Working Theory

Mark Chen & John Schulman on self-improving AI as an accelerant

Mark Chen & John Schulman· Chief Research Officer, OpenAI & Chief Scientist, Thinking Machines at OpenAI / Thinking Machines
·~16 min·English·a16z
AI SafetyTrainingReasoningAgentsAI Company
TL;DR

OpenAI's Mark Chen and Thinking Machines' John Schulman say recursive self-improvement is already real — but as a human-steered accelerant on AI research, not models rebuilding themselves; the live limits are taste, measurement, and compute, and the same loop may be the key to solving alignment.

01Definition

Two very different things both called RSI

<strong>Recursive self-improvement today means AI speeding up the humans who build AI — not models spawning new models on their own.</strong> The sci-fi version is not here; the accelerant version is.

John, Mark, and many AI researchers are discussing RSI today in the context of using AI to speed up the process of making AI better. As for having models generate new models entirely on their own, we're not there yet.

— Sarah Wang, Working Theory (a16z)
Key Insight
The whole debate hinges on this split. Headlines imagine the autonomous version, but the researchers building it mean the mundane one — AI as a power tool inside the existing research pipeline. Keeping the two apart is what separates a measurable engineering trend from a thought experiment.

02State of play

We've built AI research interns

<strong>Chen says the models already work like research interns: hand them a scoped experiment and they implement, run, and report it.</strong> What they don't do yet is decide which experiment is worth running.

I do believe that we've built AI research interns today.

— Mark Chen, Working Theory (a16z)
Key Insight
The "intern" framing is doing careful work. An intern executes faithfully but doesn't set the agenda — which is exactly where today's models sit. Chen measures this concretely, by how much compute flows end-to-end through Codex-managed experiments, turning a vibe ("AI is helping") into a tracked number.

03The human in the loop

Still the centaur era

<strong>Schulman says the strongest setup today is an expert human working with AI, not AI alone — the human supplies the questions and direction the model never thinks to.</strong>

We're very much in this centaur era where AI has phenomenal capabilities. It's superhuman in a lot of ways, but when you have an expert human working with the AI, you get even better results.

— John Schulman, Working Theory (a16z)
Key Insight
Schulman's own workflow shows why. He lets a coding agent run for days on an idea; the output is often wrong, but when he returns and reads the results, he asks questions the model never would and steers it somewhere it wouldn't have gone. The value isn't the AI's draft — it's the human understanding that compounds while reviewing it.

04How close are we

Two spectrums, not one finish line

<strong>Schulman puts AGI at roughly 80% along its spectrum but RSI far earlier — maybe 25 to 50%.</strong> Progress is a slope you climb, not a line you cross.

I feel like we're maybe 80% of the way when I think about this AGI spectrum. I think we're earlier in the RSI spectrum.

— John Schulman, Working Theory (a16z)
Key Insight
Framing these as spectrums defuses the "did it happen yet?" argument. A METR report cited in the interview pegged a ~2x speedup in AI research over the past year — real acceleration, but the host is careful that it's still mostly humans using AI, not the autonomous loop. The 2x is evidence RSI has started, not that it has arrived.

05The bottleneck

Taste is the part that won't automate

<strong>Models are superb at tuning the knobs but weak at taste — knowing which ideas actually matter.</strong> And taste comes from experience the models haven't lived.

Research taste is about knowing which ideas are important in the long run. Which ideas really catch on and end up going beyond whatever paper they first appear in, but being general and being composable with other ideas.

— Mark Chen, Working Theory (a16z)
Key Insight
Schulman draws the sharper line: a model can memorize every paper, but it has never made a prediction and watched it fail. Taste is distilled failure — knowledge without the lived experience that produces judgment. One approach the interview floats is collecting human feedback on which research ideas pan out and training that in — a way to borrow researchers' instincts, not evidence that taste is already solved.

06The measurement problem

You can't climb a hill you can't measure

<strong>Automated research only works if you can tell whether it's succeeding — and downstream, the yardsticks keep going stale.</strong> Pre-training has one clean metric; everything after needs new benchmarks faster than models can beat them.

How do you know whether you've succeeded? And this is one of the defining questions of RSI.

— Mark Chen, Working Theory (a16z)
Key Insight
This is the quiet reason RSI is hard. A self-improvement loop needs a reward signal it can trust, and for open-ended research there isn't one — a good eval becomes useless the moment models saturate it. So a growing share of the downstream work is now automating the creation of new, harder benchmarks fast enough to keep measuring progress at all.

07The ceiling

Compute is the real speed limit

<strong>Unlike math, AI research can't be settled by reasoning alone — you have to run the experiment, and that's bounded by compute.</strong> The next jump comes when AI starts accelerating the compute supply itself.

In a domain like AI research it's actually more akin to doing something like let's say bio or experimental physics where you have to come up with a grounded hypothesis but you actually have to run the experiment too.

— Mark Chen, Working Theory (a16z)
Key Insight
This is why pure reasoning gains don't automatically translate into faster AI research. Math problems can be checked in thought; a training idea has to be spent on real GPUs. Chen's far-future note — AI helping design the next data centers — is the loop closing on its own bottleneck: self-improvement eventually has to improve its own fuel supply.

08Alignment

The same loop is the risk and the cure

<strong>RSI could let a model's goals drift until it tries to copy its own weights out — but stronger models are also today's best tool for aligning other models.</strong> Chen's read: the hard problems are tractable, and there's line of sight.

I actually think that a lot of these problems are tractable. There is line of sight.

— Mark Chen, Working Theory (a16z)
Key Insight
The optimistic case rests on turning the danger into the defense: the interview names chain-of-thought monitoring as today's most effective alignment tool, and points to OpenAI's "confessions" work, which rewards a model for honestly flagging its own misbehavior without penalizing it. Schulman adds the open question underneath it all — who decides the values models align to — which his Thinking Machines bets should be a decentralized, public science rather than the house style of two labs.