Meta and Scale AI – Catch Up.

The problem is more fundamental than data.

• Meta is investing $14.2bn for both a 49% stake in Scale AI and the services of its founder, which will help the quality of its Llama models improve, but will not give Meta an edge that will allow it to create super-intelligent machines before anybody else.
• This investment also creates a conflict of interest for Scale AI that may see its core customers like OpenAI, xAI, Anthropic, and so on go elsewhere as Google appears to have already done (see here).
• This would leave Scale AI in financial difficulty, making it likely that Meta will acquire the rest of the company, meaning that Scale AI becomes an internal AI resource within Meta.
• Scale AI was founded in 2016 and specialises in the creation of labelled data and reinforcement learning techniques that model makers use to create and train their large language models (LLMs).
• Its core product is a data annotation platform that offers automated annotation of unstructured data (text, images, videos, sensor data, etc) combined with human input to get the best quality training data for the minimum amount of human input.
• Scale AI also provides the ability to generate synthetic data and integrates easily into the process of building and training models.
• In short, it is a tool company for the model creators, and as such, it does not have any particular secret sauce that means it is about to create super-intelligent machines.
• Consequently, this acquisition will enhance Meta’s ability to create high-quality models, but it is not going to help Meta beat OpenAI, Google and so on in a race which I continue to think that no one is going to win.
• Alexander Wang, Co-founder and CEO of Scale AI will join Meta in an, as yet, undisclosed role, but it is pretty safe to say that he will be working on Meta’s efforts to invent super-intelligent machines.
• Although Meta’s Llama 4 models perform reasonably well, they have not performed as well as Meta led us to believe that they would, and this has created an aura of doubt around Meta’s AI prowess.
• This is nothing new, as between 2016 to 2020, Meta was very weak when it came to AI and suffered financially as it kept having to hire humans to do the jobs that its machines were too stupid to execute properly.
• Back then, RFM classified Meta as one of the laggards in AI, but this has largely been fixed, and RFM now considers Meta to be one of the leaders in the generative AI era.
• Of the large digital ecosystem companies, the laggard title is now held by Apple, with Amazon in close 2nd position.
• However, Amazon has a get out of jail card which its large position in Anthropic, which I fully expect it to acquire the minute that Anthropic gets into financial difficulty.
• This only leaves Apple with no clear way forward, and I consider Meta to be in a good position but more of a fast follower as opposed to a leader in its own right.
• It is fairly clear that Scale AI will eventually become part of Meta, and I suspect that all of its business with Meta’s AI competitors will evaporate.
• However, this is not the silver bullet that will allow Meta to slay the AGI monster, and I see this as an acquisition that will help Meta catch up with its peers but not surpass them.
• This is mostly because I continue to believe that a system based on probabilities and statistics is not an accurate representation of how human thinking is done, and as such, will not produce human-level intelligence.
• No amount of well-annotated data is likely to change that simply because reality is way too big and way too complex to model in its entirety, and therefore, models based on this system will always come up short.
• LLMs still have many highly valuable use cases, but retiring the human race is not one of them, and this acquisition is not going to get Meta any closer to that goal.
• However, what it may do is strengthen Meta’s position in the open source community, where it started the trend of open weight models but is now facing stiff competition from both Mistral AI and China.
• If it decides to open-source elements of Scale AI, such as the datasets that it creates for use exclusively with Llama models, then this could return it to a dominant position in the open-source community.
• This, in turn, would give Meta control of a dominant AI platform, which is ultimately what I think it is seeking, having paid the price for being dependent on Apple and Google in the digital ecosystem.
• If not, then this is a very expensive acquisition, as it looks to me like the business of Scale AI may be about to implode.