Hide table of contents

(Podcast episode in link)

I had a discussion with Owain Evans about many topics around language models. We touch on his work with Truthful AI and Ought. I spent a lot of time asking how LLMS might be helpful in forecasting and where they could make the most economic impact.

I’ve known Owain for the past few years and enjoy his perspective. Before working on AI Safety, he studied cognitive science and probabilistic programming, so I think he’s well-positioned to understand the relationships between forecasting infrastructure and AI.

This is more of a research discussion than a formally planned podcast. I want to get in the habit of recording more conversations. At the same time, I think some of these threads will interest listeners.

This recording happened in late January 2023, before GPT4 or the ChatGPT Plugin System. We don’t have a full transcript (these are expensive, particularly when there is a lot of jargon), but the section titles below should summarize what we discussed.

Many thanks to Owain for participating.

Sections

00:29 AI Safety Research: Truthful AI
05:41 Challenges of Benchmarking AI Models for Truthfulness and Reliability
11:26 Use of AI for Forecasting Questions and Probabilities
14:36 Automating Prompts and Forecasting Techniques with Language Models
18:00 Exploring the Potential of Evaluation vs Generation in Improving Text Generation Models
20:44 Improving ChatGPT through User Feedback and Reinforcement Learning
24:41 Current State and Future Potential of OpenAI's AI Models and Products
26:48 Ought's Current Strategy
30:25 Trade-offs between Process-Based and Outcome-Based Systems for AI Safety and Transparency
35:14 Importance of Transparency in AI Research and its Future Implications
37:01 Integrating Large Language Models with Existing Software
44:48 Programming Assistance and Automation in Research
51:00 Exploring the Possibilities of Combining AI and Personal Assistant Services

Mentioned Papers

Iterated Decomposition: Improving Science Q&A by Supervising Reasoning Processes
Truthful AI: Developing and governing AI that does not lie
Forecasting Future World Events with Neural Networks

21

0
0

Reactions

0
0
Comments


No comments on this post yet.
Be the first to respond.
More from Ozzie Gooen
83
Ozzie Gooen
· · 9m read
Curated and popular this week
 ·  · 5m read
 · 
This work has come out of my Undergraduate dissertation. I haven't shared or discussed these results much before putting this up.  Message me if you'd like the code :) Edit: 16th April. After helpful comments, especially from Geoffrey, I now believe this method only identifies shifts in the happiness scale (not stretches). Have edited to make this clearer. TLDR * Life satisfaction (LS) appears flat over time, despite massive economic growth — the “Easterlin Paradox.” * Some argue that happiness is rising, but we’re reporting it more conservatively — a phenomenon called rescaling. * I test rescaling using long-run German panel data, looking at whether the association between reported happiness and three “get-me-out-of-here” actions (divorce, job resignation, and hospitalisation) changes over time. * If people are getting happier (and rescaling is occuring) the probability of these actions should become less linked to reported LS — but they don’t. * I find little evidence of rescaling. We should probably take self-reported happiness scores at face value. 1. Background: The Happiness Paradox Humans today live longer, richer, and healthier lives in history — yet we seem no seem for it. Self-reported life satisfaction (LS), usually measured on a 0–10 scale, has remained remarkably flatover the last few decades, even in countries like Germany, the UK, China, and India that have experienced huge GDP growth. As Michael Plant has written, the empirical evidence for this is fairly strong. This is the Easterlin Paradox. It is a paradox, because at a point in time, income is strongly linked to happiness, as I've written on the forum before. This should feel uncomfortable for anyone who believes that economic progress should make lives better — including (me) and others in the EA/Progress Studies worlds. Assuming agree on the empirical facts (i.e., self-reported happiness isn't increasing), there are a few potential explanations: * Hedonic adaptation: as life gets
 ·  · 38m read
 · 
In recent months, the CEOs of leading AI companies have grown increasingly confident about rapid progress: * OpenAI's Sam Altman: Shifted from saying in November "the rate of progress continues" to declaring in January "we are now confident we know how to build AGI" * Anthropic's Dario Amodei: Stated in January "I'm more confident than I've ever been that we're close to powerful capabilities... in the next 2-3 years" * Google DeepMind's Demis Hassabis: Changed from "as soon as 10 years" in autumn to "probably three to five years away" by January. What explains the shift? Is it just hype? Or could we really have Artificial General Intelligence (AGI)[1] by 2028? In this article, I look at what's driven recent progress, estimate how far those drivers can continue, and explain why they're likely to continue for at least four more years. In particular, while in 2024 progress in LLM chatbots seemed to slow, a new approach started to work: teaching the models to reason using reinforcement learning. In just a year, this let them surpass human PhDs at answering difficult scientific reasoning questions, and achieve expert-level performance on one-hour coding tasks. We don't know how capable AGI will become, but extrapolating the recent rate of progress suggests that, by 2028, we could reach AI models with beyond-human reasoning abilities, expert-level knowledge in every domain, and that can autonomously complete multi-week projects, and progress would likely continue from there.  On this set of software engineering & computer use tasks, in 2020 AI was only able to do tasks that would typically take a human expert a couple of seconds. By 2024, that had risen to almost an hour. If the trend continues, by 2028 it'll reach several weeks.  No longer mere chatbots, these 'agent' models might soon satisfy many people's definitions of AGI — roughly, AI systems that match human performance at most knowledge work (see definition in footnote). This means that, while the compa
 ·  · 4m read
 · 
SUMMARY:  ALLFED is launching an emergency appeal on the EA Forum due to a serious funding shortfall. Without new support, ALLFED will be forced to cut half our budget in the coming months, drastically reducing our capacity to help build global food system resilience for catastrophic scenarios like nuclear winter, a severe pandemic, or infrastructure breakdown. ALLFED is seeking $800,000 over the course of 2025 to sustain its team, continue policy-relevant research, and move forward with pilot projects that could save lives in a catastrophe. As funding priorities shift toward AI safety, we believe resilient food solutions remain a highly cost-effective way to protect the future. If you’re able to support or share this appeal, please visit allfed.info/donate. Donate to ALLFED FULL ARTICLE: I (David Denkenberger) am writing alongside two of my team-mates, as ALLFED’s co-founder, to ask for your support. This is the first time in Alliance to Feed the Earth in Disaster’s (ALLFED’s) 8 year existence that we have reached out on the EA Forum with a direct funding appeal outside of Marginal Funding Week/our annual updates. I am doing so because ALLFED’s funding situation is serious, and because so much of ALLFED’s progress to date has been made possible through the support, feedback, and collaboration of the EA community.  Read our funding appeal At ALLFED, we are deeply grateful to all our supporters, including the Survival and Flourishing Fund, which has provided the majority of our funding for years. At the end of 2024, we learned we would be receiving far less support than expected due to a shift in SFF’s strategic priorities toward AI safety. Without additional funding, ALLFED will need to shrink. I believe the marginal cost effectiveness for improving the future and saving lives of resilience is competitive with AI Safety, even if timelines are short, because of potential AI-induced catastrophes. That is why we are asking people to donate to this emergency appeal
Recent opportunities in AI safety
20
Eva
· · 1m read