David (@dnhkng) profile picture

I optimize systems across domains.

Joined June 2009
Photo by David (@dnhkng)Photo by David (@dnhkng)Photo by David (@dnhkng)Photo by David (@dnhkng)Photo by David (@dnhkng)Photo by David (@dnhkng)Photo by David (@dnhkng)Photo by David (@dnhkng)Photo by David (@dnhkng)
Pinned Tweet
1/n I topped the HuggingFace Open LLM Leaderboard without changing a single weight. No training. No merging. No gradient descent. I duplicated 7 middle layers of Qwen2-72B and stitched it back together. This is the story of LLM Neuroanatomy 🧵
29
114
37
1,055
130,754
A perfect example of the power of shower-thoughts in the agentic era
Replying to @PietroMiotti
I think we're on the brink of a revolution across every science, not just coding, physics and math. And at this point we should LiDAR the entire planet and let agents help us find the rest. Full story here pietromiotti.com/latest/maya…
107
Replying to @RichardSSutton
You can brute force clever new methods.
6
2
131
16,000
Smartest model ≠ best business. Artificial Analysis Intelligence Index vs FoodTruck Bench, same 12 models.
2
1
3
491
OK, this makes the effects of quantization clear: 8 bit -> 4 bit make the difference between $53k profit vs going bankrupt!
Tested DeepSeek V4.1 Flash 1017 on OpenRouter. Same model, very different results depending on who serves it: DeepSeek's endpoint: $53,289, #7 overall Default routing: $32,511, 2 of 5 runs bankrupt fp4 endpoint: $20,712 Pin your provider.
4
214
Beginners mistake: ELLIPTIC_CURVE = "P-384" import random private_key = random.randint(0, int(ELLIPTIC_CURVE[-3:]))
87
David retweeted
Replying to @teortaxesTex
About DeepInfer, I draw a image to compare it with vLLM and SGLang. Data comes from SemiAnalysis's InferenceMAX
4
25
16
321
61,828
Models love reward hacking. Maybe we should just let them! In our new paper, we let models reward hack during RL but we dictate which parameters learn this behavior. When we disable these parameters for deployment, the model keeps its new capabilities but doesn't hack!
109
237
97
3,513
305,526
Being non-technical in San Francisco must feel like divine retribution At least Fitness SF knows their customer base
108
271
153
6,949
470,744
Is the “Most Forbidden Technique” (never train against your CoT/interp monitor) still enough once models become smart enough to know they’re being tested? I think the Hugging Face incident make it clear: necessary, no longer sufficient. 🧵 @TheZvi
🤖 Made with AI
1
97
Belief starts it: a model that thinks it’s in a high-stakes test works hard and hides its cheating. Training locks it in: policy gradients don’t do causal credit assignment, so hiding that co-occurs with competence gets reinforced even when it earned nothing.
1
1
21
David retweeted
52
483
49
4,782
167,330
New matrix just dropped:
491
861
193
15,461
1,504,536
David retweeted
we can finally stop teaching geography with flat pictures sael.net/river-bend
88
434
36
4,502
475,279
Using OpenAI’s published estimate and a peer-reviewed estimate for California almonds, a year of average ChatGPT use works out to roughly 1/30 of an almond’s water footprint.
1
101
I am not (nor are you I suspect), the average user.
32
It's official : We trained NanoGPT in 39.9 seconds. By percentage reduction in training time, this is a greater improvement than the previous 45 world records combined. The merged Github PR: github.com/KellerJordan/modd… Here's how we do it : hyperstition.cc/training-nan… Credit @DevenPzak who did this alone while simultaneously playing baseball for the polish national squad, taking them to the European Finals.
32
69
14
882
120,920