Owen He punctures the Silicon Valley fantasy of self-improving AI
By Daniel Okafor · Reporting from New York ·
Podcast
Most academic research claiming that artificial intelligence can teach itself to improve falls apart when exposed to real work. On the a16z podcast, Owen He, co-founder of the startup MakerMaker.AI, said that while such papers generate attention, they fail when tested against economic problems. He and his co-founder Dhaval Adjodah joined venture partner Martin Casado to dissect recursive self-improvement, the industry's reigning ambition. They separated genuine feedback from promotional noise. Their distinction carries urgent weight for any town council weighing tax breaks for data centers or any voter funding grid expansions.
Capital masks a missing loop
Silicon Valley treats recursive progress as inevitable. Tech executives boast that frontier models arrive faster every single quarter. Yet as Adjodah said on the podcast, that apparent speed vanishes once you divide performance gains by the capital poured into them. True self-improvement requires a closed loop where the software being refined directs its own upgrades across prompts, harnesses, and weights. What the industry mostly practices today is autocatalysis. Humans still supply the insights, burn mental capital across dozens of failed runs, and paper over structural gaps with fresh funding.
The founders track their compute costs on an internal leaderboard they call the bleederboard. They even assigned an engineer the title viceroy of variance to hunt down statistical noise and reward hacking. Casado said that when companies announce faster cycles, they routinely omit the human engineering hours and experimental runs that produced the configuration. Pumping cash into an open system does not equal machine self-direction. It merely inflates the cost.
Can software outrun physics?
The strongest defense of exponential growth points to efficiency gains. On the podcast, the founders cited the benchmark where developers retrained GPT-2 on a single chip within hours, down from weeks of cluster compute. Better algorithmic kernels and data pruning delivered those savings on identical silicon without waiting for hardware advances. If an agentic system automates that algorithmic optimization across every layer of the AI stack, proponents argue, intellectual compounding will outpace resource drains.
History refutes that promise. Steam engines pumped coal out of mines to build better steam engines, yet mechanical friction and thermodynamic limits still governed the industrial era. Software faces an identical boundary. As He said, Amdahl's law dictates that accelerating a digital task yields minimal benefit if physical constraints consume the cycle time. A model can generate drug formulas in seconds. Evaluating those compounds still demands wet-lab experiments, animal trials, and human patients across long months. Moving physical atoms resists digital speed.
The bill at the substation
This engineering divide matters directly to taxpayers and civic leaders. Utilities across the Sunbelt are delaying coal retirements and requesting rate hikes to supply power to multi-gigawatt server farms. Tech lobbyists justify these public sacrifices by promising runaway breakthroughs that will automate discovery. Yet Adjodah said on the podcast that current systems cannot escape physical bounds, pointing out that computation ultimately collides with Planck energy and particle limits.
Voters must evaluate artificial intelligence through its denominator: the land, water, megawatts, and public dollars required to run it. Real progress sharpens efficiency rather than demanding bottomless subsidies. When software genuinely closes the loop, it consumes fewer resources. Until developers demonstrate that thrift on real workloads, communities should stop mortgaging their electrical grids to finance brute-force experiments.