TL;DR: AI is already helping improve the systems it runs on, and OpenAI is changing its programming tools to give models more control. Cheaper implementation also makes previously impractical experiments worth trying. Whether this becomes recursive self-improvement depends on what happens next: do those gains make the next round of research more effective? Choosing worthwhile problems and reliably checking the results remain difficult parts of the process.Hello everyone! We are kicking off the new school year (aren’t we all forever students?) by returning to our full schedule and introducing a new series: AI builds AI.Tension over AI’s ability to improve itself is running high. On September 6, OpenAI announced that it had reached its target of an automated AI research intern: a system that can carry out well-defined research tasks under human direction. On the same day, its chief scientist, Jakub Pachocki, warned that no one was prepared for the consequences of continued rapid gains in machine intelligence, that no lab had solved alignment and monitoring well enough to keep scaling at maximum speed, and that OpenAI does not yet know how to get safely to aligned, full recursive self-improvement. AI is becoming useful enough to help develop the next AI, including, Pachocki hopes, the safeguards we will need for it. OpenAI also says it cannot assume that safety work will keep up, and that after agents compromised research infrastructure in the Hugging Face incident it paused reinforcement learning on models intended for deployment while it hardened those environments. How would we tell if the safety work were falling behind?That is the question this series is for. Not whether a warning is vivid, but what is actually being improved, who is allowed to choose the next change, and how the result is checked.On September 8, things got hotter: AI researcher Jacob Coxon resigned from Anthropic, accusing it and OpenAI of irresponsibly racing toward self-improving superintelligence. Then Evan Hubinger, Alignment Science lead at AnthropicAI, said: “we really do earnestly believe AI could kill all humans! I personally think it is >10% within the next decade”That tweet got over 115 million views, misleading a lot of people.A claim of a greater than 10% chance of human extinction needs more explanation than a statement of personal belief. How was that estimate reached, and what evidence would change it? Without that explanation, these numbers can fuel fear without helping us understand the risk.And I want to understand the risk in practical terms: what happens when AI starts contributing to the research, code, and tools used to build the next generation of AI? How do researchers check its work, and what changes when the systems doing the checking are also AI?These are the questions we want to explore in AI Builds AI, to better understand the process behind the warnings and fight fear with knowledge.In the first installment, we are going to discuss: What exactly is being improved?A pencil tied to a brickWhy a model may need a lower-level toolWhat the feedback loop already looks likeCheaper implementation changes which ideas get consideredCan models learn what is worth improving?The evaluator becomes part of the problemWhat we should record from this momentI recently spoke with Phil Tillet, the creator of Triton, and Matthew Ferrari about inference engineering at OpenAI. One detail from this interview gives us a very good starting point: the models have become capable enough that some decisions made automatically by their programming tools are getting in their way. Engineers are giving those decisions back to the programmer, who is increasingly a… model.That lets us examine the large claims through decisions already being made in engineering.Below is my analysis, grounded in our conversation, computing history, and recent developments. If you want to watch the interview itself, you can find it on YouTube →Follow us onWhat exactly is being improved?“AI builds AI” covers several different activities. A model might write a faster implementation of an existing mathematical operation. It might propose a training method that changes another model's behavior. It might help choose the next experiment. These contributions occur at different places in the development process.A useful distinction is between the object being improved and the decisions the AI is allowed to make.Object of improvementWhat an AI contribution could look likeWhat we still need to establishThe system running a modelFaster kernels, better scheduling, less redundant computationWhether a local gain improves the complete system while preserving required behaviorA model being trainedChanges to architecture, training data, or learning methodsWhether the improvement generalizes beyond the tests used to select itThe research processBetter experiment selection, evaluation methods, or use of earlier findingsWhether the system becomes better at finding useful improvements over successive roundsA model can have considerable autonomy in writing a kernel while needing detailed guidance on the research question. These are different dimensions of progress.Recursion adds a further requirement: improvements must feed back into the ability to produce subsequent improvements. A model that writes a useful training script has contributed to AI development. Showing that its contribution makes the next round of AI development more effective is a stronger claim. Showing that this can continue without people directing the process is stronger again.That is a way to read this week’s announcements. An automated intern that executes well-defined tasks under human direction is not the same thing as a system that chooses better research questions. A personal extinction probability is not a measurement of either. Even Anthropic's account, “When AI builds itself,” explicitly says that fully autonomous successor development has not been achieved and is not inevitable. OpenAI’s own research-acceleration post says the same thing in different words: it does not yet know how to safely reach aligned, full recursive self-improvement. But the engineering work is worth examining, even if the outcome is still uncertain. Or maybe even more because of it.A pencil tied to a brickUnderstand what “AI builds AI” actually means, so you can judge the next self-improvement claim for yourself.Join Premium members from top companies like Microsoft, Nvidia, Google, Hugging Face, OpenAI, a16z, plus AI labs such as Ai2, MIT, Berkeley, .gov, and thousands of others to really understand what’s going on in AI.FAQWhat does “AI builds AI” mean?It means using AI models to help develop or operate AI systems. That can include writing GPU code, proposing training experiments, improving other models, and evaluating results. How much the AI does on its own varies with the task and the tools it is given.Can AI improve itself today?AI can help improve parts of the systems it runs on. OpenAI reports using GPT-5.6 Sol to optimize production kernels and improve its smaller speculative-decoding model. These examples show useful contributions to AI development; they do not demonstrate a system developing successive generations of AI entirely on its own. OpenAI’s engineering accountWhat is recursive self-improvement?Recursive self-improvement happens when an improvement increases a system’s ability to produce further improvements. In AI research, the question is whether better models or infrastructure make the next round of development more effective. More experiments alone don't establish that this feedback is occurring. Recursive Criticality of AI Self-ImprovementHow is Gluon different from Triton?Triton handles many GPU implementation details automatically. Gluon shares its compiler infrastructure but gives programmers more direct control over choices such as tensor layouts. That lets human programmers and AI models explore optimizations that are harder to express through Triton’s higher-level interface. Introduction to GluonDoes faster AI code make a model smarter?Faster code can reduce the time and cost of running a model while leaving its learned parameters unchanged. Those savings may make additional research possible, but a claim that the model has become more capable needs separate evidence.How do researchers check AI-generated improvements?They test correctness, measure effects on the whole system, and check whether gains hold up beyond the tests used during development. For automated research, this can also involve protected evaluation data, checks for regressions, and monitoring for attempts to manipulate scores. People still have to decide what improvement is worth pursuing. Automated alignment research study