Finbarr Taylor

Visiting Partner @ycombinator. Founder & former CTO & CEO @Shogun. Angel investor in 60+ companies. Startups, the universe and everything. 🏴󠁧󠁢󠁳󠁣󠁴󠁿

[Finbarr Taylor @finbarr]

@t_blom Awesome. Think I'm gonna get one. I bought a Hyundai IONIQ 5 and it broke down within a month. Currently going through the lemon law process 🫠

[Finbarr Taylor @finbarr]

@agupta This is going to be great! Atoms > bits.

[Finbarr Taylor @finbarr]

Oh, the tokens you'll blow!

[Finbarr Taylor @finbarr]

Looks like Mythos performs about the same as Opus 4.7 on claude code tasks:
https://t.co/ojFdjjb7KX

[Finbarr Taylor @finbarr]

@yuris Never set them to start at that time, but sometimes a 45 minute meeting is useful, e.g., for interviews. You can often stretch to an hour if needed or stop at 30 minutes without it being super rude.

[Finbarr Taylor @finbarr]

@mynameisyahia Congrats Yahia! Looking forward to working with you 😀

[Finbarr Taylor @finbarr]

@jaltma VGT has done better (~8.5x) with lower fees (0.1 vs 0.2). Has some similar exposure but purer tech composition.

[Finbarr Taylor @finbarr]

SF is undefeated.

[Finbarr Taylor @finbarr]

@albertorosasg @ycombinator Awesome! Excited to play a small part here.

[Finbarr Taylor @finbarr]

Gemini has quietly developed a way better personality than ChatGPT. Just tried it again and was pleasantly surprised by its warmth and conciseness.

[Finbarr Taylor @finbarr]

There are so many cool companies applying to YC. It's such a privilege to meet so many incredible founders and gives me great optimism about the future.

[Finbarr Taylor @finbarr]

@theo I guess if you want to use your Claude subscription to work on the OpenClaw open source project, you're SOL.

[Finbarr Taylor @finbarr]

@elonmusk @PTrubey It’s kind of broken on zipair. There’s an annoying portal that only works sometimes and on some devices.

[Finbarr Taylor @finbarr]

@abdrahmanbid @jonchu @eoghan Personally I would not. I think the entire premise of the thread is asking the wrong question. Investors aren’t a good proxy for customers. Get customers and investors will follow.

[Finbarr Taylor @finbarr]

@jonchu @eoghan Maybe you’re talking to the wrong customers. Maybe you’re talking to the wrong investors. Pretty hard to know what the quality of those conversations is. Any good founder should adapt the pitch and the audience as they go along.

[Finbarr Taylor @finbarr]

@DennisonBertram This is why I built https://t.co/vu7AeN73Kt

Have your —dangerously-skip-permissions and eat it 🎂

[Finbarr Taylor @finbarr]

@pranavmishra28 @garrytan If no relevant prior execution then the big model goes through the discovery phase and generates a new execution plan

[Finbarr Taylor @finbarr]

Well you obviously need some safeguards in the system. Error handling. Self repair mechanisms. The loop is probably like this:

- user request
- prior requests are scanned with vector search and LLM rerank with cheap model to find the most relevant ones
- first large model call is prefilled with relevant prior conversation and execution and outcome
- it can view the script and decide whether to re-execute, repair, discard, start fresh
- of course you have error handling with repair mechanism if the script fails

[Finbarr Taylor @finbarr]

@MatthewBerman @garrytan Ha, way too kind Matt - just trying to keep up with everyone else. Appreciate you 🙏

[Finbarr Taylor @finbarr]

@garrytan Also love what you’re doing here, Garry! Keep it up 😀

[Finbarr Taylor @finbarr]

@garrytan Exactly. A lot of tokens are wasted re-learning how to do the same task over and over again. Agents can maintain a library of previous requests alongside learned deterministic routines, whether skills or code files.