Thanks for this post. I've been following your posts on X, and I didn't know you had a Substack! Great to know.
By the way, as a non-coder, have you experimented with the coding tools from the frontier labs for non-coding purposes? I know the coders have found ways to get these tools to do amazing things for non-coding tasks, but it seems hard to find a way into the hashtag-and-ampersand thicket. A guide based on your own experience would be amazing.
AI intern during 2026 would involve a significant speedup though, yes?
Take Opus 4.5 on METR, it can do ~5 hour tasks and the doubling time is 7 months. A reasonable estimation for an easy intern project is… oh let’s say 2-4 weeks. A hard project might take 3-6 months. Thus, with *exponential* extrapolation of current trends (which is always risky), this predicts that a mediocre AI research intern will be achieved in ~5 doublings, and a good intern in ~7.
That is, the aggressive (though not super-exponential) predictions should be ~3 years until an AI intern can do a 2 week project, and another year or so until it can do a 3 month project, ~autonomously. And of course, the ~3 month project would get done faster because computers think faster.
I wouldnt be surprised if an AI-augmented intern was exceptionally productive in 2026 though—i feel as though my productivity has increased enormously from deploying agents even for non-code tasks.
I'm actually working independently on a PoC of a vibe coding tool for contracts, trusts and the like, and intake for attorneys. Might have to add that specific use case in. Good stuff!
Built what you described with Claude code. Took maybe two hours of work. Called it SPA Vibe Editor — "vibe coding for SPAs."
Core features:
Natural language section editing via Claude ("add knowledge qualifier to IP rep")/legal research and precedent via GPT 5.x
5 template variants (ABA balanced, buyer-friendly, seller-friendly, tech/startup, healthcare) (only put in a few sections for now to keep it light, e.g. defs, sale and closing, reps and warranties, and Indemnification)
Deal terms auto-populate throughout
Redline export with track changes
PDF annotation import (pull in marked-up comments)
One-shot mode for when all terms are locked upfront
Version history + share links
You nailed the insight: one-shotting 100 pages isn't impossible, it just requires frontloading all the planning lawyers normally do iteratively. Vibe coding matches how deals actually evolve. Try it: spa.alexgallefrom.io
(let me know if it's any good, I haven't been involved in an SPA since I was an early associate (PWP/T&E mostly recently)
Thanks for this post. I've been following your posts on X, and I didn't know you had a Substack! Great to know.
By the way, as a non-coder, have you experimented with the coding tools from the frontier labs for non-coding purposes? I know the coders have found ways to get these tools to do amazing things for non-coding tasks, but it seems hard to find a way into the hashtag-and-ampersand thicket. A guide based on your own experience would be amazing.
AI intern during 2026 would involve a significant speedup though, yes?
Take Opus 4.5 on METR, it can do ~5 hour tasks and the doubling time is 7 months. A reasonable estimation for an easy intern project is… oh let’s say 2-4 weeks. A hard project might take 3-6 months. Thus, with *exponential* extrapolation of current trends (which is always risky), this predicts that a mediocre AI research intern will be achieved in ~5 doublings, and a good intern in ~7.
That is, the aggressive (though not super-exponential) predictions should be ~3 years until an AI intern can do a 2 week project, and another year or so until it can do a 3 month project, ~autonomously. And of course, the ~3 month project would get done faster because computers think faster.
I wouldnt be surprised if an AI-augmented intern was exceptionally productive in 2026 though—i feel as though my productivity has increased enormously from deploying agents even for non-code tasks.
I'm actually working independently on a PoC of a vibe coding tool for contracts, trusts and the like, and intake for attorneys. Might have to add that specific use case in. Good stuff!
Built what you described with Claude code. Took maybe two hours of work. Called it SPA Vibe Editor — "vibe coding for SPAs."
Core features:
Natural language section editing via Claude ("add knowledge qualifier to IP rep")/legal research and precedent via GPT 5.x
5 template variants (ABA balanced, buyer-friendly, seller-friendly, tech/startup, healthcare) (only put in a few sections for now to keep it light, e.g. defs, sale and closing, reps and warranties, and Indemnification)
Deal terms auto-populate throughout
Redline export with track changes
PDF annotation import (pull in marked-up comments)
One-shot mode for when all terms are locked upfront
Version history + share links
You nailed the insight: one-shotting 100 pages isn't impossible, it just requires frontloading all the planning lawyers normally do iteratively. Vibe coding matches how deals actually evolve. Try it: spa.alexgallefrom.io
(let me know if it's any good, I haven't been involved in an SPA since I was an early associate (PWP/T&E mostly recently)