I keep getting all kinds of random requests from customers wanting us to implement the latest fad in AI:>"I want to allow AI bots to check-out & pay! implement UCP!(nevermind the customer's shop isn't US based and no one gives their credit card to an AI...)>Can you create an MCP server so our devs don't have to do any work? (nevermind that we could just ingest your API docs and have claude build the integration just as quickly)Is any of this shit actually going to be useful, or is it all crypto-type scam shit? If I build an MCP server and set it up and then OpenAI abandons the project because "lul, actually our AIs are so advanced they can build the integration on the fly so MCP is useless" I'm going to be pissed.
seriously how long do you think it'll take you to make that MCP you stupid whore.be like Nike and just do it.
>nobody gives their credit card to AII wonder if there was some kind of digital wallet rails you could use that would keep the credit card a secret and just use some sort of disposable token instead.like a dictionary. the token looks up the card.hell maybe several dumbass companies could create a myriad of the same implementations. we could call them things like. "google pay". or "apple pay". "you're a stupid dumbass pay". that last ones for you.
MCP servers are GAY.
>>109716602What fucking retard wrote that slide? The math only makes sense if you just run the sampler with no model at all.
just have the model directly call your apis man mcps is dead.
>>109716862Function calls are all you need. Just give the model curl (really it just needs a shell.)
>taking screenshots of powerpoint slides that include the squiggly linesI've seen people do this at work but it's consistently the most low IQ people I've seen this do
>>109716879>>109716862The other thing people don't realize is that if you bloat your function list you eat up context for all of those function description/parameter lists etc *every conversation turn* You really want that list to be as absolutely short and general as possible (I don't even give my agents a "write" function, they have to use shell commands to write files.)
>>109716602>(nevermind the customer's shop isn't US based and no one gives their credit card to an AI...)Even in the US this would be a PCI DSS nightmare and instantly get you banned by mastercard/visa
webMCP is a thing now and yes, your customers will be handing over their details to their clankers (securely - openai have already got this in place in the lead up to their always-on agent launch)devs should just use your api, but for consumer-facing shit it will be standard to build an agent-facing interface (via webMCP tools) going forward
>>109716881Tool defs go at the start, after the system prompt, once.A tool def of length n takes up exactly n tokens of context.
>>109716953Oh you're right actually. I don't know where I got the idea it was the other way.
>>109716852no, a good model just makes e smallerit is of course debatable that each sampled token has an independent probability of being bogus
>>109716881I'll take "What is tool search for?" for 500 Alex!
