Recent searches
Search options
I perceive OpenAI as far more precarious than popular imagination, and even their valuation, would suggest
Everything is built on first-mover advantage. The product itself is really quite corny. Just bloodless, loveless, unpolished. A prototype that escaped the lab on the strength of powerful leverage that grows easier to duplicate
Its generality will be its downfall, as specialized tools serve subsets of its customers much better.
https://www.ft.com/content/7ca691f9-e4c7-46e8-bbc2-fe95464934f8
this post brought to you by the obnoxious, two-stage login process that it constantly requires I do
@danilo after the 10000000th time getting logged out i just downloaded a tool that lets me chat with it via my API key, and lets me use Claude too (which I've mostly moved to) https://www.typingmind.com/
@keith oh no it's subtly uglier 
I wonder if there's a desktop app that also does this...
@danilo they've got a desktop app! it's absolutely uglier lol
@keith @danilo i use a self-hosted thingy called LobeChat https://chat-preview.lobehub.com/chat?session=inbox. It supports API keys for ... well a lot of providers, including OpenAI, Anthropic, Google Gemini, Perplexity, Bedrock, Mistral, and OpenRouter (also a few Chinese ones and a US-located one I won't mention here). It's wildly flexible - i self-host it using runtipi.io and basically don't think about it. for my uses it's just a webpage and could probably run on a RaspberryPi.
@cori @danilo hahah, I admittedly get it via a subscription (https://setapp.com/)
@danilo I find it pretty rude that it asks me to prove I’m a human whenever I try and log in.
@danilo I will say: the whole chatbot experience was a *superb* marketing mechanic. People could project a lot into the interaction, and the fine (un)reliability of the product was obscured by the natural meandering nature of a chat conversation.
But I find it telling how struggling the engineering of embedded applications depending on language models has been going. And that’s where I perceive more of the actual engineering happening, slower and steadier.
@Sevoris this is a great point. The chat interface is fundamentally bad long term, but makes it very easy to onboard. More fuel for the escaped prototype argument!
@danilo Yeah. I mean, the whole „unstructured“ bit is quite nice, but I see few reasons to really suppose that the whole „predict the next text token“ objective has bootstrapped an actual, direct able intelligence like some propose, and certainly nothing about this overall system design makes for *replication*, *logging* and *debugging* into the processing that is going on. Besides the data issue (keyword RAG) the obscurity of processing is just a mess.
@danilo For another example: we hyped up chatbots, and then Microsoft developed LangGraph and we basically reinvented the state graphs every chatbot design system since 2010 has had *again*, but this time with a temperamental blackbox the graph tries to restraint somewhat with a chain-of-thought -esque argument. The fact that you now need such graph structures says, I feel, more than enough about the nature of services design for language models.
@danilo "bloodless, loveless, unpolished. A prototype that escaped the lab" -- Love this! And thanks for your recent thought-provoking blog posts.
@kjrsten thanks so much for reading!
@danilo Paywall jumper: FT on Open AI's "GPT Store" https://archive.is/63dWw