[bhelkey]

Of the predictions I read, I found that the author engages in pretty heavy handed rules lawyering in order to make their predictions accurate.

For example, the author takes the stance that current self driving cars (Waymo, Zoox) do not count as self driving. The justification being that a human operator is involved some small fraction of the time.

By law, Waymo must report disengagements in California. In 2024, Waymo had ~10 thousand miles driven per disengagement, Zoox had ~28 thousand miles driven per disengagement [1]. I would say that this rate of human intervention qualifies as self driving.

[1] https://thelastdriverlicenseholder.com/2025/02/03/2024-disen...

[strange_quark]

On the contrary, it’s the companies doing the lawyering. A disengagement is when the vehicle reverts fully back to manual control. Tele-operation does not count as a disengagement, and the frequency of tele-operation intervention is a closely guarded industry secret.

[bhelkey]

Oh interesting, I had figured tele-operation would count as a disengagements.

Looking into reports you mentioned in a child comment, CNN reports Cruise needed human assistance every ~5 miles [1]. And I certainly wouldn't call a system that needs assistance every ~5-10 minutes Level 4 self driving.

Subjectively, it appeared Waymo was significantly better than Cruise in 2023 but without data it's hard know what that means in terms of human intervention.

If Waymo needed human assistance every 10-20 minutes, I would agree that it also doesn't qualify as Level 4 autonomous.

[1] https://www.cnbc.com/2023/11/06/cruise-confirms-robotaxis-re...

[brandall10]

You seem to have some deeper insight into this - in your estimation, how often does tele-operation (even a small correction) take place?

[strange_quark]

There’s been reporting on this in several mainstream publications that was accurate as far as the systems I worked with. Unfortunately I don’t want to dox myself on here, so unsatisfyingly the best I can offer is “trust me bro”.

[qingcharles]

The tele-operation is also kinda vague because as I understand it, with Waymo at least, they are not turning a steering wheel and pushing pedals at HQ, they are saying "Pull over here" etc.

[wpietri]

As somebody who has been reading them since the first year, I think you have it wrong. That self-driving prediction was always about Level 5 autonomy. What's changed between now and then is that we've basically stopped talking about that, instead accepting intervention-as-a-service companies as self driving.

[bhelkey]

The author quotes that their predictions are for Level 4 autonomy:

> The definition, or common understanding, of what self driving cars really means has changed since my post on predictions eight years ago. At that time self driving cars meant that the cars would drive themselves to wherever they were told to go with no further human control inputs. It was implicit that it meant level 4 driving. Note that there is also a higher level of autonomy, level 5, that is defined.

[wpietri]

Thanks for the correction!

[ghaff]

Well, ans we're talking about within very specific locales.

Honestly, Brooks--who has been presented and self-presented as something of a skeptic with respect to autonomous self-driving--looks like something of an optimist at this point. (In the sense that your kid won't need to learn to drive.)

[kqr]

Predicting well absent lawyering is really hard! If someone else wants to try I warmly recommend starting with e.g. the ACX 2026 prediction contest: https://www.metaculus.com/tournament/ACX2026/

[bhelkey]

> Predicting well absent lawyering is really hard!

The author engages in rules lawyering of the evaluation of the predictions. The original predictions are clear.

Another example of this is the author's prediction that no robot will be able to navigate around the clutter in a US home, "What is easy for humans is still very, very hard for robots."

The author evaluated this prediction as not being met, "...I don't count as home robots small four legged robots that flail their legs quickly to beat gravity, and are therefore unsafe to be around children, and that can't do anything at all with their form factor besides scramble".

The author added constraints not in the original prediction (safe around children, must include a form factor able to preform an action, ...) then evaluated the prediction as accurate because no home robot met the original constraint + the new constraints.

[LeifCarrotson]

Those robots are not "navigating AROUND the clutter", there are no consumer-deployed object recognition physics models that let a robot say "that's a ball that will roll, that's a shirt that will tangle, that's a sheet of paper that may slip", they're just charging chaotically through it. If you allow skittering robots, do you exclude a 90s RC car with the trigger taped down?

[bhelkey]

I would agree with you if the author wrote, "I don't count as home robots small four legged robots as they don't navigate around clutter, they just go over it" but the author didn't write that.

The author quoted two constraints (safe around children, must include a form factor able to preform an action) not specified.

The author projected that a lab demo of capabilities would not occur. I don't see safety for children as necessary for a lab demo.

[Ntrails]

I read the prediction and I see it meaning "I don't think we'll have tech for useful house robots in the average American Family Home"

The natural interpretation (for me!) was predicated on navigation - implying consideration and appropriate response to the clutter. Not merely ignoring it by being robust to the problems it engenders to movement/balance/etc.