8/11/25

See personal notes from 8/11/25 conversation. Some highlights:

She hears from colleagues that AI researchers are able to automate a lot of their coding work. Building a production-grade consumer-facing product is much harder to automate.

Her trip to Nigeria was a reminder that no amount of GPUs is going to fix people’s ability to use this in any way. Electricity and connectivity are extremely limited / intermittent. Most people on Earth aren’t eligible for decades to access this.

Could try to monitor [inertia effects] by finding highly reticent groups and monitoring adoption rates.

GPT-5 didn’t make a big impact because everyday use cases are already approaching saturation. With o3, she saw people in specialized scientific fields seeing a big jump in applicability, there will be large spiky impacts in niche applications.

She’s fairly AGI pilled but she doesn’t expect it to automate all work. Partly because of limits to adoption in many parts of the world, and also many jobs are too varied and granular. But yes in finance. Transformational to medicine and health, that’s the one area where she has a lot of optimism – personalized medicine will be highly impactful.

In the legal field, it’s difficult to say. It’ll transform access to legal services, but the infrastructure of the courts will be behind, so unclear what that will mean for legal recourse. 

Impact will be highly sector-specific; some sectors will be very resistant to automation for non-capability-related reasons. We are nowhere near figuring out how to make AI granularly effective in some sectors – hard to make AI capable in very specific sectors. For example, food service: you’d need to be good at forecasting traffic, sorting customer preferences, understanding tariff-related spikes in ingredient costs, 100 different things that go into making a restaurant successful (without even getting into physical tasks). Will AI be useful in running service businesses? Unclear. Education is another interesting example: we have wonderful tools like Khan Academy, studies show good results, but not being brought into schools.

Entrenched interests in many fields will slow adoption.

They have a whole team that’s engaged in [adoption], she’ll look into whether they have any insights to share. She hasn’t been paying close attention to their work.

It’s been clear from a lot of her projects: a large barrier to many people’s usage is language. There are 10,000s of dialects that we’re not on track for any model to speak. Need transcripts + voice recordings, and many of these languages don’t even have a written form.