Noodling on some key themes over here after reading a lot of "AI diffusion" bear cases as of late. They aren't wrong -- but does capture the hard work ahead.
1/ If code is the modality that drives AI into more categories, how do we pull these non-technical domains up into a coding domain? Seems it’s two problems: one, "transpile" the subject matter into something machines can reason about (hard), then two, deliver a trusted verdict (harder). Said differently we might need to bring vertical domains to the code, as opposed to wedging code into the vertical domains.
2/ The applied AI companies doing both appear advantaged in the long run. But rollouts are slower for sure, i.e. lots of integration work, building context, then re-modeling internal operations around a new capability. But once a customer runs its operations against your spec, switching costs go way up. You become an internal standard, and eventually an industry standard. This was Veeva, Guidewire, Avalara pre-AI (and maybe it's them post AI...)
3/ The pattern I'm liking is a split: a spec some domain expert can audit but can be executed and checked by AI. Catala does this for tax. Tax experts/lawyers can review the encoding line by line against whatever statute, then a compiler handles the rest. Cadence/Synopsys is effectively this in in chip-design, and why they're such a standard. Vanta, Drata for verification in the digital domain. Where are others?
4/ Frontier models will be strong judges over time. But high-stakes domains need the verdict as close to "human-auditable" as possible. And there are many domains where the verdicts exist in an external authority (think FDA, automotive safety, robotics/consumer electronics standards, construction permitting). We mostly got there on SOC2/ISO27000, but there's a long catalog of standards in these physical domains. Tuesday's election only strengthens those gates I'd guess.
AI diffusion has a ways to go. But big rewards in solving the above in large, complex domains.
1/ If code is the modality that drives AI into more categories, how do we pull these non-technical domains up into a coding domain? Seems it’s two problems: one, "transpile" the subject matter into something machines can reason about (hard), then two, deliver a trusted verdict (harder). Said differently we might need to bring vertical domains to the code, as opposed to wedging code into the vertical domains.
2/ The applied AI companies doing both appear advantaged in the long run. But rollouts are slower for sure, i.e. lots of integration work, building context, then re-modeling internal operations around a new capability. But once a customer runs its operations against your spec, switching costs go way up. You become an internal standard, and eventually an industry standard. This was Veeva, Guidewire, Avalara pre-AI (and maybe it's them post AI...)
3/ The pattern I'm liking is a split: a spec some domain expert can audit but can be executed and checked by AI. Catala does this for tax. Tax experts/lawyers can review the encoding line by line against whatever statute, then a compiler handles the rest. Cadence/Synopsys is effectively this in in chip-design, and why they're such a standard. Vanta, Drata for verification in the digital domain. Where are others?
4/ Frontier models will be strong judges over time. But high-stakes domains need the verdict as close to "human-auditable" as possible. And there are many domains where the verdicts exist in an external authority (think FDA, automotive safety, robotics/consumer electronics standards, construction permitting). We mostly got there on SOC2/ISO27000, but there's a long catalog of standards in these physical domains. Tuesday's election only strengthens those gates I'd guess.
AI diffusion has a ways to go. But big rewards in solving the above in large, complex domains.
