6 ms·

Overall probably a good move for Google. They were were spending 100s of millions of dollars a year to train RL agents to get good at playing Qwop for close to

Overall probably a good move for Google. They were were spending 100s of millions of dollars a year to train RL agents to get good at playing Qwop for close to a decade.

29athrowaway 2y ago
Or chess, go, shogi, arimaa, StarCraft II, etc.
- summerlight 2y ago
  In a PR perspective, at least AlphaGo did a prominent job for Google. It failed to keep its prestigious status but at least it's a leadership/product side issue, not DeepMind's.
  - hot_gril 2y ago
    That stuff gave Google the appearance of "nerds playing with their toys." Sure it's worthwhile research, but it's not great how many things Google announced that close to 0 people could use. They're trying to be more user-oriented now, and because of Cloud, their users are businesses.
    - FactKnower69 2y ago
      No. AlphaGo et al gave the appearance of "world class researchers achieving state of the art on problem classes that humans have been devoting lifetimes of study for thousands of years". the current desperate LLM scramble, including stuffing shitty, hallucinating, half-baked paraphrase spam to the top of every Google search result page, gives the appearance of "panicked flailing and tacit admission that high level decisionmaking has been captured by terrified MBA types"
      - hot_gril 2y ago
        There's a good long middle ground. Search results do seem desperate now, but they got into this position in the first place by failing to execute when they had the advantage. None of Google's customers have ever cared how well an AI can play Go. If there's one place the bragging could've translated to profits, it'd be Tensorflow dominance, which didn't happen.
    - 29athrowaway 2y ago
      That research has many applications besides just board games. Game theory is everywhere. Also, there were serious mathematicians saying a few years prior that computers would never surpass humans in Go, at least during our lifetimes.
      - hot_gril 2y ago
        Yes it does, but that didn't translate to products, so again their customers don't care.
thoi242344rerwe 2y ago
[flagged]
aprilthird2021 2y ago
I strongly disagree. Research is how you stay ahead of your competitors, yes even silly sounding research like training RL agents to play Qwop! Research is how Google invented the transformer that underlies so many current Gen AI models. Pivoting research to products is exactly the kind of short term thinking consultants or mercenaries would propose, get promoted off of, and leave just before the consequences start rearing their head
- hot_gril 2y ago
  They did research for years, but they didn't stay ahead of their competitors. They gave transformers away for free.
  - aprilthird2021 2y ago
    They did stay ahead of their competitors. In a lot of ways they still are. In a "making profit from AI" way, they're not. But that's what you hire product people for. They had the SOTA translation system of the entire world for a literal decade and didn't monetize it enough to blow their competition to dust. That's not the fault of researchers, and research shouldn't halt or be redirected to profit as a result. Rather, Google should get some better managers / product people to turn the magic their researchers create into profits.
    - hot_gril 2y ago
      They aren't ahead even in terms of AI tech now. A lot of those researchers got hired away to other companies, and ChatGPT has been ahead continuously.
      - aprilthird2021 2y ago
        They are ahead in terms of AI infra and capacity. Research it's not as easy to tell. They are behind in terms of shipping products. Yes a lot of researchers got hired away, but they still up there in the race
FactKnower69 2y ago
RNN gameplay agents and reinforcement learning are a thousand times more interesting than stochastic parrots that rephrase Google for you, you'll see once this hype wave finishes dying out in a couple years

Overall probably a good move for Google. They were were spending 100s of millions of dollars a year to train RL agents to get good at playing Qwop for close to a decade.

29athrowaway 2y ago

Or chess, go, shogi, arimaa, StarCraft II, etc.

summerlight 2y ago
In a PR perspective, at least AlphaGo did a prominent job for Google. It failed to keep its prestigious status but at least it's a leadership/product side issue, not DeepMind's.
- hot_gril 2y ago
  That stuff gave Google the appearance of "nerds playing with their toys." Sure it's worthwhile research, but it's not great how many things Google announced that close to 0 people could use. They're trying to be more user-oriented now, and because of Cloud, their users are businesses.
  - FactKnower69 2y ago
    No. AlphaGo et al gave the appearance of "world class researchers achieving state of the art on problem classes that humans have been devoting lifetimes of study for thousands of years". the current desperate LLM scramble, including stuffing shitty, hallucinating, half-baked paraphrase spam to the top of every Google search result page, gives the appearance of "panicked flailing and tacit admission that high level decisionmaking has been captured by terrified MBA types"
    - hot_gril 2y ago
      There's a good long middle ground. Search results do seem desperate now, but they got into this position in the first place by failing to execute when they had the advantage. None of Google's customers have ever cared how well an AI can play Go. If there's one place the bragging could've translated to profits, it'd be Tensorflow dominance, which didn't happen.
  - 29athrowaway 2y ago
    That research has many applications besides just board games. Game theory is everywhere. Also, there were serious mathematicians saying a few years prior that computers would never surpass humans in Go, at least during our lifetimes.
    - hot_gril 2y ago
      Yes it does, but that didn't translate to products, so again their customers don't care.

thoi242344rerwe 2y ago

[flagged]

aprilthird2021 2y ago

I strongly disagree. Research is how you stay ahead of your competitors, yes even silly sounding research like training RL agents to play Qwop! Research is how Google invented the transformer that underlies so many current Gen AI models. Pivoting research to products is exactly the kind of short term thinking consultants or mercenaries would propose, get promoted off of, and leave just before the consequences start rearing their head

hot_gril 2y ago
They did research for years, but they didn't stay ahead of their competitors. They gave transformers away for free.
- aprilthird2021 2y ago
  They did stay ahead of their competitors. In a lot of ways they still are. In a "making profit from AI" way, they're not. But that's what you hire product people for. They had the SOTA translation system of the entire world for a literal decade and didn't monetize it enough to blow their competition to dust. That's not the fault of researchers, and research shouldn't halt or be redirected to profit as a result. Rather, Google should get some better managers / product people to turn the magic their researchers create into profits.
  - hot_gril 2y ago
    They aren't ahead even in terms of AI tech now. A lot of those researchers got hired away to other companies, and ChatGPT has been ahead continuously.
    - aprilthird2021 2y ago
      They are ahead in terms of AI infra and capacity. Research it's not as easy to tell. They are behind in terms of shipping products. Yes a lot of researchers got hired away, but they still up there in the race

FactKnower69 2y ago

RNN gameplay agents and reinforcement learning are a thousand times more interesting than stochastic parrots that rephrase Google for you, you'll see once this hype wave finishes dying out in a couple years