Anthropic – Past is prologue

Mythos will be released pretty soon.

• The real danger from Mythos is that it does something really stupid, such as releasing corporate secrets or posting valuable source code online, as opposed to wiping out humanity.
• Anthropic is pumping the hype yet again by stating that its Mythos model is so good that it is too dangerous to make it generally available, and it is only allowing a few pre-vetted companies to have access to it.
• My guess is that the reality is that Mythos is in beta and is being soft-launched to a few trusted partners so that the kinks and bugs can be worked out of it before it goes on general release.
• If one casts one’s mind back to 2019, GPT-2 was considered too dangerous to release, but was then released anyway, and the same was said about the o1 model, which was also made generally available.
• Furthermore, neither of these models was the huge leap forward in performance towards artificial superintelligence that was cited as the reason for making them so dangerous, but this commentary did help stoke hype, speculation and probably the ability to raise money.
• The net result is that right before the next time Anthropic needs to raise money, Mythos will be deemed to be safe and will be made generally available to anyone who wants it.
• I am confident that Mythos is not the leap towards superintelligence that is being alluded to because, at its heart, Mythos is the same as all of the other LLMs in that it is based on statistics and probability.
• This means that it estimates the most likely next token from the one that it has generated and is not doing any reasoning from first principles.
• Hence, I expect that just like all of its predecessors, it will fail the simplest of reasoning tests despite doing mathematics and writing code like a PhD graduate.
• This is because it has no understanding of causality, and I continue to think that until someone finds a way to deal with this problem, there will be no meaningful advances to artificial superintelligence.
• That does not mean that AI is not useful, and over the last 6 months GPT, Gemini and Claude have all become far more competent and accurate at conducting fact-based research.
• However, I never ask for their opinion or to reason out a problem because this is where one ends up with either the average opinion of the internet (mostly useless) or irrational gobbledygook.
• This alone is a super-power in its own right, and I still see AI transforming the way business is conducted and contributing to productivity in many industries.
• Here, Anthropic has an edge in being more focused on the enterprise than on the consumer, as it does not have Google breathing down its neck like OpenAI does.
• Furthermore, the AI rival it does have is Microsoft, which appears to be stumbling, which gives Anthropic a pretty clear path to develop its business.
• Hence, I think this commentary is just more of the usual hype and that Mythos will be released when it is market-ready, as there is no chance of it commanding a robot army to wipe out humanity.
• In fact, it is more likely to do something irretrievably stupid that harms its user through data loss or a hack, and it is this danger that Anthropic is working on fixing.