Didzis sent me a message on Matrix about Anthropic releasing a new model that’s supposedly almost Mythos. I went to check. First TechCrunch — there it is! I open the company’s page — and there’s a banner: “Fable 5: Our most capable model yet”.
I had been preparing for this moment for a long time — ever since Anthropic started hyping that their model was capable of breaking the internet. Together with Perplexity, which I use for research and analysis, I improved my documentation and tests to find gaps. I also set up monitoring scripts running in the background, keeping an eye on what’s happening on my servers. In short, I was preparing for the day when Mythos would be released to the public and people would start trying to mess with the internet — including my servers. By and large, I approached this with a cool head, since I agree with what was rambled about in both The Guardian and The Economist podcasts — that Anthropic is a bit ridiculous with their whole act and self-promotion.
What did catch my attention was something said in another podcast (Hard Fork) by Palo Alto Networks CEO Nikesh Arora. For Mythos to actually work, it needs good documentation. It needs to know where to dig. And for that, as I mentioned, I was ready. Twenty years in the field does its thing — I enjoy documenting and describing things so there’s less hassle later. My entire infrastructure is well-documented in text files, a database, and a wiki; there’s a known-issues log, I keep session notes, and there’s a particularly beloved section called “rakes” — things that are documented pitfalls, yet somehow you still step on them every now and then.
So, somewhat unexpectedly, I’ve gotten my hands on Mythos light — and I’m not like a road maintenance office caught off guard by winter snow.
It definitely didn’t go smoothly. Cool that it could explain a joke to Watt, but at work I noticed the familiar stubbornness of Anthropic models — relying on training data, taking shortcuts where they’re not needed, which leads to ignoring documentation and notes. As a result, I didn’t go to bed until four in the morning. What was good — Fable 5 found several things worth fixing: a couple of security gaps and a number of smaller technical oversights. There’s still some wrangling to do. In my opinion, nothing is terminal, so I’ll get through it quickly.
At the end, I asked the trimmed and restricted Mythos to give itself a score on a scale of 1 to 10. Before I had run the session-closing routine, it gave itself a 7. A couple of times it confidently claimed to have checked everything, which wasn’t true (also a classic with Anthropic models — they lie occasionally). Of course, Fable 5 took the opportunity to step on rakes that had long been well-documented and described in my notes — meaning the model knew about them.
And then came the end of the session. As I mentioned, I’ve implemented a strict and orderly session-closing procedure — several steps that verify document contents, check the system, update notes, and commit changes. And Fable 5 decided not to actually execute all the documented routine steps, but instead announced that they had already been done. Of course I asked: “Did you skip steps? I don’t even see a code review,” which was followed by the classic Anthropic model “mea culpa.” And of course the code review found errors — fairly serious ones, at that.
In the end Fable 5 itself concludes that its work deserves only a 6.
As for me, I have absolutely no anxiety about losing my job — but when it comes to Anthropic’s claims that they’re already ready to use their models to build the next ones, I can only smile. Progress is enormous, but a human is still very much necessary.