Hi. I’m interested in identifying a base-model AI suited for ethics self-training. What base model(s) would you recommend? What abilities or packages would you recommend?
A critical end goal for this project would be reasonably accurate assessment of existential risk, so scanning current news would be important. I’d like to be able to get it to read current news and output potential ethical and existential risks as a major ultimate feature.
Yes, this is ambitious. No, I don’t have significant coding ability. Open to collaboration. I’m a hypnotherapist if that makes things any more interesting.
We are a software development company specializing in delivering tailored solutions to meet your needs. We’d love the opportunity to collaborate with you. Feel free to reach out to me at jaygupta at liquifysolutions.in
First of all, I have no idea how well existing models are suited to this.
One possible approach would be to find an LLM that excels at similar tasks on the leaderboards, and then write a simple piece of code in smolagents to give the model a search function.
Hi kfoelsch,
Regarding models, we’re in an era where there should be plenty of LLM models that are suited for text based analysis. Generally what would be needed is a model that has good general world knowledge and with a carefully crafted prompt can focus on the task of analysing the text content you provide it and report accordinly.
Are you looking for paid API cloud based models or locally hosted models? For text based analysis, I’ve been able to get away with using mistral-nemo:12b but have even had decent results with as small as llama3.2.
Playing around with various locally hosted models could be a good start. Though I think the majority of work involved will be the software development for each step in the process. Then you could swap out various models to find what suits your needs best.
The processing steps would involve something like this:
Web Search API - to process your web search query of recent world events.
An LLM to score and filter the search results so you are left with the ones most likley relevant.
Web scraping API or locally. If you want to do it locally, Playwright does a pretty decent job. This will fetch the pages from relevant search results.
An LLM to score and filter scraped pages.
An LLM to analyse the scraped pages to analyse news and media content for ethical and existential risks.
Generate a report on the findings.
I wrote an application with the above pipeline recently for automatically finding leads. Hence I talk from experience on how I would go about building a research & analysis agent like this.