Michael Bernadsky @OctothorpeVoid

Menlo Park, CA Joined October 2016

Tweets

976
Followers

74
Following

501
Likes

3K

Luca Ambrogioni @LucaAmb

a day ago

1/2) Have you have noticed that the forward process in a diffusion model looks a lot like the reparameterization trick in VAEs? It turns out that there is a deep connection! Curious? Watch our new vedio in the Generative Memory Lab channels (link below)

9 51 524 26K 427

Download Image

NYRE @sleenyre

2 days ago

4 year pytorch bug where all reduce operation produces INCORRECT gradients with no warning. Still not patched. Initially reported by @DrJimFan. Sharing this in case anyone is having mysterious gradient explosions. github.com/pytorch/pytorc…

12 35 390 41K 186

Simo Ryu @cloneofsimo

2 days ago

CHAT its that time of the month. Whats THE MOST interesting work youve seen recently? Ill go first with this paper.

15 23 288 20K 276

Download Image

Mike Taylor @hammer_mt

2 weeks ago

GEPA is crazy good - past typical human level prompting @DSPyOSS

3 26 390 26K 376

Download Image

Rohan Paul @rohanpaul_ai

2 weeks ago

One of the best paper of the recent week. The big takeaway: scaling up model size doesn’t just make models smarter in terms of knowledge, it makes them last longer on multi-step tasks, which is what really matters for agents. Shows that small models can usually do one step…

21 117 689 45K 506

Download Image

Yacine Mahdid @yacinelearning

2 weeks ago

the emergence of attention sinks in LLMs is so fascinating, especially the fact that some of them are useful.

15 79 883 46K 663

Download Image

Amanda Bertsch @abertsch72

2 weeks ago

.@gneubig and I are co-teaching a new class on LM inference this fall! We designed this class to give a broad view on the space, from more classical decoding algorithms to recent methods for LLMs, plus a wide range of efficiency-focused work. website: phontron.com/class/lminfere…

6 60 446 34K 468

Casper Hansen @casper_hansen_

2 weeks ago

One of the cleanest DeepResearch papers I have read in a while. They RL gpt-oss-20b into a SOTA model on Humanity's Last Exam with search, visit, code tools. LLM as a judge for rewards and REINFORCE variant was used with medium-difficulty data which required search.

7 52 348 21K 315

Download Image

Lucas Beyer (bl16) @giffmana

2 weeks ago

Very cool thread about whether LLMs can multi hop reason without CoT or not. If you're curious, read the full thread, it's well written and clearly answers.

Mikita Balesni 🇺🇦 @balesni

2 weeks ago

Very cool thread about whether LLMs can multi hop reason without CoT or not. If you're curious, read the full thread, it's well written and clearly answers.

2 8 88 50K 82

Download Image

9 26 213 41K 158

𝚐𝔪𝟾𝚡𝚡𝟾 @gm8xx8

2 weeks ago

WSM replaces learning rate decay with constant LR training and checkpoint merging, achieving the same or better results without modifying the training process. It shows that decay is just a form of weighted gradient averaging and uses this insight to merge checkpoints in a way…

2 27 200 15K 146

Download Image

anshuman @athleticKoder

2 weeks ago

The Big Fat LLM Architecture Comparison > Qwen3 Dense > Qwen3 MoE > Llama 4 > DeepSeek-V3/R1 > GPT-OSS > OLMo 2 > Gemma 3 > Mistral Small 3.1 > SmolLM3 > Kimi 2 MoE, MLA, GQA, sliding window, normalization etc. etc.

1 3 9 2K 15

Download Image

Thinking Machines @thinkymachines

2 weeks ago

Today Thinking Machines Lab is launching our research blog, Connectionism. Our first blog post is “Defeating Nondeterminism in LLM Inference” We believe that science is better when shared. Connectionism will cover topics as varied as our research is: from kernel numerics to…

240 1K 8K 3.2M 5K

Download Image

Joey (e/λ) @shxf0072

3 weeks ago

Qwen3-Next is hybrid GatedAttention (for outliers fix) GatedDelta net rnn for kv saving all new models will be either sink+swa hyprids like gpt oss or gated attn + linear rnn hybrids (mamba , gated deltanet etc) like qwen3-next age of pure attn for timemixing layer is over,

20 89 684 135K 344

Download Image

Pramod Goyal @goyal__pramod

3 weeks ago

It took me 2 days to finally grasp how fixed sparse attention works. And when it did, I realised this is one of my most genius ideas I have ever had the pleasure to encounter.

3 56 851 104K 999

Download Image

TuringPost @TheTuringPost

3 weeks ago

KV cache compression techniques ▪️KV caching (basic) – stores previously computed Keys and Values in memory and calculates attention only for new tokens. ▪️ Quantization – represents KV cache with fewer bits. ▪️ Low-rank decomposition – compresses the KV cache into smaller…

16 182 1K 87K 1K

Download Image

Avi Chawla @_avichawla

3 weeks ago

I have been fine-tuning LLMs for over two years now! Here are the top 5 LLM fine-tuning techniques, explained visually:

18 113 971 112K 2K

elvis @omarsar0

3 weeks ago

Another impressive paper by Google DeepMind. It takes a closer look at the limits of embedding-based retrieval. If you work with vector embeddings, bookmark this one. Let's break down the technical details:

35 240 2K 203K 3K

Download Image

Keenan Crane @keenanisalive

3 weeks ago

“Everyone knows” what an autoencoder is… but there's an important complementary picture missing from most introductory material. In short: we emphasize how autoencoders are implemented—but not always what they represent (and some of the implications of that representation).🧵

43 403 3K 464K 3K

Download Image

Lior Alexander @LiorOnAI

5 months ago

You can now run 100B parameter models on your local CPU without GPUs. Microsoft finally open-sourced their 1-bit LLM inference framework called bitnet.cpp: > 6.17x faster inference > 82.2% less energy on CPUs > Supports Llama3, Falcon3, and BitNet models

139 788 6K 591K 6K

Download Video

elie @eliebakouch

4 weeks ago

The technical report of @Meituan_LongCat LongCat-Flash is crazy good and full of novelty. The model is a 560B passive ~27B active MoE with adaptive number of active parameters depending on the context thanks to the Zero-Computational expert. 1) New architecture > Layers have 2…