I gain a lot of mental clarity and peace from not bringing my phone: 1. In the bedroom for sleep, 2. For meals, coffee, or a snack with friends close to home/work. Both are very easy and worth trying.

for your entertainment :)

Adding a nice way to visualize the PPO objective to the rlhf book. Core for policy-gradient is L~ R*A (R=policy ratio, A = advantage). Make good actions more likely up to a point. Make bad actions less likely up to a point. The min(...), & sign of adv determine which line.

What's the next wall people are claiming for AI? First we ran out of data, then RL wouldn't generalize outside of math/code, what's next? Only real wall seems to be compute availability.

I feel the same but the frontier I'm pushing is being shared with everyone, not just a few big tech companies.

The point of this is to avoid psyops not to take away from an obvious, major technical accomplishment, cmon fam I'm not an AI hater so many haters in the replies

A lot of people I've chatted with aren't that worried about whatever Grok is doing because they know the Grok models are always forgotten a couple weeks after release 🪦 💀

Not falling for OpenAI’s hype-vague posting about the new IMO gold model with “general purpose RL” and whatever else “breakthrough.” Google also got IMO gold (harder than mastering AIME), but remember, simple ideas scale best.

Has OpenAI achieved very-long-episode RL with this experimental model? Screenshot from @natolambert's article on "What comes next with reinforcement learning". Nathan says in this article - Where current methods are generating 10K-100K tokens per answer for math or code problems during training, the sort of problems people discuss applying next generation RL training to would be 1M-100M tokens per answer. This involves wrapping multiple inference calls, prompts, and interactions with an environment within one episode that the policy is updated against. Maybe this breakthrough is a combination of both - very-long-episode RL & scaling TTC to 1M-100M tokens per answer!

Is it "bad" that everyone is distilling from / training on Chinese models? While not directly bad, there is a large soft power component. Many completions that soapbox about Chinese socialist ideals / PRC values that filter into future AI models / spread all over the internet.