AI Judge changed title after evaluation, original title: "Miles Brundage, AI policy researcher and Executive Director of AVERI, says hardware scarcity did not shape Anthropic's Claude Mythos Preview rollout approach and attributes outages more likely to operational factors"
Logan Graham of Anthropic confirms compute has not limited rollouts.
I have not been convinced they have a huge compute problem yet. I mean everyone does to an extent in that there is scarcity etc.
But I don’t think the Mythos rollout strategy was caused by this, and there are other explanations of things like outages such as sloppiness
NEW: Chamath Palihapitiya warns Anthropic is at risk of becoming the “Friendster of the AI era” if it fails to solve its compute constraints.
Many users agreed with Miles Brundage's take that Anthropic's sloppiness reflects compute constraints, seeing it as valuable insight, while others dismissed his comments as self-serving.
No Digg Deeper questions have been answered for this story yet.
@Miles_Brundage Tried to be clear:
A lot of people have been wondering about Mythos, Glasswing, and the vulns we / our partners are fixing. Today, I’m excited for us to start sharing more. (For context, I lead Glasswing @AnthropicAI.)
Two independent evaluations this week—from XBOW and the UK AISI—confirm what we've been seeing internally: Claude Mythos Preview is a step change in autonomous cybersecurity capabilities. We need to start preparing fast for a world of models with this level of capabilities.
The UK AI Security Institute tested the model we shipped at the launch of Project Glasswing and found Mythos Preview is the first model to solve both of their end-to-end cyber ranges, including one (Cooling Tower) which no model had ever cleared. But attackers (and defenders) have sophistication & cost constraints – Mythos is also the only model that clears every one of their tasks estimated over 8 hours under their deliberately low 2.5M-token cap.
XBOW tested it on their offensive security benchmarks, finding "token-for-token, unprecedented precision." It's the only model to succeed at subtle V8 sandbox work.
Other Glasswing partners shared similar stories. In a few weeks of testing, Mythos Preview has helped them find many thousands of (estimated) high + critical severity vulnerabilities, sometimes double what they'd normally find in a year.
I don't share this to boost Mythos. In fact, this is not about Mythos. It’s about preparing for the coming world of models being better, faster, cheaper, and more creative than some of the best human experts at dual use capabilities. Clearly, we need them supporting defenders as widely as can be done safely – and especially the least resourced ones.
Within a year, Mythos will probably look quite dumb (relative to other new models). And others may release openly available or unguardrailed models of Mythos-level capabilities.
We started Project Glasswing because capabilities like Mythos Preview's won't stay rare, or stay in careful hands. We are bringing it to defenders as fast as we responsibly can, while working to figure out, for example, the right safeguards and patching & disclosure processes.
Also, to be clear, compute has never been a limiter in our rollout.
Expect a fuller update on our Glasswing work in the coming days.
XBOW report: https://xbow.com/blog/mythos-offensive-security-xbow-evaluation
UK AISI report: https://www.aisi.gov.uk/blog/how-fast-is-autonomous-ai-cyber-capability-advancing
Obviously they’d be cooked without constantly making decibillion dollar compute deals
But they are constantly making decibillion dollar compute deals
I have not been convinced they have a huge compute problem yet. I mean everyone does to an extent in that there is scarcity etc.
But I don’t think the Mythos rollout strategy was caused by this, and there are other explanations of things like outages such as sloppiness
since anthropic has materially pivoted from improving llms by a pre-training to focusing almost exclusively on post training, it's no surprise that their compute constraints would decline as a result.
additionally, you have improvements in architecture that lower the compute required per token process, which of course helps. but I think the key point and main takeaway in terms of valuing and assessing its future is the fact that Anthropic really is just going all in on a variety of low compute algorithmic constraints on the llm instead of actually improving the llm as llm.
which is fine, it's the sign of a mature company. but it's also the sign of a ceiling in terms of making llms better as tensor math-based reproducers of natural language.
https://news.ycombinator.com/item?id=48168056
@jakehalloran1 Maybe, I dunno, just saying I have not yet been convinced. They do take some heat for sure but not sure if it's that important / painful outside of the few very online ppl there and hasn't stopped them from making tons of $
@Miles_Brundage one underevaluated thing is that it’s better to double your stock price and pay a 50% premium for compute than to buy up a lot of cheap compute early when your stock price is low. cost of capital is insane if your stock is undervalued and going up in the next few years.
Obviously they’d be cooked without constantly making decibillion dollar compute deals
But they are constantly making decibillion dollar compute deals
@Miles_Brundage yeah for sure i dont think its going to make a meaningful long term harm or anything - just i think that if it were easily possible the non pay as you go corp team accounts would be a bit higher just from a dont immediately annoy new biz during growth perspective
@logangraham @Miles_Brundage
@Miles_Brundage secondly, a ton of cashflow (or cashflow equivalent) is still coming from investors. in this world, the point of revenue is to prove demand to investors, as much as it is to raise cash. so if you’re sold out due to demand - that’s mostly fine for raising the next round.
@Miles_Brundage one underevaluated thing is that it’s better to double your stock price and pay a 50% premium for compute than to buy up a lot of cheap compute early when your stock price is low. cost of capital is insane if your stock is undervalued and going up in the next few years.
@logangraham @Miles_Brundage Based on Chamath's historical ROI, it seems like he’d benefit more from focusing attention on accurately forecasting his own investments.
@Miles_Brundage as in i doubt they are turning away customers or anything but if their GTM people had a magic wand the limits would be different etc
@logangraham @Miles_Brundage Hello Logan! ☺️👋
@Miles_Brundage the behavior is worse if they don’t
@Miles_Brundage Sloppiness explains a lot more in AI than most people want to admit
kinda wild how often thats the real answer
@Miles_Brundage sloppiness is just compute problems wearing a different hat
@Miles_Brundage next big alpha
@Miles_Brundage agree with this take
@Miles_Brundage i think the level of heat they are willing to take for the usage limits outside of enterprise plans is high enough that its fair to say they likely have a pretty severe compute shortage - and - that it has nothing to do with mythos which is just an earnestly held belief
@Miles_Brundage Everything out of that guy's mouth is self-serving.
@robertwiblin @edzitron what do you think?