Mo pfp

Mo

@meb

624 Following
644 Followers
Mo pfp
Mo
@meb
How I would define an AI agent at the start of 2025: In abstract, a piece of software that can achieve a certain goal, and uses AI to navigate the fuzzy path of achieving that goal. More specifically, a combination of - Goal - Context management - Underlying model - Tools (datastores and functions with effects) - Prompting & self-prompting strategies (including delegating to other agents) Example: A geocoding agent Goal: Parse content from a sentence and geocode it Context Management: Trimmed sliding window. We only need the very most recent messages to get the job done. Underlying model: A simple model like 4o mini is more than enough Tools: A function to search cities with inbuilt query match ranking, and a function with access to a Google Maps geocoding API Prompting: Single turn. The agent should be able to get the job done in a single exchange, or ask the user to clarify the location if needed
0 reply
0 recast
8 reactions
Mo pfp
Mo
@meb
Humans write copious amounts of code which makes the LLM better at writing such code. LLM writes code that compiles to human interactable interfaces. Human, inspired by the LLM writes more code that trains the LLM. And the eternal cycle happens
0 reply
0 recast
1 reaction
Mo pfp
Mo
@meb
Is this the western LLM version of asking Deepseek R1 about what notorious historical event happened from April to June of 1989 in Beijing?
0 reply
0 recast
1 reaction
Mo pfp
Mo
@meb
I can see Typescript React as being one of the main communication languages between machines and humans for rendering rich experiences
1 reply
0 recast
0 reaction
Mo pfp
Mo
@meb
Who is buidling in GenAI on Farcaster?
1 reply
2 recasts
1 reaction
Mo pfp
Mo
@meb
Scrumban
0 reply
0 recast
1 reaction
Mo pfp
Mo
@meb
Nice, just checked out the site and that easy to use generator piece is definitely one thing they're missing. Curious on legal front, what were the areas of improvement you found?
0 reply
0 recast
0 reaction
Mo pfp
Mo
@meb
Have you heard of oneNDA? Ongoing project for a few years now that’s been worked on by many organisations across the globe. Currently though, it’s a signable word document, with a better experience locked behind some legal SaaS https://www.onenda.org
1 reply
0 recast
0 reaction
Mo pfp
Mo
@meb
Did they fix the filesystem performance issues? I remember a few years ago things like npm CI would take forever, due to having some extra transport layer in the middle
1 reply
0 recast
1 reaction
Mo pfp
Mo
@meb
Robots add in a layer of physical risk, so I'd definitely be concerned about it. That said, I'm definitely up for an AI companion that helps them in their learning journey
0 reply
0 recast
0 reaction
Mo pfp
Mo
@meb
Which app are you referring to?
0 reply
0 recast
0 reaction
Mo pfp
Mo
@meb
Fun recent observation. It’s absolutely superb at local optimisations, such as improving a negotiation email. But pretty bad at proactively taking a step back and rewriting the approach entirely without additional prompting. I think this is where the “wrapper” layer comes in, and allows us to cristiallise some of these thought patterns based on the specialised area an agent deals with
0 reply
0 recast
1 reaction
Mo pfp
Mo
@meb
The other skill devs need to learn in this age; Marketing
0 reply
0 recast
0 reaction
Mo pfp
Mo
@meb
System prompt & evals is the new “here’s your employee handbook, and examples of previous work, along with the manager that will QA your work”
1 reply
0 recast
1 reaction
Mo pfp
Mo
@meb
ORM vs raw SQL all over again
0 reply
0 recast
0 reaction
Mo pfp
Mo
@meb
Answer: Figured out langsmith is a good way to go. It provides observability on your calls, and pricing data across model providers
0 reply
0 recast
1 reaction
Mo pfp
Mo
@meb
This is one of my favourite civilisational aspects of AI. It democratises access to “pretty good” human like support that would otherwise be costly and a waste of money. Good results do require more active involvement than therapy though, but I like that I can ask ChatGPT to call me out on my BS
0 reply
0 recast
1 reaction
Mo pfp
Mo
@meb
I would have argued the opposite. A lot of therapists are just maxing sessions and replaying things. Whereas ChatGPT allows you to go very deep, guide the conversation, and get hard truths instead of veiled suggestions. There’s a great book by Richard Bandler, the OG creator of NLP, where he discusses this exact point, that his therapy consultations were aimed at reaching those brain transition moments in a single session
1 reply
0 recast
0 reaction
Mo pfp
Mo
@meb
This is a great post. People are obsessed with AI arms race. Meanwhile, gpt 4o mini or Gemini 2.0 flash are good enough already to handle an overwhelming number of use cases. People are so focused on having a phd super intelligent model, that they forget many tasks would just need a mid human
1 reply
1 recast
1 reaction
Mo pfp
Mo
@meb
My love for nature and fresh air is one of the few things that exceeds my appreciation of ChatGPT
0 reply
0 recast
1 reaction