Steven (Batman) Batchelor-Manning

@0xSero @thdxr Ive had a few side quests that agents have struggled with where changing model (even similar capability model) suddenly brings success. I regularly bounce between models now when running into sticky terrain.

These experiments will eventually culminate in a very strong work automation system that can be optimized later. Also opencode specifically allows for multi-provider multi-model. A lot of tasks (reading, running basic commands, research, docs, reviews) can be offloaded to other models. IMO opus type models will be excellent orchestrators, potentially allowing you to burn to get 2M token runs on 100k orchestrator context. If you pay for MiniMax and GLM (20$ a month) + 20$ a month on GPT you can get very far. Also local models now are becoming more and more competent, if these experiments pan out who’s to pay a 30B param model can’t be driving 90% of the token usage.

some thoughts on some of these heavily automated token burner approaches that are popping up i'm definitely not arguing that they're completely useless but i have a hard time seeing it as the general way to build things. it does let you minimize how much work you do but only via spending an amount of tokens you'd never naturally pay for and companies offering subscriptions are seeing issues with this because it's increasing average spend and breaking the economics and outside of things like porting a codebase to a different language, or large migrations if you put in 20% more effort you'd can usually get it done with way less cost and time

@0xSero Amazing work btw, your stubbornness will be the benift of many

@0xSero 🙏

The things I do out of stubbornness. huggingface.co/datasets/0xSer…

@0xSero i noticed Q3 and Q2 variants of these models turning up on HF, are they genuinly going to be pumping out better results than say Denstral 2 small q16 or q8 ?, its hard to work out where the cliff is

Model: MiniMax-M2.1-Reap-50-4bit (62gb) 1 shot prompt: “Make me a complex interesting interaction in html wow me” The hexagon shape would go wherever I click and make very cool patterns, all done in 1 html/css/js file It works well, has some issues but coding wise it’s lit

@thdxr I'd like commits before you even write them, if you live life on the edge, why not the very edge.

if you would like to help us debug issues before they hit everyone you could bun install -g opencode-ai@dev you get the latest commits before we release them and if you wanna go full bleeding edge set OPENCODE_EXPERIMENTAL=1

@simas_ch ive had an incredible amount of success with using MDB at scale for event sources systems due to the Atomic operations it offers, huge throughput capabilities because of smart choices they made with locking.

Besides horizontal scalability, is there any other reason to use MongoDB when all common databases support JSON in the database?

@WazzCrypto If it's on my current machine or my main work VM, either way I keep it on a tight leash. I've had Claudcidents before and like a number of people learned the hard way quickly. If I wanted it to be more of an exploratory task, I'll give it its own VM and set it loose

So when you guys use claude code etc you're all using it on an isolated environment like a VPS or a VM right? Or you're just letting the AI rawdog your machine with admin privileges? sounds kinda dangerous.

@peterwildeford We use it in many Cross discipline processes in our company, for qa to ops to design, it's like having a win man to get shit done

Few realize that Claude Code isn't only about code

@adamwathan We use it extensively for a number of tasks from design to operations, having the ability to setup either mcp's or api's and use CC to as the omni IDE/UI reduces context hopping and allows very quick " oh that was unexpected please document to x.md" etc ....

Is anyone using Claude Code or similar for stuff that isn't coding at all, purely because it's nice to have access to local files and stuff? I use AI as a thinking partner constantly, and not having persistent references is starting to become annoying, but worried the CLI tools are tuned too much for programming and not business strategy for example.

@tom_doerr We are using FE at scale (Geographies, nodes and performance) and its not only straight forward but keeps out the way when you need to, great library.

Alternative to Minimal APIs using REPR pattern github.com/FastEndpoints/…

@billmurphy Yh I was just informed our workspace tomorrow is now not available, so much for running a company workshop, glad I didn't fly people in ........😬

My coworking office in Bratislava just closed because of an unexploded bomb nearby. That's crazy enough. But my Aibnb is closer to the street!

@CodeRed_dev its no longer the weekend 😉 The game is looking really good btw

Trajectory dysfunction

@CodeRed_dev I in fact hope you are not working weekends and are in fact relaxing and having a great time :)

@CodeRed_dev Nope :), a service that has a similar enough name that autocomplete got in the way :)

does the @CodeRed_dev not work Saturdays? Absolute silence from the rabbit on all repos / MR's today 🙄

@soulscircuit Good to see it's CAT rated

Just started testing the LTE module for Pilet 5. It comes with a detachable SMA antenna, and it also supports an FPC antenna. Right now, it’s connected via USB and shows up as an Ethernet device. No need for extra drivers—it’s already supported by the Linux kernel.

@soulscircuit One thing id love to see is some level of dust / crap protecting the pi (because of the exposed ports) even if its just some rubber port fillers. Im not talking water proof, but I'm talking random shit in my backpack proof

Pilet 5 shown alongside its bigger sibling, Pilet 7. #raspberrypi #linux #computer

@soulscircuit Would love to see this, especially if we can see a decent power and ground line.

Thinking of adding a very simple, cartridge-like mechanism for small modules like camera, humidity sensor, gyroscope, etc to Pilet 5—nothing too fancy or complex.

🎉 #ScatterKings Launches Gold Link Series with Tiki Treasures! 🎉 Read more here : linkedin.com/feed/update/ur…

@ryanels We use both for a high-performant scaleable system, the only thing you need to decide is "is it the right tool for the job?"

Which one should I use 🤷🏻‍♀️