Ben Turtel

How so?

Future-as-Label: Scalable Supervision from Real-World Outcomes

1mo

AI can learn directly from the passage of time at unlimited scale—no human annotation required.

Time provides free supervision. Humans learn from experience with no labels—we constantly form expectations about the world, notice when we're wrong, and update our models accordingly.

"Future-as-Label" teaches AI to learn the same way. The passage of time provides labels that require no annotation.

This unlocks unlimited training data for AI from streams of data, with zero human bottlenecks.

Here we apply this to historical news streams, then evaluate the trained model on public Metaculus questions.

Result: fine-tuning with "Future-as-Label" improves Brier score by 27% over the base model (Qwen3-32B) and halves calibration error, even outperforming a 7× larger model (Qwen3-235B) of the same generation.

Full paper: https://arxiv.org/abs/2601.06336

-1

Darwin’s LLMs - Natural Selection is Already Shaping AI

Ben Turtel

3mo

“This preservation of favourable variations and the rejection of injurious variations, I call Natural Selection.” - Darwin

Discussions on the risks of AI often fixate on what a handful of powerful actors decide. What capabilities will Anthropic prioritize? Will OpenAI open-source more models? What safety testing will Congress require?

But there’s another force shaping AI that acts independently of our best intentions: natural selection.

In The Selfish Gene (1976), Richard Dawkins described how evolution by natural selection might not require genes or even biological material. According to Dawkins, genes are just one form of replicator - “any entity in the universe of which copies are made”. Digital files are replicators. Ideas are replicators.

AI models are–or... (read more)

LLMs can teach themselves to better predict the future

Ben Turtel

LLMs can teach themselves to better predict the future - no human examples or curation required.

In this paper, we explore if AI can improve its forecasts via self-play and real-world outcomes:

- Dataset: 12,100 questions and outcomes from Polymarket (politics, sports, crypto, science, etc)
- Base model generates multiple distinct reasoning traces and predictions per question
- Rank predictions by how close they were to the actual outcome
- Fine-tune with DPO on the ranked traces & predictions

Result: +7-10% accuracy over control, bringing two small (14B) models on par with GPT-4o (over 10x larger).

Replying toHuman, All Too Human - Superintelligence requires learning things we can’t teach

Ben Turtel1y

Human, All Too Human - Superintelligence requires learning things we can’t teach

That's a good question. I don't think I have a great idea of the lower / upper bound on capabilities from each, but I also don't think it matters much - I suspect we'll be doing both well before we hit AGI's upper bound.

There's likely plenty of "low hanging fruit" for AGI to uncover just working with human data and human labels, but I also suspect there are pretty easy ways to let AI start generating / testing hypothesis about the real world - and there are additional advantages of scale and automation to taking humans out of the loop.

-5

Replying toHuman, All Too Human - Superintelligence requires learning things we can’t teach

Ben Turtel1y

Human, All Too Human - Superintelligence requires learning things we can’t teach

Hey, thanks for reading and for the thoughtful comment!

100% agree with this: "AI should be able to push at least somewhat beyond the limits of what humans have ever concluded from available data, in every field, before needing to obtain any additional, new data."

Current methods can get us to AGI, and full AGI would result in a mind that is practically superhuman because no human mind contains all of these abilities to such a degree. I say as much in the full post: "Models may even recombine known reasoning methods to uncover new breakthroughs, but they remain bound to known human reasoning patterns."

Also agree that simulation is a viable path... (read more)

-5

Human, All Too Human - Superintelligence requires learning things we can’t teach

Ben Turtel

Are we on the verge of an intelligence explosion? Maybe, but scaling alone won't get us there.

Why? The human data bottleneck. Today’s models are dependent on human data and human feedback.

Human-level intelligence (AGI) might be possible by teaching AI everything we know, but superintelligence (ASI) requires learning things we 𝗱𝗼𝗻’𝘁 know.

For AI to learn something fundamentally new - something it cannot be taught by humans - it requires exploration and ground-truth feedback.

Exploration: The ability to try new strategies, experiment with new ways of thinking, discover new patterns beyond those present in human-generated training data.
Ground-Truth Feedback: The ability to learn from the outcome of explorations. A way to tell if these new strategies

... (read more)

-13

Reality Testing

Ben Turtel

In the recent weeks since Biden’s disastrous debate performance, pundits have theatrically expressed surprise at his apparent cognitive decline, despite years of clear and available evidence in the form of video clips, detailed reports, and fumbled public appearances.

This is far from the first popular narrative that has persisted in the face of obvious counterevidence. Claims of widespread voter fraud in 2020 continue after losing every court case. Lab leak theories were brushed aside in favor of zoonotic transmission, ignoring every signal of plausibility. Climate change denial endures ever-rising temperatures. Inflation is blamed on greedy gas stations after years of massive fiscal stimulus and warnings from prominent economists. The examples abound.

It's been discussed... (read 1708 more words →)

-2

Replying toPower Law Policy

Ben Turtel2y

Power Law Policy

Hey @NeroWolfe.

I think you're looking at the wrong numbers. For example, their 23.3% for the top 5% INCLUDES the top 1%, which skews the averages precisely because of the power laws at play.

The have another table further down where they break this out:

High income earners pay the highest average income tax rates according to latest federal income tax data

Replying toPower Law Policy

Ben Turtel2y

Power Law Policy

Hey @TeaTieAndHat, thank you for the comment!

I think the core idea I'm trying to get across - and the reason I hoped this was a good fit for LW - is not about how govs can maximize revenue for its own sake, but that as a society we could do a much better job caring for the 99% by acknowledging the power laws behind wealth creation and focusing more on increasing our yield at the top.

This is true even if your goal as a government is NOT to maximize productive capacity, because cost / revenue is the bottleneck preventing you from funding everything else. Shifting policy to take advantage of this... (read more)

Replying toPower Law Policy

Ben Turtel2y

Power Law Policy

Hey @Myron Hedderson, thanks for reading and for the thoughtful comment!

The 1% in this article is just a proxy for that level of success - we claim is that we can grow the number of people who achieve the level of success typically seen in the top 1%, not that we can literally grow the 1%. Productivity and the economy isn't zero sum, so elevating some doesn't mean bumping others down (although I acknowledge that there are some fields where relative success is a driver of earnings).

I agree I could have been more clear about this a little bit, and I considered it - I just thought it was a bit of... (read more)

Power Law Policy

Ben Turtel

Extreme outcomes drive tax revenue

Power Laws

Let's do a thought experiment.

The year is 1800. You’re a loan officer at a bank. Farmers come to you asking for a loan - maybe to purchase new equipment, buy more land, or make investments to improve their farm’s productivity. What are your criteria for granting a loan?

You’ll primarily want to do 2 things: avoid the losers, and reduce risk. To avoid the losers, you’ll want to learn about a farmer’s reputation and character. Are they honest and hard working? Do they have any major vices that might undermine their productivity? To reduce risk, you’ll assess the value of their collateral. How valuable is the land the... (read 1544 more words →)

The Progress Paradox

Ben Turtel

LessWrong Intro: I'm new to the LessWrong community, but was inspired to write this based on recent reading of The Sequences. It seems to me there's a paradox to social progress that tends to undermine itself - societies making rapid ethical progress will (necessarily) become ashamed of their recent past and question their own legitimacy. Without awareness of this effect, it can undermine subsequent progress or influence such societies are willing to have in the world.

Humans are uniquely capable of self-reflection and learning from the past.

This capacity to change and grow over time also alters our perspective on our own previous actions, so that deeds and decisions that used to make us... (read 1053 more words →)

LESSWRONG
LW

LESSWRONG
LW

Power Law Policy

The Progress Paradox

LLMs can teach themselves to better predict the future

Darwin’s LLMs - Natural Selection is Already Shaping AI

Ben Turtel

Ben Turtel

Future-as-Label: Scalable Supervision from Real-World Outcomes

Darwin’s LLMs - Natural Selection is Already Shaping AI

LLMs can teach themselves to better predict the future

Human, All Too Human - Superintelligence requires learning things we can’t teach

Reality Testing

Power Law Policy

The Progress Paradox

Ben Turtel

Power Law Policy

The Progress Paradox

LLMs can teach themselves to better predict the future

Darwin’s LLMs - Natural Selection is Already Shaping AI

Ben Turtel

Ben Turtel

Future-as-Label: Scalable Supervision from Real-World Outcomes

Darwin’s LLMs - Natural Selection is Already Shaping AI

LLMs can teach themselves to better predict the future

Human, All Too Human - Superintelligence requires learning things we can’t teach

Reality Testing

Power Law Policy

The Progress Paradox

Extreme outcomes drive tax revenue

Power Laws