Vectro

I am not a programmer, and I don't do AI for a living. I've spent over a year learning about LLMs because I've always liked learning new tech. I make YT videos and X posts for the community of users and developers alike.

I'm only using Hermes Agent from now on, no more OpenClaw. 3 of them are doing non-coding work for my business as I write this. My go-to LLMs are: * Kimi-K2.5 for general tasks * MiMo-V2-Pro for Linux admin Opus sub-agents on the ready for extra lifting when needed.

@BenjaminBadejo @Teknium Nous has reputation. I've been following them for years. The owner is doxed and been on interviews. Their LLMs and agent are 🔥 and it's open source anyway, so what's not to trust? They're more transparent than others.

I want to try Hermes but I’m reluctant to use something created by someone anonymous — whose IRL identity is unknown. (Not criticizing you @Teknium or casting aspersions — to each his or her own!) Just a basic diligence concern of mine. Thoughts?

I might make a video tutorial how to set up Hermes Agent and create a research bot for any industry.

Hermes Agent has been doing long-horizon research on PCB recycling for 10 days so far. • Daily source discovery (Tavily) • Deep scan every 3 days (Firecrawl) • Re-scans sources for changes • Searches for machines, new tech & research papers • Notes uses of optics, ML & AI

I'm building Vectro dot Chat for your privacy, not as an afterthought. You will be able to store chat history encrypted on the server or locally in browser. Deleted convos are gone forever. No Retention. Will use a trusted execution environment and end-to-end encrypted LLMs.

I've seen people on X set up a YOLO trading bot and tell it to just make money. I worked with GPT 5.4 to engineer a whole trading strategy. It has risk management built in. I used an AI agent along with an open source trading bot. It's in dry run mode. Will keep improving.

My assistant agent does't do coding, so it runs fine on Kimi K2.5. It has an Opus 4.6 sub-agent for when it falls short. Works great in Hermes.

@Scobleizer @Teknium The long-term memory is good, the sub-agents work. Setup is nice. So many integrations. I've been using Hermes since day 1 and now have full functioning research agent, linux admin and dry-run trading bots.

The Hermes Deep Dive. (The new hot AI agent harness). docs.google.com/document/d/1sZ… Hey @Teknium got anything to add?

Don't know if many people care about orbital space debris like I do lol youtu.be/NaFajPhl0XY

The "buy a GPU bros" Are not wrong, just narrowly focused. I like having choices. - cloud API - rent a GPU - buy a GPU My business is starting to leverage all of them for different reasons.

@TheAhmadOsman I switched to llama.cpp for local. The only OpenClaw I run are what I had pre-Hermes.

if you’re using Ollama switch to llama.cpp if you’re using OpenClaw switch to Hermes these are basics at this point

@KettlebellDan All the time, and now there's the artemis mission

ever have nights where you look up at the moon and feel like it’s calling you to explore

One cool trick I figured out for Agents... If it's not for coding and you don't want to spend on Claude tokens all the time, save money using Kimi K2.5. Then... Tell it to spawn a sub-agent with Opus 4.6 that it can only use for emergencies with your permission.

Long-term AI experiment (2 weeks so far): Agent manages an open source crypto trading bot from github (freqtrade). Dry-run only. Lots of risk management rules. Tons of logging technical and trading errors. Sticks to one strategy, collects data, then I can make adjustments.

@BenjaminDEKR I'm not even the moral authority. The compute and electricity are already stretched thin. Stop wasting it on Waifu.

Ok, then perhaps pivot away from scantily-clad AI anime girls and take this seriously

@alexinexxx I've been in the trenches since 1997

learning how linux users live is teaching me a lot. a lot about why their failure rates are so high

@0xSero I need this. Last week Claude had to build connections to 2 APIs. After some minor tweaks, it suddenly removed one entire API from the code.

Thank god for Droid's shield system. This is the 3rd time today that it decided it was to just "nuke" something. I don't understand why Claude is so happy to just nuke things, I guess that's why the DOW wanted it so bad.

@Teknium I've been trying to patch things together and run Kalshi simulations

Hmmmmm

@chutes_ai Looks like for web on PHP it needs a local container as a proxy. Going to try to make one.

AI inference should not require trust in infrastructure. chutes.ai/news/end-to-en… On March 2nd we shipped end-to-end encrypted transport on Chutes. Here's how it actually works under the hood. Your data is encrypted on your machine, directly to the GPU instance running inside a Trusted Execution Environment. It stays encrypted through our API, load balancers, and the network. Decryption only happens inside TEE-protected hardware where memory is isolated from the host. Impossible for anyone to see including us. The key exchange uses ML-KEM-768 — a NIST-standardized post-quantum key encapsulation mechanism. Every request gets a fresh ephemeral keypair. Forward secrecy by default. Resistant to future quantum attacks. Full technical breakdown in the blog: chutes.ai/news/end-to-en… If you want to try it: → Python: pip install chutes-e2ee → Any language: docker run parachutes/e2ee-proxy:latest github.com/chutesai/chute… github.com/chutesai/e2ee-…