A group of security researchers, AI entrepreneurs, and corporate executives, including myself, have released a letter calling for a return to the American ideal of transparent, predictable and evidence-based regulation when it comes to AI.
Some users criticized the push for transparent evidence-based AI regulation by arguing that related Administration actions have backfired on US defenses and that predictability remains unrealistic wishful thinking.
You can read the letter and add your name at http://freefable.org
A group of security researchers, AI entrepreneurs, and corporate executives, including myself, have released a letter calling for a return to the American ideal of transparent, predictable and evidence-based regulation when it comes to AI.
The only way to restore that trust is to return to the ideals that helped build American prosperity. A legal regime that is democratically created, transparent, fairly applied with due process, and based upon evidence. Please read the letter at http://freefable.org
Respected ex-GOOG security eng @NielsProvos, reflects the sentiment well. This action is going to drive companies to move to (Chinese) open weight models for their critical systems, lest an arbitrary decision on a Friday afternoon yanks your AI SOC: https://www.provos.org/p/case-for-open-weight-models/
Respected ex-GOOG security eng @NielsProvos, reflects the sentiment well. This action is going to drive companies to move to (Chinese) open weight models for their critical systems, lest an arbitrary decision on a Friday afternoon yanks your AI SOC: https://www.provos.org/p/case-for-open-weight-models/
AI has upended the cybersecurity world and given unprecedented capabilities to attackers. The only way to counter that is for defenders to have access to the most capable models and to maintain forward momentum in US AI research.
A number of signatories, including myself, have either seen the research that kicked off the Fable export designation or spoken to the company that did the work. We believe there were some valid findings but no unique capabilities that justify a reaction close to this.
The benefit of Mythos to an attacker would be the ability to grind away on a large amount of code for hours ($$$) and creating working exploit chains, not validating known CVEs or creating small PoCs, and other models have been shown to have the same capability with the same flaws.
Not all the signatories believe we should have any AI controls, but if we do they have to be evidence-based and fairly applied. Any fair application would catch most of the US foundation models, and also be useless as Kimi 2.7 has the same capability.
If the goal is to protect American networks and systems, then the action by the Administration backfired, by removing a key capability from US defenders and throwing the trustworthiness of the US ecosystem into doubt. They are laughing at us in Beijing.
@alexstamos Transparent and evidence-based are doable. Predictable is wishful thinking!