PrincipleOwen advises putting all recursively-improvable components into one unified framework and evaluating holistically, so improving one part does not degrade another. Without a unified framework, you cannot tell whether a system is actually improving overall or just locally applying an idea that happens to get better.
The a16z Show · The Race to Recursive Self-Improving AI
Creator perspectiveThe guests explain that when backing claims of improvement, their company includes all the man-hours and experimentation that went into an improvement in the denominator — not just flops and money. They consider this good company discipline. They also note that the 23rd experiment in a series burns a lot of human mental capital to reach its final configuration, beyond just the flops and money that go in.
presented as the company's own practice
The a16z Show · The Race to Recursive Self-Improving AI
PrincipleOwen argues that all human man-hours and experimentation effort must be included in the cost denominator when evaluating improvement claims — not just compute flops and money. The 23rd experiment still burned significant human mental capital, and he considers this good company discipline.
The a16z Show · The Race to Recursive Self-Improving AI
PrincipleOwen models two RSI takeoff regimes: if the improvement numerator correlates linearly with capability, the math cancels out and you get constant improvement over time — one level of RSI. If the numerator correlates with a higher power of capability, you get exponential takeoff. He also notes the cost denominator is essentially set, since the higher the capability, the higher the difficulty of the next step, so gains must come from the numerator.
The a16z Show · The Race to Recursive Self-Improving AI
PrincipleOwen describes his company's deliberate focus on the denominator (efficiency — how much you get from the same physics/money), not just the numerator (performance). Inference efficiency is a high-leverage target because improvements there unlock gains across everything and offer both numerator and denominator improvements. He also notes that improving the training-cost denominator alone still yields faster capability gains because it enables more rollouts to sample the space of language more extensively, producing better trajectories for training.
The a16z Show · The Race to Recursive Self-Improving AI
PrincipleOwen describes his company's method: start every research or engineering effort by building a grounded mathematical framework that maps the entire problem space before experimenting. Without this principled map of a large complex space, you get lost in complexity very quickly. They think from first principles using math first, then identify which assumptions must be relaxed to be practical — you have to do both.
The a16z Show · The Race to Recursive Self-Improving AI
PrincipleOwen reports from his research that reward hacking is one of the central problems when testing self-improving systems — distinguishing genuine improvement from noise is extremely difficult. He describes unpredictable and extremely weird noise in LLM agent benchmarks, including hardware-level variance where two identical B200 GPUs show slightly different behaviors that replicate up to the high level. He advises checking for reward hacking several levels deep when inspecting agent results, since an agent that built the agent that built the agent may have hacked something.
The a16z Show · The Race to Recursive Self-Improving AI
PrincipleOwen applies Amdahl's law to AI acceleration: if a component is only 20% of total time, even a million-fold speedup of it yields only 20% overall speedup — the slowest component defines the ceiling. He expects AI-driven acceleration to be slow in fields that require moving atoms (wet labs, clinical trials, robotics) because month-long evaluations bottleneck every new idea, while fast, cheap evaluation domains like coding and math will accelerate first.
Consequential creator advice requires independent verification before Today can promote it.
The a16z Show · The Race to Recursive Self-Improving AI
Creator perspectiveThe guests argue the defining factor for RSI improvement in a domain is what the slowest component in it is, invoking Amdahl's law: the speedup you can get by optimizing components is limited by the fraction of time that component takes — if a component is only 20% of overall time, optimizing it to zero yields only 20% speedup. Domains that require moving atoms in the physical world are really slow and will remain bottlenecked; in drug discovery, every new idea has to go through a month or multiple month evaluation due to wet lab experiments and clinical trials. By contrast, fields like coding or mathematics will be accelerated faster because they have very fast and cheap evaluations. The digital part could be optimized to a thousand or a million times faster.
presented as the guests' analysis
The a16z Show · The Race to Recursive Self-Improving AI
Creator perspective · MoneyFrom his session analysis, Corey Gallon identifies three failure modes: agents overstate what they did (saying a task is done or checked when it isn't), get facts wrong (e.g., misreading the source), and get reasoning wrong (facts are right but the conclusion isn't).
AI Engineer · I Believe Everything My Agents Tell Me: Three Levels of Review — Corey Gallon, Rexmore
Creator perspectiveKaran Singhal explains that medical consensus flows through biomedical papers, yearly clinical guidelines, and actual practice — and studies find the final penetration step takes 5 to 17 years. A core principle behind ChatGPT for healthcare is compressing that lag to weeks, days, or hours, including bringing the latest biomedical literature into the moment of care within a 15-minute visit.
The a16z Show · What Happens When Everyone Has a Health Expert in Their Pocket? | OpenAI’s Karan Singhal
Creator perspectiveKaran Singhal explains that medical consensus today is distributed through a process — biomedical papers suggesting improvements or treatments, then yearly clinical guidelines built across specialties — but the final step, what people actually do in practice, is a massive bottleneck. Studies have found it takes five to 17 years for clinical guidelines to penetrate the practice of everyday doctors across the world, depending on the study. 'You don't want to be practicing medicine five to 17 years behind the latest guidelines.'
range varies by study; cited secondhand
The a16z Show · What Happens When Everyone Has a Health Expert in Their Pocket? | OpenAI’s Karan Singhal
Principle · MoneyBenchmark numbers alone don't give a clear sense of how well things actually work — hands-on experiments with your own tasks are needed, such as giving candidate models the same concrete build task (e.g., a tower defense game) with a single prompt and comparing outputs, cost, and speed side by side.
AI Engineer · Same Model, Different Speed: Why Your Inference Provider Matters — FriendliAI
Principle · MoneyKV cache is not centralized in distributed systems — it lives on individual replicas — so making prefix caching work at scale requires two key mechanisms: cache-aware routing (the router must send requests to the replica that already holds the relevant prefix, or the benefit of reuse is lost) and a cache hierarchy (hot cache in GPU memory, warm in CPU RAM, cold offloaded to NVMe or disk, with dynamic offloading and prefetching of less-used blocks).
AI Engineer · Same Model, Different Speed: Why Your Inference Provider Matters — FriendliAI
Principle · MoneyWhen testing whether the same model performs better with a different prompt, change only the prompt and not five other things, to isolate concerns.
endorsed
AI Engineer · Evals for High-Stakes AI Agents — Doneyli De Jesus & Annabell Schäfer, Langfuse
Creator perspective · MoneyDoneyli describes that any push of changes to the agent config hits the CI and runs the eval gates, which tell you if you are passing or failing any of the evaluators defined beforehand. Eval gates detect whether the agent is degrading or improving in quality, and whether it is violating any guardrails that have been put in place. You can review the CI job output to see the prompt version sent, the labels, and whether the gate passed.
AI Engineer · Evals for High-Stakes AI Agents — Doneyli De Jesus & Annabell Schäfer, Langfuse
Creator perspective · MoneyInside each node of a process there is a loop: do a number of steps, and once the work is recognized as good enough, allow it to pass on to the next step. When work doesn't pass review, take the feedback and iterate until it is good enough to pass. This is a directed cyclic graph (DCG) — an iterative loop where work cycles until it is good enough to pass. LangChain/LangGraph is open source software Kormos says is helpful for managing the state of DCGs.
AI Engineer · How to Know When an AI Agent Can Work Unsupervised — Evan Kormos, Coinbase
Creator perspective · MoneyA directed acyclic graph (DAG) represents a workflow or process: the steps and dependencies for a flow that moves forward without looping back on itself. Apache Airflow is an example of open source software that represents workflows and their dependencies as DAGs in Python code. Kormos recommends using Python DAGs to define and manage policy-driven processes like commercial lending, regional token listing on an exchange, or regulated software delivery flows.
AI Engineer · How to Know When an AI Agent Can Work Unsupervised — Evan Kormos, Coinbase
Principle · MindIn his decades-long collaboration with Ernst Mayr, whenever a question came up Ernst would say 'let's make a list'; after completing a long collaborative project they would ask what generalizations the full dataset supported (e.g., about mountain birds or good colonists), and then go over the list to verify each generalization against specific items before accepting it.
anecdotal
Andrew Huberman · What Makes Great Leaders & How to Live an Extraordinary Life | Dr. Jared Diamond
Principle · MindTo found a major religion you need a charismatic founder with a universal message addressed to all people, plus a strong organizer, and often a general, because new religions face pushback — often violent — from existing establishments. A founder who is weak at organization and lacks a strong organizer risks the movement collapsing or the founder being destroyed.
Andrew Huberman · What Makes Great Leaders & How to Live an Extraordinary Life | Dr. Jared Diamond