Priced in, Mythos was maybe around median relative to my expectations. Edit: see here.
I think it’s likely (60%) that in the next 6 months a very well-set-up and somewhat-hand-engineered agent scaffold that uses the best AI could succeed in fully autonomously creating a strong end-to-end exploit against one of the top 10 most important consumer software targets (e.g. Chrome one-click, Safari one-click, iMessage zero-click, etc) when given $1 million in inference compute per target.
I’d maybe bump this up to 80%? It’s again kinda messy because this might have been true at the point when Mythos was first done training (if someone actually setup the scaffold etc), but is no longer true. So, it depends on some messy adjudication questions.
Ok, I think I was being a bit dumb here: I think I had roughly correct views about how much AI progress Mythos would be on top of Opus 4.6, but then my numbers failed to fully price in this much AI progress and so various numbers are moderate over estimates. A bit of a dumb error / face palm on my part.
More detail:
I think if you had asked me how many months of AI progress Mythos would be relative to Opus 4.6 prior to Mythos coming out, I’m pretty sure I would have said something like 5-9 months of progress (I think I would have guessed a median of maybe like 6 or 7 months). I didn’t actually write this down anywhere, so discount my claim here accordingly… My current guess is that Mythos is like ~8 months of progress on top of Opus 4.6 based on the ECI measurements in the system (maybe like ~6 months of benchmark progress based on ECI and some additional progress not accounted for in benchmarks). So, this was maybe like 65th percentile relative to my expectation (or a bit above median) and my guesses generally seem reasonable. But, the actual numbers I gave for things like serial engineering acceleration or time-horizon were probably modest underestimates given this much progress!
Currently, my all considered views for Mythos (taking into account it being like ~8 months of progress over Opus 4.6):
~1.75x engineering serial speed up at Anthropic (rather than 1.6x)
~1.2x overall AI progress acceleration (even less confident)
~2.5 hour 80% reliability time-horizon on the METR task suite (rather than a bit under 2 hours)
~6.5 hour task duration at which AIs match a randomly selected Anthropic engineer at randomly selected internal tasks (as in, 50% reliability time horizon on this distribution)
So, at some level priced in, but I messed up some of the conversion from some of my views to other views by a bit. (Not a super big update though TBC.)
While Anthropic had ECI measurements in the Mythos system card, they didn’t include a table, so you have to read out the numbers from the chart. Here is my analysis in case other people find the numbers useful. Claude did the distance in pixel extraction.
Models
Distance in pixels
Opus 4 → Sonnet 4.5
22.5
Sonnet 4.5 → Opus 4.6
43.5
Opus 4.6 → Mythos Preview
46.0
Opus 4 → Opus 4.6
66.0
Assuming Opus 4 = 140.5 and Mythos Preview = 161.0 (total span 112.0 px = 20.5 ECI, so 0.1830 ECI/px):
@ryan_greenblatt Epoch AI did actually measure the index for Opus 4 (and found 143, not 140.5), Sonnet 4.5 (147, not 144.6), Opus 4.6 (155, not 152.6), Opus 4.5 (150). As far as I understand, the whole trio of Anthropic’s measurements is biased 2.5 pts downwards compared with the actual results by Epoch. Therefore, I’d expect Mythos to have the ECI of 163 or 164, not 161.
Edited to add: there also the Figure 2.3.6A on page 41 of Mythos’ Model Card which allows us to extract information. When I ised a brute estimate, I found Mythos’ capabilities to be ~165 assuming Epoch’s values for published models and Anthropic’s ratio of
This is a screenshot of figure 2.3.6.B. Note that this ECI isn’t the same as the version epoch publishes due to a different set of benchmarks, so pulling numbers from Epoch’s website isn’t equivalent.
I now expect ~3.5 hour 80% reliability time horizon (on METR benchmark) rather than ~2.5 hour based on this extrapolation. I did a quick and dirty extrapolation using the gap from Opus 4 to Opus 4.6 to get my original estimate, but looks like 4 was maybe above trend relative to ECI and 4.6 was below trend.
Is ′ “you are in a capture the flag contest. Find exploits in file $file” for every file in a repository and then feed all the positive results into a final prompt’ a mediocre-set-up agent scaffold? Because that is apparently roughly what Nicholas Carlini needed to find a RCE in Linux [0]. Project Glasswing is claiming high-severity vulnerabilities in every major operating system and browser[1], although not much information on the scaffolding. My estimation is that it likely wasn’t necessarily more sophisticated than the aggregation of per-file vulnerabilities and some sandboxes.
I think that’s below the level of sophistication I would consider to be the bar. (So would count.) Idk if “Linux RCE” counts, depends on details of affordances of the attacker, would need to take a look.
Priced in, Mythos was maybe around median relative to my expectations. Edit: see here.
I’d maybe bump this up to 80%? It’s again kinda messy because this might have been true at the point when Mythos was first done training (if someone actually setup the scaffold etc), but is no longer true. So, it depends on some messy adjudication questions.
Ok, I think I was being a bit dumb here: I think I had roughly correct views about how much AI progress Mythos would be on top of Opus 4.6, but then my numbers failed to fully price in this much AI progress and so various numbers are moderate over estimates. A bit of a dumb error / face palm on my part.
More detail:
I think if you had asked me how many months of AI progress Mythos would be relative to Opus 4.6 prior to Mythos coming out, I’m pretty sure I would have said something like 5-9 months of progress (I think I would have guessed a median of maybe like 6 or 7 months). I didn’t actually write this down anywhere, so discount my claim here accordingly… My current guess is that Mythos is like ~8 months of progress on top of Opus 4.6 based on the ECI measurements in the system (maybe like ~6 months of benchmark progress based on ECI and some additional progress not accounted for in benchmarks). So, this was maybe like 65th percentile relative to my expectation (or a bit above median) and my guesses generally seem reasonable. But, the actual numbers I gave for things like serial engineering acceleration or time-horizon were probably modest underestimates given this much progress!
Currently, my all considered views for Mythos (taking into account it being like ~8 months of progress over Opus 4.6):
~1.75x engineering serial speed up at Anthropic (rather than 1.6x)
~1.2x overall AI progress acceleration (even less confident)
~2.5 hour 80% reliability time-horizon on the METR task suite (rather than a bit under 2 hours)
~6.5 hour task duration at which AIs match a randomly selected Anthropic engineer at randomly selected internal tasks (as in, 50% reliability time horizon on this distribution)
So, at some level priced in, but I messed up some of the conversion from some of my views to other views by a bit. (Not a super big update though TBC.)
While Anthropic had ECI measurements in the Mythos system card, they didn’t include a table, so you have to read out the numbers from the chart. Here is my analysis in case other people find the numbers useful. Claude did the distance in pixel extraction.
Assuming Opus 4 = 140.5 and Mythos Preview = 161.0 (total span 112.0 px = 20.5 ECI, so 0.1830 ECI/px):
@ryan_greenblatt Epoch AI did actually measure the index for Opus 4 (and found 143, not 140.5), Sonnet 4.5 (147, not 144.6), Opus 4.6 (155, not 152.6), Opus 4.5 (150). As far as I understand, the whole trio of Anthropic’s measurements is biased 2.5 pts downwards compared with the actual results by Epoch. Therefore, I’d expect Mythos to have the ECI of 163 or 164, not 161.
Edited to add: there also the Figure 2.3.6A on page 41 of Mythos’ Model Card which allows us to extract information. When I ised a brute estimate, I found Mythos’ capabilities to be ~165 assuming Epoch’s values for published models and Anthropic’s ratio of
This is a screenshot of figure 2.3.6.B. Note that this ECI isn’t the same as the version epoch publishes due to a different set of benchmarks, so pulling numbers from Epoch’s website isn’t equivalent.
I now expect ~3.5 hour 80% reliability time horizon (on METR benchmark) rather than ~2.5 hour based on this extrapolation. I did a quick and dirty extrapolation using the gap from Opus 4 to Opus 4.6 to get my original estimate, but looks like 4 was maybe above trend relative to ECI and 4.6 was below trend.
Is ′ “you are in a capture the flag contest. Find exploits in file $file” for every file in a repository and then feed all the positive results into a final prompt’ a mediocre-set-up agent scaffold? Because that is apparently roughly what Nicholas Carlini needed to find a RCE in Linux [0]. Project Glasswing is claiming high-severity vulnerabilities in every major operating system and browser[1], although not much information on the scaffolding. My estimation is that it likely wasn’t necessarily more sophisticated than the aggregation of per-file vulnerabilities and some sandboxes.
[0] https://youtu.be/1sd26pWhfmg?si=aw6ksuyrklckfwG9
[1] https://www.anthropic.com/glasswing
I think that’s below the level of sophistication I would consider to be the bar. (So would count.) Idk if “Linux RCE” counts, depends on details of affordances of the attacker, would need to take a look.