here's an update on RLing wordle, my RFT run is all setup for bootstrapping from an initial set of wordle prompts and feedback history. what's next is to intuit how different config params improve reasoning + speedups, (topK, temperature, beta, etc) but for now I have code that

link - tokenbender.com/post.html?id=a…

+1 for this you can really tell that a lot of work has gone into turning off sycophancy in gpt5, and I’m really liking it. ( i noticed emojis have reduced too ! )

<experimenting> the purple run is what loose clipping and dense rewards are doing for me, clip=1.0, lr=1e-4, weight_decay=1.0 and thanks to Anish for some generous brev compute credits!

really fun read we need evals that probe language models to output “what it sees”, not “what it thinks it sees” also very cool interp approach to visualize learned world model

some MoE reads for today:

best pipeline to learn something new is to find blogs/papers to read, implement a tiny version from scratch, document it, hack on existing oss

the only MoE project idea I can think of: start with a <2b open base model and replace ffnn layers with moe