![[object Object]](https://media.mas.to/accounts/avatars/109/287/758/133/964/366/original/0f3099c0adfb36a0.png "[object Object]")
keep this post in your back pocket for the next time some AI booster tells you that you don’t understand the technology
this is facebook’s “cutting edge agentic AI” muse, and if you check the context on the linked post you’ll see that they’ve tied up a ton of their company’s valuation in its success
because muse is fully fucked architecturally and is a fractal of unfixable security issues, it leaks much more information about its internal state than chatgpt or claude do. the latter two give the user a falsified version of the model’s internal “thinking” string; muse stores and leaks it as part of a recursive self-improvement system that doesn’t work at all
the linked post is a bunch of leaked internal state showing how a “cutting-edge agentic LLM” bypasses automated Turing tests: pure brute force and failures disguised as success. there’s nothing impressive here. this is how LLMs appear to solve problems, illustrated. https://neuromatch.social/@jonny/117483555597964151
and note carefully that Facebook considers itself a serious AI company. it has so much riding on muse succeeding. it has hired top “AI researchers” away from other AI corporations.
with the stakes ridiculously high and all the expertise in the world at their beck and call, this is the best they could do
this is just how LLMs work, laid bare by a flawed architecture, a deepfried slop implementation (fully vibed with claude code), and leadership that isn’t in on the pump end of the scam and doesn’t realize that lying about the mechanism of action of the LLM is a load-bearing part of the system