After having given Openclaw a try as my "personal assistant" for a day when traveling, I 100% want this to be one possible way I can interact with my computer going forward.
Of course it's failed hilariously in many instances, is currently not private (I want local inference before giving it access to anything material, or it'll indeed be a privacy nightmare), and crashes all the time, but the fact that a (not yet) walking, talking CVE can do a better job at this than one of the most wealthy corporations in the world after several years of trying should give them some serious pause.
The “failing hilariously” bit is critical for this road-tripping use case.
It’s only going to take one bad suggestion that leaves someone in a dangerous situation to lose faith in simply handing over a whole day’s itinerary to an LLM. Honestly that can go so bad very easily.
We are decades into the GPS navigation era and I still don’t trust the route my vehicle suggests. I have been burned so many times that we literally still compare routes from different providers for a new trip.
> I have been burned so many times that we literally still compare routes from different providers for a new trip.
I heard this often but what has been the issue in practice? The worst that happened to me is Google Maps suggesting I cross a bridge that was washed away by the last typhoon, but that's hardly Google's fault.
Only in very remote places has Google Maps failed me, at least for driving directions (for trails it's another story...)
> It’s only going to take one bad suggestion that leaves someone in a dangerous situation
I feel like if one bad suggestion can leave somebody in a dangerous situation, many other things must have failed before, such as informing oneself of the general condition of roads in a given place and the current season, having a fallback plan in case digital navigation fails or a road is unexpectedly closed etc.
> failed hilariously in many instances, is currently not private, and crashes all the time, a (not yet) walking, talking CVE
Is actually doing a better job than not doing any of that at all? This isn’t a life or death situation where something is better than nothing out of desperation. Sometimes if you can’t do it right it’s better to not do it at all. Better to wait for the full meal instead of having a “slop snack”.
I can do a terrible job at transplanting brains in robotic bodies. Terrible. Which is more than any company can do so yay?
Some things are worse than nothing in terms of quality or liability.
Yes, it's significantly better than nobody doing any of this for me, and the important thing for the purpose of this prediction is that the error rate still seems to be going down exponentially with time.
> This isn’t a life or death situation where something is better than nothing out of desperation.
That's exactly where it would make sense to try a new thing then, no?
> I can do a terrible job at transplanting brains in robotic bodies.
Sounds like a much more high stakes activity than telling me factoids around my travel itinerary, so I agree that we shouldn't have you run the neurosurgery department yet, yes.
Of course it's failed hilariously in many instances, is currently not private (I want local inference before giving it access to anything material, or it'll indeed be a privacy nightmare), and crashes all the time, but the fact that a (not yet) walking, talking CVE can do a better job at this than one of the most wealthy corporations in the world after several years of trying should give them some serious pause.