Tanjreeve

Leaving aside that Meta are selling compute. The reason I pick on Google + Meta is they are able to make their own customers + charge their own prices. Anthropic + OpenAI are both reliant on people coming up with Products powered by their AI models valuable enough to create several AWS worths of revenue and also not also not immediately moving it to open models.

with the caveat being that I came into the industry in 2014 so I may have missed the golden age of the DBA fully - I worked for a DBMS vendor on sales/solutions side for a while and most DBA’s I spoke to were slightly more knowledgeable business analysts. I met very few who really understood IO and Paging and Locks and how they’d impact a DB’s performance. Which kind of makes sense because if you were really obsessively into that you’d probably do something different or you’d work for a business dealing in storage systems. The value a DBA has been very low in my career outside of forcing process improvement just by saying no to things.

Genuinely think google + Meta might be the only companies to make a profitable AI product at this rate.

> At this rate Anthropic will exceed AWS Revenue within a year

Well that’s the point. They’re not continuing at that rate of growth.

That seems like it goes without saying that yes the rules would need to change for results to change. Its just not a hard law or even some sort of natural state of humans.

You might call it a meme but that is their own internal reporting/panicked emails. Subscriptions are offered at a large subsidy with the benefit of more training data and adoption.

> I predict that eventually the training expenditure is going to slow down. At that point, the large commercial providers are likely hoping they can keep inference demand up, and recoup their investment.

The big labs can't do this that's their edge over open models. The stickiness is critical for them to keep up. They might try again but it's going to finish badly.

"Latest frontier models" seems a bit like acting like smarter reasoning can fill in context that doesn't exist in various systems without guessing. It's convenient because it's always the users fault for not working in the platonic ideal observability setup that can feed everything on a plate. And if you're in that situation how much is the agent adding over incremental improvements?

It can't be both "so easy and hands free anyone can use it" and also "well they must be using the wrong model/holding it wrong if it's not working for them". If there weren't tradeoffs of the different models then it wouldn't be exposed to the user.

Was the main focus of your job genuinely using the aws CLI and pulling logs though? I was pretty good at it and jq but the only advantage I ever got from that was more time to deal with the problem on the ticket which has been 50/50 in my experience unless the ground is prepared with context + dev environment etc.

If people paid a profitable price for Opus 5.5 it would not be the best product in the world and would not be slotted into so many things where it's benefiting from ubiquity. That circle needs squaring. There will be an Opus 5.5 that's a 10th of the price in a few months.

Well they just need to use it properly and build a bunch of software systems for them to lever..oh hang on a minute...

The consumer product is losing them money and the possibly profitable enterprise product (API) is a smaller % of revenue and in an increasingly competitive niche. The constant disingenuous baiting and switching between the two whenever convenient is very tiresome.

If they can't raise capital then yes that will have a huge impact on a company that's a massive tower of financial engineering as coreweave have been learning.

As the conversation was about the societal trend. For the vast majority of people the answer is <= 1 month and it drops off precipitously adding months on.

Having something dumb that just cycles through the state of the art for exploits is already absolutely incredible security tooling (if people bother to make it into a process which I'm less optimistic about)

Odd behaviour that's unexplained creates conspiracy theories. That's a self created problem here.

That’s 22 million of capacity foregone for a marketing stunt. Or they had more capacity than demand which seems odd for a product that’s allegedly in constant demand etc.

There just isn’t a sane way to look at this as “investment”

Exorbitant Cloud costs and spending 22 million to solve a 1 million dollar maths problem isn’t investment though.

I think there’s some crossed wires here with the wider data centre build outs. OpenAI have their own flavours of sketchy behaviour but I’m not aware of them solely being on the hook for those amounts short of some weird accounting of Microsoft cloud costs onto them or something.

That’s mostly not the Frontier Labs. They’re definitely creating a lot of the demand but the people taking out those loans are Neoclouds and Big Tech (hence why Facebook + Google et al suddenly all have terrible cash flows after years of having more money than can be spent).

AnthropenAI are the ones driving a lot of that infrastructure spending when they’re such big consumers of compute but they’re not the ones feeling the brunt of the infra spending. Although those costs will bubble up to them indirectly eventually.

It’s like they were all trained on the same internet as the others and they all all have some sort of midpoint opinion that could be a reddit comment or one of the other websites dominated by western desk workers.

I’ve heard plenty of “tech enthusiasts” breezily talk about how close we are to natural language getting compiled down to machine code but has anyone credible got any idea if that’s even a path? Closest think I can think of is DSL type setups like a much richer version of the stuff Haskell people get excited about but that’s not close to an actual natural language "language"

Oils probably a poor metaphor but between the ubiquity of companies deriving significant value from data + analytics on larger than memory datasets and the recent mad rush for training data I don't think that is the wildest tech prediction ever.

I think LLMs are more likely to end up as something boring we don't think about like cloud computing but I don't think comparisons with crypto etc are fair as it's a much more generalisable technology.

Even if that scenario did happen (I agree I don't think most of the major providers would survive on a commodity utility model). Most of the advances in AI are driven by huge amounts of funding for research. If everyone goes to open source models/utility based pricing etc then it'll be cheaper but the current rate of improvements are being overwhelmingly driven by the frontier labs (for better or worse) and they're being driven by vast amounts of VC funding predicated on making vast amounts of money.

I'm very dubious we can have our cake and eat it of constant magnitude order improvements and also commodified utility pricing.

What semantics do you mean here? What are bad semantics in an interpretation/compilation context?

A checksum on the weights should suffice same as any other artifact.

If the original model is poisoned then yes a checksum etc doesn’t matter. If you as the end user are being diligent then you are downloading the checksum and checking the download against it for precisely the scenario you talk about.

Personally I find it a bit exotic as I’d probably just stop trusting the vendor first but checking against a checksum does stop attacks where others alter an artifact.

This is a stream of disingenuity.

> It's not actually clear that subscription users are unprofitable as a group (and the group level is the only important one).

[Open AI losing money on subscription users](https://fortune.com/2025/01/07/sam-altman-openai-chatgpt-pro...)

If anthropic were making money on subscriptions they wouldn’t be pushing people onto API + enterprise packages.

> We do know the companies are not currently profitable, because they're open and intentional about that

Why do they keep engaging in stunts like excluding infra costs then?

You are picking choosing to ignore evidence and actions when it suits you and then saying “oh we can’t possibly know” when we do absolutely know the problems of capital and debt raising on infrastructure problems (which are mostly not the Big Frontier labs either) there’s hard arithmetic for what’s needed to make these projects profitable and right now all the inputs are getting more expensive between chips and capital.

Haskell community will be overjoyed.

They do different things and sit in different parts of the stack (training vs inference). TorchTPU seems more like a play to get more people familiar with Googles TPU chips as they want to sell them to people outside of google.

If the frontier labs are struggling under the training costs and the leading open source models are making their progress by distilling from the frontier labs it sounds a bit like there’s something wrong with the economics here or that progress is going to slow quite a bit without the amount of investment.