Switched from @Paychex to @GustoHQ for payroll and our bimonthly payroll dropped by $6k. WTF? Turns out @Paychex has been collecting 5.3% for L&I tax when it should have been 0.47%. This has been going on for 3 years so they’ve cost us well over $100k. A lawsuit is possible…
@AICEOGiuliano Try my Night Build project:
https://t.co/OIXpW1m2vr
It lets you run long horizon tasks in a loop in a controlled way. Nothing to install. Just a command for your coding agent.
@AICEOGiuliano Using docs as the source of truth is a fail. You need to download the code for the library and let your coding agent just read the code.
https://t.co/HdWTI5LfAk has instructions for adding NightBuild it to your project. Nothing to download or install. The agent does everything using one simple command.
NightBuild: Hand off a coding task at bedtime. Wake up to a working artifact and a clean post-mortem.
#NightBuild is a riff on the #AutoResearch idea by @karpathy. I've expanded his idea to support performing any task autonomously over night.
#SynthOS... A vibe coding wiki that turns non-programmers into programmers in a novel way. It removes having to understand how to program.
Here's a 13 minute presentation on the Star Trek Computer built in under 2 minutes:
https://t.co/AI975KR3Yo
Engage :)
@AICEOGiuliano So we’ve worked out that you can get OSS models near frontier level using the ECW. It has to do with how the ECW breaks down the problem of reasoning.
DM me
@AICEOGiuliano There’s only a small amount of magic that we use. The cost savings are because we use a mix of fast low cost models with a single expensive final model call. Think of it as the equivalent of upscaling for reasoning
@AICEOGiuliano The benchmarks we’ve posted are very real. You can get opus 4.7 or gpt 5.5 level reasoning across hundreds of millions of tokens for a fraction of the cost.
@AICEOGiuliano So with the ECW we’ve reasoned over 400m tokens in one single task. It’s an amazing piece of tech that very few people understand the need for.
Make your OSS Library #AgentReady
Here's a new approach I came up with that uses llms.txt and a coding agent to add your library to any project in under 5 minutes.
Simply add a llms.txt file and Agent Ready Badge+Instructions to your repos README.md:
https://t.co/qtYrj3kOjn
Vectra is now #AgentReady: A new way of making your libraries consumable by Coding Agents.
https://t.co/AGxtVIwhN8
Since #vectra now supports gRPC and can be used from any programming language. You can add a rich (100% free) local VectorDB to any project in under 5 minutes.
Massive updates for my Local Vector DB #vectra
Browser & Electron support, local embeddings for offline support, Protocol Buffer support (40% smaller), multiple language bindings, pluggable storage system, and new Agent Ready support.
Vectra has ~120k downloads a month now!