Saxe: models' lack of learning caps SWE gains at ~3x, solvable to 10-100x
joshua_saxe · x · 2026-09-08
- Joshua Saxe responds to Martin Casado's claim that pushing models forward for high-skilled dev tasks no longer makes economic sense.
- Saxe identifies the real bottleneck: models don't learn over time to do what you ask, capping software engineering productivity gains at roughly 3x.
- He argues solving this could unlock 10-100x gains, so continued model progress still carries huge economic value.
More from AGI Musings
- If both VLA and WAM are wrong: three first principles for embodied AI training — qinzytech · 2026-09-08
- Yacine Predicts AI Will Wipe Out Personal Tax Accountants by Next April — yacineMTB · 2026-09-08
- Rationalists push back against 'smear campaign' article on AI risk discourse — GarrisonLovely · 2026-09-08
- Insilico's AI drug Rentosertib shows 3-4 year biological age reversal in 12 weeks — NinaDSchick · 2026-09-08
- Gary Marcus: Declaring AGI before we can build trustworthy agents is total nonsense — GaryMarcus · 2026-09-08
- If LLM scaling won't reach AGI, where exactly does the self-improving researcher loop break? — ImmuneHack · 2026-09-08