Mads

Product | AI | Applications | Builder. Curious & eclectic. MIT Sloan MBA.

[Mads @mitmads]

AI saved your team 11 hours a week. Most came back as hidden work — 6.4 hrs/wk checking and fixing the model.

87% use AI. Only 13% of orgs perform better.

The bottleneck was never adoption. It's verification.

[Mads @mitmads]

2/2
The trap: both plateaus feel identical from the inside. One means you built something real. The other means you deferred the architecture and the bill is due.
The tell is where the friction lives. The full 2x2 and how to read it: https://t.co/zAN3rLi9wj

[Mads @mitmads]

1/2
Your vibe-code velocity slowing down is not the warning sign everyone says it is.
It is the product talking back: you have built enough coherence, and enough real usage, that adding features now carries cost. Velocity was free early; nothing was load-bearing. Now everything is.

[Mads @mitmads]

[8/8] Over the next decade, only one of those is worth owning.

Full argument: https://t.co/00eMUApOkH

[Mads @mitmads]

[1/8] Most knowledge tools remember what you read. Almost none keep a track record of whether you were right.

That second gap is the expensive one.

[Mads @mitmads]

[7/8] That's the asset a knowledge graph can't hand you: a track record of your own judgment. Where your instincts were sharp, where they were reliably off, with receipts.

Build for recall, you get a library. Build for judgment, you get something that compounds.

[Mads @mitmads]

https://t.co/6CUboZbsrn

[Mads @mitmads]

https://t.co/JrocgQLwwU

[Mads @mitmads]

Three vendor pricing pivots in three weeks, at the three largest AI distribution channels.

April 3: Anthropic announced it was banning OpenClaw from Claude subscriptions and updating ToS to restrict OAuth to Claude Code and https://t.co/KMDSD4FNYf only.

April 9: OpenAI launched its $100 Pro tier, formalizing pricing convergence at $100-200 per developer per month.

April 27: GitHub Copilot moved to usage-based AI Credits, June 1 cutover. Every seat fee maps dollar-for-dollar to credits at "published API rates," meaning the seat captures zero inference margin. Fallback to lower-cost models is gone. Code review is now double-metered against both AI Credits and GitHub Actions minutes. Annual subscribers face mid-contract repricing via "model multipliers will increase June 1."

CPO Mario Rodriguez stated it plainly: "GitHub has absorbed much of the escalating inference cost behind that usage, but the current premium request model is no longer sustainable."

GitHub's pricing pivot isn't just about Copilot's economics. It's the largest enterprise software distribution channel in the world conceding that AI inference cannot be priced through a seat license. Cursor moved to usage-based last June. The labs and distribution channels are now catching up.

All-you-can-eat inference might be over soon. Every enterprise AI procurement model from this point has to be rebuilt around variable cost per task, not predictable cost per developer. The next 18 months is the unwind.

For CFOs: your AI line item is about to look like cloud spend. Pool budgets, cost-center allocation, overage policy, vendor diversification — these are now table stakes, not roadmap items.

For AI vendors: if you're still shipping flat-rate AI features inside SaaS contracts, your gross margin shows up in the next earnings cycle. The companies that figure out FinOps-for-AI now will be the procurement standard by 2027.

[Mads @mitmads]

https://t.co/n031idm7gm

[Mads @mitmads]

The lesson from Waymo's 500K rides/week is that raw pixel-to-action models get you to impressive demos, but production-grade reliability requires surrounding your end-to-end model with intermediate representations, a simulator, and an opinionated critic. https://t.co/gUDVCGhcv8

[Mads @mitmads]

The chief scientist of OpenAI (in an intvw) told builders to stop fine-tuning and start engineering context. The 2026-27 enterprise AI strategy isn't "build your own model", it's "rent the frontier and own the context layer." The defensibility shifts from weights to plumbing.

[Mads @mitmads]

Is Anthropic's Managed Agents essentially AWS Lambda for agent infrastructure ?

[Mads @mitmads]

Agents may not replace enterprise software but they'll reprogram the buying decision. When your next 'customer' is an agent evaluating backends on durability and cost parameters, Gartner magic quadrants become irrelevant.

[Mads @mitmads]

Anthropic did in four months what took Salesforce a decade: $9B to $30B ARR. But unlike Salesforce, their revenue is capped by physics — they literally can’t sell more than their GPUs can serve.