Neel Nanda ✓

@NeelNanda5

Mechanistic Interpretability lead DeepMind. Formerly @AnthropicAI, independent. In this to reduce AI X-risk. Neural networks can be understood, let's go do it!

44 286Followers

122Following

5 359Posts total

63.5KViews on collected posts

Against accounts of the same size

6 posts from the last 90 days, next to the 10K–100K follower range. below its peers on both reach and engagement.

Median views502this account1 089median for 10K–100K

Reach, %1.13%this account3.50%median for 10K–100K

Engagement, %1.32%this account2.00%median for 10K–100K

Metric	This account	Median for 10K–100K	Ratio
Median views per post	502	1 089	0.46×
Reach (views ÷ followers)	1.13%	3.50%	0.32×
Engagement rate	1.32%	2.00%	0.66×

Others in this range → Compare with another account → How these benchmarks are built →

Growth & engagement

How the posts we collected actually performed: views and reaction rate post by post, what the audience did with them, and where the follower count goes.

Views per post

57.2K3 Sep

5.1K

849

155

764 Sep

Last 6 collected posts, oldest on the left. The scale is logarithmic: one post can outrun the rest a hundred times over.

Engagement rate per post

2.71%3 Sep

3.13%

2.00%

0.65%

0.00%4 Sep

0.00%

Reactions — likes, reposts, replies and quotes — divided by views. Median for 10K–100K accounts is 2.00%.

What the audience does

Likes80.7%1 542 in total

Reposts6.9%132 in total

Replies2.4%45 in total

Quotes0.6%11 in total

Bookmarks9.5%181 in total

Share of every reaction we collected for this account. Replies mean argument, reposts mean endorsement, bookmarks mean the post was worth keeping.

The follower curve appears once this account has two daily snapshots — we take one a day, and this one is on its first.

Latest posts

@NeelNanda5 What if we project every loop iteration output onto logits or through a J-lens? 59 views · 0 likes · 0 reposts · 0 replies 04 Sep 2026 · Open on X →

@NeelNanda5 I saw a paper today about triple extremes for "looped transformers". I honestly don't understand why they think something's better than quadratics. We can form looped transformers anytime in inference, while preserving observability, simply chainning CoT and Meta rea 76 views · 0 likes · 0 reposts · 0 replies 04 Sep 2026 · Open on X →

@NeelNanda5 Would you never trade interpretability for performance? Evidence seems to suggest market prefers the latter over the former, perhaps until it's too late to change the preference 😅 155 views · 0 likes · 0 reposts · 1 replies 03 Sep 2026 · Open on X →

@NeelNanda5 agreed, we need both and losing cot monitoring would be bad. we’ve also been finding that cot and interp monitors catch different stuff, so they stack well 849 views · 17 likes · 0 reposts · 0 replies 03 Sep 2026 · Open on X →

This take is partially inspired by discourse about the OpenAI looped transformer, but worth saying in general. I'm not too worried about a model with 2x the number of effective layers of GPT-4, but the risk of a slippery slope to a large number loops seems super bad. 5.1K views · 152 likes · 5 reposts · 3 replies 03 Sep 2026 · Open on X →

A concerningly common take seems to be that keeping Chain of Thought monitorable doesn't matter because interpretability will save us, or it's already useless This is total bullshit. CoT is our best current tool for safety & interpretability, losing it would be a major trag 57.2K views · 1.4K likes · 127 reposts · 41 replies 03 Sep 2026 · Open on X →