Cognitive scientist Gary Marcus argues Trump's demand for a jailbreak-proof Anthropic Fable 5 is technically impossible

Fable ain't coming back

not surprising. to my knowledge there's a single person in the US government with experience working on frontier AI models at a company.

The interesting news is that the government is finally onboard with the exact AI threat that the safety folks talked about c.2022, "unless AI is perfect safe you can't deploy it". Cruel to do it two days after we had Fable though.

@GaryMarcus Gary love your insight but you have to understand that they are doing this to pump up evaluations and get the IPO as high as possible. These peoples don’t have morals it’s all about money

@GaryMarcus Oh well, upon reflection, I think the problem was I didn’t really start following you that closely until sometime 2025. I first started seeing your stuff in 2022/2023, less so in 2024 for some reason, then more dominantly again in 2025 and the most this year.

@AndyMasley @roanoke_gal I guess Dario got his moratorium then. What a stupid mess.

Feels like an engineering issue that’s not super hard to solve.

For eg. you can have a reviewer model that checks every output before it goes to the user and blocks what needs to be blocked. In that case it won’t matter that the initial model gets prompt injected because the reviewer model catches unwanted output.

Some version of this could probably reduce bad output by >99%.

@AndyMasley whelp i'm gonna go brush up on my chinese, have a good one everybody

@ektepenger clearly

@zephyr_z9 Fable ain't coming back

@AndyMasley I mean, isn’t that what Anthropic wanted? So long as this is also applied to similarly capable models, isn’t this the “Pause” that all the lab leaders have said they’d prefer?

@zephyr_z9 My strategy plan.

🔻↩️↩️

@ShakeelHashim I think its sadly what was predicted for in http://europe2031.ai

This is why Mistral is so important, but even with Mistral, Europe needs to really pick it up with how much compute it has.

@zephyr_z9 My Next Plan!

👇👇 👇 👇

@RighttoTryGuy @AndyMasley Yes except every time it's allegedly the end of the world, the model turns out to not be able to count the r's in strawberry, the world continues not ending, and everyone's code gets 10% better with 5% less effort, which appears to have stopped with this latest order.

@GaryMarcus Huh whenever you said LLMs would never be reliable enough I thought you were referring to their ability to complete tasks not in terms of jailbreaking

@GaryMarcus Of course this can be done, is called cLLM. It just requires really high brain power to find the mechanism, but otherwise it's brutally simple once you know it.

@AndyMasley Anthropic has won

@zephyr_z9 My market analysis is as follows:

👇Strategy as follows

👇

👇

👇

@GaryMarcus