Revolut says that in less than a year it’s managed to slash its [code generation] costs by roughly 90 percent by making self-hosted models its default instead of relying primarily on commercial services like Claude.
Oh that’s why an hacker managed to phish them 680 times to get all details on his victims.
Passports. Documents, transactions, and so on: the hacker pretended to be an Italian police officer and asked to get info on millionaires, an LLM believed that without question 680 separate times.
Self-hosting is illegal in 5 … 4 … 3 …
This is going to be a trend, and why I think the IPO for Anthropic is ridiculous. Competition is massive, there is no moat.
Any sensible company, that has the captial, is hedging their bets. My place has plenty of Claude and Copilot licenses, but we also have healthy local projects going on to.
Personally I can use Claude if I want, but I tend to use the self-hosted platform (Qwen and GLM on DGX Sparks before anyone asks). Big Boss told everyone to be sensible with expenditure, so this part of me doing my bit.
Before anyone comments about “is Big Boss also saving money?”. Yes he is - the guy comes from a French farming family, he is flies economy out of choice and is well-known in our industry for sacrificing not just his bonus, but also his salary when the company has been in rocky waters over the years.
It still makes sense for very small teams where the burden of maintaining the infra is too high, but if you have 2k engineers it’s a no brainer to self host. It’s only a matter of time before open weight models can process vision as well.
I think the doomerism was all to get that moat. There’s no vendor lockin either, even if you’re using the API. You can switch between providers really easily.
You can run litellm as a proxy to not even worry about changes in the API (much).
Not to mention there isn’t public pushback against an office with a server room like there is against dedicated data centers. Assuming they are running servers and not just some desktops with GPUs.
If AI isnʻt going to go away, it needs to be cheap, sustainable, open source, and self hosted/ hostable.
And ethically trained. People should be financially compensated if their data was used to train AI without their consent.
I don’t think it’s really possible to do this.
I mean speaking just logistically, how would you determine who gets how much? If it’s based on quality, how do you determine what is of high quality? Otherwise, does everyone who has every written anything that finds its way into a training dataset get like a tenth of a cent?
Alternative proposal: all AI models must be released as open weights, available to everyone.
AI, through its use of training data, has benefited from the sum total of human knowledge, therefore the results of this training should be given back for the use of all humanity.
That seems fair to me.
You are right if we look at it short term, long term llm will be replaced with the next thing and who knows maybe it can be trained ethically, at least we should require research into it.
I don’t think it’s really possible to do this.
Why not? Both parties have to figure out how much they want to pay or to get payed. Similar like it works for any other license like on Netflix or Spotify. It is not the AI companies job to determine how much they will get payed but how much the data is worth to them and the data creator how much they want.
If i am at a bakery and the bread they offer does not look like its for that much to me then i have not to buy it.
How do you suggest an individual should go about this?
Do I go down to my local courthouse and file a suit against Anthropic so I can ask the judge to subpoena them to determine if my Reddit shitpost from 2012 is in their training set?
Class action lawsuits. How would something like this work for anything else? Like Netflix subscription violations?
And we are talking about meaningful data not your shitpost, like licensed code or a book you wrote and anthropic pirated.
And we are talking about meaningful data not your shitpost, like licensed code or a book you wrote and anthropic pirated.
Ah ha! You see, that’s the problem. My shitpost is meaningful data. This is something people seem to have a really hard time understanding. Scaling laws and the size of the search space in AI training means that the datum is not valuable. It’s only the data (in conglomerate) that is value for training, and my shitpost would be a part of that.
Once you understand this, you should see what I meant when I said it’s not feasible. How do you determine and then fairly distribute the payout to everyone who has ever written or recorded anything in digital format? This is not a simple supply vs demand equation, like your bakery example assumes.
Also I have no idea what Netflix has to do with this discussion.
Ah ha! You see, that’s the problem.
No. Meaningful in the sense of copyright. Data you get copyright
There you got your distinction.
About time this hit enterprise.
Been saying the downfall of OpenAI and Anthropic will be local models that can accomplish the same thing without needing a dstacenter worth of GPUs to run.
Plus you have no total token limit on self hosted models. You can let a slop agent run for weeks without ballooning costs or getting rate limited.
And no model censoring, so you can go build that nuclear bomb you always wanted lol.
Both AI giants are already losing a ton of money because their pricing isn’t viable for the amount of training they’re doing trying to be on top. If they readjust, they’re gonna suffocate against much cheaper solutions.
You totally have a limit. It’s what the cards can output.
Revolut has 54x B200 cards (according to their video on this). So their limit is whatever that can output each month
The data center is still needed, just not the same scale. You don’t run “small” local models to compete with Claude’s size. They don’t publish numbers, but estimates were like 1-2T params for the models like early Opus. You need a lot of hardware to run those efficiently at a scale an enterprise would need.
Or they just spent enough to upgrade developer desktops with RTX Pro’s to run quantized ~100B param models at their desks. Those are very good for coding.
According to them they host 54 B200 cards to do code generation. While claude etc does the high level planning and design
