5 ms·

This is so indicative of Google culture missing the point. The idea of spending $10M training a single model is treated as a casual reality. But “tHaNk GoOdNeSs

by keenon 3y ago

This is so indicative of Google culture missing the point. The idea of spending $10M training a single model is treated as a casual reality. But “tHaNk GoOdNeSs those generous open source people published their HiGh QuAlItY datasets of ten thousand examples each. Otherwise we’d have no way of creating datasets like that…” :| the sustainable competitive advantage has been and will continue to be HUGE PROPRIETARY DATASETS. (Duh - this is as true for new AI as it was for old AI = ad targeting). It was the _query+click pairs_ that kept Google dominant all these years, not the brilliant engineers. They had all of humanity labeling the entire internet with “when I click on this page/ad for this query I do/don’t search again” a billion+ times a day for a decade. For good measure they’ve also been collecting your email, your calendar, and your browsing habits for nearly as long. The fact that they’ve managed to erase that historic advantage from their collective consciousness (presumably because AI researchers would rather not spend time debugging data labeling UI) is strange to me. It at least deserves a mention in a strategy memo like this. Not vague platitudes about “influence through innovation.” Spend that $10M you were going to spend on a training run as $9.9999M on a private dataset, then the remaining $100 on training. Better still, build products that gets user behavior to train your models for you. Obviously. We’re going to watch the biggest face plant in recent economic history if they can’t get this one together. I can’t decide if that makes me happy about an overdue changing of the guard in the Valley or sad about the fall of a once great company. It’s not about the models! Model training is a commodity! It’s about the data! Come on guys.

bionhoward 3y ago
One way to push back on the data argument is to consider the progress DeepMind made with self play. Perhaps Bard can self-dialogue and achieve superhuman results. I won’t be surprised. Plus the underlying architecture is dense. Sparse transformers are a major upgrade. That’s only one of many upgrades you can make. There is still a lot of headroom and IMHO GPT-4 already implements AGI if you give it the right context
- jxmorris12 3y ago
  Self-play works for eg Go because there’s a perfect simulator of the game – which gives at least one very clear training signal, winning. There’s no simulator for conversation, no winning, and no training signal. Self-dialog doesn’t make any sense
  - lucubratory 3y ago
    I could think of ways you could make a good attempt at it. You could have a generator/discriminator relationship where you have a model whose purpose is to evaluate model outputs for things other than just toxicity (basically RLHF for capabilities), then use that to train. You could have code generation tasks where the code is actually executed and a success/failure signal sent based on code performance. You could do logical puzzle generator/logical puzzle solver pairs and have a separate system evaluate answer correctness based on a human dataset baseline, or maybe a model programmed to be able to turn natural language logic puzzles into API calls to a formal logical engine solver to get the answer for comparison. You could make a simple program to turn randomly generated mathematical problems into word problems, use an AI to add extraneous detail while protecting the core problem description, then give the resulting word problems to an AI and use a separate AI to extract out the final answer or conclusion. Then compare that answer to what the calculator says for the original math problem. All of those have problems and would be very compute expensive, plus the limitation I struggle to see around where if you're using a model to train another model you maybe can't get better than that model. But I think we could build architectures which provide large labelled training datasets to LLMs for any problem that can be deterministically solved by us using more traditional computing methods, like maths and some logic puzzles. Maybe if we use those datasets we can make it so that LLMs are able to do maths and difficult logic problems natively, and maybe the internal machinery they develop to do that could help them in other areas. Would be a fun research project.

by keenon 3y ago

bionhoward 3y ago

One way to push back on the data argument is to consider the progress DeepMind made with self play. Perhaps Bard can self-dialogue and achieve superhuman results. I won’t be surprised. Plus the underlying architecture is dense. Sparse transformers are a major upgrade. That’s only one of many upgrades you can make. There is still a lot of headroom and IMHO GPT-4 already implements AGI if you give it the right context

jxmorris12 3y ago
Self-play works for eg Go because there’s a perfect simulator of the game – which gives at least one very clear training signal, winning. There’s no simulator for conversation, no winning, and no training signal. Self-dialog doesn’t make any sense
- lucubratory 3y ago
  I could think of ways you could make a good attempt at it. You could have a generator/discriminator relationship where you have a model whose purpose is to evaluate model outputs for things other than just toxicity (basically RLHF for capabilities), then use that to train. You could have code generation tasks where the code is actually executed and a success/failure signal sent based on code performance. You could do logical puzzle generator/logical puzzle solver pairs and have a separate system evaluate answer correctness based on a human dataset baseline, or maybe a model programmed to be able to turn natural language logic puzzles into API calls to a formal logical engine solver to get the answer for comparison. You could make a simple program to turn randomly generated mathematical problems into word problems, use an AI to add extraneous detail while protecting the core problem description, then give the resulting word problems to an AI and use a separate AI to extract out the final answer or conclusion. Then compare that answer to what the calculator says for the original math problem. All of those have problems and would be very compute expensive, plus the limitation I struggle to see around where if you're using a model to train another model you maybe can't get better than that model. But I think we could build architectures which provide large labelled training datasets to LLMs for any problem that can be deterministically solved by us using more traditional computing methods, like maths and some logic puzzles. Maybe if we use those datasets we can make it so that LLMs are able to do maths and difficult logic problems natively, and maybe the internal machinery they develop to do that could help them in other areas. Would be a fun research project.