Even if this is not 100% serious, it is really starting to feel like the ship computer from star trek is not too far away.
15 years later and we are perpetually “5 years out”. Yes you can take a taxi ride in a closed circuit, but that’s much closer to where we were in 2007 than where we thought we’d be today, and it took 15 years to get here.
This we could actually do 20 years earlier. [1]
A first culmination point was achieved in 1994, when their twin robot vehicles VaMP and VITA-2 drove more than 1,000 kilometres (620 mi) on a Paris multi-lane highway in standard heavy traffic at speeds up to 130 kilometres per hour (81 mph). They demonstrated autonomous driving in free lanes, convoy driving, automatic tracking of other vehicles, and lane changes left and right with autonomous passing of other cars.
You can literally watch cars self-driving in all kinds of places and conditions. Yes, they make mistakes. So do humans.
That's a very strong statement with not much to back it up.
They drive fine in straight, wide, sunny, south US roads (and even there not always), they struggle even in US cities, put them in any European country and it's game over. Mountain roads in swizerland during a snow storm ? Foggy twisty roads in the woods ? These won't be solved easily, even Waymo's ceo acknowledged that fully autonomous cars won't be able to drive everywhere.
The fact that the two cases that work best have a similar climate to the 2007 DUC really highlights the reality that these methods haven’t been proven to scale generally. The industry is still chasing that 2007 success, and it’s not surprising that over 15 years they’ve improved that. But do I need to link to all the promises from CEOs about where they thought we’d be today? Those predictions were based on the idea that the DUC prototypes would be more generally applicable. The successes since then have shown we can make the experience better, but don’t show we can solve autonomous driving in the general case.
- https://www.youtube.com/watch?v=kx7fHEhnIZk (snow, like the google demo)
- https://www.youtube.com/watch?v=LSX3qdy0dFg | https://youtu.be/LSX3qdy0dFg?t=2229 (snow, but not the one I was thinking of)
- at current traffic levels, accidents are rare
- but fatality rates are high (20% of helicopter accidents)
- the commercial carriers have much better accident statistics than general aviation
- commuter and on-demand flights are much worse than commercial scheduled flights
- rather more than half of all accidents have a root incident near an airport - taxi-ing, departure, initial climb, approach, landing.
My conclusion is that mass adoption of flying cars (as in, millions of people piloting small aircraft with varying levels of maintenance, safety inspections, training, and traffic control) would be a terrifyingly foreseeable disaster.
On the other hand, I hold out real hope for fully autonomous vehicles being potentially safer than a distracted teenager on the road.
I might feel like intervening, 1 out of 10 times at this point. I might not be the typical driver, but I definitely feel like its ready for early adopters now.
However, even though I'm a big fan, I don't see how these can easily transition to "mass consumption", because as we get into the uncanny valley where the auto drive is good enough to take over, the masses are going to completely check out of their responsibility to be a good backup driver.
So I feel like we are going to be stuck in the current space for a long time, maybe 10 years. Until you Auto Drive is so good, you can ride one without getting a drivers license.
This rings me a lot. It feels like the current generation AI companies/projects have been rewarded for making people believe the future is near. In reality, we're just driving towards the top of a local maxima for possible big money. We clearly won't reach AGI with the current LLM approaches, for example. (Perhaps, there might be a breakthrough in computer hardware that might make it possible, but only in significantly inefficient ways.)
Have any evidence to back this up? Scaling laws seem to show we aren't near a plateau and it's not clear what kind of capability GPT-4,5 or 6 may have.
Actually ChatGPT has an IQ of ~83, so that is quite close to average human intelligence.
Furthermore, it was trained only on digital text, arguably that would be it's only "sensory organ". It had no other senses with which to correlate terms and concepts it inferred from text, and look how amazing it is just from that.
As the other poster said, multimodal training is the next step and people are not going to be prepared for it.
I asked about a specific Dutch book, ChatGPT was wrong about the author (it was another author born a century later). I corrected it but got told that the two authors were the same and it was a pseudonym.
I ask the birthdate of the correct author. It gave me relatively correct answer with date of birth and death.
I then asked about the birthdate of the wrong author. It told me again a, relatively correct answer, indeed he was born long after the other author died.
I asked ChatGPT how it could be that the dates differed. It told me that it is very usual for an author to go by a pseudonym.
I told it it was wrong. They are different authors living in different centuries . But it stubbornly refused to accept it, teaching me again that it is perfectly common for authors to go by two different names.
edit: Just to add when asked for a description of the book it gave me a very believable summary, which was total nonsense. This is what really disturbed me about ChatGPT. Though I am very impressed by the fact that we now have a system that is very good at parsing human language. Something which was long thought to be impossible. Combining that strength with an, actual, datasource would be the only way forward in my opinion.
ChatGPT doesn't do the getting really angry part because it can't feel shame or insecurity about not knowing things.
Then I asked it "Could you adapt the function so that it works on Venus, where years have 224 days?"
It offered me a new version of the function, which simply checks if the year is a multiple of 224. Apparently on Venus the number of days in a year and the frequency of leap years are the same number. It qualified the answer: "It's worth noting that this function is based on current knowledge and understanding of Venus..."
I asked it "What if we want the function to use Venus days as well as Venus years?"
It offered me the same function, except that a) the variable 'years' was now called 'days', and b) the modulus was changed from 224 to 224.701.
So I asked "Should the argument to the last function be a float or an integer?"
It gave me 3 pars of complete nonsense about how the difference between floats and integers affects the precision of calculating leap years (while again warning that the exact value of the Venus year might change).
ChatGPT does a very good imitation of a certain type of candidate I've occasionally interviewed, who knows almost nothing but is keen to try various tricks to bluff you out, including confidently being wrong, providing lots of meaningless explanation, and sometimes telling you that you are wrong about something. I have never hired anyone like this, but I've occasionally come close.
I have been trying various interview questions on ChatGPT, originally because my colleagues warned me that a candidate who was surreptitiously using it could ace almost any interview. I was skeptical and I have not been convinced.
But I think it's actually a great exercise to practice interviewing on it. If ChatGPT can answer your questions accurately (try to be fair and ignore its slightly uncanny tone), then you probably need better questions. If you are quite technical and put some thought into it, you should be able to come up with things which are both novel enough and hard enough that ChatGPT will simply flounder catastrophically. (I'm not referring to 'tricks' like the Venus question, but real questions on how to achieve something moderately complicated using code.) It's a really good reminder too that when we ask candidates to write code, we should examine and debug it in detail, then ask decent follow-up questions, rather than just accepting something that looks right.
I typed: did you know that you can cross the cavern by just saying fly away
GPT Said: In Colossal Cave Adventure, "fly away" is indeed one of the possible commands to cross the cavern.
I felt like I was talking to a kid pretending to know more about the topic than they really do.
In fairness, I had given several correct alternatives before this so maybe it was the whole interaction that led it to the conclusion that "fly away" was a legitimate solution.
This is an incredibly bold prediction that isn't supported by the opinions of the majority of people in the field and certainly doesn't have any real backing other than your gut.
Well, DUH!
"It is difficult to get a man to understand something when his salary depends upon his not understanding it."
- Upton Sinclair.
If Astronomers were predicting a mass-extinction level asteroid impact for the year 2050 with 50% probability I doubt you would be so cavalier.
I haven't seen that prediction. What I have seen is "AGI is 2 years out", and I have been seeing that for 4 years.
Much like the self-driving cars that were (according to the experts in the industry) 5 years out since 2012, and still not here in 2023.
Maybe if the experts in the industry were more vocal about how far off they are, you wouldn't be reading comments like mine.
> If Astronomers were predicting a mass-extinction level asteroid impact for the year 2050 with 50% probability I doubt you would be so cavalier.
if they had been saying, since 2012, that it's five years away, I won't be the only one laughing at them.
When it comes to AI, though, the world is a lot more forgiving, and collectively more forgetful of the predictions.
Take your pick:
https://nickbostrom.com/papers/survey.pdf
https://aiimpacts.org/2022-expert-survey-on-progress-in-ai/
https://research.aimultiple.com/artificial-general-intellige...
https://forum.effectivealtruism.org/posts/7JxsXYDuqnKMqa6Eq/...
https://www.metaculus.com/questions/5121/date-of-artificial-...
https://www.lesswrong.com/posts/hQysqfSEzciRazx8k/forecastin...
>What I have seen is "AGI is 2 years out", and I have been seeing that for 4 years.
We obviously don't travel in the same circles because I don't know anyone credible saying that.
“Asteroid extinction” hasn't gone through several rounds of hype of imminence from people working in the field with a financial interest in that perception to extended “winters” as the basis of the last imminence cycle bursts in my lifetime, so... maybe the two things aren't analogous.
Even the idea that LLMs can eventually get there isn't taken seriously.
You literally have no way to make that determination.
Any strong declarative statements require justification, period, whether that is an assertion of existence or non-existence.
> unless you can prove that our current approaches are on the right track
How anyone can look at the progress in machine learning in audio, video and written expression, and not think "we're on the right track" is honestly beyond me. You can start here:
https://www.lesswrong.com/posts/K4urTDkBbtNuLivJx/why-i-thin...
My theory on this is that it would confuse the dataset having to both transcribe and then "understand" what was asked. By reducing this single variable [which we all know is technically already possible: audio transcription], the dataset is allowing itself to be trained with less initial noise.