"Do X" is great. But "Now change X to Y in the style of Z" is what blows my mind.
I initially put in a prompt like "tell me a story about the hunting dog and the guard dog." and the results were pretty meh. I was interested to see if it was actually pulling in data from FB so I put her name in instead (she had an FB page yes). It did not seem to do much. So I put in some things like "the little hunting dog is always hungry." "the hunting dog is always looking to hunt" and then did a fuller prompt of "the hunting dog and the guard dog are going to the park. tell me a story about it" and finally got a very good fully fleshed out story about how the little dog was on the prowl for a squirrel while the guard dog watched out and it was really well written and made about as much sense as a fictional tale could. There was no awkwardness or other tells that would have made you think that this was written by anything other than a human.
Anyway, that is huge that it can do this. I do remember when I used to play with other chatbots that remembering past context was something they were unable to do, at least not well, but in most cases it seems they didn't even attempt to. ChatGPT has blown me away in many ways, but this is a very specific leap forward.
I was showing it off to my wife. I asked it to write a lecture for an intro to Latin American poetry with a focus on Pablo Neruda, including a background on Chilean history.
The first result was just 5-6 paragraphs with the most rudimentary 101 on all the above topics.
Then I told it to rewrite it as if the lecture was going to be given to grad students. Next, I told it to include examples of Neruda's poetry along with some analysis. I also asked it to flesh out the segment on Chilean history, with a focus on the independence movement.
The final lecture was good enough that you could walk into a classroom of college seniors and deliver it without any edits. The first one might have been good enough for 9th graders at most.
In fact, I'd submit that phenomenon as its own entry in the Weirdest or Most Surreal Tech Trend contest. People who, when confronted with a talking dog, say things like, "This talking dog is an idiot. It thinks 429 is prime, its painting looks like something from Bob Ross on acid, and the C++ code it wrote for me is full of bugs. I don't see what all the hype is about."
Some of this dismissiveness is just whistling past the graveyard, but much of it comes from people who genuinely think the human brain is something that operates by means of ethereal, ineffable, and unreproducible mystic forces.
Its just a mindset I don't understand. And on places like reddit there is definitely some astroturfing going on or at least some definite "camps" that don't like SDC technology, but on HN AFAICT its been genuine disdain. So weird.
Apparently Cruise and Waymo are doing much more impressive things, but they only get a fraction of the press.
Tesla refuses to publish this data.
Tesla operates nearly everywhere. The scale of what's happening isn't even comparable.
https://sites.google.com/view/fsdbetacommunitytracker/home?p...
Even the best self reported areas are less than 100 miles between disengagement.
Waymo and Cruise are at around 30,000 miles between disengagement.
Tesla has released nothing.
No one serious in the self driving space is using disengagements to pick winners or leaders on the path to commercialization. There is no standard about what qualifies as a disengagement and companies will interpret and report their disengagements differently. Those interpretations also change over time, making it even more difficult to actually understand how a company has progressed in its technology.
And again you're missing the point. Waymo/Cruise are operating in an extremely different and much more narrow context than Tesla's FSD. Let's go put Waymo on a random road in a random city and see how it fares compared to Tesla FSD. I guarantee it will be tremendously worse and no one who knows anything about the space would argue otherwise.
Also, why is Tesla so adamant about how many accidents occured simply with FSD disabled? Shouldn't we know how long prior to the accident FSD was engaged? I mean, hitting the brakes in panic only technically makes it "not FSD".
I'd be happy if you showed me any reported data, self or Tesla, of 30k miles before disengagement with FSD. Even within one order of magnitude instead of two.
You pretty clearly don't know what you're talking about.
> Apparently Cruise and Waymo are doing much more impressive things, but they only get a fraction of the press.
They are doing objectively much less impressive things by relying on sensors that require much less intelligent analysis and will lead them to dead ends at the higher end of abilities that Tesla's approach won't.
This is one I take the side of the skeptic, and I can explain why. In college ~1994 a friend entered an autonomous ground vehicle contest. It was a big deal and our university hosted it each year - this was the 2nd year. He had cobbled together an unusual electric wheelchair base, a 486 PC and a tripod with an old camcorder looking down and forward with a frame-grabber card. The course was about one lane wide, bounded by 2 white lines painted in a grass field. There were some other obstacles and stuff (later years they had sections of dashed white lines and other complexity). We had talked about line following algorithms and such, and he had tried some things with limited success.
One day I said I had an idea, and he said OK but I need the hardware back in an hour. I don't recall if it was C or Pascal, but it was familiar enough that I coded something in an hour - with a visual display showing the 15 horizontal lines of pixels my code analyzed (at about 10 fps). We added some stuff later to incorporate sonar data, but that hour of code got us second place. I'm not here to brag about my fabulous code, I'm here to tell you that "faking it" with lane following and simple driving skills is really easy. There was zero "understanding" in what it was doing, zero ability to recover when it went wrong, zero of just anything but the simplest of code to enable moving forward without crossing the lines. I could write pseudo-code for it here in a comment in 10 minutes. When you think about it, driving safely on real roads requires semantic understanding of a lot of things and it's impossible to observe a vehicle driving and have any idea what level of comprehension it has. This combined with personal experience knowing how easy it is to do rudimentary lane following causes me to set the bar very high indeed. Parlor tricks are not safe, and whatever they have on the road today surely falls somewhere between that and fully sentient AI. But where is it exactly? What is good enough for your spouse and kids? And what is good enough for someone else to be letting drive on the road with them? Or while they cross the street on foot?
When a car can't "see" a semi crossing the road, or one doesn't identify something as human because it's not in a crosswalk, or another runs into and pushes an airplane, no I'm not impressed in the slightest at what these things CAN do reasonably well. The instances of what they get wrong tell me so much more about how far they have to go.
GPT wasn't hyped much at all, in comparison. It just appeared. The next couple of iterations will be interesting to watch, since right now we don't really know what the rate of improvement is going to be. We might be fighting Skynet next year, or we might be in for another 50-year-long AI winter.
I think it's a good parable. It's something expensive and man made. Even though it's on average dumber than a typical human, people try to ascribe godliness to it. And so many judge whether it has achieved omniscience. For many of us, it's just a better tool, an IDE, but for some, it's an oracle or divination tool that they use to decide whether they should marry someone.
I think the term "artificial intelligence" makes it even more confusing.
1. We've learned that relu is a really good function to use for neural networks.
2. We've learned that regularization is essential to having large networks without overfitting.
3. We've learned how to scale models up and down so that you can train something on a giant cluster then run it on your phone.
4. We've learned how to apply this to lots of specific kinds of problems.
etc, etc, etc.
We do NOT simply do what we did in the 80s then throw more hardware at it. Instead we are getting better at making the hardware do something useful.
ChatGPT3 isn't really the definitive sign of the latter. Its just a stepping stone on the way. As soon as humans discovered differential calculus, that was the genesis to the eventual development of a general AI. Everything else along the way is an issue of optimization and spending raw compute time, whether its done by humans or machines.
I've seen so many pieces of incredible technology struggle to find a use-case. By contrast, ChatGPT almost has too many use-cases to catalog, and while it's far from perfect, it seems obviously, to me, to be a tipping point of sorts.
Maybe it will just peter out at this level like self driving car technology did, but I really doubt it, and it seems to me that the building blocks are already there to commercialize this and use it in a real way. The future is going to be interesting...