Colin Flaherty

As people think about the future of the internet, I hear lots of complaints about the omnibox and how text based chat is too limited. Comparisons are made between the drab ChatGPT UI and the more colorful experiences one can find in the pinnacle of Web 2.0 websites and apps. I…

I think default RCS on iOS is so underpriced rn. Seeing typing awareness and read receipts when I text green numbers is quickly neutralizing my emotional attachment to blue messages, faster than I imagined. And I'm usually a luddite about these things

IN NEWS: @slingshotai_inc raises $93M for AI Therapy. We asked founders @neilparikh and @danielreidcahn about how they're reinventing therapy. "We're inventing a whole new modality of therapy. Crises might happen at 2 AM, right? Your therapist isn't going to answer the phone."…

Let's go daniel!

I see a world in which any startup can submit their RL environment to a big lab to fine tune a custom agent on. This would unlock a lot of user experiences for users that foundation labs are currently structurally unable to deliver on

building interactive AI can be challenging because there is a long-tail of interactions the user will try, and you need to think through enough of them to effectively guide the user onto the right path(s) without degrading the overall product experience. I remember back in the…

it's so crazy to me that OpenAI was only at $30M ARR in 2022. Times change so fast

If everyone has their own AI generated content engine (e.g. their own TV shows to watch), we’ll need mechanisms to enable “crossover episodes” between your content and mine. Maybe it’s a shared storyline, shared characters, shared aesthetics, or a subtle allusion

the proliferation of ai content across the web could manifest in two ways (likely both): - just-in-time generation - reccomendation algorithms that need to sort through 10,000x more content

For many applications, the future of RL training environments is just releasing agents on real users in a real-world RL loop, not creating simulated environments. Subsidize with VC dollars and make it really cheap to use. Users won't care if it only works N% of the time then.…

the soham joke stopped being funny when every startup and vc under the sun started using it farm impressions :(

innovator's dilemma is fundamentally an analysis of managerially run companies, not founder-led companies. live player theory is more useful when forecasting the moves of incumbents like openai and meta. innovator's dilemma is useful for thinking about apple though

converting threadpools to async code is a cathartic experience

I'm wondering if I should be pre-emptively setting up my fine-tuning stack for when OAI drops their open source model

I’m very convinced we are heading towards a world where we are changing the substance of media itself. It's not static. It's alive. Stories that recast themselves, that rewrite their own scripts in the moment, just for you. A story that learns and grows with you. A conversation.

notes from @colin__flaherty's talk on rag x agents