Mechanistic interpretability often relies on *interventions* to study how DNNs work. Are these interventions enough to guarantee the features we find are not spurious? No!⚠️ In our new paper, we show many mech int methods implicitly rely on the linear representation hypothesis🧵

3 months ago, I discovered DeltaNet. I spent hours trying to understand it. Feeling amazed, I shared the blog here on my x, which had less than 20 followers back then. Then, @Songlin replied. And that simple reply ended up shifting the trajectory of a 16 y/o's life. (a thread)

Dimitris Papailiopoulos Soham Daga I feel that these papers (from my group) are examples of what you are nominally asking for: 1. arxiv.org/abs/2505.20809 2. arxiv.org/abs/2505.15105 3. arxiv.org/abs/2505.13898 4. arxiv.org/abs/2501.17148 5. arxiv.org/abs/2505.11770 6. aclanthology.org/2024.emnlp-mai… 7.

Represent! ✌️

Andrew Ng’s piece on 🇺🇸 vs 🇨🇳 competition in AI: “Because many US companies have taken a secretive approach to developing foundation models—a reasonable business strategy—the leading companies spend huge…to recruit key team members from each other who might know the ‘secret

OpenAI's OSS model possible breakdown: 1. 120B MoE 5B active + 20B text only 2. Trained with Float4 maybe Blackwell chips 3. SwiGLU clip (-7,7) like ReLU6 4. 128K context via YaRN from 4K 5. Sliding window 128 + attention sinks 6. Llama/Mixtral arch + biases Details: 1. 120B MoE

For researchers, scientists, and academics tackling hard problems: Gemini 2.5 Deep Think is here. 🤯 It doesn't just answer, it brainstorms using parallel thinking and reinforcement learning techniques. We put it into the hands of mathematicians who explored what it can do ↓

Maia Wasn't actually Habermas, but a Freudian social theorist named Karola Brede. Really fascinating article connecting Karp's PhD dissertation to Palantir boundary2.org/2020/07/moira-…

New Anthropic research: Persona vectors. Language models sometimes go haywire and slip into weird and unsettling personas. Why? In a new paper, we find “persona vectors"—neural activity patterns controlling traits like evil, sycophancy, or hallucination.

It is wild to me how little Deep Learning researchers know about basic statistical theory. Everyone acts like all to all attention is a free lunch while basic stats has shown many better ways to capture long range dependencies instead of comparing every token to each other.

Four TSMC 2nm fabs will be in mass production next year and monthly capacity over 60,000 wafers-per-month (wpm), media report, citing unnamed supply chain sources. 2nm wafers cost US$30,000 each, 50% more expensive than 3nm. 1/2 $TSM $SSNLF $INTC #semiconductors #2nm

Excited about our new work: Language models develop computational circuits that are reusable AND TRANSFER across tasks. Over a year ago, I tested GPT-4 on 200 digit addition, and the model managed to do it (without CoT!). Someone from OpenAI even clarified they NEVER trained

gpt-oss is probably the most standard MoE transformer that ever was. Couple of details worth noting: - Uses attention sinks (a.k.a. registers) - Sliding window attention in every second layer - YaRN context window extension - RMSNorm without biases - No QK norm, no attn. softcap

I noticed that OpenAI added learnable bias to attention logits before softmax. After softmax, they deleted the bias. This is similar to what I have done in my ICLR2025 paper: openreview.net/forum?id=78Nn4…. I used learnable key bias and set corresponding value bias zero. In this way,

Deep dive into Sink Value in GPT-OSS models! Analyzed 20B (24 layers) and 120B (36 layers) models and found (correct me if I'm wrong) Key Findings: 1. 20B model has larger sink value, 20B: mean=2.45, 120B: mean=1.93, 2. Clear swa/full-attn layer alternation: full-attn layers

just remembered that time Noam Shazeer dropped the hardest line ever written in an ML paper

I've written the full story of Attention Sinks — a technical deep-dive into how the mechanism was developed and how our research ended up being used in OpenAI's new OSS models. For those interested in the details: hanlab.mit.edu/blog/streaming…