"AI Safety" Regulatory Compliance

[chadbrewbaker]

There needs to be something in the README about regulatory compliance.

Real "AI Safety" is traditional cybersecurity.

Hopefully in the near future there should be a @tinygrad library - simple formally verified parsers around the LLM process. That is true AI safety. Bonus aside you can use Valiant's result from 1970s that CFG == MATMUL, and use TinyGrad for parsing. Boolean MATMUL is bottleneck: https://www.cs.cornell.edu/home/llee/talks/bmmtalk.pdf

How should this messaging be reflected in the README to help users defend themselves from predatory "AI Safety" companies extorting them?

This is what LLM proposed:

The Truth About "AI Safety"

• LLMs Are Autocomplete: See PicoGPT - you can train a LLM in one page of Python given the US Library of Congress on USB stick. They are shocking because of their speed at searching library knowledge - putting advanced concepts in reach of the common man.
• Direct Threat Acknowledgement: All hardware is vulnerable to attacks, whether from humans or automated bots. Real safety is traditional cybersecurity best practices.
• AI Safety as Censorship: Corporations and governments misuse "AI safety" as pretext for technology hoarding, censorship, and maintaining power.
• Extortion in Disguise: Many companies claim to sell "AI safety" while manipulating fears for profit, effectively extortion rackets on lawful users of LLMs.

@oliviazhu - your input would be helpful. I want this project to have FedRAMP certified artifacts using already approved AWSLinux/RHEL packages.

[hwchase17]

not sure what exactly this is asking for. if the goal is to help users defend themselves from predatory "AI Safety" companies extorting them im not sure the readme of this package is the best place to reflect that. we have a section in langchain documentation on security, perhaps more applicable there