Comments
Trying this out for using for various Unjournal.org processes (like prioritizing research, finding potential evaluators, linking research to pivotal questions) and projects (assessing LLM vs human research evaluations). Some initial forays (comming from a conversation with Xyra). I still need to human-check it.
~prompt to Claude code about @Toby_Ord and How Well Does RL Scale?
``Toby Ord's writing -- what do the clusters look like? What other research/experts come closest to his post .... https://forum.effectivealtruism.org/posts/TysuCdgwDnQjH3LyY/how-well-does-rl-scale``
This interactive visual (with some extra prompts)
Key Takeaways
- For “How Well Does RL Scale?”, the closest intellectual neighbors are:
- ML researchers focused on compute efficiency and scaling laws (e.g. 1a3orn, Pablo Villalobos), and
- AI safety researchers analyzing inference scaling implications (e.g. Ryan Kidd, ryan_greenblatt).
- Stuart Armstrong appears to be the only researcher in this corpus writing substantively across all three of Toby Ord’s main intellectual themes:
- AI scaling
- Existential risk / longtermism
- Infinite ethics and decision theory
Most Similar Researchers to“How Well Does RL Scale?”
Top Authors (by Semantic Similarity)
| Author | Distance | Key Work |
|---|---|---|
| 1a3orn | 0.143 | Parameter Scaling Comes for RL, Maybe; New Scaling Laws for LLMs |
| Pablo Villalobos | 0.150 | Trading off compute in training and inference |
| Matrice Jacobine | 0.168 | Does RL Really Incentivize Reasoning Capacity in LLMs? |
| Lukas Finnveden | 0.169 | Before smart AI, there will be many mediocre or specialized AIs |
| ryan_greenblatt | 0.179 | What’s going on with AI progress and trends? |
| Ryan Kidd | 0.188 | Implications of the inference scaling paradigm for AI safety |
@1a3orn @Pablo Villalobos @Matrice Jacobine🔸🏳️⚧️ @Lukas Finnveden @Ryan Greenblatt @Ryan Kidd -- if you have a chance, let me know if this is accurate/relevant.
Most Similar arXiv Paper
| Paper | Authors | Distance |
|---|---|---|
| The Art of Scaling RL Compute for LLMs | Khatri et al. | 0.125 |
| Webscale-RL: Automated Data Pipeline for Scaling RL | Cen et al. | 0.142 |
| AReaL: Large-Scale Asynchronous RL for Language Reasoning | Fu et al. | 0.142 |
| Does RLHF Scale? | Hou et al. | 0.162 |
Toby Ord’s Writing Clusters
Based on semantic analysis, Toby Ord’s writing appears to fall into three main thematic clusters.
Cluster 1: AI Scaling & Timelines (Recent)
- “How Well Does RL Scale?” (Oct 2025)
- Core claim: RL training scales poorly; most gains come from inference scaling (longer chains of thought), not training compute.
Key implication: This likely lengthens AI timelines and affects governance and safety strategies.
Cluster 2: Existential Risk & Longtermism
- The Precipice and related work
- “The Precipice Revisited” (Jul 2024)
- “Shaping Humanity’s Longterm Trajectory” (Jul 2023)
- “Robust Longterm Comparisons” (May 2024)
Topics include:
- Existential risk estimation
- Discounting
- Anthropic effects / anthropic shadow
Cluster 3: Infinite Ethics & Decision Theory
- “Evaluating the Infinite” (Sep 2025) — hyperreal-based approach to divergent sums and integrals
- “What’s wrong with the dogmatic solution to recklessness?” (Feb 2023)
- “The Ethics of Copying Conscious States” (Aug 2025)
Topics include:
- Pascal’s Mugging
- Unbounded utility
- Decision-making under infinity
I'm curious about how representative the EA forum is to the EA community, particularly in regards to worldview. One thing you could try is to take the EA community surveys done by RP, and estimate how a random representative of "the forum" would answer, where the random representative might be a random user or a random comment/post weighted by karma or amount of text
I fed your message to Scry! Seems the forum is significantly more longtermist.
99% karma-weighted of tagged posts about AI seems wrong
if you check the top 4 posts of all time, the 1st and 3rd are about FTX, the 2nd about earning to give and the 4th about health, totalling > 2k karma
might want to check for bugs
You're so right, I should have caught that. The database data is good, but Opus 4.5 did a rare error. I've since updated the response.
Im still in awe of current AI capabilities
It's hard to fathom what's straightforwardly possible already.