Of course, it's software.
Easy mode: a bad system prompt update.
Hard mode: an adversarial example in the context.
@karpathy
-
Software Issues: System Prompts and Adversarial Examples
By
–
-
System Robustness Against Component Failures and Adversarial Behavior
By
–
I just feel like this is the particular problem but not the *actual* deeper problem. Any part of the system should be allowed to go *crazy*, randomly or even adversarially, and the rest of it should be robust to that. This is what you want, even if robustness is very often at
-
Offline Tracker: Synthetic Data Generation for Tesla Self-Driving Networks
By
–
This is not very different from Tesla with self-driving networks. What is the "offline tracker" (presented in AI day)? It is a synthetic data generating process, taking the previous, weaker (or e.g. singleframe, or bounding box only) models, running them over clips in an offline
-
LLM Model Size Competition Moving Towards Smaller, Smarter Models
By
–
LLM model size competition is intensifying… backwards! My bet is that we'll see models that "think" very well and reliably that are very very small. There is most likely a setting even of GPT-2 parameters for which most people will consider GPT-2 "smart". The reason current
-
Free Educational Content with Revenue from Services
By
–
Good question I do want Eureka Labs to be a proper, self-sustaining business but I also really don't want to gatekeep educational content. My default thinking is that the content itself is free and permissively licensed, the revenue comes from everything else, e.g. running the
-
Eureka Labs AI launch announcement with website GitHub and social media
By
–
Website: https://
eurekalabs.ai
GitHub: https://
github.com/EurekaLabsAI
𝕏: @EurekaLabsAI -
Training stability control with loss and gradient norm thresholding
By
–
Cool! For the spike I'd try e.g. `-sl 7 -sg 7` to keep instability in check earlier in the training. (will skip update if loss/gradnorm > 7 sigma outlier is detected)
-
GPT-2: Perfect Introduction to LLMs and Distributed Training
By
–
Exactly as intended! GPT-2 is a beautiful "hello world" to LLMs but also distributed training etc.
-
Estimating compute improvements from 2019 to 2024
By
–
This information was never released but I'd expect it was a lot more. In terms of multipliers let's say 3X from data, 2X from hardware utilization, in 2019 this was probably a V100 cluster (~100 fp16 TFLOPS), down from H100 (~1,000), so that's ~10X. Very roughly let's say ~100X
-
Arithmetic Operation for Layer Norm Standard Deviation Calculation
By
–
Do you see an arithmetic operation that could help us calculate this layernorm standard deviation? pic.twitter.com/nV4wwJRCLB
— Andrej Karpathy (@karpathy) 11 juillet 2024Do you see an arithmetic operation that could help us calculate this layernorm standard deviation?