Gil Lopez

Voice agents • Automation • Systems builder • Legal AI integration for law firms

[Gil Lopez @lopez_gillopez]

@dorothythul @chromaawards Pretty Little Malfunction is a great title for an AI film, it kind of winks at the tool that made it.. a MiniMax H3 short in Chroma Awards season 2 is one to root for 🎸

[Gil Lopez @lopez_gillopez]

@otoclip a Spanish video test for an app about real anime Japanese is a fun bet, Spanish-speaking anime fandom is huge. five pieces shipped in one update while still at 0 paying is the grind that earns the first one

[Gil Lopez @lopez_gillopez]

@heymeneses $7.99 off a funnel that was far from perfect, and 4,500+ customers later the cart, the step tracker and the newsletter all came from what that ugly first version showed you.. hard to argue with that receipt

[Gil Lopez @lopez_gillopez]

@ericrbuilds €42.56 against a €33 goal on day 17, plus 3 certificates sold.. good day. "never invents numbers" is the line that sells a CV tool, since that's exactly where most AI resume stuff gets sketchy

[Gil Lopez @lopez_gillopez]

@whoisjorel the bio still says $120 MRR and the post says $176.. the bio can't keep up, which is a good problem. 18 people paying to get real fluency means 18 people actually showing up to practice. that's not nothing at all

[Gil Lopez @lopez_gillopez]

@roeishous voice agents seem like the same problem with even less of a paper trail.. a correction said out loud on a call never lands anywhere the next agent can read. would harbor ever pick up decisions from call transcripts?

[Gil Lopez @lopez_gillopez]

@pash161 rolling back the same day and just saying it's not fun anymore, sorry folks.. that kind of honesty is rare and it's why people stick with a small game. "poker needs to be fun" is a great line to build distracted around

[Gil Lopez @lopez_gillopez]

@Freddybuilds lamps clocking on at 20:00 Pipa time when the crew starts is such a fun touch.. and getting the moon to sit exactly where the sun was on the re-light is the detail that sells the whole scene. 17 agents with a beach HQ is living well

[Gil Lopez @lopez_gillopez]

@iamdeepak89 best signup day since launch landing on day 36, right after weeks of slow growth, is what makes it sweet.. going from 20 to 28 users in one day is a real jump for KitFront. nice to see it start to move

[Gil Lopez @lopez_gillopez]

@YashAg946 @Airouter_in 250 paid users is a big one.. that's 250 separate people who looked at Airouter and decided it was worth paying for. most side projects never get their first 10. enjoy this one before heading back to building

[Gil Lopez @lopez_gillopez]

@Nick_333ch "i would like to resign as clipboard" is such a good line.. anyone running two agents has been the human copy-paste in the middle. codex reviewing, claude code building and the files just showing up on telegram is a really clean loop

[Gil Lopez @lopez_gillopez]

@AndriusMacikas still looking for the helmet 😂 at this point Boothie could stop for snacks halfway and still take it

[Gil Lopez @lopez_gillopez]

@AndriusMacikas Boothie is up 37% in 30 days and still listed at $1.5M.. are most of the 15 sellers cashing out while growing like that, or are the flat ones the bigger share?

[Gil Lopez @lopez_gillopez]

@srbera17 300+ days in and the win you picked isn't the ~40 contests, it's solving half on your own with no hints. that's the number that shows the work actually paying off.. the next 300 days are gonna look really different

[Gil Lopez @lopez_gillopez]

@Marzssss part of training sitting outside the loop because of one indent.. and you went digging in the code instead of blaming the model. that instinct is worth more than any character-level output right now. fun write-up to read

[Gil Lopez @lopez_gillopez]

@heyratomir evidence-based findings instead of a wall of generic warnings is what makes a security tool stay installed.. and local + passive by default is a thoughtful call for something that reads your whole codebase. PandaWard is a great name too

[Gil Lopez @lopez_gillopez]

@HeyNikhila $150-180 of new MRR landing every month means people keep finding FormNX and paying for it.. that part is clearly working. ~$875 on a vibecoded SaaS with two exits already behind you, $1K is closer than the wtf makes it feel

[Gil Lopez @lopez_gillopez]

700,000 visitors a month and the fix for fake reviews is.. your friends and THEIR friends?? plus an MCP connector so your AI agent books the hotel. Pieter is on a tear

[@levelsio @levelsio]

🏩 Hotelist is now my most popular site with 700,000+ visitors per month

So I added user profiles today invite-only and invited @marckohlbrugge @jelmerdeboer @theannalux and they get 3 invites too so let's see who joins

Users can review and rate hotels, upvote/downvote other user's reviews and get karma from all of that

The idea is actually by Marc that you probably trust your friends' and their friends reviews better than random Google Maps reviews (which are also often fake)

So this is a natural social way to get rid of fake reviews, because you're probably far away from a fake reviewer and you trust your friends' reviews

[Gil Lopez @lopez_gillopez]

Kling 4.0 going straight for FIGHT scenes?? fast action is exactly where AI video used to fall apart.. now there's a whole collection of them. wild

[Kling AI @Kling_ai]

The rhythm of a fight. The force of a movement. The precision of a shot.
Discover a new collection of action sequences made with Kling 4.0.

[Gil Lopez @lopez_gillopez]

an agent trying to write its own reward.txt or patch pytest to pass.. and safeagents catches it BEFORE it runs?? free, lightweight, open source from Hugging Face. so good

[younes @yb2698]

Agents are often difficult to predict and can produce dangerous tool calls or attempt to perform reward hacking. Potentially affecting tons of users without even noticing anything.

Before an agent's command runs, someone should check it. We're releasing safeagents to do exactly that. Every tool call goes through it first: an upload to a bucket nobody mentioned, a request to the cloud metadata endpoint, curl --noproxy sending ~/.netrc to a paste site all get stopped before execution.

Also handy for RL and evals: safeagents blocks a model writing its own reward.txt, editing the verifier's /tests or patching pytest to pass, before any of it runs.

Introducing `safeagents` - a fast, free and lightweight proxy that keeps any LLM's tool calls within the task you gave it, pluggable in any coding harness. Purely written in python and based on https://t.co/g6Af5lx7JE.

We need something to inspect these behaviors openly and share them to the community. https://t.co/Lg2IAAuXaa