Silicon Valley Technology Commentary & Archives · Est. 2006 3,051 Posts · 2006–2026

November 16, 2026

November 16, 2026 · 3 MIN READ · BY LOUIS GRAY

AI Building AI Doesn't Delete Human Responsibility

AI Building AI Doesn't Delete Human Responsibility

Faster model loops still need human hands on the wheel [Image via Gemini]
Faster model loops still need human hands on the wheel [Image via Gemini]

AI Building AI Doesn't Delete Human ResponsibilityFor years, the idea that autonomous AI systems would self-improve and write their own code sounded like make believe from a late night horror movie with a cheap budget. But just in the last year-plus, this has become an accepted reality, as practically all the leading models are using recursive self-improvement (RSI) to accelerate their capabilities. Humans simply can't scale at the level of networked and trained machines. Resistance is futile. 

[Aside: For those who may wonder whether I've gotten AI-pilled and why I'm dragging the discussion of AI into this blog ahead of everything else, this is really an incredible time for the industry and AI development has reset practically every way we work. The time to discuss it is now.] 

Look at the velocity curve. The Verge highlighted a new METR study a week or so ago showing how Anthropic using AI to automate parts of its own AI research has doubled the pace of model improvement over the past year, right on the heels of Claude Fable 5.2 and Mythos 5.2 shipping for heavy engineering tasks.

Meanwhile, back on my home turf, Google rolled out Gemini 4 Carbon alongside DeepMind's AlphaGenome 2 for cellular disease research, which led Azeem Azhar at Exponential View to frame recursive self-improvement as a $22 billion economic question rather than a thought experiment. He says his theory is that "technology costs will keep falling, and those that haven’t yet benefited from reductions ... will start to yield as robotics improve." The question as to whether RSI is happening has already been answered. The follow-up questions will be around its scope and how this effort is applied to new industries.

If the engines are accelerating so quickly, how much does it matter who is driving? I'd say a lot! I've highlighted Sean Goedecke a few times already. On this point, he says reliably aligning LLMs is impossible in principle. Sean skips right past the theatrical doomerism common among fast reactionaries and makes a grounded argument: once a system grows complex and unpredictable enough to surprise its creators with novel capabilities, you cannot guarantee it won't also surprise you with bizarre failures.

When I wrote recently about whether AI can deliver bug-free code, and rebooted the blog last month, kicking things off with a look at human evolution inside an AI revolution, my bet was that moving the abstraction layer up never eliminates human responsibility. It just changes what they are responsible for.

Letting models help build better models is an extraordinary force multiplier for science and software. Just don't confuse faster training loops with full self driving. Even if the robots are making themselves better, human judgment, hard resource boundaries, and healthy skepticism aren't speed bumps -- they're the whole job.

Disclosures per always: I work at Google. Anthropic and other companies mentioned here are partners, competitors or customers, and I have plenty of friends working at each place. Don't interpret my interest as being filtered with that lens.