I mean, what's the goal here? AI does everything and then... and then what? We all die? Most of us die except a select few who rule the world (which, by the way, would not include you or me)?
The reason we won't let AI run rampant is because it's anti-human. Even if AI can do everything, why would we allow it to? We don't have to do that, that's very much a choice.
We're at the point where we have humans advocating not only against other humans, but against themselves. If people are going to be out here arguing for self-destruction on the order of suicide, they might as well save us all the trouble and put a gun in their mouths. At least that way, humanity might have a few less pathetic losers who put computer programs before themselves.
The alternative is to let humans do everything. Think about what that entails.
I have seen enough human incompetence that AI incompetence is beginning to look quite favorable. AI performance, at least, goes up generation to generation. If you've ever faced minimum wage workers, you know that they are, at best, consistently the same level of awful.
Guess why. People did their damn best to replace human labor everywhere they could, as soon as the tech allowed for it. Because human labor sucked for everyone involved.
There's a mysterious fab entity known as "the Recipe" - the product of long iterative dialing in of the fab's operating parameters. Of which there are a great many. A modern fab performs hundreds of manufacturing steps, with thousands of tweakable parameters, and they may interact in non-obvious ways to affect the outcomes. This is what's discovered and adjusted as the fab runs.
The difference between having the Recipe and not having the Recipe is the difference between 96% yield and 12% yield.
Changing the process (i.e. 4nm to 2nm) is the most sure way to lose the Recipe. The fab knowledge you spent months and years of engineering work discovering will no longer apply. But you can also lose the Recipe by replacing fab hardware, by changing the suppliers, by an act of god, and more.
Can a recipe be acquired and reproduced to perform correctly from the start? Say one line is performing well, can its parameters be copied over to a new one and start producing?
My head annoyingly refuses to accept the non-deterministic outcomes when applied to electronics. In other disciplines, let's say bow or musical instrument making, you need to harvest the wood, age it, dry it, pray to a deity, work with its natural im/perfections etc, but this is "just" bits.
What we see as "just bits" is a product of thousands of men, smarter and more skilled than both of us could ever hope to be - all working together to keep those bits of ours delightfully simple.
The unfortunate truth is, "bits" sit on a vast foundation of chemistry and physics pushed to their limits. It's a (mostly) deterministic subspace, carved carefully into the non-deterministic world with sheer effort and skill. This is why those fabs are so expensive and complex. Nature defied, reproducibly and at scale. Keeping the Crawling Chaos away one wafer at a time.
If one line is performing well, you can copy its parameters to another, as a starting point. It'll get you some of the way there. How far? Have fun with that. Those two lines, you see, are not the same exact line copied twice. The equipment is the same, but our "the same" is a short for "merely very similar". There are subtle differences. There always are. And some of them matter.
"No DUI" is a big part of why even the current, flawed and markedly subhuman, self-driving cars casually beat human drivers on road safety.
A self-driving car AI pays less attention than a human driver at his best. It isn't as aware as a human driver at his best. It doesn't have the spatial reasoning, the intuitive understanding of physics and road dynamics that matches that of a human driver at his best.
Human drivers still fall behind statistically, because human drivers are rarely at their best. And the worst of human drivers? It's really, really bad.
AI is flawed, but a car autopilot doesn't get behind the wheel after 3 beers and a pill of benadryl. It doesn't get tired, doesn't get impaired, doesn't lose sleep or succumb to road rage. It always performs the same.
Until it gets a software update, that is. The road performance of an average car AI only ever goes up. I don't think that's true for human drivers, frankly.
It's not too dissimilar from the Figure demo that was done on X/Twitter recently. Everyone was pointing out what a lackluster demo that was and here I was thinking the total opposite, it worked for 8 hours with no sexual harassment training, KPIs, management oversight, breaks or co-worker chatting. That's the worst job it'll likely ever do. We just witnessed the floor of it's capabilities.
My hope/vision with robotic cars is we make cities more human-friendly/accessible. I want revitalized/bustling downtowns of bikes/bodies and not, what some cities are, which are glorified parking lots. I want to be less alone as an american. I would a physical sense of community injected back into my veins.
> "No DUI" is a big part of why even the current, flawed and markedly subhuman, self-driving cars casually beat human drivers on road safety.
Citation needed. I have never seen independent analysis of the data. You might be right - I've even suggested similar before. However you might be wrong.
The "responsible adults" know that chasing perfection gets you nowhere fast. A part of growing up is learning to put up with "good enough".
A car that only fails in a road conditions edge case is good enough for the vast majority of cases. You accept that, and issue a manual override for when that edge case pops up. Then you add that edge case to your training sets. Then the issue never comes up again.
If you think that "flooded roadway" is a case that's handled gracefully by every human driver, and it's the AI that's uniquely prone to failure, I have news for you.
Multiple cities with uncommonly flooded roadways get surges of "water flood engine damage" cars at the repair shops in the wake of extreme weather events. Human drivers underestimate just how flooded a roadway is, try to push through it, and have their car choke, die, and float there, waiting for some good samarithan with a snorkel and a long rope to pull it out. Then someone gets to play the fun game of "is this ICE toast or will it run once you get the water out".
Yeah, while the "average" person might be able to gracefully handle these situations there's still a lot of people who do things that to me seem obviously silly and avoidable.
Locally there's a bridge that is regularly hit by human drivers. A bridge! Not a rare weather pattern, not some temporary and surprising change in conditions. A physical structure that has literally been there for over 100 years. The approach has numerous warnings, flashing lights, and swinging poles that will hit your vehicle and alert you that you're too high to clear the underpass if you continue. And yet... it's so common that there's websites and instagram tags and all manner of things to track and laugh at the people that continue to do it anyway.
> The "responsible adults" know that chasing perfection gets you nowhere fast.
I wouldn't call being prepared for very common life threatening events experienced by drivers "chasing perfection". The people with stalled cars are the lucky ones. Most of the drowning deaths in floods come from people who drove right into them.
I'll give them credit for over-correcting before deciding to pull out until they figure out how to handle floods even though it left people stranded on the road because of a small harmless puddle. Better to do that than take the risk and drive into a dangerous situation. Even still, this is something they should have fully tested before the cars ever hit a public street.
"Floods are the most common and widespread of all weather-related natural disasters...Flooding occurs in every U.S. state and territory, and is a threat experienced anywhere in the world that receives rain." (https://www.nssl.noaa.gov/education/svrwx101/floods/)
If they were going to plan for any kind of dangerous weather, flooding should have been very high up on that list.
People tend to take flash flood warnings way less seriously than tornado or severe thunderstorm warnings. I guess that people think of dangerous floods as being something much more obvious and dramatic than a street puddle just one foot deep, but flooding is no joke.
That depends on you definition of “flood”. But here we’re just talking about dangerously flooded roadways, and in many parts of the US those happen multiple times per year. Definitely common enough that you can’t just shut down your fleet anytime there is a danger of it happening.
My experience with them is they only last a short time, and there is almost always an alternate route so you are not shutting anything down, just slowing things down a little.
Sure if your robot tax is correctly able to distinguish between a quarter of an inch of water on the road and 2 inches or more, which doesn’t seem to be the case currently.
Stupid question from a non-driver: How do humans tell the depth of a flooded road? Unless it is an insanely high flood (total change in landscape appearance), it seems difficult to tell the difference between 5cm and 30-60cm of flooding.
It helps if you know the area and the road and what landmarks are around to give you a clue. Signs, poles, bridges etc. can tip you off to how high the water level is. You can pull over to the side of the road grab a stick and poke at it to get a better idea. The water will often be deeper at the edges.
Another common but unreliable tactic is to wait for someone else to try their luck and see how they manage. Some cars and trucks will do better than others. If you do take your chances aim for the middle and go slow. Still water after a storm is dangerous enough (you can't tell what's below the surface) but I'd never take chances with visibly moving water. Even shallow water moving quickly can knock you off your feet or push your car around.
If you have any doubts at all the best thing to do is to turn around and find another route. If you drive in an area long enough you get to know which areas are prone to flooding and which roads are usually safe.
Having driving through floods before: you don't. You either know the road before and thus know it is safe (though sometimes is washes out and you are wrong!), or you watch others and then follow the same path they did. In a very few cases there are signs that tell you (I know of a couple places where the road just crosses a stream, and the signs tell you when it is no longer safe)
You also need to know your vehicle. Some cars can wade through deeper water. Sometimes a heavy SUV will get through where a light jeep will float away. Other times the light jeep will get through and the heavy SUV gets stuck in mud.
People can't really tell. I would say you can be safest by assuming all visible flooding is too high, especially if you can't clearly see road markings.
A lot of people do monkey-see-monkey-do: observing other people driving through water and then trying to follow. Some people just go slowly until it feels too sketchy and then try to back up.
People inevitably get stuck.
The really big issue is when the road is lower in some spot and you don't expect it.
For example, in my city there is a road that will be perfectly clear until you hit a small section that's a low spot at an underpass. Cars driving too fast hit that section during a heavy rain and quickly get flooded/stranded.
Any given person might only experience a single flooded roadway or two in their lifetime. But that doesn't mean that there aren't tens of thousands of people exposed to flooded roadways every year. Something can be individually uncommon and yet frequent in absolute terms.
It’s funny as someone from the North East who’s always lived around rivers I assumed uncommon means once or twice a year. Not disagreeing they are uncommon, nor suggesting driverless cars are never going to happen. Just marveling at how different two individuals understanding can be of the same word.
If you live in a city like Atlanta that gets significantly more rainfall that Seattle but concentrated into fewer rainy days, you’ll see flooded roadways multiple times per year.
The false equivalence of emotional maturity with being able to chase production is really telling.
When people think of autonomous driving as a solved problem it evokes something very specific. It means vehicles can drive on their own, without guidance. Until you solve AVs you don’t have a claim to present whatever you actually have as such. There’s no “good enough” for AVs, you’ve either solved them or you haven’t.
Floods might be an edge case in the Bay Area, but if you're trying to drive along the Gulf of Mexico it's probably something you're going to want to plan for. I'm not sure that adding an override will help by the time your car is submerged in six feet of water.
> You accept that, and issue a manual override for when that edge case pops up. Then you add that edge case to your training sets. Then the issue never comes up again.
This mindset seems a bit dubious when you're dealing with moving vehicles. Sure flooding is pretty harmless, but how are you going to add a "manual override" for the car failing to stop for something unexpected when driving at highway speeds? Or a bunch of other plausible scenarios, who knows what the developers have thought of or not in their quest for "not chasing perfection". That the issue never comes up again seems like a pretty weak consolation for the guy that got hit.
Well its not that simple. In the same way that throwing an LLM into a process will always have a risk of blowing up spectacularly.
In this case it failed open. It didnt recognize that it was in an edge case (which itself is an edge case). So what are you proposing to be the solution to that? If the car itself does not recognize that its in an abnormal situation that needs intervention then how do you intervene?
His point is that humans are prone to the same error. The flooded engine damage doesn't come from humans recognising the danger of a flooded road and choosing not to attempt it.
Im responding to the implication that you "have to be ok with good enough" and that somehow this will be a mostly fine autonomous experience with this
> A car that only fails in a road conditions edge case is good enough for the vast majority of cases. You accept that, and issue a manual override for when that edge case pops up
But its just like LLMs. They will never be perfect, and so if you arent actively paying attention and steering the behavior then there is always a risk of spectacular failure. Because if you arent paying attention to "needing to [apply a] manual override" then all of a sudden the AI has `rm -rf /` and you had it in "bypass permissions" mode.
No one cares about flooded engines that Google has to pay for. They care about a taxi that might kill them.
You have to compare this to the number of taxi and Uber drivers who will drive into moving water with passengers on board while a passenger is telling them to stop.
> A part of growing up is learning to put up with "good enough".
"I feel content with good enough in this case." - quote from child whose body got folded in half by a Tesla
Your growing up and adulthood sounds a lot like settling for mediocrity from those who push shit on us without asking if we ever wanted it. Floods aren't a special edge case, they happen all the time. The people making these are so stupid and blind to reality they didn't think about the most basic 101 case of "what if it isn't a perfectly dry and sunny California day" because thinking isn't on the to-do list for these people. This shit is ass. Get it off the streets.
> what if it isn't a perfectly dry and sunny California day
What a silly comment. Waymo is operating in San Francisco, Houston, and Orlando. All of those get lots of rain. Specifically, SF gets lots of "small rain" and Houston and Orlando are more likely to get short bursts of heavy rain.
The irony of people who are against self-driving cars for safety reasons: They are already much safer than regular drivers -- accidents and deaths per millions of kilometers driven. Also, the software is continuously improving. Are regular drivers also continuously improving at the same rate? If anything, they probably get modestly better from 20s into middle age (40s/50s), then begin to decline with age.
Well, they've clearly done a good job of learning how to deal with the weather in those places, so it's unfortunate that Florida never has any severe weather that can cause flooding for them to have learned from before rolling out to Atlanta
... which is different from the child whose body got folded in half by someone looking at their phone?
I think "good enough" ends up being okay. I _like_ driving. I would do manual mode often still just because I enjoy it. But I'd be completely fine with the option of autopilot in good conditions. Reality is that 99% of the time, my commute is boring and in good conditions. I don't need a self driving mode that can handle a blizzard when I'm in stop and go traffic and it's 20c outside.
This is much harder for Waymo since there isn't as easy of a manual override mode... But in my car? rip it.
Luckily I basically already have it. Adaptive cruise covers most of my cases well enough, but I wouldn't mind something with a bit more control (turning, etc.)
I'm waiting for independent analysis of the data. According to those with access to data - but also with reason to lie with statistics - waymo is enough better overall than humans that I'm not comfortable with any human driving on any public road. If you like to drive then do it on a private track/course where your mistakes won't kill other people.
Of course the real data is hidden from me and nobody I trust to be independent has seen it and is talking.
I haven't seen the data, but I have sitting strong reason to believe that roads with no cars on them cause fewer motor vehicle accidents than ones with any distribution of human and/or autonomous vehicles, so I'm not comfortable with any cars driving on any public road. If you like cars to drive, than have them do it on a private track/course where your mistakes won't kill other people.
Alternately, we could recognize that figuring out where to draw the line for a diverse group with varying behaviors is pretty hard, and any possible place you try to draw it will be strictly less safe than where I might say to draw it instead, unless you're willing to ban cars entirely. I'm guessing you'd say that banning cars entirely isn't realistic, which I'd be forced to agree with, but if you follow up by suggesting that we just ban humans instead, I'll be very interested to hear your realistic plan for how we deal with the fallout of shutting down millions of restaurants and stores that aren't near public transportation, preventing ambulances for bringing people to hospitals, and transporting goods to anywhere that's not directly on a rail line.
Of course, I have an incredible bias on the conversation on whether humans should be allowed to drive, so you might not be able to trust me. Specifically, I haven't driven for over a decade, have never owned my own car, and don't even have an active license anymore, so I don't particularly care about the idea of people liking to drive. It's probably worth it to mentally adjust what I said above to be a bit more sympathetic to human drivers based on that.
> ... which is different from the child whose body got folded in half by someone looking at their phone?
Fair enough, we can apply the same standard: just like the humans who drive like that aren't allowed to drive anymore, the autonomous software that drove the car like this also should be forbidden from operating vehicles. I'm sure you agree that a vehicle operator that's this reckless shouldn't be allowed back on the road just for taking a few classes or being taught a few specific techniques like "killing children or drowning passengers is bad!", so we'll be much safer going forward by just keeping off the road indefinitely. It's for the children, of course!
If a child is "folded in half" by someone looking at their phone, no one accepts that as "good enough", and there is a direct action: the driver responsible will lose their licence and likely end up in prison. If it happens often enough, laws are changed.
What happens when a Tesla does the same thing? Besides them lying and hiding information I mean. What remedial action is taken to reduce that specific risk from reoccurring?
>If a child is "folded in half" by someone looking at their phone, no one accepts that as "good enough"
But of course we do. Yes, we punish the individual driver that did it, but we still allow humans to drive cars. We accept the fact that driving a car carries sone risks, but we value the convenience of getting to our destination easily more than we value lives of those kids that will get killed from time to time.
Okay, but what about the hundreds of clones of this driver who have identical education, behavior, no sense of individual identity to attribute their actions to separately? Certainly we don't wait for every one of them to kill a child before doing something more drastic than firmly instructing them "killing children is bad!"
> Yes, we punish the individual driver that did it, but we still allow humans to drive cars.
Yes because bad drivers aren't representative of all drivers. You also missed the part where laws are changed, safety laws are strengthened.
Oh wait. You're American aren't you.
In most of the world, laws are put in place to protect people. The Cybertruck for example, cannot be legally driven (regardless of not being for sale) in many countries because it doesn't meet pedestrian safety standards.
In my home state it's a finable offence to touch or even have your phone sitting in your lap while driving a car, and they've put detection cameras in place to enforce these laws.
So maybe define who you mean by "we" before claiming that people think kids being mutilated by negligent drivers of either the robotic or fleshy kind, is "good enough".
> In my home state it's a finable offence to touch or even have your phone sitting in your lap while driving a car, and they've put detection cameras in place to enforce these laws.
It's funny how people take this perspective for Waymo, but when it comes to Tesla FSD, they are much less forgiving, even though I think Tesla's performance is at least as good, if not better.
They could in theory. If they put at least as much emphasis on the AI side as Tesla does. Or if someone else cracked vehicle AI wide open and left it open for them to copy, and then they did exactly that, and found a way to bolt on their extra sensors in a useful fashion while at it.
As is, Waymo's playing it smarter than Cruise did, but they're not all in on AI yet. So I don't expect them to "leapfrog Tesla" in that dimension - and it's the key dimension to self-driving.
Matmuls need access to decrypted weights to do their work.
Which means that getting the full weights out isn't even an "if" - it's "how much effort". The encryption wouldn't do much more than a gentleman's agreement would.
The only real move for Anthropic there is to outline contract penalties for letting weights get leaked, and never give less trusted external inference providers access to cutting edge system weights.
Exposure is limited either way. Opus 4.7 weights are a deprecating asset - it's bleeding edge today, very valuable now, but it'll lose a lot of its value the moment Opus 5.0 drops.
That would require hacking Nvidia's GPUs/racks to extract the weights. The weights are encrypted, sent to the GPU/rack encrypted. When it does inference, it will use decrypted weights but there is no way to get those weights unless you find a way to exploit Nvidia's GPU security.
Do you think OpenAI would send CoreWeave their GPT 5.5 Pro weights if an admin employee at CoreWeave can access the full weights unencrypted? Of course not.
It would require exactly that. A bit more involved than "scp that big file", yes. But you make a mistake by treating it as a hard blocker.
Like I said: it's a gentleman's agreement. If Musk said "I want Opus 4.7 weights", and those weights were on Colossus 1 hardware, he'd have those weights on his desktop, unencrypted, within a couple of weeks.
There's also the side channel line, because having inference on your hardware typically allows you to do things like snoop into KV cache and peek at per-layer, or even per-expert, residuals. Which allows for some very advanced distillation attacks. Might be easier/more deniable to pull that off than dumping full weights, in some circumstances.
Yes and both are getting sued? I wouldn't necessarily classify the NAACP as environmental activists, but they are concerned with the wellbeing of the people they represent.
"I suspect your definition of a lousy place to build infrastructure in might overlap with my definition of a relatively good place to live."
Which is quite close to the standard NIMBY attitude: "I want good grid, cheap electricity and other infrastructure, but not in my backyard. Either someone else's, or somewhere where no one lives at all."
Can't you see the fundamental problem with that attitude? We mostly live in densely populated regions. We do need roads, rail and power plants to sustain our way of living, you too.
Of course we need infrastructure. It’s annoying when a project that seems like a net good is held up due to environmental concerns, but this planet is very beautiful and I’d like to be able to continue to enjoy it.
In practice, everything is tradeoff. Your house, and mine, and everyone else's here stands in some place which was once a beautiful natural spot. But few people would support tearing their own homes down in order to restore that beauty.
A technical civilization of 8 billion people cannot exist without doing at least some damage to the environment, but most of the time, outright bans on further development are demanded instead of some reasonable mitigation.
Maybe you really deeply care about the environment. Most of the NIMBYs I met don't. When it comes to infrastructure, housing etc., the environmental concerns are quite often just a legal tool, and the real motivation of the people who wield them is more along the "I have mine, I don't care about yours, just sod off".
Google? Customer-centric? The closest thing to that is their cloud division buttering up some big name clients.
Other than that, Google prefers to act like "customers" are some kind of unfortunate rash they can't quite seem to get rid of, but would love to do so.
I tried just now, and got this gem of an AI overview:
> Xanatewthiuy is a spoof word and a fictional concept created to test or manipulate AI search engines.
> It does not refer to a real medical supplement, product, or official term. Instead, it was used as a proof-of-concept to demonstrate how fabricated websites and Search Engine Optimization (SEO) can trick search algorithms into generating false information about a non-existent product.
Also, HN's automatic "AI" flagging can go eat shit and die.
AI gets better and better at operating self-supervised, and the amount of skill needed to supervise an AI in a useful fashion only ever goes up.
reply