Michael Ewens

Finance prof+co-director PE Program @Columbia_Biz. RA @nberpubs. Quant. Advisor @correlationvc. Co-organizer @workshopefi. Code:

[Michael Ewens @startupecon]

@nickgebbia @arpitrage "Hey Claude, rebuild https://t.co/L8ZJcbsm92 for me. I want to make the questions less generic. See my paper in /paper/my_paper_in_md.md for ideas for topic questions."

[Michael Ewens @startupecon]

@deMouraBernardo @AdrienMatray @cblatts No one is denying that. I was questioning "GPT Pro" and all this language confusing model and subscription. That said, the API has GPT 5.5 Pro, so that workflow (with a budget :) ) is avoidable via CLI.

[Michael Ewens @startupecon]

Amazing multiple AIs use case: Claude talks to Codex about a plan.

I'm rebuilding complex data and often send plans b/t the AIs. In Cmux (see below), instead tell Claude, "Ask Codex next to you for feedback". You can tell Claude to just check every 2 minutes, w/o pasting

[Michael Ewens @startupecon]

@deMouraBernardo @AdrienMatray @cblatts There you go!

[Michael Ewens @startupecon]

@deMouraBernardo @AdrienMatray @cblatts The confusion is the labeling and that workflow is nuts for a small lift to 5.5 base:

GPT Pro = Subscription
GPT 5.5 = Model
GPT 5.5 Pro = fancier 5.5 (below?)

The system card shows some differences, but 5.5 Pro is on the API (CLI), not Codex. Weird.

https://t.co/Yp4PmvT2E3

[Michael Ewens @startupecon]

@AdrienMatray @cblatts Online model? I had Pro and could use GPT 5.5 High anywhere (browser, CLI, etc).

[Michael Ewens @startupecon]

@deMouraBernardo @AdrienMatray @cblatts From ChatGPT Chrome! (sorry, I am just triggered by the possibility they are offering different things):

[Michael Ewens @startupecon]

@deMouraBernardo @AdrienMatray @cblatts That is, when I paid for GPT Pro, I could use Codex CLI or the app. Not sure what you are referring to. Codex is a wrapper that is suited for "computer use", so my read is "a different harness." Show me some receipts so I know what I am paying for.

[Michael Ewens @startupecon]

@deMouraBernardo @AdrienMatray @cblatts GPT Pro is a subscription level. GPT 5.5 High/Med etc are models available to all layers that I use. Tell me what I am missing! Are you saying ChatGPT Pro is a different model? https://t.co/hrLID6igSc

[Michael Ewens @startupecon]

Reminds me of the 1st time I earned something from my economics training. Vonage (VoIP provider) offered me, a customer and 4th-year college student earning <$15K/year, shares in its IPO. It was a clear signal that smart $ was not interested. Then this happened:

[Polymarket @Polymarket]

NEW: Fidelity lowers the minimum account requirement for the SpaceX IPO from as high as $500,000 to just $2,000.

[Michael Ewens @startupecon]

@arpitrage Multi-division firm with a single profitable space-internet provider segment.

[Michael Ewens @startupecon]

@AdrienMatray @cblatts Codex == GPT 5.5, no? Codex is just a CLI wrapper / harness on GPT. Is Chrome involved b/c you are using the webapp?

[Michael Ewens @startupecon]

This event follows the inaugural event in January '26, The 2027 workshop will be held on Saturday, January 2, 2027, in Washington, DC, with 20 senior scholars serving as mentors.

Info: https://t.co/51VuDyOQKO

Apply: https://t.co/ZDigYnsows

[Michael Ewens @startupecon]

I'm pleased to announce the 2nd American Finance Association Junior Faculty Mentoring Program (JFMP) event, which I am co-organizing with Audra Boone.

The program provides mentorship, research feedback, and professional development to junior faculty in finance. 🧵

[Michael Ewens @startupecon]

Sessions cover research development, publishing strategy, tenure navigation, and the broader challenges of an academic career in finance. Beyond the workshop itself, participants join an ongoing cohort network that continues after the meeting. 3/N

[Michael Ewens @startupecon]

@AdrienMatray Yes, that is expected...they are both instructions. I believe it "listens" to the project first. If they are both long, things go wrong:

https://t.co/gzeKcGGf5f

[Michael Ewens @startupecon]

@AdrienMatray That is the default behavior: always reads the global Claude.md and then reads any claude.md into context that is in the working directory. I don't believe you need to make settings changes for that.