Artificial Intelligence – Dirty Laundry

I don’t expect this fight to last long.

• The blocking of Anthropic’s best model to non-US actors looks more like an argument that should be held behind closed doors than a long-term hold-up, as Anthropic’s own statements and actions are deeply conflicted, and a pro-business White House has no interest in stopping US companies making money from foreigners.
• There is no doubt that the Mythos model is very good at detecting vulnerabilities in software, meaning that keeping it out of the hands of bad actors is certainly a good idea at least until everyone has had time to make their systems safe.
• This time, the disagreement is over Fable 5 (Mythos with guard rails), where a 3^rd party has found a way around the guard rails that Anthropic has said is not serious and appeared unwilling to pull Fable 5 until the flaw is fixed.
• The White House’s response is equally petulant, and this move could easily give OpenAI time to catch up and damage Anthropic’s ability to go public.
• However, Anthropic’s statements and its actions are often contradictory, making it difficult to take anything it says very seriously.
  • First, AI slowdown: where Anthropic has repeatedly called for the industry to slow down the development of super intelligence but then failed to do so itself.
  • In fact, I think it has done the opposite and produced better models than its competitors, which has put it in a position to file for an IPO at a valuation probably north of $1tn.
  • Hence, all of its statements in this area look disingenuous.  
  • Second, Regulation: Anthropic has repeatedly called for AI regulation (in what critics view as an attempt to achieve regulatory capture) but then makes a fuss when it gets regulated in a manner that hinders its business.
  • Third AI Safety: where Anthropic has repeatedly stated that AI safety is a top priority but then fails to live up to this standard.
  • In this situation where a vulnerability has been found that could be serious, the safest thing to do is close access down until the vulnerability is fully assessed and fixed if necessary.
  • By this standard, Anthropic should welcome the export control, which it should have implemented itself and not left the government to do it.  
• The real problem here is that Anthropic and the White House are at ideological loggerheads, which unfortunately has created a lot of animosity leading to arguments being held in the public square rather than behind closed doors.
• I am pretty sure that all of the AI producers are dealing with these sorts of issues, and I continue to think that self-regulation is the best option as long as the industry is mature enough to do it effectively.
• One only has to look at what has happened in Europe, where the over-regulation of AI is killing its hopes of being relevant, which is becoming a cautionary tale.
• The net result is that this argument is likely to be resolved like the ones before it, the ban will be lifted, and the industry will move on.
• I still think that Mythos (Fable-5 minus the cybersecurity guard rails) will be made available at some point and that it will turn out not to be as powerful or as terrifying as we have been led to believe.
• By this time, the critical infrastructure will have been hardened against it, meaning that the risk of it falling into the wrong hands will have largely been addressed.
• The fact that many companies in the developed world rely on software from 3^rd party vendors makes it much easier to deal with this threat as opposed to China, where many companies write all of their software themselves.
• This represents just the next stage of the ongoing process of figuring out how much regulation AI needs to make it safe but not so much regulation that its ability to create wealth is impacted.
• Hence, I think that the Anthropic IPO will go ahead as envisaged and that there will be significant demand for it, as I think that its financials are likely to make surprisingly good reading.
• However, this means that the valuation will be very high, meaning that the value in the AI trade will remain elsewhere and mostly in the picks and shovels.