HackerTrans
トップ新着トレンドコメント過去質問紹介求人

diyer22

no profile record

投稿

Open model StepFun-3.5 is #1 on MathArena, an uncheatable math benchmark

twitter.com
3 ポイント·投稿者 diyer22·5 か月前·2 コメント

A new local LLM king: Step-3.5-Flash-int4

old.reddit.com
2 ポイント·投稿者 diyer22·5 か月前·1 コメント

Got burned by an Apple ICLR paper – it was withdrawn after my Public Comment

twitter.com
6 ポイント·投稿者 diyer22·8 か月前·3 コメント

Show HN: I invented a new generative model and got accepted to ICLR

discrete-distribution-networks.github.io
656 ポイント·投稿者 diyer22·9 か月前·91 コメント

コメント

diyer22
·5 か月前·議論
MathArena uses newly released competition sets and evaluates models close to the event. They also mark models released after the competition date as potential contamination.

On Feb 6, the just-concluded AIME 2026 I, Step 3.5 Flash take first place. Step 3.5 Flash was released on Feb 1, making cheating impossible.
diyer22
·5 か月前·議論
StepFun has open-sourced Step-3.5-Flash: 196 B total parameters, 11 B active, 256 K context length. Strong performance, with speed as the highlight—blazing fast, peaking at 350 tokens/s. It’s currently in promotion and free on OpenRouter `step-3.5-flash:free`.

More detials: https://static.stepfun.com/blog/step-3.5-flash/
diyer22
·8 か月前·議論
A record of getting burned by a “sloppy, low-quality, and irresponsible” paper from Big Tech.
diyer22
·9 か月前·議論
Thanks for all the great feedback! I've created a Twitter thread to discuss future development and share updates. Would love to connect with you all there:

https://x.com/diyerxx/status/1978531040068321766

Getting started on Twitter is so tough—engaging with my posts would really help me out a lot!
diyer22
·9 か月前·議論
I believe DDN is capable of handling TTS (text-to-speech) tasks, because with the text condition, the generation space is significantly reduced.

And it's recommended to combine it with an autoregressive model (GPT) for more powerful modeling capabilities.
diyer22
·9 か月前·議論
During neural network training, the ground truth (GT) must be known to compute the loss.

In DDN, the GT is only used to calculate the loss and guide sampling; it never becomes an input to the model.
diyer22
·9 か月前·議論
Exactly what i think!

- The DDN single-shot generator architecture is more efficient than diffusion.

- DDN is fully end-to-end differentiable, allowing for more efficient optimization when integrated with discriminative models or reinforcement learning.

- Moreover, DDN inherently avoids mode collapse.

These points are all mentioned in the blog: https://github.com/Discrete-Distribution-Networks/Discrete-D...
diyer22
·9 か月前·議論
Thanks for the idea, but DDN and flow can’t be flipped into each other that easily.

1. DDN doesn’t need to be invertible. 2. Its latent is discrete, not continuous. 3. As far as I know, flow keeps input and output the same size so it can compute log|detJ|. DDN’s latent is 1-D and discrete, so that condition fails. 4. To me, “hierarchical many-shot generation + split-and-prune” is simpler and more general than “invertible design + log|detJ|.” 5. Your design seems to have abandoned the characteristics of DDN. (ZSCG, 1D tree latent, lossy compression)

The two designs start from different premises and are built differently. Your proposal would change so much that whatever came out wouldn’t be DDN any more.
diyer22
·9 か月前·議論
It does seem that way — we’re both trying to overcome the limitations imposed by LLM tokenization to achieve a truly end-to-end model.

And, their work is far more polished; I’ve only put together a quick GPT+DDN proof-of-concept.

Thank you for sharing.
diyer22
·9 か月前·議論
Yes, there is a transform that make final size of stem features remains unchanged
diyer22
·9 か月前·議論
Even in the earliest stages of the DDN concept, we had already decided to pass features down to the next layer.

I never even ran an ablation that disabled the stem features; I assume the network would still train without them, but since the previous layer has already computed the features, it would be wasteful not to reuse them. Retaining the stem features also lets DDN adopt the more efficient single-shot-generator architecture.

Another deeper reason is that, unlike diffusion models, DDN does not need the Markov-chain property between adjacent layers.
diyer22
·9 か月前·議論
It's just a coincidence—the guided images used for ZSCG all come from Celeb-A, whereas the DDN model was trained only on FFHQ.

Besides, I feel the red shoulder strap/blob is reconstructed rather poorly.
diyer22
·9 か月前·議論
This understanding is incorrect. The video samples all the leaf nodes of the entire tree only to visualize the distribution in latent space. In normal use, only the L outputs along a single path are generated.
diyer22
·9 か月前·議論
1x1 convolution is the most lightweight operator for transforming features into outputs.

3x3 convolution is the most common operator used to provide basic computational power.
diyer22
·9 か月前·議論
I understand that by "discrete number" you mean the selected output of each layer.

Both the "feature" and the "selected output" are designed to be passed to the next layer.
diyer22
·9 か月前·議論
That's right! The second paragraph on OPTIMIZATION WITH SPLIT-AND-PRUNE in the original paper:

> Inspired by the theory of *evolution and genetic algorithms*, we propose the Split-and-Prune algorithm to address the above issues, as outlined in algorithm 1.
diyer22
·9 か月前·議論
I agree with @ActivePattern and thank you for your help in answering.

Supplement for @f_devd:

During training, the K outputs share the stem feature from the NN blocks, so generating the K outputs costs only a small amount of extra computation. After L2-distance sampling, discarding the other K-1 outputs therefore incurs a negligible cost and is not comparable to discarding K-1 MoE experts (which would be very expensive).
diyer22
·9 か月前·議論
In DDN, 1×1 convolutions are used only in the output layers of the Discrete Distribution Layer (DDL). The NN blocks between DDLs, which supply the fundamental computational power and parameter count, adopt standard 3×3 convolutions.
diyer22
·9 か月前·議論
We provide the source code and weights along with a Docker environment to facilitate reproducing the experimental results. The original paper’s EXPERIMENTS section mentions the hardware configuration (8× RTX 2080 Ti).
diyer22
·9 か月前·議論
Exactly! The paragraph on Efficient Data Compression Capability in the original paper also highlights:

> To our knowledge, Taiji-DDN is the first generative model capable of directly transforming data into a semantically meaningful binary string which represents a leaf node on a balanced binary tree.

This property excites me just as much.