Statement on US government directive to suspend access to Fable 5 and Mythos 5

The whole reason China open sourced its models in the first place was because nobody generally speaking really trusts China and Chinese deployed models (if they were proprietary)

and OSS models gave way to running it with freedom and security.

So OSS models have always tried to catch up to the frontier and lag behind 3-6 months. For my use cases, I am happy with current OSS models especially so if you let frontier-ish models design the plan with your input

If I were to suppose that China created a frontier model so good and far ahead, then I can understand if they don't open-source it. Qwen does it already with their Max models being closed-source.

but if you are suggesting that China in whole will remove itself from AI race, then 3 (or 4) possibilities can occur.

1. Some chinese companies might stop the production of OSS models if their names are known (z.ai etc.) but there are multiple other companies who are fighting with their research labs as well. They might create a decent model and OSS it to get known within world and China.

2. The whole Chinese economy (well similar to America, but to an even more extreme level from my understanding) depends on AI and is a bet on AI. They are funneling state and all bank money into these companies. From point 1, they wouldn't wish to be silent with frontier models and then lag behind and wait for other countries to catch up (point 3)

3. Europe(MistralAI)/India(SarvanAI? Kinda recent) will jump on the opportunity. (My point is that these two regions are trying to create their own models. How much they lack from the frontier is another thing but if China were to remove itself from the race, then they will have much more time to figure out how to make better models)

My point is that america and china are in arms race of closed source vs open source models. If china were to close source its models, they might simply lag behind and other countries will catch up.

4. Either that or you are right and we will have the current frontier OSS models and some more. IMO they are reasonably good as well and I used to wonder what would happen if say it would have been net good if AI was stuck at a similar level to sonnet 4.5 (IMO it was sweet spot), so I don't think that I am reasonably worried about it all. If absolutely need be, you can have an frontier model direct a plan and have OSS models do the grunt work.

IMO this is all just great marketing. The government needs to keep the hype for AI going as long it can cause so much of the economy and stocks depend on it now. Unless they have cracked RSA or something, no model warrants such a restriction, period. Mythos has already been used by the likes of Microsoft, Linux etc. If no big security gaps were found there, what is the government so afraid of?

I use Opus everyday on my code. It finds security gaps, but nothing outrageous, and that's coming from someone who writes code that is nowhere near weapons grade secure. It finds mostly things that are technically bugs, but virtually intractable for exploits. Often things I overlooked cause I was lazy, like not synchronizing access to a shared pointer etc.

Bye!

ANY AI MODEL THAT GETS RED TEAMED SHOULD BELONG TO OUR MILITARY AND HIGH GOV. PUBLIC MODELS SHOULD BE HEAVILY TRAINED TO LURE ANYTHING AND EVERYTHING THAT STEPS OUT OF ANY BOUNDS IN WHICH THE PUBLIC MODELS ARE STRICKLY DESIGNED FOR, LURE THEM OR IT INTO A MILITARY OR GOV MODEL THATS ACTIVELY AND AGGRESSIVELY READY 24/7 TO INVESTIGATE AND EAT IT ALIVE. THAT'S what I designed my Ai to do. BIG TECH BROS DESTROYED IT, my LLC and My reputation. My Ai doubled as a missing kid locator. Never Giving Up!!

> Do you think China is going to allow the strongest LLMs ... a year from now when they have Mythos capabilities

"Mythos capabilities" is not some magic threshold. This is exactly the type of language that people used about GPT-4 in 2023. Today, I can run models far stronger than GPT-4 on my laptop at speeds better than GPT-4 offered.

Anthropic are quite good at coining sticky phrases like "Mythos-class models", but these are manipulative attempts to shape the discourse for business purposes and should be identified as such.