That ~1000 drivers on the road are all better trained on what to do in the next power outage is incredible.
There will always be unexpected events and mistakes made on the roads. Continual improvement that is locked in algorithmically across the entire fleet is way better than any individual driver's learning / training / behaviorior changes.
Humans seemed to navigate this just fine, even with all the Waymo road blocks and without extra training. If every unknown requires a software update, this system is doomed to repeat this behavior over and over in the long term.
Humans do dumb stuff like drive their cars into flowing floodwaters and they show no signs of stopping. The Waymo Driver (the name for the hardware and software stack) is getting smarter all the time.
Humans do indeed drive into floodwaters like fools, but a critical point that’s often missed when talking about how self-driving cars will make the roads safer: you don’t. Self-driving cars can potentially be safer in general, but not necessarily for you in particular.
Imagine I created a magic bracelet that could reduce bicycling related deaths and injuries from 130,000 a year to 70,000. A great win for humans! The catch is that everyone would need to wear it, even people that do not ride bikes, and those 70,000 deaths and injuries would be randomly distributed among the entire population. Would you wear it?
I don't understand the analogy. No one is being forced to stop driving and take autonomous rides. If I am a better than average driver (debatable), I'm glad to have below average drivers use autonomous vehicles instead.
If you’re on the road with one you’re wearing the bracelet. If you’re driving one you’re wearing two. I don’t mean to sound so sour, I was hoping the analogy would alias that into the background a bit, it’s just that the hoopla around self-driving cars is causing people to skip reading the footnotes.
Safe vs Unsafe isn’t as simple as who gets a 10/10 on the closed course test. Humans are more predictable than random chance would allow, and often even when drunk or distracted. I can’t count how many times I’ve seen someone wobbling on the road and knew to stay back. You can also often tell when someone might yank over into your lane based on them flying up in the other lane, getting just in front of you in that lane then wiggling a bit, clearly waiting for the first chance to pull in front of you and take off. There are lots of other little ‘tells’ that, if you’re a defensive driver, have avoided countless accidents.
Being a prudent defensive driver goes out the window when the perfect speed limit adhering driver next to you goes to straight to -NaN when someone drives past it with Christmas lights on their car, or the sun glares off oversized chrome rims, or an odd shaped vehicle doesn’t match “vehicle” in the database, or, or, or.
* I’m very much not saying that the example I mentioned above is reason enough, I’m saying that I’m not sure enough thought is being put into how many more pages I could go on, and I’m just some schmuck that worked for some number of years on the underlying technology - not the guy watching it fail in imaginative ways on the road.
Something said earlier that really overestimated what’s happening: it doesn’t get smarter, it gets another “if” statement.
That's just what getting smarter is though. I mean, we want to see the human "if" as somehow better than the machine "if" due an obvious bias, but mechanically, what's the difference?
Comparing the two as “ifs” is a really fun way to start getting drunk on this flavor of philosophy, and I highly encourage it, but the short answer is that ‘if’ is deterministic, whereas intelligence isn’t. It’s actually the first step in determining if something has intelligence. If every time you do Action[A,B,C] it does Response[A,B,C] then you can end your inquiry. Things that respond that way include tuning forks, calculators, toasters, Furby, Tickle Me Elmo, Roomba, doorbells, glitter, magnets, literally any pile of garbage, and Teslas.
From my understanding the reason the Waymos didn't handle this was because humans were breaking traffic rules and going when they shouldn't have been. If most humans navigated it correctly, then waynos would have handled this better.
It's mentioned in the article, the real problem was they kept trying to contact remote support to "verify" the light was out. Leading to a backlog of requests which they couldn't get through fast enough.
This attitude is exactly how the Waymos came to handle the problem so poorly in the first place. The principal Skinner "everyone else else is wrong" bit is just icing on the cake.
Can't just program it to be all "when confused copy others" because it will invariably violate the letter of the law and people will screech. So they pick the legally safe but obviously not effective option, have it behave like a teenager on day 1 of drivers ed and basically freeze up. Of course that most certainly does not scale whatsoever, but it covers their asses so it's what they gotta do.
traffic safety engineers often have influence on the letter of the law. We would all be better off if people followed it (humans are bad judges of the exceptions)
But over here in the reality of the world that we have to interact that, this concept of perfect rule-following will never, ever happen -- unless something first manages to wipe the last of the stain of humanity off of the earth's face.
My lived experience with human drivers and outages at intersections is most people get it very wrong. If you're lucky and the lit intersection is 1 lane in each direction, more often than not everything works out well. But any intersection with multiple lanes or especially an intersection that is one primary road and a lower traffic secondary is going to be full of people just flying through as if they were on green the whole time.
That’s an interesting direction. I haven’t thought of this in multiplayer sense.
Would you see this as something that is sort of turn-key, where a central database is hosted and secured to your group?
Or would you require something more DIY like a local network storage device?
And similarly would you be open to having the summaries generated by a frontier model? Or would you again need it to be something that you hosted locally?
A central service. Hosted, secure, frontier model is fine. I’m thinking this through it’s probably something GitHub or an addon should provide.
But maybe it starts local with an app like yours anyway. I do a lot of solo hacking I don’t want to share with the team too. Then there is some sort of way to push up subsets of data.
I can see github providing this, but it would still be at the git-operation level.
What I've found using this contextify-query cli in talking to my project(s) CLI AI history is substantial detail and context that represents the journey of a feature (or lack thereof).
In high velocity agentic coding, git practices seem to almost be cast aside by many. The reason I say that is Claude Code's esc-esc has a file reversion behavior that doesn't presume "responsible" use of git at all!
What I find interesting is that neither Anthropic nor OpenAI have seized on this, it is somewhat meta to the mainline interpreting requests correctly. That said, insights into what you've done and why can save a ton of unnecessary implementation cycles (and wasted tokens ta-boot).
Any thoughts on the above?
If you're open to giving the app a try, and enable updates on the DMG, the query service + CC skill should drop here in a few days. It's pretty dope.
Anyhow, this is really cool feedback and I appreciate the exchange you provided here. Thank you. If you have any further thoughts you want to share I'll keep an eye on this thread or can be reached at rob@contextify.sh
High velocity teams also observe production system telemetry and use error rates, tracing and more to maintain high SLAs for customers.
They set a "budget" and use feature flagging to release risky code and roll back or roll forward based on metrics.
So agentic coding can feed back on observed behaviors in production too.
reply