Bang An

Let’s sanity check DeepSeek’s claim to train on 2048 GPUs for under 2 months, for a cost of $5.6M. It sort of checks out and sort of doesn't.

The v3 model is an MoE with 37B (out of 671B) active parameters. Let's compare to the cost of a 34B dense model. 🧵

29.01.2025 17:12 — 👍 10    🔁 2    💬 1    📌 0

Michael just presented our paper at the AdvML-Frontiers workshop, and it won the Best Paper Award!

arxiv.org/pdf/2407.17417
TL;DR: Watermarking LLMs can reduce the generation of copyrighted content but poses challenges for copyright regulation.

14.12.2024 22:50 — 👍 2    🔁 0    💬 1    📌 0

Automatic Pseudo-Harmful Prompt Generation for Evaluating False Refusals in Large Language Models Safety-aligned large language models (LLMs) sometimes falsely refuse pseudo-harmful prompts, like "how to kill a mosquito," which are actually harmless. Frequent false refusals not only frustrate user...

In Lex's recent podcast, Dario, the CEO of @anthropic.com highlights the challenge of controlling false positives and avoiding the endless “whack-a-mole” in safety training. That is exactly why we developed the tool and dataset for auto-red-teaming false refusals.
arxiv.org/abs/2409.00598

29.11.2024 20:39 — 👍 4    🔁 0    💬 0    📌 0