Comment by madflo
1 day ago
putting my old man cap on and I would like to weigh in on the US admin export control on Mythos.
It does remind me of the mid-1990s when suddenly asymmetric cryptographic tools such as PGP became a reality and a wide usage possible due to the growing base of internet users.
Governments (US, France…) did not understand how to regulate and banned export (and asked users to apply for a licence).
I do see a strong parallel with the situation that we are currently living.
What’s interesting is what’s happened out of the few years where regulations were strong enough to reduce innovation.
Well, open source won for the common and everyday uses, and even more powerful crypto has been developed and used by corporations and governments.
I can certainly imagine LLMs taking a similar path.
How? Nothing stopped cypherpunks in the 90s+ from meaningfully innovating in the space. Table stakes entry for competitive LLMs runs to nation-state budget-levels. There will be no film about 3 genius kids up-ending the trillion dollar oligopolies - they're already working for A\ or OAI.
Capital has eaten software.
The really scary thing now is that these models are totally controlled by private companies. With encryption the smartest people in the room worked in academia and open source tooling quickly caught up with commercial offerings.
We need a real effort to get these technologies free as in beer and to model ourselves on that movement.
Interesting comparison, thanks for sharing! It reminds me of this post about how machine learning and encryption have some fundamental similarities: https://reiner.org/neural-net-ciphers
> I can certainly imagine LLMs taking a similar path. Maybe it's useful to think about what fundamental differences could contribute to LLMs taking a very different path. What comes to mind is the scaling hypothesis, implying that the best LLMs will require enormous capital investment.
That seems largely incompatible with open source barring a fundamental change. There's open weights, but I can't think of a clean historical analogy there and find it extremely difficult to even guess how the future will go
Following the PGP example, I wonder how long until "illegal" tshirts with weights printed on them start popping up.
If they ban GPUs we can always multiply the matrices on paper
You’re going to need a big shirt, right?
Whatever size is most popular at Warhammer conventions, that should be big enough for a 10T parameter model
That would be a good one! Count me in!
You'd need a mountain's weight in t-shirts.
Whether or not you think it's a good thing, controlling LLMs is arguably a lot more effective than cryptography. For crypto as long as you have at least one currently unbreakable algo out there (which fits on a single page of code/math), there's not really much point to regulating it. And to use crypto. the client has to know the algorithm details.
LLMs are already being kept closed weight/source by default. On the client side it's just a generic API client. The underlying technology (weights) wasn't going to be exported even if allowed.
But what's more interesting isn't binary access or not--it's monitoring the chat content, and potentially influencing its replies. (Perhaps the old GPS SA is a better analogy than encryption export.) For example, model providers could be required to allow the government to detect suspected foreign government users and silently degrade performance. They could be required to flag potential exploit discoveries and then send them to CISA for remediation. Or, they could be required to inject disinformation about sensitive topics so that even if you jailbreak, the model is incapable of discussing topics like, say, the presidential motorcade or the design of military bases.
I think what's also very similar between that situation and this one is the technology is not understood at all by the people in government. They've just been told by certain people it's powerful and dangerous