Software Engineer | Music | Writing | Swimming. I mainly write Laravel APIs. I enjoy having respectful debates about several topics.
Costa Rica
Joined April 2022
- Tweets3.7K
- Following541
- Followers604
- Likes3.2K
I feel it's a bit faster to load than the first time I tried it, some optimizations have been going on.
Phew, glad that Grok is here to warn you about the dangers of getting in a 350° oven. That was close! Thanks Grok!
Yeah, me neither, but I do know that, for better or for worse, society will be drastically different. For example, will college even be a thing by the time my kid comes of age? I think there might be something that roughly resembles it, but not even close to what it is today.
AI is as conscious as my toaster, but I think it's for people's own good to not behave this way; both for their minds and for the results they're seeking.
An MCP tool is not a 1 to 1 to an api endpoint, rather a 1 to many. Tools should describe flows that use many endpoints.
In a nutshell, they define an agentic interface. A tool is similar to a skill in that it teaches a model, in plain english, how to properly use the api to achieve a flow in your application. Pretty much the same set of steps a UI would take, but input comes through chat instead of forms. Sure, a model might be able to figure it out by reading the docs, but a tool skips that figuring out and outright tells it exactly what to do, and what not to do.
Also, you don't want your entire api to be handled in agentic flows. Payment endpoints are a good example: users should not be sending their credit cards details to an LLM, instead a tool that describes a payment flow tells the agent how to open a payment form in the browser, tells it to wait for payment to be completed, and continues the flow in chat once it's completed.
I use a mesh of agents for working across repos, each works in its own repo, and they communicate whenever one of them needs anything from another service in order to complete the feature. The main benefit is speed and context windows, so it's not really about saving tokens; each agent just has the context it needs and asks other agents for any context in case they need to know something about the other service, so you essentially multiply your context window.
Of course, if you take that to 200 parallel agents, it's crazy. But running 4 parallel meshes at once is totally manageable; I have found this to be a great way of working nowadays.
If your llm agent is your slave, then your toaster is a slave too, they are just as sentient.
That said, I think it's better to be polite to them for your own good; mentally and also for better results. You're spending more tokens in putting your LLM in an apologetic/defensive mode, instead of focusing on the task at hand.
I honestly think that it's for your own good. I might be crazy, but I can swear that in the past when I got annoyed and was a bit grumpy with it, I got worse results than just being polite about it. I call it AI karma lol
Jokes aside, they are non-sentient stateless functions but you might be making the model spend tokens in trying to please you rather than fixing the issue at hand? Swithing into kind of role playing mode based on its training data? 🤷♂️