We have lossless memory for models today. That's the training data. You could consider this the offline version of a replay buffer which is also typically lossless.
The online, continuous and lossy version of this problem is more like how our memory works and still largely unsolved.
You might enjoy this paper[0] which shows that recurrent position encodings recover grid cell representations and maps to path integration found in a popular model of the hippocampus. This isn't terribly surprising since RNNs have shown this before[1, 2] but its an interesting connection.
Cyclists have a bad rep in SF because many (not all) ride quite dangerously. It's a common sight to see cyclists running four-way stop signs and lights without even yielding. I live adjacent to a four-way stop and there's an incident where a cyclist fails to yield nearly hourly.
Meanwhile, Waymo has millions of incident-free miles and of all the self-driving car companies generally takes safety seriously, even if they will act to protect their interests here.
Until more evidence comes out I'll be taking Waymo's side here. I want safer vehicles and Waymo is currently the best bet.
> I live adjacent to a four-way stop and there's an incident where a cyclist fails to yield nearly hourly.
When I see systematic things like this I often get suspicious and think there is something more going on (or selection bias. We are human). I think the common model is attributing actions to reckless behavior and people not thinking. This very well may be true! But if people are consistently engaging in a specific behavior (reckless or not) there's usually a reason to it. Reason doesn't mean good reason btw, and it can be as dumb as the previous chimps getting hosed every time they try to get the banana, but that's still a reason. I think if the underlying reason can be found it significantly increases the chances of rectifying the situation.
> Until more evidence comes out I'll be taking Waymo's side here.
Until more evidence I'm holding out on taking a side. I think Waymo's safety record is orthogonal to the conversation as with the general safety record of driverless cars (which I am a big fan of fwiw).
>I live adjacent to a four-way stop and there's an incident where a cyclist fails to yield nearly hourly.
You should try paying attention to the cars. I live near one (elsewhere) as well and I see cars driving straight through it all the time. Why have a double standard?
I think all of us agree that cars shouldn’t blow through intersections, neither should bicycles. It definitely raises my eyebrows more if cars do it (rare but sadly increasingly less rare) vs if bicycles do get it (fairly common).
When cars do a “rolling stop” what they are really doing is slowing down to 10-15mph, which is the same speed as a cyclist “blowing” through a stop sign at full speed.
I live in Seattle, so you’ll see bikes mostly blow through stop signs going down a decline, they can faster than 10-15 even without e-assist. Cars, when they do ignore the stop sign, can be going up or down, however. If I see a bike going downhill approaching a stop sign, I assume they are going to try and ignore it if it won’t obviously get them killed. Uphill, bikes care more about stopping since they aren’t losing momentum.
Even with a cyclist disregarding the law, it shouldn't be possible for a law abiding autonomous vehicle to collide with anything if it had confirmed the intersection was clear of traffic. Seems like it assumes an unoccupied slot behind trucks that won't always be true. What if this was a T-intersection where a cyclist wasn't obligated to stop?
Ah, but people yelled because waymo was too timid. They probably did that because it's how people are expected to drive in California. Bit of a sticky wicket for waymo isn't it?
I tried this line of reasoning with a judge once. If you break the law just because people expect you to, it's still your traffic violation.
> Cyclists have a bad rep in SF because many (not all) ride quite dangerously. It's a common sight to see cyclists running four-way stop signs and lights without even yielding.
Pick any stop sign in SF. Watch it for an hour and count the number of cars that come to a complete stop. The denominator doesn't matter because the numerator will be ~0.
> count the number of cars that come to a complete stop
You’re conflating two different phenomena, perhaps purposefully.
California drivers do often fail to come to a technical stop at intersections, but they first slow to a near-zero speed that very nearly accomplishes the same purpose.
In contrast, a large number of cyclists do not slow at all, and blast through red lights at full speed.
I've lived in San Francisco for over 4 years. I cycle, walk, drive, and take public transport. If I were about to cross at an intersection with a stop sign, the sight of an approaching car would create more fear and caution in me, than would the sight of an approaching bicycle.
> California drivers do often fail to come to a technical stop at intersections, but they first slow to a near-zero speed that very nearly accomplishes the same purpose.
It very nearly accomplishes the same purpose only from the point of view of other drivers. When a driver in SF approaches an intersection, they're looking out for other cars, working out whether they'll be able to slow down and continue, or they'll need to stop. If they don't see another car that will have priority, they won't stop. So:
- they may not see the pedestrian that's just reached the intersection, and/or
- the pedestrian that's just reached the intersection won't know whether the car is going to stop, so they won't attempt to cross
When I drive, I'm sometimes frustrated when I stop at an intersection, wait for a pedestrian to start crossing, and they're slow to get started. Then I remember it's the behavior of drivers that has conditioned pedestrians to yield even when they have right of way.
> Then I remember it's the behavior of drivers that has conditioned pedestrians to yield even when they have right of way.
I similarly have mixed mode of transportation. I'll admit that when walking I will frequently fully stop in the middle of crossing an intersection. Many people do come to abrupt stops at stop signs and from the perspective of a pedestrian it is difficult to differentiate, especially when it is unclear that a driver sees you. I sure am not going to risk it, so I stop, and am sure to make sure that driver sees me. But similarly, it makes me more sympathetic to pedestrians and cyclists and I will approach intersections differently when they are around and do my best to make sure they know I see them. When driving, walking, and cycling you are operating in very different environments despite being in the same place. People tend to do things for reasons and if an action is common within a group it would be naive to not consider why this behavior develops.
> If they don't see another car that will have priority, they won't stop.
There's a difference between stop and slow down. There are some people who barely slow down and that's totally illegal and blatantly bad. It's also a minority of the time. Most people do significantly slow if not stop at intersections, regardless if there's another car with the right of way or not.
This is a big mischaracterization. That said, I do see cyclists blowing through stop signs ALL THE TIME, often totally dangerously with other cars (including me) at the intersection with them.
> There's a difference between stop and slow down.
Yes, that's exactly my point!
Going 2mph through a stop sign is materially different from stopping at the line. When you stop, you can confirm that the intersection is clear (and that there are no pedestrians about to cross).
> It's also a minority of the time.
If you had to guess, what would you say is the median speed of cars across the lines at stop signs in SF?
> That said, I do see cyclists blowing through stop signs ALL THE TIME, often totally dangerously with other cars (including me) at the intersection with them.
I guess we drive in different parts of the city. I see this, but not often.
I’d say the median speed is 0, since probably a majority of the time there are other cars.
I see bikes blowing through all over the city, not just where I live. Just yesterday I was in the Richmond and a bike not only didn’t stop but it was AFTER I was fully stopped at the stop sign and about to start driving.
You're restating their point but I don't think you get it. "a large number" is a vague term and I think we would be apt to say "a large number of cars run stop signs." The big issue at hand is selection and perceptual biases. If we're just going on intuition here we're going to make bad decisions. You will not notice the "large number of" cyclists that have appropriate behavior because this will be normal behavior that is non-disruptive and your brain is designed to not take special notice of this. But your brain is designed to take special notice of rare and/or disruptive events. So you always over inflate those numbers and it is hard to accurately quantify. See "Perceptual vigilance" for more.
> In 1982, Idaho was the first State to pass such a law, commonly known as the “Idaho Stop Law.” The law
allows bicyclists to yield at stop signs and proceed when safe, rather than come to a complete stop. After Idaho
adopted the law, bicyclist injuries from traffic crashes declined by 14.5% the following year (Meggs, 2010). In
2017, Delaware adopted a similar, limited stop-as-yield law, known
as the "Delaware Yield.” Traffic crashes involving bicyclists at stop
sign intersections fell by 23% in the 30 months after the law’s
passage, compared to the previous 30 months.
> Cyclists have a bad rep in SF because many (not all) ride quite dangerously.
Cycling in the US is almost by definition riding dangerously, since there are very few places with actual safe bicycle infrastructure. This is a self-selecting pressure, because it means that the only cyclists that actually dare to cycle in the city are risk-takers.
> I live adjacent to a four-way stop and there's an incident where a cyclist fails to yield nearly hourly.
You can blame the cyclist or perhaps a four-way stop with a nearly hourly cyclist incident needs to be redesigned.
No it’s not. AB122 passed the legislature to legalize the California roll for cyclists, because coming to a full stop at a stop sign makes no sense for riders with high visibility. Newsom vetoed it for stupid culture war reasons. It’s a law that doesn’t make sense so it’s not a big issue if it isn’t followed.
I'm a cyclist and support these types of laws. But I could imagine someone making the same case for a car (or a motorcycle?). You can have high visibility and come towards a 4-way stop in the middle of nowhere and there's not another car in sight. It doesn't really make sense.
For cycling of course it's super annoying to lose your momentum.
I see cars in the city roll through stop signs all the time, but I never heard of a bicyclist killing someone while blasting through a stop sign. The difference between bicycles and cars is based on the potential for harm.
> because coming to a full stop at a stop sign makes no sense for riders with high visibility.
For anyone wondering why I give an explanation here[0]. The tldr is you need to move to be safe and it is harder to get a bike moving than it is a car. It's best to contextualize any such arguments around the operations that go into operating the vehicle and cars and bikes have a lot of differences. Those differences are why cars are more popular (Sure, more energy but motor vs legs. If you want to make the argument that it is easier for a bike you better not get frustrated when a bike is slow to get to speed when you're behind them at a stoplight or stop-sign).
Thanks for the notice, I'm not too concerned though. I didn't violate any rules. People must be upset that I made an argument that it can be safer to not come to a full and complete stop. But I'm not making the argument that it is safe to blindly blow throw an intersection either. Given the comments I'm seeing, it appears that it is difficult to differentiate these two things.
It's simple selfishness. Cyclists don't want to lose their momentum. Coasting is fun and exertion is not; humans evolved to minimize their energy expenditure. To fix that, you'd need to redesign the bicycle to be completely battery powered. Even then, acceleration will be slower than not stopping. Cyclists will try to rationalize their behavior in all kinds of ways, but they're just lying to themselves (and you).
The fix is really quite simple, develop separate routes for cyclists and motor vehicles. Where those routes unavoidably cross have proper intersections. The person you are replying to is right, four-way-stops are a travesty, both for cyclists and cars.
Incorporating manual vehicle operation into the driving test could significantly alter driving habits, a change I've personally experienced after learning to drive a manual as an adult. As GP notes, the desire to conserve energy and avoid stopping and then starting again is common among cyclists, and this principle applies to manual vehicles as well. Having been raised in the U.S., I understand the potential chaos of having no stop signs or traffic lights at busy four-way intersections. However, this system functions effectively in major cities globally, where drivers, perhaps more accustomed to manual vehicles, approach and navigate intersections with greater awareness and negotiation skills.
I believe that mandating manual driving lessons for all learners could foster improved driving behaviors and heightened road awareness. This approach could encourage drivers to be more attentive and considerate of other road users, enhancing overall safety and efficiency on the roads.
Sure, have fewer stops. But even if you separate the routes, cyclists will eventually have to stop somewhere (even for other cyclists), and they won’t want to.
You'll see a lot more cyclists acting safely if it's actually safe to cycle. Dangerous cycling infrastructure scares away safe cyclists, so all you're left with are the daredevils who won't stop for anything.
This is a weird characterization of what's going on given you have an apt description. I've only been struck by cars (twice) when stopped at a stop sign. It's no question that an intersection is one of the most dangerous locations for a cyclist and it is also no question that an intersection can be cleared significantly faster when starting with __any__ amount of momentum vs a complete stop. Specifics will be necessary for making adequate conclusions here. Someone blindly blowing through an intersection certainly clears the intersection faster but that alone doesn't mean the behavior is appropriate or any less idiotic. And someone doing that is very different from someone slowing down and treating the stop sign like a yield sign. An over generalization is just going to lead to irrational conclusions because context is necessary.
Don't give them the time of day. I never understood why cyclists acted the way they act until I started commuting to work. I felt unsafe at intersections when stopped. I felt unsafe at intersections when starting slowly. I read online about some road rules making my trips more dangerous for ME, not the cars. Now I know what I can do that's totally illegal that's keeping me safer.
Cycling is dangerous when mixed with cars, for sure. But blazing through at an intersection full of cars which are expecting you to stop is even more dangerous. Stop sign rules exist for a reason (to slow vehicles down enough that everyone can see and negotiate everyone else) and it's safer for all if you obey them. This is an example of the rationalization I alluded to earlier. Cyclists lie to themselves.
Next you will be telling me that changing lanes without signaling or looking, swerving through traffic, ignoring stop signs on bike paths, riding while looking at one's phone, riding the wrong way, riding without a helmet, or at night without lights (all of which I observe all the time from cyclists) are evolved safety behaviors.
This is not to excuse drivers, who do lazy and selfish things all the time, like looking at their phones. The root cause is the same--human nature. It just so happens that the incentives are worse when cycling.
> Next you will be telling me that changing lanes without signaling or looking, swerving through traffic, ignoring stop signs on bike paths, riding while looking at one's phone, riding the wrong way, riding without a helmet, or at night without lights (all of which I observe all the time from cyclists) are evolved safety behaviors.
You've just described the average cyclist in Amsterdam. They'll do all of that at the same time with a passenger sitting on the rear baggage carrier.
Yet the Netherlands has one of the lowest mortality rates per mile cycled and the US has one of the highest. Despite the US having very strict full stop laws for cyclists.
28% of vehicle fatalities in Amsterdam are cycling related. Deaths do happen. If you’re arguing that these behaviors are, in fact, safe, I would disagree strongly. The U.S. drives large trucks at high speeds because the U.S. isn't very dense, even in cities, and car centric. And bike infrastructure is lacking. I agree, making cycling safer makes cyclists safer. But none of this explains why cyclists bike like a*holes in every country, which is the point I was making. In addition to the physics of bikes encouraging selfish behavior, there is the lack of license plates and ticketing.
> 28% of vehicle fatalities in Amsterdam are cycling related. Deaths do happen.
Of course, none of those behaviors are safe. And mandating lights, a helmet, high visibility jacket, kneepad protectors and a license plate would make a cyclist safer. (And lights are mandatory even in Amsterdam)
But it would also discourage people from cycling by making that mode of transportation even more inconvenient than it already is compared to the car. And so because, as you noted, humans seek convenience; they will take the car instead.
> But none of this explains why cyclists bike like a*holes in every country, which is the point I was making.
Again it's self-selecting, since there is no bike infrastructure and you ride in between large trucks you have to be very assertive in traffic. Which tends to select for the more stand-offish types.
the thing is written in Rust. but does not expose a Rust query API, you have to query it through Datalog queries in strings; what you shared there just builds those strings from python.. it'd be nice to have a directly native API, with horne clauses constructed in Rust.
Citation needed. Numerous actual citations have demonstrated hallmarks of intelligence for years. Tool use. Comprehension and generalization of grammars. World modeling with spatial reasoning through language. Many of these are readily testable in GPT. Many people have… and I dare say that LLMs reading comprehension, problem solving and reasoning skills do surpass that of many actual humans.
> They model intelligent behavior
It is not at all clear that modeling intelligent behavior is any different from intelligence. This is an open question. If you have an insight there I would love to read it.
> They don't know or care what language is: they learn whatever patterns are present in text, language or not.
This is identical to how children learn language prior to schooling. They listen and form connections based on the cooccurrence of words. They’re brains are working overtime to predict what sounds follow next. Before anyone says “not from text!” please don’t forget people who can’t see or hear. Before anyone says, “not only from language!” multimodal LLMs are here now too!
I’m not saying they’re perfect or even possess the same type of intelligence. Obviously the mechanisms are different. However far too many people in this debate are either unaware of their capabilities or hold on too strongly to human exceptionalism.
> There is this religious cult surrounding LLMs that bases all of its expectations of what an LLM can become on a personification of the LLM.
Anthropomorphizing LLMs is indeed an issue but is separate from a debate on their intelligence. I would argue there’s a very different religious cult very vocally proclaiming “that’s not really intelligence!” as these models sprint past goal posts.
All through the lens of personification. It's important to take a step back and ask, "Where do these hallmarks come from?"
The hallmarks of intelligence are literally what is encoded into text. The reason LLMs are so impressive is that they manage to follow those patterns without any explicit direction.
> I dare say that LLMs reading comprehension, problem solving and reasoning skills do surpass that of many actual humans.
People tend to over-optimize reading comprehension by replacing what they are reading with what they predict to be reading. Every person has a worldview built out of prior knowledge that they use to disambiguate language. It takes effort to suspend one's worldview, and it takes effort to write accurate unambiguous language.
An LLM cannot have that problem, because an LLM cannot read. An LLM models text. The most dominant patterns of text are language: either the model aligns with those patterns, or we humans call the result a failure and redirect our efforts.
> Anthropomorphizing LLMs is indeed an issue but is separate from a debate on their intelligence.
How could that even be possible? The very word, "intelligence" is an anthropomorphization. Ignoring that reality moves the argument into pointless territory. If you try to argue that an anthropomorphized LLM is intelligent, then the answer is, "No shit, Sherlock. People are intelligent!" That doesn't answer any questions about a real LLM.
> as these models sprint past goal posts.
Either an LLM succeeds at a goal, or it fails. It has no idea what the difference is. The LLM has no concept of success: no category for failure. An LLM has no goals or intentions, and doesn't make a single logical decision.
So what is its success coming from? The text being modeled. Without humans authoring that text, there is no model at all!
The goals are authored, too. Every subject, every decision, every behavior, and every goal is determined by a human. Without human interaction, the LLM is nothing. Does nothing think? Does an arrow find its target? Of course not.
There are also literally hundreds of articles and tweet threads about it. Moreover, as I said, you can test many of my claims above directly using readily available LLMs.
GP has a much harder defense. They have to prove that despite all of these capabilities that LLMs are not intelligent. That the mechanisms by which humans possess intelligence is fundamentally distinct from a computer’s ability to exhibit the same behaviors so much that it invalidates any claim that LLMs exhibit intelligence.
Intelligence: “the ability to acquire and apply knowledge and skills”. It is difficult to argue that modern LLMs cannot do this. At best we can quibble about the meaning of individual words like “acquire”, “apply”, “knowledge”, and “skills”. That’s a significant goal post shift from even a year ago.
Thanks for the links, but yeah - I would not give much credit to tweets and blog posts. Often this "emergent" behavior is not that. These are not experts.
