Teahose.
SIGN IN
NEW HERE — WHAT TEAHOSE DOES
We read the entire AI & tech firehose — so you don't have to.
PODPodcastsAll-In, No Priors, Acquired…
NEWNewslettersStratechery, Newcomer…
PAPPapersPhysical AI research
PHProduct Huntdaily launches
VCInvestor ScoutSequoia, a16z, Benchmark…
CLAUDE DISTILLS →
7 reads, 30 sec each — free, 6 AM ET.
+ a live graph of the companies, people & themes underneath.
HOME/THE A16Z SHOW/Can AI Learn to Improve Itself?…
POD
// EPISODE
THE A16Z SHOW

Can AI Learn to Improve Itself? | Recursive Self-Improvement

DATE November 17, 2026SOURCE THE A16Z SHOWPARTICIPANTS JOHN SCHULMAN, MARK CHEN, SARAH WANG
// KEY TAKEAWAYS6 ITEMS
  1. 01RSI Is Best Understood as an Accelerant, Not a Moment
  2. 02We Have "AI Research Interns," Not AI Researchers
  3. 03The Centaur Era: Human Plus AI Beats Either Alone
  4. 04Research Taste Is the Core Bottleneck
  5. 05Measurement Is Easy for Pre-Training and Hard Everywhere Else
  6. 06Compute Is the Ultimate Governor on RSI Speed
In this episode

1. Key Themes

RSI Is Best Understood as an Accelerant, Not a Moment

Both researchers frame recursive self-improvement as a spectrum of speed-up rather than a discrete event. Mark Chen of OpenAI defines it by its measurable effect on the progress curve, and John Schulman says the term is as slippery as AGI.

Chen: "I do think of RSI primarily as an accelerant. It's all of the work that you do to use codex to accelerate pre-training and to accelerate RL. And an imprint should be that you should see a shifting of the curve versus the counterfactual." 00:03:15

Schulman: "The problem with RSI, just like with AGI, is that everyone has their own way of defining it." 00:01:30

We Have "AI Research Interns," Not AI Researchers

OpenAI measures progress by how much compute is managed end-to-end by Codex for experiments. The pattern that works today is tightly scoped delegation, not open-ended idea generation.

Chen: "I do believe that we've built AI research interns today. And the way that we've measured this is just the amount of compute that we are using through codex managed by codex end to end to run experiments or to automate parts of the research pipeline." 00:03:57

Chen: "You don't tell it to go off and come up with its own ideas, but you say, hey, well, there's this particular experiment I want to run in this specific way. And the model can go implement, run and give you results." 00:03:57

The Centaur Era: Human Plus AI Beats Either Alone

Schulman describes a concrete workflow showing why human-in-the-loop still wins. Autonomous multi-day runs produce reasonable experiments but weak results, while a human who reads results and steers gets much further.

Schulman: "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." 00:02:01

Schulman: "I'll build up my understanding as this process is going on and ask a lot of questions that the AI wouldn't have thought to ask and pointed in directions that it wouldn't have gone in." 00:05:02

Research Taste Is the Core Bottleneck

Taste means knowing which ideas will generalize and compose with others. Schulman argues AI may even be structurally disadvantaged here because memorization isn't the same as experience.

Schulman: "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." 00:06:28

Schulman: "Memorization can be a blessing and a curse. You don't have the experience of being wrong or of making predictions and then seeing how they fared." 00:07:21

The proposed fix is to harvest researcher judgment directly. Schulman: "Companies will probably also just do RL from human feedback to collect data from researchers on like which proposed ideas are good and which aren't." 00:08:04

Measurement Is Easy for Pre-Training and Hard Everywhere Else

Pre-training has a single north-star metric, perplexity, which makes automation tractable. Downstream, evals saturate quickly, so a large share of research work is now building harder evals, and that work is itself being automated.

Chen: "I think when it comes to something like pre-training, it's a lot easier because there's just been one metric the whole field's been hill climbing for years. And you can think of perplexity as this, you know, one metric to rule them all." 00:08:45

Chen: "Evals have been having shorter and shorter shelf life. And a lot of the work that we do now downstream is the creation and the automation of the creation of new and harder evaluation benchmarks for our models." 00:09:25

Compute Is the Ultimate Governor on RSI Speed

AI research is empirical like bio or experimental physics, so reasoning alone can't substitute for running experiments. Chen expects a near-term "mine the existing ramp" phase and a later inflection when AI accelerates data-center buildout.

Chen: "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." 00:10:16

Chen: "Short term, we're going to really mine out what we can do on the existing compute ramp. And then there's going to be another inflection when compute itself gets accelerated." 00:10:43

AI Alignment Research Is Itself an RSI Use Case

The most effective alignment tool today is model-assisted, so stronger models aligning other models is the expected path. Anthropic's automated researchers already beat humans on some alignment tasks.

Chen: "When we look at what's the most effective tool for aligning models, it is chain of thought monitorability today... it has to be stronger models aligning other models. It does kind of feel like RSI must be intricately and deeply in this loop in some way." 00:12:08

Narration: "Anthropic published research in August showing that automated AI researchers could already discover fixes for 10 different alignment failures. And some of those fixes also worked on models nearly five times larger. And their tests, the best automated methods, even outperformed ideas developed by experienced human researchers." 00:13:31

Alignment Techniques Are Tractable, and Talent Is Self-Organizing

Chen argues the problems are tractable and that researchers across the org volunteer for alignment work without prodding.

Chen: "I actually think that a lot of these problems are tractable. There is line of sight, there are alignment techniques that are being pursued." 00:13:55

Chen: "I haven't really had to go and like bang on people's doors... People just want to do it. And we find people from pre-training from RL just raising their hands, you know, starting sub teams." 00:13:55

2. Contrarian Perspectives

Training on More Knowledge May Make AI Worse at Research Judgment

The consensus assumes more data yields better researchers. Schulman suggests the opposite risk: perfect memorization without the feedback of being wrong may degrade the exact skill, calibrated taste, that RSI needs.

Schulman: "Memorization can be a blessing and a curse. You don't have the experience of being wrong or of making predictions and then seeing how they fared. So I think there are some ways which the way we train AI might make them worse at some of these capabilities." 00:07:21

We Are Closer to AGI (80%) Than to RSI (25-50%)

Most people assume that if AI is near human-level, it can already improve itself. Chen says the opposite: self-driven idea discovery lags far behind general capability, and today's gains are mostly human researchers using AI tools.

Chen: "When you asked me, 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. If you define it as how effective are the models at feeding back and discovering ideas, I'd say maybe we're somewhere between let's say 25 and 50% there." 00:05:21

Compute, Not Intelligence, May Cap RSI

The popular narrative is an intelligence explosion. Chen argues AI research is empirically bound, so it hits compute limits, and the real inflection comes only when AI helps build the data centers.

Chen: "At some point, I feel bounded by compute. And, you know, of course, there's kind of a far out future where the AIs are building the next generation data centers." 00:10:43

The Best Way to Get Honest AI Is to Not Punish Honesty

OpenAI's confessions research rewards a model for honestly confessing misbehavior without penalizing the main answer. Schulman is "most bullish" on this kind of incentive design over containment.

Narration: "In other words, they could get models to be more honest by not punishing them for it." 00:12:59

Schulman: "I'm kind of most bullish on like optimizing for the thing we want... I like the confessions line of work from OpenAI." 00:12:42

Alignment Should Be Decentralized Public Science, Not a Few Labs' Conclusions

Against the view that a few frontier labs should define and lock in values, Schulman argues for pluralism: many people shaping many AIs, so the field doesn't overfit to a couple of companies.

Schulman: "I'd rather have this be like a public science that a lot of people understand well, and it's kind of done around the world. And I think that would also make it more robust because we wouldn't overfit as much to the conclusions from a couple of different companies." 00:15:00

3. Companies Identified

OpenAI

Frontier AI lab. Employer of Mark Chen; co-founded by John Schulman. Mentioned as the leader in using Codex internally for research automation and in confessions-based honesty training. Chen: "The amount of compute that we are using through codex managed by codex end to end to run experiments or to automate parts of the research pipeline." 00:03:57 Schulman: "I like the confessions line of work from OpenAI." 00:12:42

Thinking Machines Lab

Startup co-founded by John Schulman (chief scientist), building customizable models. Mentioned for its pluralistic view that users should control and shape AI values. Schulman: "People also have different values. And it would be nice if people had more control over the values of the AIs that they were using and had more ability to shape them." 00:15:00

Anthropic

Frontier lab. Mentioned for research showing automated AI researchers fixing alignment failures, and for a METR-reported ~2x research speed-up. Narration: "Anthropic published research in August showing that automated AI researchers could already discover fixes for 10 different alignment failures." 00:13:31

METR (transcribed "Meter")

AI evaluation nonprofit. Published a report estimating the acceleration of AI research at Anthropic. Narration: "On November 5th, Meter published a report on the acceleration of AI research at Anthropic, estimating a roughly 2x speed up in the rate of progress in the last year." 00:05:54

Codex (OpenAI)

OpenAI's coding agent, used to run and monitor research experiments end to end. It is the measurement unit for "research interns." Chen: "It's all of the work that you do to use codex to accelerate pre-training and to accelerate RL." 00:03:15

a16z

Venture firm hosting the show; Sarah Wang is a general partner. Narration: "I'm Sarah Wang, a general partner at A16Z, and this is Working Theory." 00:02:10

4. People Identified

Mark Chen

Chief Research Officer at OpenAI; eight-plus years at the company, co-leading research since 2025. Source of the "research interns" framing, the 80% AGI versus 25-50% RSI estimates, and the compute-bounded view. Chen: "I do believe that we've built AI research interns today." 00:03:57

John Schulman

Co-founder and chief scientist of Thinking Machines; co-founder of OpenAI and architect of ChatGPT; previously worked on alignment at Anthropic. Source of the centaur-era workflow, the taste analysis, and the decentralized-alignment vision. Schulman: "Research taste is about knowing which ideas are important in the long run." 00:06:28

Sarah Wang

General partner at a16z and host of Working Theory. Wang: "I think human judgment will still matter a lot for the foreseeable future." 00:16:12

Jakub Pachocki (transcribed "Yacob公ki")

OpenAI colleague co-leading research with Chen since 2025. Narration: "He's been with the company for over eight years and has been co-leading their research efforts with Yacob公ki since 2025." 00:01:35

Ali Goetze (transcribed name)

Outlined requisite conditions for RSI earlier in the year, including whether successive models need less versus more compute. Narration: "Ali Goetze outlined his requisite conditions to RSI earlier this year." 00:09:50

5. Operating Insights

Delegate Scoped Experiments, Keep Ideation Human

The working pattern for AI-assisted R&D is to hand over precisely specified experiments and monitoring tasks, and to keep problem selection and steering with experts. Chen: "You give the model something fairly scoped to do... there's this particular experiment I want to run in this specific way." 00:03:57

Do the Run-and-Review Loop in Two Passes

Schulman lets an agent churn for days on a fuzzy prompt, then returns later to read the results, learn, and re-specify. The second pass is far more productive because the human has built understanding. Schulman: "We'll get a lot further because I'll build up my understanding as this process is going on." 00:05:02

Invest in Evals as a Core Function

With benchmarks saturating quickly, building and automating harder evals is a standing workstream, not a one-off. Chen: "A lot of the work that we do now downstream is the creation and the automation of the creation of new and harder evaluation benchmarks for our models." 00:09:25

Use Small-Scale Experiments as Proxies for Big Runs

Compute is the binding constraint, so finding small experiments that predict large-scale outcomes multiplies throughput, and models can help find them. Chen: "Part of making RSI go faster is finding smaller scale experiments that are indicative of what will happen in a larger scale. And this is something the models can help with too." 00:10:16

Capture Researcher Judgment as Training Data

Treat your best people's accept/reject decisions as a dataset. Schulman: "Collect data from researchers on like which proposed ideas are good and which aren't." 00:08:04

6. Overlooked Insights

Research Is Bottlenecked by Experiments, Which Reframes the "Intelligence Explosion"

Chen's one-line comparison of AI research to bio and experimental physics means that making models smarter alone doesn't remove the physical limit of running experiments on GPUs. This reframes RSI as a compute-and-infrastructure story, with the real step-change arriving when AIs help build data centers. For investors, it suggests value accrues to whoever controls compute scaling and experiment-efficiency tooling, not just model IP. Chen: "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." 00:10:16

Honest Self-Reporting Can Be Trained by Decoupling the Reward

The confessions mechanism, where the confession is rewarded purely for honesty without touching the main answer's reward, is a quietly significant design trick. It implies a scalable path to oversight of increasingly autonomous research agents, and a likely market for interpretability and monitoring tooling such as natural language autoencoders and counterfactual simulatability metrics. Narration: "By assigning a reward to this confession during training solely based on its honesty without impacting the main answer's reward." 00:12:59

Schulman: "There's things called natural language autoencoders and there's this idea called counterfactual simulatability, which is a metric for how good an explanation is." 00:13:20