[MikeB]

This is a Transformer Neural Network model where the same network & weights can play Atari, caption images, chat, stack blocks with a real robot arm and more, depending on context.

https://www.deepmind.com/publications/a-generalist-agent

It scores well in most tests, however it still suffers from the same problems as other Blackbox ML models, acknowledged in the PDF (Model Card/Appendix A).

• Organisation: DeepMind
• Model Date: May 2022

...

• Data: The vision and language datasets used include racist, sexist, and otherwise harmful context
• Risks and Harms: In addition to the potential harms of toxic image and language training data, Gato’s real world embodiment introduces physical safety harms due to misuse or malfunctioning.
• Mitigations: No mitigation of bias introduced by vision and language data beyond the filtering of sexually explicit content, as in Alayrac et al. (2022). Physical risk is mitigated through safety measures implemented by robotics environment designers.
• Future work: The interaction of diverse training data domains and the different affordances faced in evaluation is poorly understood, and potential ethical and safety risks arise as the generalist’s capabilities grow.

Conclusion:
Transformer sequence models are effective as multi-task multi-embodiment policies, including for real-world text, vision and robotics tasks. They show promise as well in few-shot out-of-distribution task learning. In the future, such models could be used as a default starting point via prompting or fine-tuning to learn new behaviors, rather than training from scratch.

Given scaling law trends, the performance across all tasks including dialogue will increase with scale in parameters, data and compute. Better hardware and network architectures will allow training bigger models while maintaining real-time robot control capability. By scaling up and iterating on this same basic approach, we can build a useful general-purpose agent.

[ivan.moony]

If I understood correctly, in addition to training what to do, this AI is also trained on what *not* to do. If this is a case, I think the idea is very good.

[MikeB]

The datasets they like to use contain everything...

There's no set structure or Whitebox portion. Symbols can't be extracted and understood either before output. If anything, a second, blackbox AGI could check output for sensitivity purposes.