Carolyn Meinel's Research Project

BestWorld: Aiding AI Alignment with a Human/AI System that Generates True, Believable, and Beneficial Information

AI alignment can be considered as the combination of subsets of the disciplines of computer security, quality control, and forecasting of alternate conditional outcomes such as those my colleagues and I experimented with under IARPA's FOCUS program. I have experience and in many cases refereed research papers on these topics. I am viewing AI Alignment through the lenses of these disciplines.

Bottom line: Hybrid Human/Many-Federated-LLMs Systems could play a role in AI Alignment

In part for this project, four colleagues and I, as team leader, are building a human/AI system that:

• Creates knowledge that is true, believable, and can enable us to create a better world for all.
• Pairing forecasts with news stories can build believability, but only if we are good at forecasting.
• My colleagues and I have considerable evidence that the important factors in high accuracy forecasting include:
  • Use of historical data either directly relevant (time series) or comparisons (lessons learned).
  • Integrative complexity.
  • High quality forecasters as humans-in-the-loop to create a pipeline to high quality analyses.
  • We can use semantic analysis to extract this informatio, as embedded in their rationales, into reports and a newsfeed.
• In general, from where will factors of AI governance emanate? Over time, we expect to get a better handle on this.
• All these factors are essential for AI alignment/governance. Why?
• This also has commercial potential for a hybrid news/social media/forecasting system.
• Thus it is likely to become self sustaining without continued donations.

Our planned knowledge product:

• A hybrid news/social media/forecasting system integrated via, in part, the evolving Activitypub protocol:
• Its central news generation activity: hybrid human/AI forecasting, producing predictions and rationales, fed into:
• Newsletters
• Websites
• Live online and in person events
• Videos
• Social media
• other?

We believe that a hybrid human/AI system such as what we are preparing to test:

• ... could enable peace and prosperity and avert most existential risks. We may as well think big!
• Flowchart here.
• A less ambitious flowchart here, from one of our GFC2 hybrid forecasting models.
• Spreadsheets that partially substantiate the BestWorld opportunity.

Inspiration for this research includes my participation in:

• The Existential Risk Persuasion Tournament
• The AI Safety Fundamentals course
• My use of Semantic Search and Rosette in IARPA's GFC2
• Collaboration with colleagues to write this research paper (I'm second author)
• Eugene Reyes' and Bruce Lawhorn's collaboration with me, see our webinar here
• Eight years of replicability experiments
• And especially our team at BestWorld!

So far, we have:

• Developed a basic plan.
• April 2024, launched The BestWorld, Inc. as the nonprofit entity to pursue this research.
• Recruited a full C-Suite so the project can grow (see an ambitious flowchart here).
• Raised enough money to cover computing costs of our first hybrid human/AI project.
• Reanalyzed my Semantic Search results from IARPA GFC2. See here and here for details.
• Evaluated in depth the AI/federated-LLM provider Linguistic Systems, Inc.
• I foolishly signed the NDA Linguistic Systems required before I could see the live system:
• Architecture of federated LLMs with multiple feedback loops including humans.
• It also is capable of extreme accuracy, no hallucinations. Details here --->
• It incorporates Semantic Search, which I had beta tested 2017 -- 2020.
• Provided a live demonstration of its Ai Translate system.
• They gave me a report with screenshots from this live demonstration...
• But marked it "Confidential" and so far has refused my request to post it publicly.
• I'm hoping they will change their minds. Sigh. Stay tuned. But we have other options.

Plans B, C, and D we have been evaluating for proof of concept experiments:

• Cultivate Labs, a subcontractor to INFER.
  • It is using ANTHROP\C as a forecasting aid there, details here.
  • Its forecasting platform is excellent. I know this because...
• I'm a "Pro Forecaster" at INFER, just a few hours/week.
• Good Judgment, Inc. which is using ChatGPT as a summarization aid.
• A BestWorld homebrew system using open source elements, under development. (I'm excited!)
• We have evaluated the Quorum Research forecasting platform -- details here-->

We launched a BestWorld newsletter, with two more newsletters in development.

• Developed specialized email lists of
• professional forecasters
• epidemiologists
• AI alignment professionals.

Data sources: Evaluated and tested some for both newsletters and hybrid forecasting:

• Built spreadsheets of Sahelian and Georgian local news sources.
• Evaluated Sahelian news sites for scraping, looks easy!
• Yes, I know scraping is rude. Some sites will impose permanent blocks on scrapers.
• Many news sites are now also blocking search engine spiders, lawsuits underway.
• Later, we may need large contributions to pay for data APIs, e.g. X, Facebook.

Opened a bank account for our nonprofit.

Performed reanalyses of my prior AI/LLM research:

• Studied recent research on problems with MTurk workers, which informed my reanalysis of...
• LLM detection of logical constructions in rationales by MTurkers provided by IARPA's GFC2.
• Detection of semantic signatures of accurate forecasting using those MTurk rationales.

Reviewed Christopher Karvetski's and my 2019 hybrid human/AI system

Pending tasks:

• Wire transfer funds into our nonprofit bank account from those who have pledged.
• AI Alignment Jobs newsletter? should we do it?--->
• Run a live test of what we hope will be an improved human/AI forecasting system.

So now I'm the stage of ... Move slowly and don't break things... until I get feedback on this report from BlueDot judges and my teammates at BestWorld..

P.S. In case you are wondering who I am and why I'm here --->

Site map

© 2024 Carolyn Meinel. All rights reserved.