I'm getting a bit tired of these disguised adverts.
Here's how non robotics engineers used AI to do a short robot integration task faster than other non robotics engineers without AI.
Where "better" mostly means faster, and who knows what happens on longer horizons, with actual robotics experts, robustness requirements, or tasks where the hard part is control rather than API spelunking.
> I'm getting a bit tired of these disguised adverts.
Its not disguised. Corporate blogs exist overtly to promote the company and its work.
Disguised promotions where notionally independent media publish promotional pieces as news concealing that they were fed to them by party whose products they promote area thing, but this is just the most overt undisguised promotion.
> Its not disguised. Corporate blogs exist overtly to promote the company and its work.
It is. That makes the "research" heavily biased. If xAI did the same thing, with Elon Musk screaming about that it is "AGI", you would not believe them at all.
Given that the work is not independent, such articles of this "research" can easily be manipulated or the results being massaged to promote the company positively.
But when others outside of the company try out the work or reproduce it, they get different results. So of course we continue to hear unverified research especially in AI when the frontier labs do not release their architecture, weights at all.
So I would not straight up believe results from the first party source unless multiple sources outside of the company have verified it.
> Preliminary trials with Claude Mythos Preview showed that it would not provide an apples-to-apples comparison with other models because of how we had set up the experiment and how the model was served.
What does this mean? My guess is they couldn’t co-locate Mythos close enough to reduce latency?
(I’m assuming this experiment pre-dates the export controls)
> My guess is they couldn’t co-locate Mythos close enough to reduce latency?
I doubt network latency is the reason. Even when connecting from literally across the world network latency is lost in the noise of overall response latency of even fast models.
The overall response latency of the model very well could have been the difference, though. AFAIK Mythos is structured to do relatively slow "deep thinking".
Depending on the timeline, it could be that they're not allowed to access Mythos because of something like non-US citizens on the team or the lack of some way for them to meet the constraint DOD has them under.
I strongly suspect if that was the case they would have just directly mentioned that Mythos couldn't be used because of that reason, it would be less confusing and less suspect messaging than saying it wasn't an "apples-to-apples comparsion".
This mostly reads as a comparison between Opus 4.7 and 4.1 it would be more interesting if they reran the experiment against a team of humans with 4.7 and see how much the humans still improve the results today.
I'm getting a bit tired of these disguised adverts.
Here's how non robotics engineers used AI to do a short robot integration task faster than other non robotics engineers without AI.
Where "better" mostly means faster, and who knows what happens on longer horizons, with actual robotics experts, robustness requirements, or tasks where the hard part is control rather than API spelunking.
> I'm getting a bit tired of these disguised adverts.
Its not disguised. Corporate blogs exist overtly to promote the company and its work.
Disguised promotions where notionally independent media publish promotional pieces as news concealing that they were fed to them by party whose products they promote area thing, but this is just the most overt undisguised promotion.
> Its not disguised. Corporate blogs exist overtly to promote the company and its work.
It is. That makes the "research" heavily biased. If xAI did the same thing, with Elon Musk screaming about that it is "AGI", you would not believe them at all.
Given that the work is not independent, such articles of this "research" can easily be manipulated or the results being massaged to promote the company positively.
But when others outside of the company try out the work or reproduce it, they get different results. So of course we continue to hear unverified research especially in AI when the frontier labs do not release their architecture, weights at all.
So I would not straight up believe results from the first party source unless multiple sources outside of the company have verified it.
> Preliminary trials with Claude Mythos Preview showed that it would not provide an apples-to-apples comparison with other models because of how we had set up the experiment and how the model was served.
What does this mean? My guess is they couldn’t co-locate Mythos close enough to reduce latency?
(I’m assuming this experiment pre-dates the export controls)
> My guess is they couldn’t co-locate Mythos close enough to reduce latency?
I doubt network latency is the reason. Even when connecting from literally across the world network latency is lost in the noise of overall response latency of even fast models.
The overall response latency of the model very well could have been the difference, though. AFAIK Mythos is structured to do relatively slow "deep thinking".
Depending on the timeline, it could be that they're not allowed to access Mythos because of something like non-US citizens on the team or the lack of some way for them to meet the constraint DOD has them under.
I strongly suspect if that was the case they would have just directly mentioned that Mythos couldn't be used because of that reason, it would be less confusing and less suspect messaging than saying it wasn't an "apples-to-apples comparsion".
This mostly reads as a comparison between Opus 4.7 and 4.1 it would be more interesting if they reran the experiment against a team of humans with 4.7 and see how much the humans still improve the results today.
stop trying to make fetch happen
Do you want Terminators? Because this is how you get Terminators.