernobyl]
The commie professor fails to say how expects to constrain Russia and China.
Do fuck off, professor.
I can’t think of anything more likely to retard and stifle the growth of AI than a scheme of ‘sensible’ government regulation. Such schemes are always and inevitably co-opted by a panorama of special interests, all pushing their own ends, many of which have little or nothing to do with the progress of the thing being regulated. Plus, of course, the pressures of the ‘lanyard’ regulator class, which will always favour more regulation = more work for regulators. Rent-seekers gonna rent-seek.
llater,
llamas
Yep. As the world scrambles to implement AI, he’s talking about slowing it down.
And it’s exactly like the climate/zero energy cause, where the real substantive players (China) are not even in the game. UK Net Zero? China destroys your efforts twice a week. Dumbasses.
China isn’t going to be constrained by any such AI regulation or license requirement. But go ahead and constrain yourself.
Hugo Rifkind (a nepo-journo on The Times, for those that don’t know) was wanging on today about the need for government to control the Internet (because children). A bunch of people with PPE degrees controlling the Internet? No, I can’t think of a single thing that could go wrong.
You’re right, not gonna happen. If they could stop China nicking IP, they might have a leg to stand on with controlling AI.
Note how we don’t have regulation to stop dodgy regimes putting nukes together. That’s mostly done through arm-twisting over trade and such, and the occasional bombing run.
Rather different from a country setting rules for the kind of reactors are allowed to be built. Though we suck at those, too.
This is how we handle nuclear power
So? AI is NOT nuclear power.
Not sure what they think can be ‘improved.’
You ask a question, it gives an answer. Does it decide its answer was crap, and then figures out a way to give a ‘better’ answer? And this is a problem? That’s going to TAKE OVER THE WORLD !?!?!?
Recognize a problem, figure out how to fix it/prevent recurrence.
Controlled RSI? Don’t recognize problems? Don’t fix them? You can only fix 3 probles a year, else it would be ‘uncontrolled?’