I am working with Nuño Sempere on a project to extract latent probabilities from GPT-3.
Primary outcomes:
improve on the state of the art in anti-hallucination and truthful question answering using LLMs.
measure information retrieval + architecture tweaks vs crowd performance on prediction markets.
elicit explanations for the reasoning behind the model's decisions, both directly and indirectly
Our team is currently #7/#61 on the Autocast Competition (forecasting.mlsafety.org). We're prioritizing understandable, legible, and safe behavior above optimizing for capabilities.
Nuño is an expert on forecasting at the Quantified Research Uncertainty Institute. He is the author of forecasting.substack.com, was a summer fellow at FHI, created the "Estimated Value" sequence, made metaforecast.org, and is a founding member of samotsvety.org.
I've been at Microsoft for ~3 years, have a bit of experience with LLMs, did 5 internships, won multiple awards in international competitions (including a $35k prize in the HITB AI Challenge), was invited to speak at an IEEE conference, got into Stanford, and met Geoff Hinton once.
Paying rent for experimentation and testing on cloud GPUs. Only so much you can do with APIs.
We'll apply for a second round of funding to scale up our approach if initial results are promising.
Updates:
• The Autocast Competition (mlsafety.org) was closed due to the FTX collapse, so we decided to scrap the paper and reorient towards eventually selling the project to Anthropic instead.
• No outputs on the development side in the last two weeks because I needed a break after pushing to wrap up work prior to my vacation and continuous exhaustion isn't sustainable.
• Applied to SERI MATS to get more time to work on this, got an informal accept from the mentor we targeted, but waiting for official decisions to be out.
@Austin thanks! Quick answers:
Deliverables: We'll open source our methods, code, models, data, animations, and any additional information needed to reproduce the experimental results. We aim to submit a paper to NeurIPS 2023 within the next 8-9 weeks. Public release date is currently 14 weeks from now.
Commitment: I am taking 4 weeks off (starting late April) to focus primarily on this project. As far as when to scale: it's hard to give a firm date since the field moves so fast, but this is really a function of how much we raise. Some parts of our architecture are scale invariant, others plug into publicly available LLMs, and some components of the system are traditional software. On the margin, dollars spent on inference and evaluation (for e.g ablation studies/prompt testing) are more useful than dollars spent on training, at least until you get pretty far down the list of ideas. We'll make the decision to scale when we think it's a good idea, and we don't yet know precisely when that will be.
Hi Sheikh! This seems like a neat project - it's awesome to hear that Nuno is involved here too. A couple questions that might help investors evaluating this:
What are the deliverables if experimentation goes well -- eg published paper? Blog post? Interactive website?
Roughly how much time do you and Nuno expect to put into this before deciding whether to scale up?
@aaronl yep! Mostly along the lines of [2206.15474] Forecasting Future World Events with Neural Networks (arxiv.org)
I'm curious to learn more about the second primary outcome, "measure information retrieval + architecture tweaks vs crowd performance on prediction markets". This sounds like the main tie-in to forecasting. Is the idea to predict the probability of an event using GPT-3 (either by asking directly or extracting probabilities in a lower-level way) and compare the accuracy of these predictions to prediction markets?