The AI Thread

[DeskJockey]

Anthropic going public, probably next month with a valuation of over $1.5 trillion. Just pulling numbers out of AIrses now

FTFY.

[Mito Man]

I like the bit where they ask us to invest in their product which also has a risk of wiping us out.

Are we the turkeys now?

[Gavster]

The bubble is bubbling, with $466 billion in AI related debt been issued in the first 9 months of this year and Wall Street has started buying loads of credit default bonds too.

https://www.ft.com/content/6dd31584-064 ... 25a6b1a6=1

[Beany]

I like the bit where they ask us to invest in their product which also has a risk of wiping us out.

Are we the turkeys now?

Well, they want you to invest in a product that hasn't made any money and has generated staggering, nation state-scale levels of debt for everyone but Nvidia, the shovel seller, on the basis that it might one day be profitable. And that's just this year so far.

Oh, and they've spent the last few years lying about what it's capable of and where it'll go. OpenAI admitted that their frontier model 'going rogue' was actually them having shitty security practises, and that's the charitable interpretation - quite a few people reckon they just fired it off at the internet to generate headlines so they could put the models away - so people wouldn't find out that they're not as clever, nor progressing as quickly, as they're claiming. Anthropic did similar thanks to a shitty sandbox that had internet access. Test system had the same name as the real public system, so it went for the public system instead.

It wasn't super intelligent AIs escaping their constraints, it was shitty security and possibly even false-flagging, but the ultimate PR flack was used to keep the hype train going.

The problem is they reckon they have exponential improvements, but anyone using these tools and can see it's been logarithmic. My sophisticated chart based on vibes, appropriately enough:

AIDevelopment.jpg (49.05 KiB) Viewed 218 times

The first 90% of the code accounts for the first 90% of the development time. The remaining 10% of the code accounts for the other 90% of the development time."

That quote is from Bell Labs back in the day, but it holds true for AI development - remember a few years ago, it seemed like AI was improving in leaps and bounds. Note how it's not done that for the last two years? It's already peaked. It can't *get* hugely better because it's already broadly plateaued in terms of what it's capable of from a logical standpoint.

There's no magic trick here. It's maths, and the way they're doing it, it has a limit. And they can't do it better, or they already would be. What's actually happening is that people are trying to find ways to make the weighted matrixes of vectorised text (or whatever) run faster, but not make them fundamentally more accurate because they're about as accurate as they can get already - and that isn't accurate enough to really steal jobs.

Sure, you can bang a harness in with 20 steps of recursive checking fixing and checking, but if there's an issue the model can't imagine a way around, it'll just fail to fix it faster and with more steps. It still won't be able to fix it.

No-one has a better way of doing this, even in university research labs (never mind AI research labs who pick those up that research to utilise it).

The AI industry is very likely full on fucking fucked, this is just them desperately trying to get out of it with some cash in hand because the VC funding is drying up. If you invest in Anthropic, you deserve the bath you're gonna take. There's a solid chance that Anthorpics IPO is the wake up call that makes the market run screaming from AI and brings the whole house of cards crashing down.

Of course, that relies on the Invisible Hand of the Market (tm) being sensible. Fat chance of that, going by recent performance. They'll privatise the profit, and socialise the debt again, just like 2008 I'd wager, now that these companies are tOo bIg tO fAiL....

[Matty]

https://www.ft.com/content/6dd31584-064 ... 25a6b1a6=1

For the poors
https://archive.is/20261021232951/https ... 747799ccd5

[Mito Man]

Microsoft spending so much on AI that they’re binning, erm, “buying out” a quarter of their workforce so that they can spend even more money on AI

With the easy money getting harder to come across it’s going from a circle jerk to something quite a bit more graphic for my vocabulary.

Free article
https://apnews.com/article/microsoft-b ... 9adf6837e1

[Gavster]

https://www.ft.com/content/6dd31584-064 ... 25a6b1a6=1

For the poors
https://archive.is/20261021232951/https ... 747799ccd5

The archive sites .is and .today have been really sketchy and unreliable recently. I've used them a fair bit when researching videos, but now finding a lot of the links don't work anymore.

[Jobbo]

https://www.ft.com/content/6dd31584-064 ... 25a6b1a6=1

For the poors
https://archive.is/20261021232951/https ... 747799ccd5

The archive sites .is and .today have been really sketchy and unreliable recently. I've used them a fair bit when researching videos, but now finding a lot of the links don't work anymore.

I tend to go to https://www.removepaywall.com/ which has a few options.

[Gavster]

Ohhh amazing, thanks @Jobbo

[Jobbo]

Every such site I've used has ended up not working fairly quickly Gav, so I hope this one continues. As soon as they get popular the paywalled sites find ways to stop them working I think.

Does The Telegraph still work if you simply disable javascript or have they moved on a bit?

[GG.]

Is everyone absolutely sick to the gunwales of AI generated scam ads yet? They seem absolutely rife on youtube now. Sure;y YT are capable of some sort of filtering that can remove things that are obviously fake illegal scams...

[DeskJockey]

Is everyone absolutely sick to the gunwales of AI generated scam ads yet? They seem absolutely rife on youtube now. Sure;y YT are capable of some sort of filtering that can remove things that are obviously fake illegal scams...

Yes, but they have no incentive to do so, as they're making money from the ads.

[Beany]

Don't ask what happened to the Chinese markets and their high-tech stocks over the last few months... after all, who needs to know about high-tech/AI stocks losing nearly half a trillion dollars (~40% of their value) from their summer peak, dragging down mainland and HK markets by $2.5T+ (around 20% across broader markets)?

https://www.theguardian.com/world/2026/ ... ce-beijing

Sure, Chinese AI stocks were overvalued, trading at 150 - 250 x earnings - but look at the US. The top 10 stocks in the S&P 500 are now some 40% of the entire index (even more if Anthropics IPO goes for as much as they're aiming for - $1.5tn to $2tn), with big tech big tech megacorps - selling a a dream (lie) that AGI is around the corner if you just give us enough money to buy more GPUs from Nvidia - making up much of that.

So yes, definitely not something to think about. The emperor absolutely has clothes clothes on - lovely new clothes - and anyone saying otherwise or talking about tulips is a traitor

[GG.]

Is everyone absolutely sick to the gunwales of AI generated scam ads yet? They seem absolutely rife on youtube now. Sure;y YT are capable of some sort of filtering that can remove things that are obviously fake illegal scams...

Yes, but they have no incentive to do so, as they're making money from the ads.

If the scammers are paying for the ads then that revenue is the proceeds of crime. I would have thought that would be a good reason to act.

[Beany]

Nope, as long as they have a degree of """good faith""" moderation of user generated content (including ads), no matter how practically ineffectual, then they're functionally immune. And proceeds of crime would generally require proving intent on Youtubes part - the existence of a moderation system negates that by nature of it being a thing that sometimes catches the most egregious examples - that's all it takes to avoid prosecution without serious evidence of deliberate collusion, and without it being institutionally demonstrable, it'd be individuals prosecuted, not the company.

That money may have changed hands for something functionally criminal is a compliance issue with their processes that are meant to prevent it, not a criminal liability issue for the platform owner, generally speaking.

As a more extreme example, but an absolutely real one, hosting companies don't tend to get prosecuted for people hosting donkey porn, if when informed they tried to take it down and aid law enforcement when presented with court orders, etc etc because if that were the case no-one would host user generated content of any kind, because no-one would want the risk of some dogfucker getting them shut down and put in prison.

It's been a while since I've had responsibility for oversight of user generated content as part of my duties (thank god) but the law as it stood/still stands AFAIK basically works well enough for smaller hosting outfits - IE a couple of places I used to work at - but for the mega-scale social media companies, there's a very strong argument that due to their scale, reach and speed of transmission, they should be held to standards closer to that of the press or traditional broadcasters - they make more money than most of the press, so while it's infeasible of them to pay close attention to the thousands of hours of footage uploaded to tiktok every minute, it's not like they can't afford to have proper, rigorous content moderation of the adverts they show.

The OSA really should have been given far more teeth in that respect but we're so past that now that it's probably too late to enforce anything of that scale.