[avx56]

It's not much of a competitor with this performance. From the OpenChatKit model they claim to be using:

> Write a rhyming poem about ray tracing

OpenChatKit: The sun is shining

The birds are singing, and the flowers are blooming. I'm so glad that I'm alive, and I'm so glad that I'm here. I'm so glad that I'm alive, and I'm so glad that I'm here. I'm so glad that I'm alive, and I'm so glad that I'm here. [repeated several more times]

ChatGPT: In a world of graphics, sharp and bright, Where realism is the ultimate sight, There's a technique that takes the lead, It's called ray tracing, indeed!

Ray tracing is a visual art, That sets the scene right from the start, By tracing rays of light in the air, It creates an image beyond compare.

[DanyWin]

To me OpenChatKit is just a first step towards better and better open-source models. Other actors like AWS and Hugging Face are also working on that and Hugging Face has already proved its ability to train and make available LLMs on a huge scale like Bloom.

I think it's just the beginning and the open-source community will provide very competitive LLMs.

[rvz]

> I think it's just the beginning and the open-source community will provide very competitive LLMs.

It is indeed. The way to challenge OpenAI's offerings is with open-source AI models. Even better when they achieve and surpass GPT-4 level capabilities.

They (OpenAI) cannot win the race to the bottom or $0. Stable Diffusion (and even the leaked Facebook LLaMa) is already at the finish line and more alternatives will also be there to surpass GPT-3 and 4 and will release them for free in the open.

Eventually, Open source AI models will eventually disrupt closed ones. Just like how DALLE-2 has been disrupted quickly by Stable Diffusion.

[rlt]

I think one caveat is access to training data. If proprietary models can be trained on useful data from private sources, or worse, if there are successful legal challenges against using public but copyrighted data for training, then it will be difficult for open-source models to compete with proprietary models.

[bootsmann]

1) In a lot of western countries (EU, UK too I think) the hammer has already come down in favor of using public but copyrighted data.

2) Wouldn’t that cause open source models to be favored? A big company has lawyers that ensure that the internal practices comply with the law while on the other hand, good luck suing some random guy from 4chan who made a model that may or may not incorporate copyrighted data.

[rlt]

Sure there will be "bootleg" models available on BitTorrent or whatever, but generally "open source" refers to legitimately licensed code a big company with lawyers would be ok with incorporating into their own business.

[KyeRussell]

Your incorrect assertions about the current state of play was what turned this from a bit of a hypothesis to “open source always wins” delusion.

[dragonwriter]

It actually wasn’t that long ago that GPT based models (including Copilot) frequently got stuck in loops like that; I'm... not taking that as a big negative sign. Yeajz its behind the closed SOTA. That’s where you expect an open competitor to start.

[claar]

ChatGPT v4:

  In pixeled realms where photons race,
  A wondrous dance of light's embrace,
  Ray tracing weaves its subtle art,
  Illuminating shadows' heart.
  With every bounce and each reflection,
  It births a scene of pure perfection.

  A cascade of color, depth, and tone,
  In virtual worlds, it carves its throne,
  From mirrored glass to shimmering streams,
  Ray tracing paints our digital dreams.
  A masterstroke in the canvas vast,
  Its legacy through ages cast.

[xvector]

Legitimately better than what 99% of humans could write, and done in seconds at that. I'm amazed.