Lots of New AI Models Released, Lot’s More Hype, and Some Interesting Stuff

We checked out the new AI model news from Anthropic, Google and Meta/Facebook so you don’t have to.

Meta flop

Meta launched their comeback AI model and nobody really cared. The lack of coverage is, in a strange way, quite newsworthy, though to be fair, there was a lot of more important news last week. The new Meta AI model, called Muse Spark and is from Meta’s Superintelligence Labs. We can tell you right now, it ain’t super-intelligent. Ask it to guess a random number and just like the old Llama 4, it will mostly say “27”. It apparently has a proprietary mode called Contemplating Mode, but in the version we tried it appeared to have been disabled.

Anyway, Meta says:

Muse Spark is our most powerful model yet. It currently powers the Meta AI app and website, and will be rolling out to WhatsApp, Instagram, Facebook, Messenger, and AI glasses in the coming weeks.

Anthropic creates lots of news by not releasing their new model

Meta’s AI Hype Department marketing must be so jealous of Anthropic. Anthropic got all the news last week by saying they would not release the new Claude Mythos model because it was too intelligent. Anthropic claimed it was too dangerous to publicly release because it could break software security and, as the Guardian put it “help hackers disrupt much of the world’s most important software”. Oddly, few people, except The Register, noticed Anthropic revealed that their run rate is already at $30 billion. That’s a big number, possibly the fastest revenue growth of any company ever.

Anyway, could Anthropic’s Mythos claim be true? In our opinion, it might be true. This is not AI super-intelligence either, rather just the result of an AI model that has been highly trained on software. On the other hand this could be just more AI hype, remember the GPT-2 that was too dangerous to release or OpenAI’s “Q” model from 2 years ago that was “so powerful it alarmed staff”? People that know more about security software than we do were sceptical about Anthropic’s claims.

Google releases an AI model that certainly deserves attention, Gemma 4.

Gemma 4 is quite powerful, quite intelligent and all round very useful. The interesting part is it runs on really cheap computer hardware. This is possibly a tipping point. If you and your business can run Gemma 4 privately and at zero costs, and it can do most of the tasks, then why are you paying OpenAI or Anthropic? The Register calls AI models like Gemma 4 “serious enterprise platforms”. This may well be where AI really makes the biggest impact, not chasing the illusion of super-human intelligence, but doing lots of day-to-day tasks privately and competently on cheap local hardware that you keep inside your own business.