Real Alpha
Watch modeSubscribed

MTS · 2 min · Oct 19, 2026

Build for AI’s future capabilities, not today’s limits

Felix Rieseberg

Engineering Manager, Anthropic

Leads engineering for Claude.ai, Cowork and Claude Code Desktop at Anthropic; Electron co-maintainer.

  • Theo JaffeeHost · Co-founder & host, MTS

    One of the first members of a16z's new media team; earlier hosted the Emergent Ventures-backed Theo Jaffee Podcast.

  • Sophia DewHost · Host, MTS

About this segment

Felix recalls Boris showing an early Claude Code when the model was still a weak software engineer. The lesson: anticipate where models are headed instead of designing only around current limits.

From “Use AI Like It's Already Superhuman | Felix Rieseberg”, Oct 19, 2026. Watch the full episode

“whatever the capabilities are today just like completely ignore them what would you do if the thing was superhuman”
Felix Riesebergat 11:01 · MTS11:01

Takeaways

  1. Claude Code was explored before the model was a strong coder.
  2. Early products focused more on autocomplete than autonomous work.
  3. Design for what models could do if they were superhuman.
Watch 2:14 from 9:01

In this episode

Full episode
  1. 0:19AI is most valuable when it expands what people can attempt3:42

  2. 9:01Build for AI’s future capabilities, not today’s limits2:14

  3. 12:26AI shifts managers from directing work to enabling it2:55Up next
  4. 3 more moments in the full episode

Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

  1. 9:04

    so i guess like looking zooming out a little bit um like with claude code how did you guys sort of choose i think it was always very like impressive and admirable that there was such a high degree of focused focus on coding ananthropic um so what was like the moment when that was decided to be the direction that was going to change everything? And then what are some of the, I guess, like a timeline of how Cloud Code has advanced?

  2. 9:33

    I think I personally really love that. I think everyone takes different stories, like different lessons from the story of Cloud Code. But Boris first started showing a version of Cloud Code internally, I think September 2024. I think that's right, roughly. And it is remarkable to me because at the time, obviously, the model was not there in any way. it was like a pretty I think it was like Sonic 3.5 it was among most benchmarks a pretty terrible software engineer I think it was like in the 30s for Sweebench and like you know this but like Sweebench right we kind of moved on from Sweebench since and the products at the time were really more focused on sort of this like human in the loop slightly smarter autocomplete like the first versions of cursor those were really the products we were making at the time but I think what

  3. 10:24

    What Boris sort of saw was how the model is going to get there, like a pretty optimistic view of how the model is going to get there. And I think Sonnet 3.7 was already like doubling this revenge score and it started being, still by today's standards, wasn't a great software engineer, but you could see the move from I don't need to start writing the line anymore. The model can autocomplete this function or whatever for me, I don't need to do this.

  4. 10:49

    and the the biggest lesson that I take from this like early time is really so many people when they use AI today it's like really hard to internalize the progress and the speed of the progress yes right it's like so difficult for like most people so the the lesson that I try to teach and learn from cloud code is whatever the capabilities are today just like completely ignore them what would you do if the thing was superhuman exactly yeah yeah that I mean that's