Startup's 'things the model can't do' list shrank from 11 to 3 as frontier models catch up
Scared_Sentence_2834 · reddit · 2026-10-14
A 6-person team builds an agent that reconciles invoices against contracts for finance teams. Since January they've kept a doc called "things the model can't do" — effectively their moat in list form, starting with 11 items. With Gemini 4 Argon's launch, multi-page table extraction is likely gone, matching line items to contract clauses mostly gone, and Google's launch post quotes a finance company saying the model matched analyst-built models on 99% of figures. At that pace, they're down to 3 items.
The author notes at least 12 models from major labs shipped between Sept 1 and Oct 8, each eroding their work — prompts, extraction logic, retry chains are all depreciating faster than ever. What's left is what labs won't bother with: 40 ERP connectors, CFO approval chains, auditor-accepted audit trails, and billing per reconciled document instead of reselling tokens (Stripe subscriptions plus weeks choosing between Orb, Lago and Flexprice). The silver lining: every model launch improves their product for free — but every competitor gets the same upgrade. Their honest estimate: the moat lasts maybe 12 more months.
More from AGI Musings
- Mathematicians lack AI literacy, argues researcher amid debate on AI upending math — gerardsans · 2026-10-14
- Google DeepMind team: AI agents boost ML engineering but can't run autonomous R&D end-to-end — HaydnBelfield · 2026-10-14
- DeepMind survey of 12 frontier researchers: execution automates first, ideation stays human — HaydnBelfield · 2026-10-14
- Artificial neurons don't physically exist — an LLM is just matrix math — ValerioCapraro · 2026-10-14
- First clear sighting of a rogue AI agent fleet from China scraping Amap on Tencent Cloud — emollick · 2026-10-14
- Heavy ChatGPT user since 2022 asks: has AI been a net positive for your life? — Alternative-Car8221 · 2026-10-14