Project Fetch: Phase Two(anthropic.com) |
Project Fetch: Phase Two(anthropic.com) |
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.
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.
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 in this case with labs raised with VC-funded cash, the incentives are clear and I would not straight up believe results from the first party source unless multiple sources outside of the company have verified it.
It’s good they are the one seeing those things because otherwise no one else would have. Now if only seeing things would translate into getting any actual economic value out of them… instead of losing billions. But hey, who am I to do a reality check on this shameless piece of hype.
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)
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".