Marco Di Cesare

Founder of @luminafrontier, the AI marketing platform for luxury hotels. Zurich.

[Marco Di Cesare @marcodice_ai]

OPUS DOESN'T NEED TO MAKE EVERY DECISION IN YOUR AGENT

Jev found 51 of 54 instructions buried in 30,143 lines of a call transcript for $0.32.

I published the tests and a prompt to audit your agent. https://t.co/cjQUagj9ue

[Marco Di Cesare @marcodice_ai]

this is dope

[Marco Di Cesare @marcodice_ai]

I put Jev in front of an AI writer to save tokens.

The writing got worse

Then I gave it the small decision *before* the writing. If you're building your first agent, this is the job I'd test first https://t.co/ZbU4ioAxro

[Marco Di Cesare @marcodice_ai]

@higgsfield_ai @gregisenberg This is amazing! Can we also get “same building all the way through” button? I make videos of real places and that's still the thing I fight most 😅

[Marco Di Cesare @marcodice_ai]

@seroundtable Imagine spending months deciding exactly what goes on your hotel's website, then Google puts an AI Overview beside your name and writes the answer for you

[Marco Di Cesare @marcodice_ai]

@karrisaarinen @linear Can it switch models mid-task? A lot of my work starts as a messy research question and ends in code. Sometimes the model I’d pick at the start isn’t always the one I want finishing it

[Marco Di Cesare @marcodice_ai]

@vral @mrhaikuswan Are you calling TypeSafe directly? I hit a 30 requests/min limit through LLM API tonight. Cost wasn’t the bottleneck at all

[Marco Di Cesare @marcodice_ai]

Testing Jev tonight on 936 multilingual intent-routing examples.

It got 65% right, while assigning its top answer 80% probability on average.

It’s fast as hell. That gap is what I’d want to understand before letting it make decisions inside my agents.

[Marco Di Cesare @marcodice_ai]

@marclou Very true, was about to buy some hyped jev domain last night. l’m looking at Jev right now but want to find one place in the agents I already run where it actually replaces a full LLM call. Seems like a better starting point than building a new wrapper

[Marco Di Cesare @marcodice_ai]

@lilyraynyc I can see that someone landed on a page from ChatGPT. I still can’t see what they asked or what answer made them click. That’s the report I’d actually want!

[Marco Di Cesare @marcodice_ai]

@muratcan Did Opus see ground-truth frames of the location, or only the prompt and output? I’m making videos of real hotels; a shot can look physically plausible and still be the wrong building!

[Marco Di Cesare @marcodice_ai]

@tiangolo I posted about this today. The v1 can be impressive and still be useless for the final user. I’m much prouder of v13, when it finally survives the actual constraints.

[Marco Di Cesare @marcodice_ai]

Can someone explain me what is this "one shotted" or "one prompt" obsession?

So proud of my "_v13" iteration that is actually working and perfect for my prod environment or for my clients

[Marco Di Cesare @marcodice_ai]

@joyrap Taking all of it! Thanks Bill

[Marco Di Cesare @marcodice_ai]

Still can't believe I left my 9-5 very well paid job 3 months ago to found a startup, and I am still with no salary, spending my savings and working nonstop.

back to X and want to start seriously building in public

[Marco Di Cesare @marcodice_ai]

@akbuilds_ Thanks mate!

[Marco Di Cesare @marcodice_ai]

@BabaDev_ amen

[Marco Di Cesare @marcodice_ai]

I used to be a GPT and Codex fanboy. Now I am just a reset button fanboy

Thank you Opus 5.5

[Marco Di Cesare @marcodice_ai]

@gauravsbuilding very true, still a part of me is skeptical in running them too early

[Marco Di Cesare @marcodice_ai]

Opened my Codex stats: 35.1B tokens.

I can get AI to execute almost anything when I already know what I want. But I sat down to write a video and got stuck on the part before the prompt: what the hell is it actually about?