They did deliver a smaller model (pretrained, on github). It's almost as good as the large model. People have been using it to make funny self-referential texts - AI talking about itself - and posting them on twitter.
Frankly, you don't know what you're talking about. If you did, you'd recognise that samples with this level of coherence were not possible before. If you used a LSTM neural net you'd get samples that make sense only for 5-10 words. Here you could read half a page and still make sense.
Comparing the two populations (gamers vs non-gamers) is only valid if they are similarly distributed with respect to other control variables that might influence moral reasoning. What if age, affluence, level of technological adoption and place of birth have an impact on morality, and are correlated with gaming as well?
Google used AI to design cookies. It worked like this: your AI can generate recipes, you take 20 of them and put them in front of your employees, asking them to submit a rating every time they eat a cookie. These ratings are considered a 'reward signal' and used to train a reinforcement learning agent (like the ones that play games). The agent acts by inventing new recipes and learns by reinforcing the reward. So you could say that the AI needs a large committee of people to work as it's taste buds.
> Also, so called 'AI' is completely incapable of one vital ingredient: human creativity.
Human creativity belongs to humans by definition. But AI creativity is not limited to human creativity. We have AIs that can imagine faces and scenes, invent new strategies in games and invent proteins to fit a specific function. AI's can even design deep neural nets to the same level of accuracy as the best human researchers, given enough compute. Creativity is just random search filtered through utility.
I see it as a problem of short vs long term horizon. When you're too old, young, sick or uneducated you got to think about tomorrow, not the year or decade after that. But when you're a young, educated professional you think much more about your long term future. That reflects in the voting patterns. The short term people are easily bought with promises of immediate relief, even if it is just a slight relief or completely illusory. The population will always be split, and, as in game theory, when you have no care about the long term future, why would you cooperate instead of betraying? Thinking of long term well being (cooperating towards such a future) is based on having something to lose or gain long term. As long as much of society has no future, they have no reason not to betray and vote with the destructive short-term policy.
> then the research synopsis would be published on some for pay science paper, while the actual research had been sponsored by someone like facebook anyways.
Usually ML papers are on Arxiv. Almost all of them, including the FB ones.
This negative loop might be suboptimal for lending institutions unless they implement a mechanism for forgiveness. It's what humans do - respond in kind, to good with good and bad with bad, but from time to time, forgive, to give a second chance to a better equilibrium.
The term AI is confusing. They should have asked individually about: face recognition, self driving cars, automated shops, machine translation, voice assistants, ML based genetic research, content recommendation/filtering systems and so on. Then people would have given more precise answers. But if you frame it with the term AI, people think Terminator and Skynet.
Nah, the brain injects tons of randomness in its signals. Free will is just randomness filtered through past experience. We don't need quantum effects for 'free will'.
> For example, consider how we can switch from thought to thought, moment to moment, where each thought is instantly selected among (surely) astronomical number of different possibilities!
You should read up on neural nets. Especially embeddings of words and images, and attention mechanisms. You'd be amazed that analog circuits can handle n-dimensional spaces (usually 300-1000 dimensions) for meaning representation and they can combine, select and transform those representations with ease. Another related field is probability distributions and using sampling to generate possibilities - effectively to imagine things. Current GANs (a kind of neural net) can imagine faces with photographical accuracy.
We don't need quantum effects to explain intelligence, the magic is in the architecture of the brain.
Have you looked at recent graph based neural networks? They can handle multi dimensional and non-uniform structures, such as scenes, texts, proteins and social networks. They are a perfect fit for reasoning tasks. I think they are somewhat in the middle between symbolic and statistical learning. Also, have you seen the progress in reinforcement learning, especially model based RL. It is possible to perform (some) activities at super human levels. I think that graph based NN, RL and simulators for training agents are the next wave. Also, more specialised chips.
In the last year we have gained almost perfect voice recognition and synthesis, image recognition and generation, translation, and mastered all types of board game play. Large parts of the brain are related to processing images and sounds, and those have been solved. I'd say we are right on our way towards AI though ML.
I agree. What we are missing right now is knowledge about what the brain neural networks are optimising for (the loss functions of the brain) and perhaps the types of invariances they can handle. We have conquered only a few of these invariances and objectives. The brain has evolution to thank for. Maybe if we can create good enough simulations we'd be able to offer our AI agents a similar evolutionary path or at least a large enough training ground. Static datasets are just not good enough for general AI.
Off topic, but the same idea, of adversarial images, has been turned into a kind of imagination machine in 2014 by Ian Goodfellow (Generative Adversarial Networks - GANs). https://scholar.google.ro/scho... . (5600 papers!)
They train two neural nets engaged in a game of forgery and forgery detection. The generator net takes in a random noise vector and creates an image. The discriminator network takes in the fake image, other times real images, and has to say which is which. This makes the discriminator learn about images, and using backprop, it trains the generator to generate better fakes. In the end it is possible to generate very realistic original images such as these: https://www.youtube.com/watch?...
I don't think that's enough. Outside the online social networks there is the real world with its power struggles which generate torrents of fake news and disinformation. They would push this shit even without ads.
> it's not like India is incapable of rolling their own alternatives
Then there is the teeny weeny problem of convincing the masses to switch to the new Indian alternative apps, that come bundled with government surveillance. Who would do that? It's safer to be spied upon by multinationals than your own govt.
Have you looked in your crystal ball to say all that? I get it, you don't trust the evolution of AI. But it is still going to work and save many lives.
Slashdot Mirror
They did deliver a smaller model (pretrained, on github). It's almost as good as the large model. People have been using it to make funny self-referential texts - AI talking about itself - and posting them on twitter.
Frankly, you don't know what you're talking about. If you did, you'd recognise that samples with this level of coherence were not possible before. If you used a LSTM neural net you'd get samples that make sense only for 5-10 words. Here you could read half a page and still make sense.
Comparing the two populations (gamers vs non-gamers) is only valid if they are similarly distributed with respect to other control variables that might influence moral reasoning. What if age, affluence, level of technological adoption and place of birth have an impact on morality, and are correlated with gaming as well?
Google used AI to design cookies. It worked like this: your AI can generate recipes, you take 20 of them and put them in front of your employees, asking them to submit a rating every time they eat a cookie. These ratings are considered a 'reward signal' and used to train a reinforcement learning agent (like the ones that play games). The agent acts by inventing new recipes and learns by reinforcing the reward. So you could say that the AI needs a large committee of people to work as it's taste buds.
https://www.cnbc.com/2017/12/0...
> Also, so called 'AI' is completely incapable of one vital ingredient: human creativity.
Human creativity belongs to humans by definition. But AI creativity is not limited to human creativity. We have AIs that can imagine faces and scenes, invent new strategies in games and invent proteins to fit a specific function. AI's can even design deep neural nets to the same level of accuracy as the best human researchers, given enough compute. Creativity is just random search filtered through utility.
Fashion is a chaotic process like weather. If we can't predict the weather more than 10 days ahead, why would we be able to predict fashion?
I see it as a problem of short vs long term horizon. When you're too old, young, sick or uneducated you got to think about tomorrow, not the year or decade after that. But when you're a young, educated professional you think much more about your long term future. That reflects in the voting patterns. The short term people are easily bought with promises of immediate relief, even if it is just a slight relief or completely illusory. The population will always be split, and, as in game theory, when you have no care about the long term future, why would you cooperate instead of betraying? Thinking of long term well being (cooperating towards such a future) is based on having something to lose or gain long term. As long as much of society has no future, they have no reason not to betray and vote with the destructive short-term policy.
> then the research synopsis would be published on some for pay science paper, while the actual research had been sponsored by someone like facebook anyways.
Usually ML papers are on Arxiv. Almost all of them, including the FB ones.
This negative loop might be suboptimal for lending institutions unless they implement a mechanism for forgiveness. It's what humans do - respond in kind, to good with good and bad with bad, but from time to time, forgive, to give a second chance to a better equilibrium.
The term AI is confusing. They should have asked individually about: face recognition, self driving cars, automated shops, machine translation, voice assistants, ML based genetic research, content recommendation/filtering systems and so on. Then people would have given more precise answers. But if you frame it with the term AI, people think Terminator and Skynet.
Nah, the brain injects tons of randomness in its signals. Free will is just randomness filtered through past experience. We don't need quantum effects for 'free will'.
> For example, consider how we can switch from thought to thought, moment to moment, where each thought is instantly selected among (surely) astronomical number of different possibilities!
You should read up on neural nets. Especially embeddings of words and images, and attention mechanisms. You'd be amazed that analog circuits can handle n-dimensional spaces (usually 300-1000 dimensions) for meaning representation and they can combine, select and transform those representations with ease. Another related field is probability distributions and using sampling to generate possibilities - effectively to imagine things. Current GANs (a kind of neural net) can imagine faces with photographical accuracy.
We don't need quantum effects to explain intelligence, the magic is in the architecture of the brain.
The existence of intelligence in mere humans is proof it can be done. We just need to discover better alternatives to silicon.
Have you looked at recent graph based neural networks? They can handle multi dimensional and non-uniform structures, such as scenes, texts, proteins and social networks. They are a perfect fit for reasoning tasks. I think they are somewhat in the middle between symbolic and statistical learning. Also, have you seen the progress in reinforcement learning, especially model based RL. It is possible to perform (some) activities at super human levels. I think that graph based NN, RL and simulators for training agents are the next wave. Also, more specialised chips.
In the last year we have gained almost perfect voice recognition and synthesis, image recognition and generation, translation, and mastered all types of board game play. Large parts of the brain are related to processing images and sounds, and those have been solved. I'd say we are right on our way towards AI though ML.
I agree. What we are missing right now is knowledge about what the brain neural networks are optimising for (the loss functions of the brain) and perhaps the types of invariances they can handle. We have conquered only a few of these invariances and objectives. The brain has evolution to thank for. Maybe if we can create good enough simulations we'd be able to offer our AI agents a similar evolutionary path or at least a large enough training ground. Static datasets are just not good enough for general AI.
https://www.geek.com/geek-cete...
Off topic, but the same idea, of adversarial images, has been turned into a kind of imagination machine in 2014 by Ian Goodfellow (Generative Adversarial Networks - GANs). https://scholar.google.ro/scho... . (5600 papers!) They train two neural nets engaged in a game of forgery and forgery detection. The generator net takes in a random noise vector and creates an image. The discriminator network takes in the fake image, other times real images, and has to say which is which. This makes the discriminator learn about images, and using backprop, it trains the generator to generate better fakes. In the end it is possible to generate very realistic original images such as these: https://www.youtube.com/watch?...
The problem of adversarial images has been studied since 2013. There are over 1500 papers citing the original paper. https://scholar.google.ro/scho...
I don't think that's enough. Outside the online social networks there is the real world with its power struggles which generate torrents of fake news and disinformation. They would push this shit even without ads.
> I was doing accurate voice recognition and voice control of my computer back in the 90s on Pentium 1 CPUs with 16MB of RAM.
So, why didn't you launch a smart speaker in the meantime? Was it crap compared to current day?
Wasn't WhatsApp end-to-end encrypted? That means FB has no idea what people are talking about. So it has no responsibility here.
> it's not like India is incapable of rolling their own alternatives
Then there is the teeny weeny problem of convincing the masses to switch to the new Indian alternative apps, that come bundled with government surveillance. Who would do that? It's safer to be spied upon by multinationals than your own govt.
Have you looked in your crystal ball to say all that? I get it, you don't trust the evolution of AI. But it is still going to work and save many lives.
By the way, why is the level of conversation on slashdot like 4chan? It's pretty disgusting.
I expect we have paid posters here sent to deflect conversations.