Skip to content

← Writing

engineering

OpenAI DevDay, GPTs, and the Five-Day Firing Saga

· Jerwin Arnado

Archive note: this is a backdated post, written years later while rebuilding this site. It’s dated to the moment it covers, but the hindsight is real.

November at OpenAI, in two acts that should not belong to the same month.

Act one, November 6 — DevDay. OpenAI’s first developer conference ships the platform play: GPT-4 Turbo (128K context — an entire codebase’s worth — at a fraction of the old price), the Assistants API (threads, retrieval, code execution — half the agent plumbing everyone hand-rolled after AutoGPT, productized), and custom GPTs with a store coming — app-store economics aimed at chatbots.

Act two, November 17 — the firing. The board removes Sam Altman with a cryptic statement (“not consistently candid”). Then, across one surreal weekend: the interim CEO declines to stay quiet, a second interim CEO arrives, Microsoft announces it will hire Altman and anyone who follows, and over 700 of ~770 employees sign a letter threatening to walk. By November 21, Altman is back, the board that fired him mostly gone. Five days, start to finish.

What DevDay means for builders

The price-and-context improvements are straightforwardly great. The strategic layer deserves more care: GPTs and the Assistants API are platformization — OpenAI absorbing the wrapper layer. Every startup whose product was “ChatGPT plus a system prompt plus your documents” watched its moat get keynoted out of existence. The pattern is as old as platforms (Twitter’s API massacre has its own post): build adjacent to the platform’s obvious roadmap and you are doing their R&D for free. Build something with independent data, distribution, or domain depth, and the platform’s growth is your tailwind instead.

What the saga means for everyone else

I care less about the palace intrigue than about what the week demonstrated:

  1. The governance was the product claim, and it failed visibly. OpenAI’s whole unusual structure — nonprofit board over capped-profit company — existed to prove safety could outrank money. The one time the board exercised that power, it was reversed in five days by employee equity and a cloud contract. Whatever your doomer-vs-builder priors, the empirical result is in: the steering wheel everyone pointed to is decorative.
  2. Microsoft won without moving. Hosting the models, holding the compute, and credibly offering to absorb the entire staff mid-crisis — that’s not partnership, that’s gravity. The week revealed who actually holds the keys, and it isn’t the board, and it isn’t the lab.
  3. For those of us building on the API: vendor risk got a face. For one weekend, the platform under thousands of products had no CEO, a mutinous staff, and an unclear future. My supply-chain rules apply unmodified: abstraction layer over the model API, a tested fallback (the local-model floor exists for exactly this reason), and no architecture that dies if one vendor has a bad week. November was the fire drill; it cost nothing. The real fire will not.

Filed for the record

The frontier of computing spent a weekend being decided by group chats and a sign-off from Satya Nadella, and on Monday everyone went back to shipping. The technology remains remarkable; the institutions around it remain improvised. Both facts in the ledger, as usual — and my integration code now has a MODEL_PROVIDER env var it didn’t have in October.