Here's an animation of a @PyTorch DataLoader.
It turns your dataset into a shuffled, batched tensors iterator.
(This is my first animation using @manim_community, the community fork of @3blue1brown's manim)
Here's a little summary of the different parts for those curious:
1/5
My current personal agent + app setup and use cases:
Setup:
- Agent: Hermes from @NousResearch
- Client: Hermex on iOS from @uzairansar
- Model: Claude for most, @CoreWeave inference for open models to do cheaper scraping tasks
- VM: @exedev
- App tech: @Tailscale with little web apps, some OSS projects
Side point: Hexmex has a very nice iOS experience - live widgets, persistent Bots, dynamic island + notifications, and just generally a place to chat to agents.
Use cases:
- vibe-coding hyper personal games with friends
- content for my LearnHarness by @amosblomqvist to learn distributed systems
- pull bookmarks from X so I don't use the app on my phone, and instead read longer form content and get addicted to short videos
- general life organization stuff, like tracking TODOs and goals
- built an app to make watching live rugby matches easier on my phone
Hermex 2.0 is out 🚀
Hermex now talks to Hermes directly, no webui in between.
Chat, Sessions, Tasks, Kanban, Skills, Memory, Usage, Files and Git, all on your Hermes host. Move your chats over from Settings.
Free and open source on the App Store: apps.apple.com/us/app/hermex…
Every run brings the next discovery closer, as long as what you learn in production makes it back into the next experiment. But is that always the case?
CoreWeave Forge connects the whole loop in one place: Run, Observe, Evaluate, Curate and Improve. It stays open to any model, framework or cloud, so every run compounds.
Discovery is a human pursuit. We’re built for yours.
CoreWeave Database is now generally available.
Managed PostgreSQL now runs within your CoreWeave Kubernetes Service (CKS) cluster, so agents and RAG pipelines no longer cross clouds to access their data.
Built on CloudNativePG, with pgvector included. crwv.co/uuHsu
new post-training example → SFT on agent traces with TRL
> a big model works as a coding agent in @opencode via OpenEnv
> the tasks it solves become the training data
> 2B model goes from 5.5% to 34.4% on held-out tasks
full guide with everything open: github.com/huggingface/trl/t…
my vocal synth models the vocal tract as a network of branching acoustic tubes, unlike vocaloid which stitches together and pitch-shifts phoneme recordings. still working on improving naturalness by tuning it against recordings of real singers, but I'm already impressed at how cool it sounds!
Systems is games is a perfect use case for weight space programming
These would be difficult to justify training a model on, want to be custom small models for easy distribution, and yield diverse, emergent gameplay dynamics if done right.
Some of my favourite gameplay comes from emergent complexity - some that come to mind: red stone in Minecraft and zonai devices (Zelda botw and totk in general)
A Rube Goldberg machine maker where some of the parts have little brains
A self-aiming catapult, a homing drone, a paddle that juggles…
Dragology was used to make it feel satisfying to connect parts and the models were generated straight into weights using the technique from the Weight Space Programming post.
A Rube Goldberg machine maker where some of the parts have little brains
A self-aiming catapult, a homing drone, a paddle that juggles…
Dragology was used to make it feel satisfying to connect parts and the models were generated straight into weights using the technique from the Weight Space Programming post.
Drag-and-drop makes the digital graspable, but building it well is fiddly enough that many don’t bother. Declarative Dragging (with Josh Horowitz & Jeff Heer, #uist2026 best paper award), lets you make an element draggable just by listing the states dragging it can reach.
Drag-and-drop makes the digital graspable, but building it well is fiddly enough that many don’t bother. Declarative Dragging (with Josh Horowitz & Jeff Heer, #uist2026 best paper award), lets you make an element draggable just by listing the states dragging it can reach.
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.
We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵
Blog: rikiyat.github.io/blog/weigh…
If you're looking for more ways to create models from prompts, you can use programasweights.com/ from @yuntiandeng - you prompt it and it compiles a lora adapter on the fly for classification tasks.
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.
We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵
Blog: rikiyat.github.io/blog/weigh…
i've reproduced the results from the MIT blog post today on weight-space programming. but one additional result i have to add: this is a new, step-change emerging capability that's appeared in just the latest set of models.
Questions:
Is it a form of distillation?
How the hell!?
What types of models can it build reliably?
Is this real life?
When should you use it instead of prompt/harness engineering or training?
Wtf!?
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.
We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵
Blog: rikiyat.github.io/blog/weigh…
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.
We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵
Blog: rikiyat.github.io/blog/weigh…
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.
We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵
Blog: rikiyat.github.io/blog/weigh…