Prakash states that concerns about non-frontier AI models are no longer valid after Mythos, as cybersecurity and financial trading now require frontier models to compete effectively
Robin Hanson sees little evidence for frontier models dominating financial trading.
@8teAPi I don't see much evidence that financial trading will be dominated by frontier models.
Prakash@8teAPiThese concerns are no longer valid post Mythos. It is now clear that a number of verticals are going to be winner take all games that require the frontier model to win. Cybersecurity, financial trading both fall into this category, but I imagine we will discover more. Using a cheap 12 month old non frontier model in a world where attackers are using a frontier model will doom you.
@8teAPi There is a quantity quality tradeoff here, and it just isn't clear where quality wins big.
Prakash@8teAPiI’m not arguing that capabilities are here already (after all some are still disputing Mythos capabilities as Anthropic have been unable to provide sufficient evidence until the bugs have been patched, so even in cybersecurity there is no evidence). I am arguing that Mythos reveals that any zero sum game with adversarial participants will reward the frontier model user over the cheaper older model user. That resolves the business case for “why fight to be a frontier lab if you can just wait six months and use a much cheaper equivalent model”.
#1332Prakash@8TEAPII’m not arguing that capabilities are here already (after all some are still disputing Mythos capabilities as Anthropic have been unable to provide sufficient evidence until the bugs have been patched, so even in cybersecurity there is no evidence).
I am arguing that Mythos reveals that any zero sum game with adversarial participants will reward the frontier model user over the cheaper older model user.
That resolves the business case for “why fight to be a frontier lab if you can just wait six months and use a much cheaper equivalent model”.
Robin Hanson@robinhanson@8teAPi I don't see much evidence that financial trading will be dominated by frontier models.
#1332Prakash@8TEAPIVery true. Does running a billion instances of Llama-3 outweigh running a single instance of Mythos at the same task?
I think you’d have to judge on a compute equivalent basis. Given that, if the same amount of compute applied to an older model would have produced the same Mythos equivalent results we would have already seen it obviously because older models had already been applied widely to the same problems while Mythos instances have been limited to a small group of users for security reasons.
Robin Hanson@robinhanson@8teAPi There is a quantity quality tradeoff here, and it just isn't clear where quality wins big.