Implement reflection

[mijhaels]

Context: Implement reflection, a technique that allows generating more coherent and natural texts using pre-trained language models.

Problem or idea: Reflection is based on two articles that propose different methods to incorporate world knowledge and causal reasoning in text generation. The articles are:

1. ArXiv Article
2. GitHub Repository

Solution or next step: I would like the Auto-GPT project to include reflection as an option to improve the quality of the generated texts. @Torantulino, what do you think of this idea?

[Torantulino]

That it VERY interesting indeed. I've been playing about with getting Auto-GPT to improve itself and write it's own code, this might just be the ticket!

Executing unreviewed AI Generated code is a security risk though, so we'd have to think of the safest way to do this.

Submit a pull request!

[mijhaels]

I code but only know about web programming, and I'm a junior too... sorry Haha

[Torantulino]

No worries at all, and never let being a junior stop you, the absolute best way to learn is by trying and failing!

[NotoriousPyro]

Executing unreviewed AI Generated code is a security risk though, so we'd have to think of the safest way to do this.

Probably best to have it raise pull requests so that they can be reviewed manually.

[algopapi]

I kinda implemented this. In my implementation I let the agent reflect every N steps which can be quite interesting but requires more testing.

[Bearnardd]

Around weekend I will do a little bit of research around this topic. @algopapi would you mind sharing your implementation? Have you pushed it on your fork?

[LeonardoLGDS]

Also, probably will have to create some kind of benchmarks so that the AI can know the direction it needs to go. "Improve itself" can be a bit vague

[Bearnardd]

@LeonardoLGDS I had no time to check the paper yet but I assumed there are some guidelines on benchmarking provided in paper

[aristotaloss]

If anyone can get me a GPT4 key, either directly via your org with prepayment or by getting me in touch with someone at OpenAI, I'm willing to implement it. I think it could be a lot of fun, and to add an edge, I'll do it on a fresh Linux w/ firewalled access to only the OpenAI API, and a USB chainsaw.

Jk, not the chainsaw.

[younghuman]

I'm planning to implement this over the weekend. Another best paper for self-reflection is this: https://arxiv.org/abs/2304.03442

They also have a way of evaluating the agents with some questions.

@Torantulino @Andythem23

[Boostrix]

That it VERY interesting indeed. I've been playing about with getting Auto-GPT to improve itself and write it's own code, this might just be the ticket!

I guess, a starting point would be accepting actual constraints/restrictions, aka:

• context window size (restrict to 8k, ideally much less, probably 50% of that)
• restrict changes to a single isolated module (which would mean commands or even better plugins)
• accept that the underlying architecture would then need to work analogous to message passing, and possibly via pipes - anything else won't scale or would change too many places in the source tree at once, basically any agent would consist of a list of other sub-agents, all of which would fork each other as needed (think makefiles/clustering)
• alternatively, come up with a module that can deal with patches/diffs for features spanning multiple files/contexts initially
• to provide sufficient surrounding context, freely use Python docstrings - basically reject PRs that don't have a ton of surrounding context in the form of comments and a ton of unit tests/test coverage
• extend git_operations.py to add support for traversing commit logs (patches + log messages)
• come up with a new plugin to handle github API integration, as per: Auto-GPT Recursive Self Improvement #15 (comment)
• have unit tests and benchmarks for regression testing purposes
• improve docker integration for CI
• use a custom local LLM to come up with a heatmap (query/idea -> location) that helps identify path/file name and line number based on querying the LLM for an idea/change, so that the agent can determine what set of files/places is likely to be relevant for certain changes, and then narrow down via git logs and commit history

[github-actions]

This issue was closed automatically because it has been stale for 10 days with no activity.