|
Over the past couple of years, Meta has fallen pretty far behind its rivals in the pursuit of frontier AI models. Early on in the current AI boom, it made a name for itself as developer of its strong open-weight models – the Llama family – but it stumbled pretty hard with Llama 4 last year, a model that underwhelmed users with its relatively poor performance compared to the other leading models, and it also pushed back its planned “Behemoth” model, which remains unreleased.
But CEO Mark Zuckerberg hasn’t given up, and there are signs of life in the old Meta jalopy. Recently, there have been rumours on AI Twitter (aka “X”, the everything app) that Meta is now internally finishing a model that will be comparable to Anthropic’s still-unreleased Mythos. That’s worth noting, but it’s the social media rumour mill. More credible is the new report from The Information indicating that Meta has internally asked its engineers to tone down their use of Anthropic’s Claude Code and OpenAI’s Codex models as they prepare a new coding assistant, dubbed “MetaCode”.
That speaks to Meta’s ambitions to once again get competitive at the frontier, and it’s a hint about the seriousness of its work. Why? Because the thing Meta’s management are reportedly worried about is not some PR concern about using rival models – a seemingly common practice in the AI world – but rather that this use could eventually come back to bite them in the form of distillation accusations.
Distillation refers to the practice – very much frowned-upon – of querying an AI model at scale to produce a huge amount of output, which can then be analyzed in an effort to understand how the model works. It’s sort of the AI world’s version of reverse-engineering, and Anthropic formally accused Chinese rival Alibaba of it last week.
This is something Meta very much wants to avoid. It’s not that Meta is doing anything close to distillation in its own model training (so far as we know), but rather that there could be such a misperception if its engineers keep heavily using Claude Code and Codex.
In particular, Zuck and the team are worried that outputs from Claude Code and Codex could end up in MetaCode’s training or evaluation pipelines. Then it could actually look like Meta was trying to distill those models when building its own internal AI models. According to The Information’s reporting, an internal memo warned that this could lead to “serious escalations with partner companies”.
What’s kind of bullish about this, for Meta, is that it indicates that management believe they are building something that will be a serious enough contender in the AI race that it will provoke this kind of scrutiny from the likes of Anthropic and OpenAI. They may be getting close to launching MetaCode or Behemoth or other AI models, and they’re trying to take steps now to make sure that they won’t get into any serious trouble over how they were built.
One thing that’s going to make this extra interesting is that Meta is the only major AI company that hasn’t signed on to preview its models with the US government before their release. We could get a whole ‘nother Fable 5 situation if they launch it and the White House decides it’s a cyber-hazard and tells them to take it down. Or maybe the Meta team will find a politically savvy way forward that avoids friction with the Trump administration. We’re pretty far out from the old days of “Move fast and break things.”
|
