Microsoft and OpenAI's close partnership shows signs of fraying

Let that sink in for anyone that has incorporated Chatgpt in their work routines to the point their normal skills start to atrophy. Imagine in 2 years time OpenAI goes bust and MS gets all the IP. Now you can't really do your work without ChatGPT, but it cost has been brought up to how much it really costs to run. Maybe $2k per month per person? And you get about 1h of use per day for the money too...

I've been saying for ages, being a luditite and abstaining from using AI is not the answer (no one is tiling the fields with oxen anymore either). But it is crucial to at the very least retain 50% of capability hosted models like Chatgpt offer locally.

Our Thought Leaders think like that at least. They also pretty much told us to use AI or get fired

It's more important to find a problem and see if this is a fit for the solution, not throw the technology at everything and see if it sticks.

I have had no needs where it's an appropriate solution myself. In some areas it represents a net risk.

This perspective makes zero sense.

What makes sense is to extract as much value as possible as soon as possible and for as long as possible.

Power requirements will drop too.

As well, as people adopt, the output of training costs will be averaged over an ever increasing market of licensing sales.

Looking at the cost today, and sales today in a massively, rapidly expanding market, is not how to assess costs tomorrow.

I will say one thing, those that need gpt to code will be the first to go. Becoming a click-click, just passing on chatgpt output, will relegate those people to minimum wage.

We already have some of this sort, those that cannot write a loop in their primary coding language without stackoverflow, or those that need an IDE to fill in correct function usage.

Those who code in vi, while reading manpages need not worry.

Unlike most Gen AI shops, OpenAI also incurs a heavy cost for traning base models gunning for SoTA, which involves drawing power from a literal nuclear reactor inside data centers.

There is real competition now that plenty of big box stores' websites also list things you won't see in the stores themselves*, but then also Amazon is also making a profit now.

I think the current situation with LLMs is a dollar auction, where everyone is incentivised to pay increasing costs to outbid the others, even though this has gone from "maximise reward" to "minimise losses": https://en.wikipedia.org/wiki/Dollar_auction

* One of my local supermarkets in Germany sells 4-room "garden sheds" that are substantially larger than the apartment I own in the UK: https://www.kaufland.de/product/396861369/

Most of their revenue is the subscription stuff, which makes it highly likely they lose money per token on the api (not surprising as they are are in price war with Google et al)

If you have an enterprise ChatGPT sub you have to consume around 5mln tokens a month to match the cost of using the api on GPT4o. At 100 words per minute that’s 35 days on continuous typing which shows how ridiculous the costs of api vs subscription are.

This is generally true but seems to be, if anything, inverted for AI. These models cost billions to train in compute, and OpenAI thus far has needed to put out a brand new one roughly annually in order to stay relevant. This would be akin to Apple putting out a new iPhone that costed billions to engineer year over year, but was giving the things away for free on the corner and only asking for money for the versions with more storage and what have you.

The vast majority of AI adjacent companies too are just repackaging OpenAI's LLMs, the exceptions being ones like Meta, which certainly has a more solid basis what with being tied to an incredibly profitable product in Facebook, but also... it's Meta and I'm sure as shit not using their AI for anything, because it's Meta.

I did some back of napkin math in a comment a ways back and landed on that in order to break even merely on training costs, not including the rest of the expenditure of the company, they would need to charge all of their current subscribers $150 per month, up from... I think the most expensive right now is about $20? So nearly an 8 fold price increase, with no attrition, to again break even. And I'm guessing all these investors they've had are not interested in a 0 sum.

No.

> Those who code in vi, while reading manpages need not worry

I think that's the wrong dichotomy: LLMs are fine at turning man pages into working code. In huge codebases, LLMs do indeed lose track and make stuff up… but that's also where IDEs giving correct function usage is really useful for humans.

The way I think we're going to change, is that "LGTM" will no longer be sufficient depth of code review: LLMs can attend to more than we can, but they can't attend as well as we can.

And, of course, we will be getting a lot of LLM-generated code, and having to make sure that it really does what we want, without surprise side-effects.

That sounds silly at first read, but there are indeed people who are so stubborn to still use numbered zip files on a usb flash drive in stead of source control systems, or prefer to use their own scheduler over an RTOS.

They will survive, they fill a niche, but I would not say they can do full stack development or be even easy to collaborate with.

But Nowadays GPT has been quantized and cost-optimized to hell that it's no longer as useful as it was and with Claude or Gemini or whatever it's no longer noticeably better than any of them so it doesn't really matter what happens with their pricing.

This is fascinating to think about. Wonder what kind of shielding/environmental controls/all other kinds of changes you'd need for this to actually work. Would rack-sized SMR be contained enough not to impact anything? Would datacenter operators/workers need to follow NRC guidance?

It would not make sense to reduce output quality only to save on compute at inference, why not offer a premium (and perhaps perhaps slower) tier?

Unless the cost is at training time, maybe it would not be cost-effective for them to keep a model like that up to date.

As you can tell I am a bit uninformed on the topic.

Your C-suite execs are paid to skate where that particular puck is going. If they didn't, people would complain about their unhealthy fixation on the next quarter's revenue.

Of course, if the junior-developer role is on the chopping block, then more experienced developers will be next. Finally, the so-called "thought leaders" will find themselves outcompeted by AI. The ability to process very large amounts of data in real time, leveraging it to draw useful conclusions and make profitable predictions based on ridiculously-large historical models, is, again, already past the proof-of-concept stage.

And rarely can you or the model/intern can tell ahead of time which tasks are in each of those categories.

The difference is, interns grow and become useful in months: the current rate of improvements in those tools isn't even close to that of most interns.

Hum... The judge is still out on that one, but the evidence is piling up into the "yes, not using it is what works best" here. Personally, my experience is strongly negative, and I've seen other people get very negative results from it too.

Maybe it will improve so much that at some point people actually get positive value from it. My best guess is that we are not there yet.

99% of the work I do happens in a large codebase, far bigger than anything that you can feed into an AI. Tickets come in that say something like, “Users should be able to select multiple receipts to associate with their reports so long as they have the management role.”

That ticket will involve digging through a whole bunch of files to figure out what needs to be done. The resolution will ultimately involve changes to multiple models, the database schema, a few controllers, a bunch of React components, and even a few changes in a micro service that’s not inside this repo. Then the AI is going to fail over and over again because it’s not familiar with the APIs for our internal libraries and tools, etc.

AI is useful, but I don’t feel like we’re any closer to replacing software developers now than we were a few years ago. All of the same showstoppers remain.

This says 506 tokens/second for Llama 405B on a machine with 8x H200s which you can rent for $4/GPU so probably $40/hour for a server with enough GPUs. And so it can do ~1.8M tokens per hour. OpenAI charges $10/1M output tokens for GPT4o. (input tokens and cached tokens are cheaper, but this is just ballpark estimates.) So if it were 405B it might cost $20/1M output tokens.

Now, OpenAI is a little vague, but they have implied that GPT4o is actually only 60B-80B parameters. So they're probably selling it with a reasonable profit margin assuming it can do $5/1M output tokens at approximately 100B parameters.

And even if they were selling it at cost, I wouldn't be worried because a couple years from now Nvidia will release H300s that are at least 30% more efficient and that will cause a profit margin to materialize without raising prices. So if I have a use case that works with today's models, I will be able to rent the same thing a year or two from now for roughly the same price.

With LLMs, if you have a use case which can run on an H100 or whatever and costs $4/hour, and the LLM has acceptable performance, it's going to be cheaper in a couple years.

Now, all these companies are improving their models but they're doing that in search of magical new applications the $4/hour model I'm using today can't do. If the $4/hour model works today, you don't have to worry about the cost going up. It will work at the same price or cheaper in the future.

The price of a model capable of 4o mini level performance used to be 100x higher.

Yes, literally 100x. The original "davinci model" (and I paid $5 figures for using it throughout 2021-2022) cost $0.06/1k tokens.

So it's not inverting in running costs (which are the thing that will kill a company). Struggling with training costs (which is where you correctly identify OpenAI is spending) will stop growth perhaps, but won't kill you if you have to pull the plug.

I suspect subscription prices are based on market capture and perceived customer value, plus plans for training, not running costs.

One other difference from Bitcoin is that the price of Bitcoin rises to make it all worth it, but we have the opposite expectation with AI where users will eventually need to pay much more than now to use it, but people only use it now because it is free or heavily subsidized. I agree that current models are pretty good and the price of those may go down with time but that should be even more concerning to OpenAI.

> The ability to process very large amounts of data in real time, leveraging it to draw useful conclusions and make profitable predictions based on ridiculously-large historical models, is, again, already past the proof-of-concept stage.

[citation needed]

I wrote "inside" to mean that those mini reactors (300MW+) are meant to be used solely for the DCs.

(noun: https://www.collinsdictionary.com/dictionary/english-thesaur... / https://en.wikipedia.org/wiki/Heterosemy)

Replace it with nearby if that's makes you feel good about anyone's username.

So when maintenance is required, it will be done by adding phrases like "Users should be able to select multiple receipts" to the existing script, and re-running it to regenerate the code from scratch.

Don't confuse the practical limitations of current models with conceptual ones. The latter exist, certainly, but they will either be overcome or worked around. People are just not as good at writing code as machines are, just as they are not as good at playing strategy games. The models will continue to improve, but we will not.

There is also the active and passive efforts to poison the well. As LLMs are used to output more content and displace people, the LLMs will be trained on the limited regurgitation available to the public (passive). Then there’s the people intentionally creating bad content to be ingested. It really is a lose for big service llm companies as the local models become more and more good enough (active).

I will observe that there have been at least three natural-language attempts in the past, none of which succeeded in being "just write it down". COBOL is just as code-y as any other programming language. SQL is similar, although I know a fair amount of non-programmers who can write SQL (but then, back in the day my Mom taught be about autoexec.bat, and she could care less about programming). Anyway, SQL is definitely not just adding phrases and it just works. Finally, Donald Knuth's WEB is a mixture, more like a software blog entry, where you put the pieces of the software inamongst the explanatory writeup. It has caught on even less, unless you count software blogs.

Even if they were to only gouge the current ~11 million paying subscribers, that's around $40/person/month over current fees to break even. Not chump change, but nowhere close to $2k/person/month.

No biz collapse will remove llama from the world, so if you're worried about tools disappearing then just only use tools that can't disappear

The reactor does not need to be in the datacenter. It can be a couple hundreds meters away, bog-standard cables would be perfectly able to move the electrons. The cables being 20m or 200m long does not matter much.

You’re right though, putting them in the same building as a datacenter still makes no sense.

At some point it does make sense to have a small reactor powering a local datacenter or two, however. Licensing would still be not trivial.

It's not clear that OpenAI has any moat to build

There seems to be some renewed interest for smaller, possibly better-designed LLMs. I don’t know if this really lowers training costs, but it makes inference cheaper. I suspect at some point we’ll have clusters of smaller models, possibly activated when needed like in MoE LLMs, rather than ever-increasing humongous models with 3T parameters.

Ours told us not to use LLMs because they are worried about leaking IP and confidential data.

Sorry for being rude.

I think we're done here.

That’s users, not subscribers. Apparently they have around 10 million ChatGPT Plus subscribers plus 1 million business-tier users: https://www.theinformation.com/articles/openai-coo-says-chat...

To break even, that means that ChatGPT Plus would have to cost around $50 per month, if not more because less people will be willing to pay that.

From my view chatbots are still in the "selling dollars for 90 cents" category of product, of course it sells like discounted hotcakes...

That feels like saying that using spell check or autocomplete will make one's spelling abilities atrophy.

OpenAI’s potential issue is that if Google offers tokens at a 10% gross margin, OpenAI won’t be able to offer api tokens at a positive gross margin at all. Their only chance really is building a big subscription business. No way they can compete with a hyperscaler on api cost long run

So 3x the fees, if they're currently at $20/user/month. That's a big jump, and puts the tool in a different spending category as it goes from just another subscription to more like another utility bill in users' minds. The amount of value you're getting out of it is hard to quantify for most people, so I imagine they'd lose customers.

Also there's a clear market trend, and that is that AI services are $20 for the good version, or free. $60 is not a great price to compete in that market at unless you're clearly better.

They make a loss overall because they spend a ton on R&D.

I don’t think it’s that cut and dried though. Many users run into similar issues as other issues with things like reasoning (which is (allegedly) being addressed) and hallucinations (less so) both of which in turn become core reasons for subsequent better versions of the tech. Whether the subsequent versions deliver on those promises is irrelevant (though they often don’t) to that, at least IMHO, being a core reason to “stay on board” with the product. I have to think if they announced tomorrow they couldn’t afford to train the next one that there would be a pretty substantial attrition of paying users, which then makes it even harder to resume training in the future, no?

I can absolutely not go on the internet and download the source code for windows, because microsoft's security is impeccable.

From your comments, it's clear you've already made up your mind that it can't possibly be true and you're just trying to find rationalisations to support your narrative. I don't understand why you feel the need to be rude about it though.

Every single person who claimed AI was a great help to their job in writing software that I've encountered was either inexperienced (regardless of age) or working solely on very simple tasks.

When it comes to stuff like this, the state of the art at any given time is irrelevant. Only the first couple of time derivatives matter. How much room for growth do you have over the next 5-10 years?

As I said, I'm taking bets…

Show me a Markov generator that can explain how it works.