3 ms·

It would be great if some of these datasets were free and opened up for public use. Otherwise it seems like you end up duplicating a lot of busywork just for mu

It would be great if some of these datasets were free and opened up for public use. Otherwise it seems like you end up duplicating a lot of busywork just for multiple companies to farm more money. Maybe some of the European initiatives related to AI will end up including the creation of more open datasets. Then again, maybe we're still operating from a framework where the dataset is part of your moat. It seems like such a way of thinking will severely limit the sources of innovation to just a few big labs.

KaiserPro 1y ago
> operating from a framework where the dataset is part of your moat Very much this. Its the dataset that shapes the model, the model is a product of the dataset, rather than the other way around (mind you, synthetic datasets are different...)
andy_ppp 1y ago
Why would companies paying top dollar to refine and create high quality datasets give them away for free?
- flir 1y ago
  Same reason they give open source contributions away for free. Hardware companies attempting to commoditize their complement. I think the org best placed to get strategic advantage from releasing high quality data sets might be Nvidia.
  - mh- 1y ago
    Indeed, and they do. https://huggingface.co/nvidia/datasets?sort=most_rows https://huggingface.co/nvidia/datasets?sort=most_rows
- charlieyu1 1y ago
  There are some good datasets for free though, eg HLE. Although I’m sure if they are marketing gimmicks
- murukesh_s 1y ago
  Well that was the idea of "open"AI isn't it [1]? [1] https://web.archive.org/web/20190224031626/https://blog.openai.com/introducing-openai/ https://web.archive.org/web/20190224031626/https://blog.open...
  - delfinom 1y ago
    ClosedAI gonna ClosedAI
  - sumedh 1y ago
    Check the date. This was published before anyone knew it running an AI company would be very very expensive.
    - some_random 1y ago
      I feel like that was by far the most predictable part of running an AI company.
      - sumedh 1y ago
        Hindsight is 20/20.
- azemetre 1y ago
  Because we can make the government force them to.
  - andy_ppp 1y ago
    The US government is going to allow that right?
    - azemetre 1y ago
      If you vote for enough people of a certain party yes, we all currently see how quickly things can change in this country if those in power want change to happen.
gexla 1y ago
Right, and they pay a lot of money for this data. I know someone who does this, and one prompt evaluation could go through multiple rounds and reviews that could end up generating $150+ in payouts, and that's just what the workers receive. But that's not quite what the article is talking about. Each of these companies do things a bit different.
NitpickLawyer 1y ago
> Maybe some of the European initiatives related to AI will end up including the creation of more open datasets. The EU has started the process of opening discussions aiming to set the stage for opportunities to arise on facilitating talks looking forward to identify key strategies of initiating cooperation between member states that will enable vast and encompassing meetings generating avenues of reaching top level multi-lateral accords on passing legislation covering the process of processing processes while preparing for the moment when such processes will become processable in the process of processing such processes. #justeuthings :)
illegalmemory 1y ago
This could work with a Wikipedia-like model. It's very difficult to pull off, but a next-generation Wikipedia would look like this.
- yorwba 1y ago
  I think it would be difficult to make that work, because Wikipedia has a direct way of converting users into contributors: you see something wrong, you edit the article, it's not wrong anymore. Whereas if you do the same with machine learning training data, the influence is much more indirect and you may have to add a lot of data to fix one particular case, which is not very motivating.
ripped_britches 1y ago
Don’t worry - the labs will train based on this expert data and then everyone will just distill their models. Or, now that model itself can be an expert annotater.

KaiserPro 1y ago

> operating from a framework where the dataset is part of your moat Very much this. Its the dataset that shapes the model, the model is a product of the dataset, rather than the other way around (mind you, synthetic datasets are different...)

andy_ppp 1y ago

Why would companies paying top dollar to refine and create high quality datasets give them away for free?

flir 1y ago
Same reason they give open source contributions away for free. Hardware companies attempting to commoditize their complement. I think the org best placed to get strategic advantage from releasing high quality data sets might be Nvidia.
- mh- 1y ago
  Indeed, and they do. https://huggingface.co/nvidia/datasets?sort=most_rows https://huggingface.co/nvidia/datasets?sort=most_rows
charlieyu1 1y ago
There are some good datasets for free though, eg HLE. Although I’m sure if they are marketing gimmicks
murukesh_s 1y ago
Well that was the idea of "open"AI isn't it [1]? [1] https://web.archive.org/web/20190224031626/https://blog.openai.com/introducing-openai/ https://web.archive.org/web/20190224031626/https://blog.open...
- delfinom 1y ago
  ClosedAI gonna ClosedAI
- sumedh 1y ago
  Check the date. This was published before anyone knew it running an AI company would be very very expensive.
  - some_random 1y ago
    I feel like that was by far the most predictable part of running an AI company.
    - sumedh 1y ago
      Hindsight is 20/20.
azemetre 1y ago
Because we can make the government force them to.
- andy_ppp 1y ago
  The US government is going to allow that right?
  - azemetre 1y ago
    If you vote for enough people of a certain party yes, we all currently see how quickly things can change in this country if those in power want change to happen.

gexla 1y ago

Right, and they pay a lot of money for this data. I know someone who does this, and one prompt evaluation could go through multiple rounds and reviews that could end up generating $150+ in payouts, and that's just what the workers receive. But that's not quite what the article is talking about. Each of these companies do things a bit different.

NitpickLawyer 1y ago

> Maybe some of the European initiatives related to AI will end up including the creation of more open datasets. The EU has started the process of opening discussions aiming to set the stage for opportunities to arise on facilitating talks looking forward to identify key strategies of initiating cooperation between member states that will enable vast and encompassing meetings generating avenues of reaching top level multi-lateral accords on passing legislation covering the process of processing processes while preparing for the moment when such processes will become processable in the process of processing such processes. #justeuthings :)

illegalmemory 1y ago

This could work with a Wikipedia-like model. It's very difficult to pull off, but a next-generation Wikipedia would look like this.

yorwba 1y ago
I think it would be difficult to make that work, because Wikipedia has a direct way of converting users into contributors: you see something wrong, you edit the article, it's not wrong anymore. Whereas if you do the same with machine learning training data, the influence is much more indirect and you may have to add a lot of data to fix one particular case, which is not very motivating.

ripped_britches 1y ago

Don’t worry - the labs will train based on this expert data and then everyone will just distill their models. Or, now that model itself can be an expert annotater.