chris - senior product engineer at @retailbook_ - cgym.dev
London
- Tweets 2,408
- Following 1,089
- Followers 581
- Likes 82,754
the game changers are back
Introducing Prime Sandboxes:
MicroVM sandboxes purpose-built for RL training.
Model training requires running tens of thousands of concurrent sandboxes, leading to complex and costly configuration. We built Prime Sandboxes for our own team. Today we're releasing them publicly.
6
396
this era is where i was rapidly progressing and learning about how computers worked
eventually turning to macbooks and iphones to get real work done - but could not have got here without it
using arch linux in uni was a real blast when trying to debug why hdmi was not working pre presentation
126
these tiny models are going to go crazy in manufacturing and robotics
We release Needle 3: A Sliceable 8-29MB automation foundation model that can match DeepSeek V4 Flash.
One set of weights, every depth from 2 to 20 layers a model of its own, 25-121M parameters at CQ2-bit, built on our Simple Attention Networks and running locally at up to 4k tokens/sec decode speed on a Raspberry Pi 5.
Needle does not chat. Every turn is a function call: give it the tools your app exposes and it picks the right ones and fills every argument from what the user said, or hand it a schema and it returns a typed record. Ask for something no tool covers and you get an empty list, not a guess.
That trade is lets 121M parameters trained on 360B tokens of structured data beat models 10x their size on mobile tool calls and match 2-3x bigger models on structured JSON extraction.
It runs on mobiles, wearables, smart home devices, small robots and microcontrollers, with prebuilt engines for macOS, Linux, Windows, Android, iOS, watchOS, tvOS, the browser and WASI hosts. Try it in your browser: cactuscompute.com/needle
1
118
all my employed friends outside of sf in the normal world use codex claude and opencode
everything else is more for fun and experimental
2
8
350
this is my obsession as of late and its sooo good to wind down and play this after prompting for 12 hours at the job
RuneScape: Dragonwilds 1.0 is OUT NOW!
Master survival in a new RuneScape world. Grind skills from 1 to 99, craft powerful gear and take down Kuldra, the Dragon Queen, alone or with friends!
Available on PC, PS5, XBOX Series X|S and Nintendo Switch 2. dragonwilds.runescape.com/
81
new meta right now on best harness and coding agent is
- @opencode with go subscription
- deepseek v4.1 flash
working on my game engine and a port of dwarf fortress to macos and it is currently cheaper and faster then running gpt 5.6 astra on medium
opencode v2 is also a god send in terms of ui and agent state management - best in the scene right now
if your hitting limits on codex or claude - drop everything and setup a opencode go subscription
1
5
233
new method of producing high quality code with agents - read the code and work on a single agent thread
pov: me
1
59
gpt 5.6 astra is one of the most jagged models i have ever used
some times the code is perfect, other times it introduces weird coding patterns that dont exist anywhere else in the repo
odd
59
Everybody should slow down how we use agents. Take your time to craft good architecture and product requirement documents.
Then get agents to ship them in sequential order while skimming over the code to make sure everything looks right.
Slow is smooth and smooth is fast.
I am merging in higher quality code that I dont have to go back and touch on. Less PR's in total but the quality is significantly higher and scales better.
2
67
plan mode is has come back into fashion but we just call it architecture / product requirement docs
1
2
165
Intelligence is shockingly cheap now.
Working with @opencode opencode v2 running deepseek v.4.1 flash and we really have reached state of the art models for $10 a month.
This is the best price to performance ratio that you can get right now and have a model that executes and derives beautiful code.
With a deep product requirements and architecture document - @deepseek_ai v4.1 flash model is executing this perfectly.
137
easy change that makes my game operate much better is by decoupling the engine's tick system to the renderer
so fps doesn't affect physics in my engine.
1
37
started applying best practices to my shitty side projects and its making things actually move forward in a structured way - who would've known
proper linear tickets and plans go miles in a game:
@linear
1
56