Zuckerberg undercuts an industry.
- Meta Platforms has just dropped a bomb on the generative AI industry by offering more than OpenAI does and doing so for free.
- Combine this with the substantial cash burn going on at many of these startups and one can rapidly see how there is going to be a serious problem.
- Meta Platforms offering models to the open-source community is nothing new as Llama has been available for 15 months but, as always, the devil is in the details.
- Llama is available in different sizes with the flagship at around 400bn parameters and smaller brothers ranging from 7bn – 80bn or so.
- The smaller ones are much easier and cheaper to train but their performance level is meaningfully less than the big models from OpenAI, Copilot, Anthropic and so on.
- For the last 15 months, only the smaller brothers have been made available meaning that open source could never really compete head-to-head with the industry’s leading lights.
- What has changed now is that with Llama 3.1, Meta Platforms is making the biggest and best model available to anyone who wants it for free.
- This does not mean that your average hobbyist with a powerful laptop and graphics card is going to start competing with OpenAI, but anyone who has enough money to rent capacity from Nvidia or others can.
- This is because Nvidia’s AI Foundry which launched earlier this week (see here) is offering access to this model as well as the tools required to customise, package and run it.
- AWS, Databricks and Groq are also supporting the model within their infrastructure and so the barriers to entry for cutting-edge model training and service creation have just been reduced to almost nothing.
- Initially, this model is being offered for synthetic data generation, but I don’t think that there is anything to stop someone from customising the model to do pretty much anything that OpenAI, Anthropic, Gemini, and so on are capable of and at a similar level of quality.
- This is because the initial tests of Llama 3.1 405B (see here) indicate that it is just as good as the best of the rest.
- Furthermore, I think that there is a strong possibility that anyone who uses Llama 3.1 405B may be able to create a generative AI service and run it at a fraction of the cost of the other players.
- This is because GPT-4, Gemini, Claude and so on are by all accounts all larger than 1tn parameters meaning that they are 2-4x the size of Llama 3.1 405B.
- While Llama 3.1 405B, probably needed a lot more training cycles than GPT-4 etc to get it to its current state, it will be much cheaper to execute inference because of its smaller size.
- This implies an upfront fine-tuning cost that could be equivalent to OpenAI etc but a much lower inference cost.
- RFM Research (see here) has concluded that for models that run at scale (millions of users), the cost of inference dwarfs the cost of training and this is where the size of the model could become a big competitive factor.
- When one considers the financial condition of two of the leading start-ups in this space, things get even messier.
- The Information managed to get access to some financial data on OpenAI and concluded that with even $3.5bn in revenues in 2024, its compute spend of $7bn and its staff cost of $1.5bn would leave it burning $5bn this year alone, meaning that another raise will be coming pretty shortly.
- Anthropic is also expected to burn around $2.7bn in 2024 meaning that it too will soon be going back to Amazon cap in hand.
- These business models are all predicated on getting millions of users to pay $20 per month for their services which, thanks to Meta, may soon be available elsewhere for a fraction of this price or even free.
- This means that there is going to be even more price pressure, and I suspect that PwC which just signed a huge deal with OpenAI as a user and reseller of ChatGPT Enterprise will be feeling a little silly and looking to renegotiate the price.
- Should Llama 3.1 405B live up to its promise, then I think that this could be the event that signals that the bursting of the AI hype bubble is fairly close.
- This may take some months, but unlike superintelligent machines or artificial general intelligence, the bursting of the AI bubble is now much closer than it was a week ago.
