AI's power gap widens between elite and everyone else

Just in case you were in any doubt about the power and potential of AI technologies, you might consider taking a bit of time to read about Anthropic's latest Mythos model. "

It's not being released to the general public at the moment (or, perhaps, ever!) based on the overview posted by the company yesterday (link below, "Project Glasswing").

Anthropic are making the model available only to selected partners to help them identify and secure serious holes in their software. This is because Mythos has been able to find completely undiscovered vulnerabilities that, until now, had been totally undetectable by standard processes (i.e. automated testing, human review).

As an insight on costs to run the model, it's an eye-watering $25 per million input tokens and $125 per million output tokens. That's going to add up pretty quickly, but from an enterprise (or governmental) standpoint, I imagine these costs will be readily accepted. I can well imagine vendors highlighting that their software has been 'passed by Mythos' or 'Mythos checked' in due course.

This is one of the first public examples of the big LLM labs having highly capable models that you and I will not (ever / readily) get access to. We've known for quite a while that obviously these labs have models running that are generations better than what we're able to use publicly. The interesting point with this announcement is that we are now entering the "them vs us" stage. Mythos is available, especially if you're the US Government, or the likes of Microsoft.

But if you're a nation-state or an enterprise wondering if Mythos is already freely wandering about your networks and systems, what is your next step?

If you've got a cup of coffee, it's also worthwhile reading the 'system card' (i.e. the 'release notes') of Mythos. If you were unsettled by some of the previous frontier model behaviour, make sure it's a large cup of coffee.

Project Glasswing and the information that Anthropic published about the model is very helpful for the marketplace - as long as the marketplace is listening. I've no doubt many will be pouring over the implications of this announcement for weeks. (I've already had two 'what could this mean?' conversations yesterday evening with clients.) The strong risk is that this news will be missed by many, especially those in senior positions who are responsible for AI governance but perhaps aren't appreciating the speed to which the industry is evolving.

I also read that this week on Bloomberg that Anthropic has hit a $30 billion annualised revenue milestone. The Mythos announcement(s) will be very helpful with valuations in due course.

"Me? I'm looking forward to the next version of Opus-with-bits-of-Mythos included. No pressure, Anthropic - but when is that coming? ;-)