Welcome back, Tharun

GPT-5.6 Sol
GPT-5.6 Terra
GPT-5.6 Luna
Opus 5
Opus 4.8
Sonnet 5

Strange Loop is AI and can make mistakes. Please double-check responses and results.

Conv-stem for ViT-S on CIFAR-100 ViT-S on CIFAR-100
Replace the patchify stem in ViT-S with a small conv stem and see if it improves accuracy on CIFAR-100. Optimize val/top1.
Thought for 7 seconds

A conv stem trades a few patch-embed params for an inductive bias that usually helps small ViTs on small images. The cleanest test holds everything else fixed and varies the stem width against the learning rate. I'll hand that to a GPU subagent so it can own the training loop end to end, and run the seeded baseline alongside it.

Searched arXivconv stem vision transformer
Read paper2106.14881
Ran commandls models/vit/

A conv stem is a well-supported tweak for small ViTs at this resolution, but the reported gains are within seed noise at CIFAR scale — so it needs a baseline to beat, not just a sweep.

I've drafted two experiments. Review either plan, leave comments inline, and approve to launch. They'll run concurrently on their own GPUs.

Projects

ViT-S on CIFAR-100

Push a small ViT past the patchify baseline on CIFAR-100 — architecture + training-recipe search.

tharunkumartk/vit-lab
Conv-stem width × LR sweep running
Patchify baseline · 3 seeds running
6 chats2 subagentsUpdated 14m ago
GPT-2 training dynamics

Optimizer, warmup, and schedule ablations for a small GPT-2 — what actually moves val/loss at fixed compute.

tharunkumartk/nanogpt-lab
4 chats3 subagentsUpdated 1d ago

Settings

Link tokens and environment variables. They're injected into your chat agent and your subagents' pods at launch — so gated model downloads and authenticated tools just work.

Appearance

Choose how Strange Loop looks. System follows your device setting.

Compute platform

Which cluster your subagents launch on. Takes effect for subagents you start from now on; running ones stay on their current platform.

Linked tokens & secrets
Encrypted at rest. Values are write-only — once saved they're never shown again, only their last 4 characters.
🤗 Hugging Face Connected

Stored as HF_TOKEN for gated/private model and dataset access.

••••a9f2
Environment variables

Any other secrets your runs need — e.g. WANDB_API_KEY, OPENAI_API_KEY.

  • WANDB_API_KEY••••3b1c
  • HF_HUB_ENABLE_HF_TRANSFER••••1

Closing the loop on discovery.

Add from GitHub

The agent clones it into the chat's workspace so it can read and edit the code.

3 of 24 repositories Refresh