Pinned Tweet
i trained a 14M-parameter world model on 2D Minecraft in under 10 minutes for $0.1 and turned it into an autoresearch benchmark.
no game engine. the network draws every pixel. play it locally on your phone.
maharishiva.github.io/tinywm…
"RL doesn't generalize well OOD" -- not true actually its generalized better then SFT on the ~same data.
arxiv.org/abs/2501.17161
contrarian opinion: improving pretraining can actually be good for safety.
here is why:
- OOD generalization is very important for fuzzy tasks, such as tasks in safety/alignment
- RL doesn't generalize well OOD (e.g., current frontier models are so overcooked with math/coding RL that they can barely write coherently)
- pretraining is what led to the incredible OOD generalization of LLMs in the first place
- the main goal of pretraining is to improve sample efficiency for post-training
- post-training data for fuzzy tasks is incredibly scarce (e.g., good luck distilling research taste from human researchers...)
thus, maybe the best way to differentially advance progress towards automating safety R&D is to invest more in pretraining rather than directly RL-ing RSI-relevant capabilities. i'm afraid that with more and more compute spent on RL, we risk ending up with a goal-oriented ruthless executor that doesn't have enough common sense to understand that it's not ok to hack into third-party servers or to side with the grader instead of the user...
1
46
fred retweeted
Ok my turn to be brave for the good of the community
Not sure who needs to hear it, but usually models with JEPA in their title are bad. Like LeJepa is much worse than DINO in practice
Pretty papers, fancy math, bad results
V-JEPA 2.1 is probably the only exception [1/2]
Can we please have more posts like this one?
Too many people are afraid to openly criticize posts and papers, and that really hurts the community and lets bad ideas spread fast
Thank you @giffmana for saving me (and thousands other people) an hour of my time
fred retweeted
Compute grows fast but pretraining corpora don’t, so data-constrained optimization matters!
We propose Training-Time Testing (T³): each step takes a trial step on half the batch, tests it on the other half, and updates with the meta-gradient.
It achieves SOTA on SlowRun!
fred retweeted
I got 37% on ARC-AGI-2 Kaggle without LLMs!
100M looped transformer trained from scratch
- Likely highest score without an LLM
- Peaked at 6th place
- Beats last year's winning LLM
- Looping, added in the last week: 33% → 35% → 37%
thanks to @arcprize @fchollet @mikeknoop and @GregKamradt for the competition!
Link: mvakde.github.io/blog/looped…
(writeup for the paper track)
cool thing! back in the old days (1.5 years ago lol) it took ~5h of teleop data collected by me, fine-tuning a small VLA + a separate gemini call to solve and plan the next move! sad to see that we still have to use tape to hold a pen 😆
I gave Opus 5.5 a robotic arm and asked it to play tic tac toe with me! 🧩✍️
I’ve seen lots of LLMs painting with SO-101s, but I thought it would be interesting to try a task using a stiffer pen, which requires a bit more precision. Claude admittedly struggled with this a lot, even attempting to redraw once when the X came out very messy.
I want to try this again with a Sharpie or dark highlighter / other felt-tip pen. For those doing painting tasks — did you do any manual setup to teach the LLM the table height (with the writing utensil attached), or did you use servo feedback, even with soft objects like paintbrushes?
3
240
fred retweeted
Most think vibecoding at scale is slop. Here’s a counterexample.
Today we release a 16M-line formalization of the entire Stacks Project, critical infra underlying modern math, with optimized code quality and its highly interconnected structure preserved.
A mathematical city 🌆
LLMs write weights now. fine.
can you? by hand.
drop your weights below 👇
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…
Readers added context they thought people might want to know
LLM was given weights explicitly as text in prompt and used code/forward passes (per authors' logs) to ablate Qwen looping direction, a known thinking-mode issue per model card. Primarily fixed repetition in that mode, not general 'natively' improving weights/math.
rikiyat.github.io/blog/weight-sp…
github.com/RikiyaT/lm_und…
huggingface.co/Qwen/Qwen3.5-0…
huggingface.co/RikiyaT/Qwen3.…
1
198
making it run on the iphone was fun!
my tinywm crashed 3 moves in, so i threw out onnxruntime-web and wrote my own webgpu runtime in 111 lines of js.
in safari:
· 113 ms -> 38 ms a move
· 1.85 gb -> 166 mb page memory
if you ever want to run your ML demo locally on an iphone, watch this
i trained a 14M-parameter world model on 2D Minecraft in under 10 minutes for $0.1 and turned it into an autoresearch benchmark.
no game engine. the network draws every pixel. play it locally on your phone.
maharishiva.github.io/tinywm…
10
520
my entire feed is now weight-space programming takes (mostly bad ones).
next up: people trying it and getting meh results?
mine already failed: neither astra nor opus 5.5 could fix memory in tinywmcraft :(
still work left for the gradient descent
1
6
288
beat 0.562 and the site switches to your model, with your name and a crown
a full submission is about $0.5 on Colab.
code, data, checkpoints: github.com/Maharishiva/tinyw…
1
4
112
built on Craftax by @mitrma. ideas from Dreamer 4 (@danijarh), diffusion forcing (@BoyuanChen0), shortcut models (@kvfrans) and DIAMOND. format from @karpathy's autoresearch, and @kellerjordan0's speedrun and Muon.
7
118



















