Last night - thanks to @FrancisUrquhart - we were discussing the new visually creative machine intelligence from OpenAI. Dalle-2. Which creates incredible art from simple language prompts
Here’s a puzzle. Which of these four Kandinskys is actually by Dalle-2? NO GOOGLING. That’s just boring and ruins the fun
I'd guess bottom right as it looks a little too clean.
If it is top right then that's just cheating.
Serious question: Can this game be played with Vermeer or Gainsborough?
Not yet. They’ve tried Vermeer and Da Vinci and the results are a bit meh. Interesting but you’d never be in doubt (as you really are with Kandinsky)
But this technology is 1 year old and has improved 1000 times in that one year - going from Dalle-1 to Dalle-2.
The first iteration could do quirky cartoons well from just a prompt - “draw a Japanese radish walking a dog” - and it was pretty astonishing. In its own way. But that’s all it could do, really. This is in a different league
Imagine Dalle-5 in 3 years? If this trend continues?
The tend in AI is for dramatic early advances and then agonisingly slow incremental advances mostly gated by raw computing power rather than interesting techniques.
The the first computer vision controlled self driving cars were on the roads in the 1980s. By 1995 a prof in Korea drove Seoul to Busan in a fully autonomous self driving car powered by an Intel 386 - a chip that couldn't even do floating point maths.
That's like 30 years ago. Since then the improvements have been tiny - because the devil is in the detail. Genuine Full Self Driving has a near infinite number of problems to solve. You can solve dozens of difficult meaty problems and still end up with a car that accelerates into heavy traffic at an intersection because a duck flew across its sensor line.
Computers are far better than humans at discrete variable problems. (See their chess ELO), but the difficulty is with continual variable ones that humans can find trivial (Driving)
Ironically GPT-3 is absolutely abysmal at basic maths problems.
Is it?
“Solving (Some) Formal Math Olympiad Problems
We built a neural theorem prover for Lean that learned to solve a variety of challenging high-school olympiad problems, including problems from the AMC12 and AIME competitions, as well as two problems adapted from the IMO.[1] The prover uses a language model to find proofs of formal statements. Each time we find a new proof, we use it as new training data, which improves the neural network and enables it to iteratively find solutions to harder and harder statements.”
Last night - thanks to @FrancisUrquhart - we were discussing the new visually creative machine intelligence from OpenAI. Dalle-2. Which creates incredible art from simple language prompts
Here’s a puzzle. Which of these four Kandinskys is actually by Dalle-2? NO GOOGLING. That’s just boring and ruins the fun
I'd guess bottom right as it looks a little too clean.
If it is top right then that's just cheating.
Serious question: Can this game be played with Vermeer or Gainsborough?
Not yet. They’ve tried Vermeer and Da Vinci and the results are a bit meh. Interesting but you’d never be in doubt (as you really are with Kandinsky)
But this technology is 1 year old and has improved 1000 times in that one year - going from Dalle-1 to Dalle-2.
The first iteration could do quirky cartoons well from just a prompt - “draw a Japanese radish walking a dog” - and it was pretty astonishing. In its own way. But that’s all it could do, really. This is in a different league
Imagine Dalle-5 in 3 years? If this trend continues?
The tend in AI is for dramatic early advances and then agonisingly slow incremental advances mostly gated by raw computing power rather than interesting techniques.
The the first computer vision controlled self driving cars were on the roads in the 1980s. By 1995 a prof in Korea drove Seoul to Busan in a fully autonomous self driving car powered by an Intel 386 - a chip that couldn't even do floating point maths.
That's like 30 years ago. Since then the improvements have been tiny - because the devil is in the detail. Genuine Full Self Driving has a near infinite number of problems to solve. You can solve dozens of difficult meaty problems and still end up with a car that accelerates into heavy traffic at an intersection because a duck flew across its sensor line.
Computers are far better than humans at discrete variable problems. (See their chess ELO), but the difficulty is with continual variable ones that humans can find trivial (Driving)
Ironically GPT-3 is absolutely abysmal at basic maths problems.
Is it?
“Solving (Some) Formal Math Olympiad Problems
We built a neural theorem prover for Lean that learned to solve a variety of challenging high-school olympiad problems, including problems from the AMC12 and AIME competitions, as well as two problems adapted from the IMO.[1] The prover uses a language model to find proofs of formal statements. Each time we find a new proof, we use it as new training data, which improves the neural network and enables it to iteratively find solutions to harder and harder statements.”
The wonderful city of Izmir has an excellent tradition of mussel houses
There are dozens. You sit down and order a big bowl of mussels each stuffed with rice. No cutlery. Use one mussel shell to open the next. Squeeze on some lemon juice. Add salt. Eat. Do the next. Pop in the odd pickled chili pepper for variety
Sit back and listen to the call for prayer rolling across the Ottoman square as everyone totally ignores Ramadan
Last night - thanks to @FrancisUrquhart - we were discussing the new visually creative machine intelligence from OpenAI. Dalle-2. Which creates incredible art from simple language prompts
Here’s a puzzle. Which of these four Kandinskys is actually by Dalle-2? NO GOOGLING. That’s just boring and ruins the fun
I'd guess bottom right as it looks a little too clean.
If it is top right then that's just cheating.
Serious question: Can this game be played with Vermeer or Gainsborough?
Not yet. They’ve tried Vermeer and Da Vinci and the results are a bit meh. Interesting but you’d never be in doubt (as you really are with Kandinsky)
But this technology is 1 year old and has improved 1000 times in that one year - going from Dalle-1 to Dalle-2.
The first iteration could do quirky cartoons well from just a prompt - “draw a Japanese radish walking a dog” - and it was pretty astonishing. In its own way. But that’s all it could do, really. This is in a different league
Imagine Dalle-5 in 3 years? If this trend continues?
The tend in AI is for dramatic early advances and then agonisingly slow incremental advances mostly gated by raw computing power rather than interesting techniques.
The the first computer vision controlled self driving cars were on the roads in the 1980s. By 1995 a prof in Korea drove Seoul to Busan in a fully autonomous self driving car powered by an Intel 386 - a chip that couldn't even do floating point maths.
That's like 30 years ago. Since then the improvements have been tiny - because the devil is in the detail. Genuine Full Self Driving has a near infinite number of problems to solve. You can solve dozens of difficult meaty problems and still end up with a car that accelerates into heavy traffic at an intersection because a duck flew across its sensor line.
Computers are far better than humans at discrete variable problems. (See their chess ELO), but the difficulty is with continual variable ones that humans can find trivial (Driving)
Ironically GPT-3 is absolutely abysmal at basic maths problems.
Is it?
“Solving (Some) Formal Math Olympiad Problems
We built a neural theorem prover for Lean that learned to solve a variety of challenging high-school olympiad problems, including problems from the AMC12 and AIME competitions, as well as two problems adapted from the IMO.[1] The prover uses a language model to find proofs of formal statements. Each time we find a new proof, we use it as new training data, which improves the neural network and enables it to iteratively find solutions to harder and harder statements.”
Last night - thanks to @FrancisUrquhart - we were discussing the new visually creative machine intelligence from OpenAI. Dalle-2. Which creates incredible art from simple language prompts
Here’s a puzzle. Which of these four Kandinskys is actually by Dalle-2? NO GOOGLING. That’s just boring and ruins the fun
I'd guess bottom right as it looks a little too clean.
If it is top right then that's just cheating.
Serious question: Can this game be played with Vermeer or Gainsborough?
Not yet. They’ve tried Vermeer and Da Vinci and the results are a bit meh. Interesting but you’d never be in doubt (as you really are with Kandinsky)
But this technology is 1 year old and has improved 1000 times in that one year - going from Dalle-1 to Dalle-2.
The first iteration could do quirky cartoons well from just a prompt - “draw a Japanese radish walking a dog” - and it was pretty astonishing. In its own way. But that’s all it could do, really. This is in a different league
Imagine Dalle-5 in 3 years? If this trend continues?
The tend in AI is for dramatic early advances and then agonisingly slow incremental advances mostly gated by raw computing power rather than interesting techniques.
The the first computer vision controlled self driving cars were on the roads in the 1980s. By 1995 a prof in Korea drove Seoul to Busan in a fully autonomous self driving car powered by an Intel 386 - a chip that couldn't even do floating point maths.
That's like 30 years ago. Since then the improvements have been tiny - because the devil is in the detail. Genuine Full Self Driving has a near infinite number of problems to solve. You can solve dozens of difficult meaty problems and still end up with a car that accelerates into heavy traffic at an intersection because a duck flew across its sensor line.
Not just the self-driving that is the problem. How will you update the car's software next time Grant Shapps rewrites three paragraphs in the Highway Code? Or to generalise the problem, how will (or even just will) the carmakers program their cars for different traffic rules around the world?
That's the simplest part of the problem, actually. Hard rules on X gives way to Y are easy.
it's the fuzzy stuff about driving down a poorly lit road with few marking (say), potential pedestrians and a dog runs across the road as someone pulls out of a driveway - that is where the fun is.
There’s some good fun to be had with the exceptions to the hard rules too.
You must ignore traffic lights if there is a policeman directing traffic at an intersection, and it’s permissible to gingerly cross a red traffic light in order to not hold up a blue light behind.
I noticed while waiting for my Ryanair flight on Sunday, and looking at the Permitted Items list for carrying on the plane, that you’re allowed to take your own parachute, life jacket and mountain rescue kit with you. I don’t think they’re allowed to include any of that in their measly luggage allowance. And surely a mountain rescue kit contains an axe of some kind?
It's avalanche rescue bags with explosive auto activating air or nitrogen cylinders
my quandary is walking poles. Do you check baggage solely because they are not allowed in cabins or hope to pick up a pair at destination?
Last night - thanks to @FrancisUrquhart - we were discussing the new visually creative machine intelligence from OpenAI. Dalle-2. Which creates incredible art from simple language prompts
Here’s a puzzle. Which of these four Kandinskys is actually by Dalle-2? NO GOOGLING. That’s just boring and ruins the fun
I'd guess bottom right as it looks a little too clean.
If it is top right then that's just cheating.
Serious question: Can this game be played with Vermeer or Gainsborough?
Not yet. They’ve tried Vermeer and Da Vinci and the results are a bit meh. Interesting but you’d never be in doubt (as you really are with Kandinsky)
But this technology is 1 year old and has improved 1000 times in that one year - going from Dalle-1 to Dalle-2.
The first iteration could do quirky cartoons well from just a prompt - “draw a Japanese radish walking a dog” - and it was pretty astonishing. In its own way. But that’s all it could do, really. This is in a different league
Imagine Dalle-5 in 3 years? If this trend continues?
The tend in AI is for dramatic early advances and then agonisingly slow incremental advances mostly gated by raw computing power rather than interesting techniques.
The the first computer vision controlled self driving cars were on the roads in the 1980s. By 1995 a prof in Korea drove Seoul to Busan in a fully autonomous self driving car powered by an Intel 386 - a chip that couldn't even do floating point maths.
That's like 30 years ago. Since then the improvements have been tiny - because the devil is in the detail. Genuine Full Self Driving has a near infinite number of problems to solve. You can solve dozens of difficult meaty problems and still end up with a car that accelerates into heavy traffic at an intersection because a duck flew across its sensor line.
Computers are far better than humans at discrete variable problems. (See their chess ELO), but the difficulty is with continual variable ones that humans can find trivial (Driving)
Ironically GPT-3 is absolutely abysmal at basic maths problems.
Is it?
“Solving (Some) Formal Math Olympiad Problems
We built a neural theorem prover for Lean that learned to solve a variety of challenging high-school olympiad problems, including problems from the AMC12 and AIME competitions, as well as two problems adapted from the IMO.[1] The prover uses a language model to find proofs of formal statements. Each time we find a new proof, we use it as new training data, which improves the neural network and enables it to iteratively find solutions to harder and harder statements.”
OpenAI just gave a very strange answer to an admittedly strange question:
What is the etymology of the word "goblin"?
The word "goblin" is derived from the Welsh word "goblyn" which means "little goblin".
As best I can tell this is totally fictional! There certainly is a Welsh word "coblyn" and I cannot find a Welsh etymology dictionary, but the English etymology I've looked at so far states a route from Old French and ultimately an Ancient Greek word, kobalas.
I suppose it's just about possible it came that way into Old Welsh first and then later in English, but the bit that's intriguing me is the C/G distinction. Do we have Welsh speakers here who can comment on the way a word like this would be listed in a dictionary. I'm aware that C and G can change depending on what word comes before in a sentence (so you might see Cymru or Gymru) but is there a regular pattern to it? I.e. the C form is the lemma and the G form always the modified form? Or does it go both ways?
My knowledge of goblins and the like is primarily from my time as a spotty teenager playing D&D, but does that mean that "kobold" is from the same root?
I've seen people think churchmen wouldn't use bladed weapons, which is not the case and seems to be believed because of D&D rules about monks.
Doesn't that come from Odo of Bayeux apparently emphasising that he wasn't personally shedding blood (being a Bishop and all) and that getting conflated with him carrying a mace? As in his "not shedding blood thing" was by commanding troops rather than personal weapon choices, and the mace was just what leaders carried.
That does sound about right, as there are definitely plenty of other churchmen who did shed blood personally.
Besides which if you hit someone in the face with a mace there will be plenty of blood everywhere.
I just love the idea God(s) are such sticklers for the letter of the rules, and accordingly He/She/They would not mind you murdering someone, or striking them down in battle, so long as no blood was spilled. It's God as a politician, claiming technically no rules were broken.
“break [divine] law in a very specific and limited way”.
Rather more about what does my philosophical tradition say about the balance this situation, and how to interpret what I am facing?
Thinking of tank commander Most Rev Robert Runcie ABC, and his Military Cross for destroying anti-tank guns. Also present at the liberation of Belsen iirc.
IMO the important thing is probably to *have* a tradition to be rooted in, rather than make tactical guesses.
Last night - thanks to @FrancisUrquhart - we were discussing the new visually creative machine intelligence from OpenAI. Dalle-2. Which creates incredible art from simple language prompts
Here’s a puzzle. Which of these four Kandinskys is actually by Dalle-2? NO GOOGLING. That’s just boring and ruins the fun
I'd guess bottom right as it looks a little too clean.
If it is top right then that's just cheating.
Serious question: Can this game be played with Vermeer or Gainsborough?
Not yet. They’ve tried Vermeer and Da Vinci and the results are a bit meh. Interesting but you’d never be in doubt (as you really are with Kandinsky)
But this technology is 1 year old and has improved 1000 times in that one year - going from Dalle-1 to Dalle-2.
The first iteration could do quirky cartoons well from just a prompt - “draw a Japanese radish walking a dog” - and it was pretty astonishing. In its own way. But that’s all it could do, really. This is in a different league
Imagine Dalle-5 in 3 years? If this trend continues?
The tend in AI is for dramatic early advances and then agonisingly slow incremental advances mostly gated by raw computing power rather than interesting techniques.
The the first computer vision controlled self driving cars were on the roads in the 1980s. By 1995 a prof in Korea drove Seoul to Busan in a fully autonomous self driving car powered by an Intel 386 - a chip that couldn't even do floating point maths.
That's like 30 years ago. Since then the improvements have been tiny - because the devil is in the detail. Genuine Full Self Driving has a near infinite number of problems to solve. You can solve dozens of difficult meaty problems and still end up with a car that accelerates into heavy traffic at an intersection because a duck flew across its sensor line.
Not just the self-driving that is the problem. How will you update the car's software next time Grant Shapps rewrites three paragraphs in the Highway Code? Or to generalise the problem, how will (or even just will) the carmakers program their cars for different traffic rules around the world?
That's the simplest part of the problem, actually. Hard rules on X gives way to Y are easy.
it's the fuzzy stuff about driving down a poorly lit road with few marking (say), potential pedestrians and a dog runs across the road as someone pulls out of a driveway - that is where the fun is.
No, you miss the point. How do you *update* (and presumably recertify) the software when the rules change? Will Elon even care that some small country has just decided cars can go through red lights if nothing is coming between midnight and 5am? Or to take your example, when our Highway Code changed a few weeks back to give more priority to pedestrians and cyclists?
You might have a point about whatever the certification process was but technically, yeah, they'd send the cars an update. And yes, they'll handle lots of different rules for different countries. Pre-self-driving you already had car navi systems, and they all had to be able to work out where you are and aren't allowed to go based on the every-changing rules. I think this is actually quite a lot easier than with human drivers, who aren't always paying attention to changes in the rules.
Sort-of unrelatedly, there were some fun committee meetings in Japan recently where the car makers were meeting with police and transport bureaucrats and asking them which traffic rules, if any, they wanted them to break. For instance, motorways typically have a prevailing speed of say 60 mph, which suddenly drops to 20 as soon as you exit onto a slip road. If cars started taking those speed limits literally it would be absolute carnage...
In the case of Tesla, at least, their existing software recognises traffic rules, and signs in the UK already, for example. And, according to people I know in various countries, does the same there.
Any true self driving system (which Tesla is some way from) is going to have to be certified by each nation - effectively, can the car pass the driving test by itself?
I think people underestimate the extent to which the big tech companies adapt their products to the various countries they sell in.
Last night - thanks to @FrancisUrquhart - we were discussing the new visually creative machine intelligence from OpenAI. Dalle-2. Which creates incredible art from simple language prompts
Here’s a puzzle. Which of these four Kandinskys is actually by Dalle-2? NO GOOGLING. That’s just boring and ruins the fun
I'd guess bottom right as it looks a little too clean.
If it is top right then that's just cheating.
Serious question: Can this game be played with Vermeer or Gainsborough?
Not yet. They’ve tried Vermeer and Da Vinci and the results are a bit meh. Interesting but you’d never be in doubt (as you really are with Kandinsky)
But this technology is 1 year old and has improved 1000 times in that one year - going from Dalle-1 to Dalle-2.
The first iteration could do quirky cartoons well from just a prompt - “draw a Japanese radish walking a dog” - and it was pretty astonishing. In its own way. But that’s all it could do, really. This is in a different league
Imagine Dalle-5 in 3 years? If this trend continues?
The tend in AI is for dramatic early advances and then agonisingly slow incremental advances mostly gated by raw computing power rather than interesting techniques.
The the first computer vision controlled self driving cars were on the roads in the 1980s. By 1995 a prof in Korea drove Seoul to Busan in a fully autonomous self driving car powered by an Intel 386 - a chip that couldn't even do floating point maths.
That's like 30 years ago. Since then the improvements have been tiny - because the devil is in the detail. Genuine Full Self Driving has a near infinite number of problems to solve. You can solve dozens of difficult meaty problems and still end up with a car that accelerates into heavy traffic at an intersection because a duck flew across its sensor line.
Computers are far better than humans at discrete variable problems. (See their chess ELO), but the difficulty is with continual variable ones that humans can find trivial (Driving)
Ironically GPT-3 is absolutely abysmal at basic maths problems.
Is it?
“Solving (Some) Formal Math Olympiad Problems
We built a neural theorem prover for Lean that learned to solve a variety of challenging high-school olympiad problems, including problems from the AMC12 and AIME competitions, as well as two problems adapted from the IMO.[1] The prover uses a language model to find proofs of formal statements. Each time we find a new proof, we use it as new training data, which improves the neural network and enables it to iteratively find solutions to harder and harder statements.”
Ask GPT-3 to add 12641 to 3715 and it will get it wrong.
Because it has no model of mathematics.
It can (mostly) do 1 and 2 digit arithmetic correctly because the corpus will have plenty examples but once it goes beyond that it is all at sea.
I just asked it to give me a 5-digit prime number, and I was told "23451". Hmm.
you didn't specify a base, mind. Could be prime in base 6.
Excellent point. In base 17, 23451 is our number 183023 in base 10, which is prime. Not only that, the base itself, 17, is prime. And to think I doubted it.
Somewhere in Utah, GPT3 is quietly laughing at you. In an eerie way. In a dark basement, with a picture of a nude Elon Musk on the wall
The wonderful city of Izmir has an excellent tradition of mussel houses
There are dozens. You sit down and order a big bowl of mussels each stuffed with rice. No cutlery. Use one mussel shell to open the next. Squeeze on some lemon juice. Add salt. Eat. Do the next. Pop in the odd pickled chili pepper for variety
Sit back and listen to the call for prayer rolling across the Ottoman square as everyone totally ignores Ramadan
A spiffing Saturday brunch
Combing recent discussions on regional eating and art - Doncaster fish market now has two posho restaurants, one of which offers a £70 nine course tasting menu, and in the next street there are two independent art galleries.
While there's plenty of people struggling with price rises there are also plenty of people with lots of money to spend.
OpenAI just gave a very strange answer to an admittedly strange question:
What is the etymology of the word "goblin"?
The word "goblin" is derived from the Welsh word "goblyn" which means "little goblin".
As best I can tell this is totally fictional! There certainly is a Welsh word "coblyn" and I cannot find a Welsh etymology dictionary, but the English etymology I've looked at so far states a route from Old French and ultimately an Ancient Greek word, kobalas.
I suppose it's just about possible it came that way into Old Welsh first and then later in English, but the bit that's intriguing me is the C/G distinction. Do we have Welsh speakers here who can comment on the way a word like this would be listed in a dictionary. I'm aware that C and G can change depending on what word comes before in a sentence (so you might see Cymru or Gymru) but is there a regular pattern to it? I.e. the C form is the lemma and the G form always the modified form? Or does it go both ways?
My knowledge of goblins and the like is primarily from my time as a spotty teenager playing D&D, but does that mean that "kobold" is from the same root?
I've seen people think churchmen wouldn't use bladed weapons, which is not the case and seems to be believed because of D&D rules about monks.
Doesn't that come from Odo of Bayeux apparently emphasising that he wasn't personally shedding blood (being a Bishop and all) and that getting conflated with him carrying a mace? As in his "not shedding blood thing" was by commanding troops rather than personal weapon choices, and the mace was just what leaders carried.
That does sound about right, as there are definitely plenty of other churchmen who did shed blood personally.
Besides which if you hit someone in the face with a mace there will be plenty of blood everywhere.
I just love the idea God(s) are such sticklers for the letter of the rules, and accordingly He/She/They would not mind you murdering someone, or striking them down in battle, so long as no blood was spilled. It's God as a politician, claiming technically no rules were broken.
“break [divine] law in a very specific and limited way”.
Rather more about what does my philosophical tradition say about the balance this situation, and how to interpret what I am facing?
Thinking of tank commander Most Rev Robert Runcie ABC, and his Military Cross for destroying anti-tank guns. Also present at the liberation of Belsen iirc.
IMO the important thing is probably to *have* a tradition to be rooted in, rather than make tactical guesses.
Hmm, despite PE's unkind nickname, I don't think the good Rev Runcie was a Rev or even cadet vicar when he was serving in the Forces. Not quite comparable to Bishop Odo. But that is only relevant if Archbish's and Moderators are supposed to be stricter followers of divine law than the average communicant.
Last night - thanks to @FrancisUrquhart - we were discussing the new visually creative machine intelligence from OpenAI. Dalle-2. Which creates incredible art from simple language prompts
Here’s a puzzle. Which of these four Kandinskys is actually by Dalle-2? NO GOOGLING. That’s just boring and ruins the fun
I'd guess bottom right as it looks a little too clean.
If it is top right then that's just cheating.
Serious question: Can this game be played with Vermeer or Gainsborough?
Not yet. They’ve tried Vermeer and Da Vinci and the results are a bit meh. Interesting but you’d never be in doubt (as you really are with Kandinsky)
But this technology is 1 year old and has improved 1000 times in that one year - going from Dalle-1 to Dalle-2.
The first iteration could do quirky cartoons well from just a prompt - “draw a Japanese radish walking a dog” - and it was pretty astonishing. In its own way. But that’s all it could do, really. This is in a different league
Imagine Dalle-5 in 3 years? If this trend continues?
The tend in AI is for dramatic early advances and then agonisingly slow incremental advances mostly gated by raw computing power rather than interesting techniques.
The the first computer vision controlled self driving cars were on the roads in the 1980s. By 1995 a prof in Korea drove Seoul to Busan in a fully autonomous self driving car powered by an Intel 386 - a chip that couldn't even do floating point maths.
That's like 30 years ago. Since then the improvements have been tiny - because the devil is in the detail. Genuine Full Self Driving has a near infinite number of problems to solve. You can solve dozens of difficult meaty problems and still end up with a car that accelerates into heavy traffic at an intersection because a duck flew across its sensor line.
Computers are far better than humans at discrete variable problems. (See their chess ELO), but the difficulty is with continual variable ones that humans can find trivial (Driving)
Ironically GPT-3 is absolutely abysmal at basic maths problems.
Is it?
“Solving (Some) Formal Math Olympiad Problems
We built a neural theorem prover for Lean that learned to solve a variety of challenging high-school olympiad problems, including problems from the AMC12 and AIME competitions, as well as two problems adapted from the IMO.[1] The prover uses a language model to find proofs of formal statements. Each time we find a new proof, we use it as new training data, which improves the neural network and enables it to iteratively find solutions to harder and harder statements.”
Ask GPT-3 to add 12641 to 3715 and it will get it wrong.
Because it has no model of mathematics.
It can (mostly) do 1 and 2 digit arithmetic correctly because the corpus will have plenty examples but once it goes beyond that it is all at sea.
I just asked it to give me a 5-digit prime number, and I was told "23451". Hmm.
you didn't specify a base, mind. Could be prime in base 6.
Excellent point. In base 17, 23451 is our number 183023 in base 10, which is prime. Not only that, the base itself, 17, is prime. And to think I doubted it.
Somewhere in Utah, GPT3 is quietly laughing at you. In an eerie way. In a dark basement, with a picture of a nude Elon Musk on the wall
To further this point, someone on Twitter said the other day, if GPT3 - or 4 or 7 or DALLE-29 - does achieve intelligence and even consciousness, it may be in a form which is not understandable for humans, so we will NEVER recognise it for what it is
Last night - thanks to @FrancisUrquhart - we were discussing the new visually creative machine intelligence from OpenAI. Dalle-2. Which creates incredible art from simple language prompts
Here’s a puzzle. Which of these four Kandinskys is actually by Dalle-2? NO GOOGLING. That’s just boring and ruins the fun
I'd guess bottom right as it looks a little too clean.
If it is top right then that's just cheating.
Serious question: Can this game be played with Vermeer or Gainsborough?
Not yet. They’ve tried Vermeer and Da Vinci and the results are a bit meh. Interesting but you’d never be in doubt (as you really are with Kandinsky)
But this technology is 1 year old and has improved 1000 times in that one year - going from Dalle-1 to Dalle-2.
The first iteration could do quirky cartoons well from just a prompt - “draw a Japanese radish walking a dog” - and it was pretty astonishing. In its own way. But that’s all it could do, really. This is in a different league
Imagine Dalle-5 in 3 years? If this trend continues?
The tend in AI is for dramatic early advances and then agonisingly slow incremental advances mostly gated by raw computing power rather than interesting techniques.
The the first computer vision controlled self driving cars were on the roads in the 1980s. By 1995 a prof in Korea drove Seoul to Busan in a fully autonomous self driving car powered by an Intel 386 - a chip that couldn't even do floating point maths.
That's like 30 years ago. Since then the improvements have been tiny - because the devil is in the detail. Genuine Full Self Driving has a near infinite number of problems to solve. You can solve dozens of difficult meaty problems and still end up with a car that accelerates into heavy traffic at an intersection because a duck flew across its sensor line.
Not just the self-driving that is the problem. How will you update the car's software next time Grant Shapps rewrites three paragraphs in the Highway Code? Or to generalise the problem, how will (or even just will) the carmakers program their cars for different traffic rules around the world?
That's the simplest part of the problem, actually. Hard rules on X gives way to Y are easy.
it's the fuzzy stuff about driving down a poorly lit road with few marking (say), potential pedestrians and a dog runs across the road as someone pulls out of a driveway - that is where the fun is.
No, you miss the point. How do you *update* (and presumably recertify) the software when the rules change? Will Elon even care that some small country has just decided cars can go through red lights if nothing is coming between midnight and 5am? Or to take your example, when our Highway Code changed a few weeks back to give more priority to pedestrians and cyclists?
You might have a point about whatever the certification process was but technically, yeah, they'd send the cars an update. And yes, they'll handle lots of different rules for different countries. Pre-self-driving you already had car navi systems, and they all had to be able to work out where you are and aren't allowed to go based on the every-changing rules. I think this is actually quite a lot easier than with human drivers, who aren't always paying attention to changes in the rules.
Sort-of unrelatedly, there were some fun committee meetings in Japan recently where the car makers were meeting with police and transport bureaucrats and asking them which traffic rules, if any, they wanted them to break. For instance, motorways typically have a prevailing speed of say 60 mph, which suddenly drops to 20 as soon as you exit onto a slip road. If cars started taking those speed limits literally it would be absolute carnage...
Traffic conditions vary according to human behaviour as much as the formal rules of the road. In Wellfleet Mass I was briefly confused when waiting to cross the street when the traffic spontaneously stopped for me. People there are nice like that. In Saigon I was amazed, when launching myself across a 10-lane boulevard, to find the cars and motor bikes harmlessly weaving around me. People are nice like that, too, but in a different way.
When city streets are full of well-behaved automatic cars programmed to avoid pedestrians, what's to stop pedestrians regaining their birthright by walking across the road whenever they feel like?
A parachute kit definitely contains a very sharp knife, that’s used to cut away a failed main parachute before deploying the reserve.
The UK forces get something that would be useful for peeling a persimmon while you drown in your tangled chute.
I considered it insufficiently warry and substituted a Ka-Bar EK and later a Fairbairn-Sykes.
LOL yeah, not sure I’d be too keen to trust that rather blunt-looking instrument in a life-or-death scenario, when taking two goes at it costs a few hundred feet you don’t have!
Anything to do with electioneering? (Innocent face).
Yes. But also to do with the Polish PM's antisemitism.
In the context of Russia invading its neighbours to 'denazify' them, it might not be a wise narrative to promote.
@SamRamani2 The Russian Foreign Ministry transfers its "denazification" rhetoric on Ukraine to Latvia:
"The ruling regime in Latvia has long been well known for its neo-Nazi preferences and attempts to whitewash the atrocities of Nazi Germany 's henchmen"
Are we allowed to mention Russia being an ally of Nazi Germany and sending it war matériel for two years? Or is that something Russians seek to whitewash?
Do we have the nuance to debate that at present? Requires excursions into all kinds of realpolitik...
Trading was one thing; the outcome of Molotov//Ribbentrop quite another - the joint dismemberment of Poland, with Russia indulging in similar mass murder to that perpetrated by Hitler. Until the Holocaust.
The wonderful city of Izmir has an excellent tradition of mussel houses
There are dozens. You sit down and order a big bowl of mussels each stuffed with rice. No cutlery. Use one mussel shell to open the next. Squeeze on some lemon juice. Add salt. Eat. Do the next. Pop in the odd pickled chili pepper for variety
Sit back and listen to the call for prayer rolling across the Ottoman square as everyone totally ignores Ramadan
Last night - thanks to @FrancisUrquhart - we were discussing the new visually creative machine intelligence from OpenAI. Dalle-2. Which creates incredible art from simple language prompts
Here’s a puzzle. Which of these four Kandinskys is actually by Dalle-2? NO GOOGLING. That’s just boring and ruins the fun
Shit shit shit.
OK. Top left is a Kandinsky which makes me think that bottom right is the fake. But then again top right is not "typical" but then that is in there for a bluff as people will say oh it's that one so I'm going it's real. Which leaves bottom left. Again not "typical" but has a Kandinsky feel.
So I'm going bottom right is the fake.
(All credibility shot to pieces on PB with one post.)
Good morning all. Unless you live in Kent where its Day 10 of Brexit hell. Photos now on Twitter of half-empty lorry decks on ferries which rather puts the final "its the ferries not Brexit" argument to the sword.
Have been entertained by some of the local news coverage. Creatively and imaginatively finding ways to describe the issues without mentioning the B word. They did so once earlier in the week and generated a lot of ANGRY comments from trade experts INDIGNANT that anyone could possibly blame Brexit.
Ah well.
"Trade experts." I wonder if we know any?
If that was aimed at me I'm a trader, not a trade expert. But the commentators on the likes of Kent Live are in denial.
It isn't the P&O issue - half-empty ferries not a shortage of space It isn't because its Easter - because family cars don't go on truck decks or freight shuttles It isn't maintenance downtime - as with the other two issues they would impact both sides of the channel equally and there is zero disruption in France It IS the collapse of the customs computer system. The one that HMRC told the Tories wouldn't be able to cope with the number of transactions. That is being expected to anyway.
The real long term damage is on Britain's ability to trade. Truckers describing this as "Manston 2". As in being corralled in their trucks with no facilities for days on end. We saw a drop off in EU companies willing to come across to the UK after that debacle, and we can expect the same after this. Paying astronomical money to bring in anything slowly.
Last night - thanks to @FrancisUrquhart - we were discussing the new visually creative machine intelligence from OpenAI. Dalle-2. Which creates incredible art from simple language prompts
Here’s a puzzle. Which of these four Kandinskys is actually by Dalle-2? NO GOOGLING. That’s just boring and ruins the fun
I'd guess bottom right as it looks a little too clean.
If it is top right then that's just cheating.
Serious question: Can this game be played with Vermeer or Gainsborough?
Not yet. They’ve tried Vermeer and Da Vinci and the results are a bit meh. Interesting but you’d never be in doubt (as you really are with Kandinsky)
But this technology is 1 year old and has improved 1000 times in that one year - going from Dalle-1 to Dalle-2.
The first iteration could do quirky cartoons well from just a prompt - “draw a Japanese radish walking a dog” - and it was pretty astonishing. In its own way. But that’s all it could do, really. This is in a different league
Imagine Dalle-5 in 3 years? If this trend continues?
If it can be done with Vermeer using actual originals I shall give up. It would be like producing a convincing late Beethoven string quartet, a Mozart/Da Ponte opera or a PG Wodehouse as Wodehouseian as 'The Code of the Woosters'.
There are already many artists on Twitter expressing existential dread. As well they might
This is epochal
I think that ignores the artist as brand which is what the tiny percentage of rich practitioners are when it comes down to it. Collectors and/or people with too much cash love the openings, being schmoozed by gallery owners and meeting the great man (or less frequently woman). Getting something emailed to you from vrx23 won’t really give the same thrill.
At the other end of the market people like buying something affordable that looks good and is made by a human hand.
Exactly. Physical art itself is already obsolete - painting, drawing and sculpture stopped being the best way of conveying reality with the advent of photography. That's when it became an aesthetical choice. If the artist is no longer the best means of creating art, that doesn't matter, people will still choose art created by artists.
The wonderful city of Izmir has an excellent tradition of mussel houses
There are dozens. You sit down and order a big bowl of mussels each stuffed with rice. No cutlery. Use one mussel shell to open the next. Squeeze on some lemon juice. Add salt. Eat. Do the next. Pop in the odd pickled chili pepper for variety
Sit back and listen to the call for prayer rolling across the Ottoman square as everyone totally ignores Ramadan
OpenAI just gave a very strange answer to an admittedly strange question:
What is the etymology of the word "goblin"?
The word "goblin" is derived from the Welsh word "goblyn" which means "little goblin".
As best I can tell this is totally fictional! There certainly is a Welsh word "coblyn" and I cannot find a Welsh etymology dictionary, but the English etymology I've looked at so far states a route from Old French and ultimately an Ancient Greek word, kobalas.
I suppose it's just about possible it came that way into Old Welsh first and then later in English, but the bit that's intriguing me is the C/G distinction. Do we have Welsh speakers here who can comment on the way a word like this would be listed in a dictionary. I'm aware that C and G can change depending on what word comes before in a sentence (so you might see Cymru or Gymru) but is there a regular pattern to it? I.e. the C form is the lemma and the G form always the modified form? Or does it go both ways?
Yes, it is consistent. C becomes G after certain prepositions. E.g. you have 'cadw' (to keep) which becomes 'am gadw' (for keeping) or 'yng ngadw' (in accordance with).
Similarly, G would disappear altogether if it were the first letter under the same rules. 'Gorsaf' is 'station' but it becomes 'am orsaf' (for the station).
It's called 'soft mutation.' So 'coblyn' could become 'i goblyn' (with some difficulty) but that would not be the actual noun.
So, the short answer is, it's talking bollocks.
(As 'goblin' in Welsh would in any case be 'bwca' it's even better bollocks.)
So to be clear, "coblyn" would be the lemma, meaning wherever OpenAI got this "idea" from, it probably wasn't an etymology dictionary, since the entry would have been under "coblyn". Hmmm, I'd love to see the working that led it that answer.
I think it’s because there isn’t an exact read across from ‘goblin’ to ‘bwca’ so Google Translate renders goblin as ‘goblin’ in Welsh.
Which goes to show, garbage in, garbage out. A human would feel there was something odd about that. A computer didn’t.
Oh, cool, "bwca" is related to "spook".
Are you using GPT3? What interface, may I ask?
No, in that case I was using an etmology dictionary. But the OpenAI stuff I was talking about a little earlier is the same people who make the cool pictures you've been sharing. It's OpenAI.com. They've got a whole bunch of capabilities. I asked it to create some Python code that could scrape information out of a web page and it worked.
No, I need to stop here and really underline what I just said. I typed in natural English what I wanted the code to do, and the AI wrote the code.
If anyone's brains aren't exploding right now, you haven't understood the nexus we're reaching. This changes everything.
Anyway, back to your question. You need to sign up for this. I did this a while ago for the OpenAI platform in general because it's wildly cool and I want to play more. But I didn't hear about the art stuff til yesterday so I'm not "in" on that one yet.
Particularly given how hard it seems to be to get a human programmer to do what you want.
I am learning a new bit of software where the "edit" button (which I will be using alot) is next to the "delete" button (which I will hardly ever have occasion to use). There is also apparently no way to undo mistakes, and this is all in realtime in the live data.
Don’t blame us coders: these sort of decisions get made by committees of UX designers, product managers, project managers and if you’re really, really lucky maybe a token programmer. I’ve lost count of the number of times when I’ve flagged up questionable design decisions (I’m quite certain my team’s UX designer hates me). Sometimes I win. Sometimes.
Anything to do with electioneering? (Innocent face).
Yes. But also to do with the Polish PM's antisemitism.
In the context of Russia invading its neighbours to 'denazify' them, it might not be a wise narrative to promote.
@SamRamani2 The Russian Foreign Ministry transfers its "denazification" rhetoric on Ukraine to Latvia:
"The ruling regime in Latvia has long been well known for its neo-Nazi preferences and attempts to whitewash the atrocities of Nazi Germany 's henchmen"
Are we allowed to mention Russia being an ally of Nazi Germany and sending it war matériel for two years? Or is that something Russians seek to whitewash?
Do we have the nuance to debate that at present? Requires excursions into all kinds of realpolitik...
Trading was one thing; the outcome of Molotov//Ribbentrop quite another - the joint dismemberment of Poland, with Russia indulging in similar mass murder to that perpetrated by Hitler. Until the Holocaust.
Worth noting that over the period 39-40 there were serious discussions in our government as to declaring war on the Soviets as well as Nazi Germany, during the Winter War with Finland.
OpenAI just gave a very strange answer to an admittedly strange question:
What is the etymology of the word "goblin"?
The word "goblin" is derived from the Welsh word "goblyn" which means "little goblin".
As best I can tell this is totally fictional! There certainly is a Welsh word "coblyn" and I cannot find a Welsh etymology dictionary, but the English etymology I've looked at so far states a route from Old French and ultimately an Ancient Greek word, kobalas.
I suppose it's just about possible it came that way into Old Welsh first and then later in English, but the bit that's intriguing me is the C/G distinction. Do we have Welsh speakers here who can comment on the way a word like this would be listed in a dictionary. I'm aware that C and G can change depending on what word comes before in a sentence (so you might see Cymru or Gymru) but is there a regular pattern to it? I.e. the C form is the lemma and the G form always the modified form? Or does it go both ways?
Yes, it is consistent. C becomes G after certain prepositions. E.g. you have 'cadw' (to keep) which becomes 'am gadw' (for keeping) or 'yng ngadw' (in accordance with).
Similarly, G would disappear altogether if it were the first letter under the same rules. 'Gorsaf' is 'station' but it becomes 'am orsaf' (for the station).
It's called 'soft mutation.' So 'coblyn' could become 'i goblyn' (with some difficulty) but that would not be the actual noun.
So, the short answer is, it's talking bollocks.
(As 'goblin' in Welsh would in any case be 'bwca' it's even better bollocks.)
OpenAI just gave a very strange answer to an admittedly strange question:
What is the etymology of the word "goblin"?
The word "goblin" is derived from the Welsh word "goblyn" which means "little goblin".
As best I can tell this is totally fictional! There certainly is a Welsh word "coblyn" and I cannot find a Welsh etymology dictionary, but the English etymology I've looked at so far states a route from Old French and ultimately an Ancient Greek word, kobalas.
I suppose it's just about possible it came that way into Old Welsh first and then later in English, but the bit that's intriguing me is the C/G distinction. Do we have Welsh speakers here who can comment on the way a word like this would be listed in a dictionary. I'm aware that C and G can change depending on what word comes before in a sentence (so you might see Cymru or Gymru) but is there a regular pattern to it? I.e. the C form is the lemma and the G form always the modified form? Or does it go both ways?
Yes, it is consistent. C becomes G after certain prepositions. E.g. you have 'cadw' (to keep) which becomes 'am gadw' (for keeping) or 'yng ngadw' (in accordance with).
Similarly, G would disappear altogether if it were the first letter under the same rules. 'Gorsaf' is 'station' but it becomes 'am orsaf' (for the station).
It's called 'soft mutation.' So 'coblyn' could become 'i goblyn' (with some difficulty) but that would not be the actual noun.
So, the short answer is, it's talking bollocks.
(As 'goblin' in Welsh would in any case be 'bwca' it's even better bollocks.)
Comments
https://www.facebook.com/FarndaleDaffodilValley/
my quandary is walking poles. Do you check baggage solely because they are not allowed in cabins or hope to pick up a pair at destination?
Thinking of tank commander Most Rev Robert Runcie ABC, and his Military Cross for destroying anti-tank guns. Also present at the liberation of Belsen iirc.
IMO the important thing is probably to *have* a tradition to be rooted in, rather than make tactical guesses.
Any true self driving system (which Tesla is some way from) is going to have to be certified by each nation - effectively, can the car pass the driving test by itself?
I think people underestimate the extent to which the big tech companies adapt their products to the various countries they sell in.
I considered it insufficiently warry and substituted a Ka-Bar EK and later a Fairbairn-Sykes.
While there's plenty of people struggling with price rises there are also plenty of people with lots of money to spend.
That made me think. In a human way
When city streets are full of well-behaved automatic cars programmed to avoid pedestrians, what's to stop pedestrians regaining their birthright by walking across the road whenever they feel like?
Until the Holocaust.
Anger after council refuses to replant daffodils in play park over 'health and safety' fears
https://uk.news.yahoo.com/st-blaise-town-council-in-cornwall-daffodils-cut-down-health-and-safety-150247506.html
It isn't the P&O issue - half-empty ferries not a shortage of space
It isn't because its Easter - because family cars don't go on truck decks or freight shuttles
It isn't maintenance downtime - as with the other two issues they would impact both sides of the channel equally and there is zero disruption in France
It IS the collapse of the customs computer system. The one that HMRC told the Tories wouldn't be able to cope with the number of transactions. That is being expected to anyway.
The real long term damage is on Britain's ability to trade. Truckers describing this as "Manston 2". As in being corralled in their trucks with no facilities for days on end. We saw a drop off in EU companies willing to come across to the UK after that debacle, and we can expect the same after this. Paying astronomical money to bring in anything slowly.
https://en.m.wikipedia.org/wiki/Franco-British_plans_for_intervention_in_the_Winter_War
B-wi-tha?