Fun fact: the wedding industry in just the United States is 4-5x larger than the global cyber insurance industry (~70-80bn USD vs ~15-20bn USD)

Talking to @nfcodes about going from Next.js -> LiveView. His biggest complaint is around colocation of client/server-side stuff. His words: it's much easier to write front-end code with React compared to doing it with LiveView. One complaint is the pattern of having client-side hooks all in app.js - even if they're used in a single LiveView. Would be nicer to colocate them with the LV they're used in. (I am just the messenger)

Probably not the "real" Q* but interesting concept nonetheless. My understanding/summary: First, I think most people understand that chain-of-thought prompting improves an LLMs ability to reason. But, why? For a "non-optimal" language model, chain-of-thought prompting increases the likelihood of a model producing "correct" continuations. I think this probably arises naturally as a function of training data. Correct answers to complex reasoning problems are likely to show up next to the work/reasoning that produces that answer. This observation led us to: 1. Chain-of-thought prompting techniques to elicit this reasoning in LLMs 2. Generation of datasets with intermediate rationales for fine-tuning LLMs to produce these rationales naturally without prompting The challenge with both is that inevitably there are rationales which are out-of-distribution. Additionally, generating valid rationales is challenging and requires expert annotators. So how can we build reasoning into these models implicitly? This work kind of builds on the idea of introducing a scratchpad, pause tokens, or other meta tokens to a language model "force" a model to think before generating a response. They do this in 3 steps: 1. Think. Generate rationales. Rationales are inserted between <|startofthought|> and <|endofthought|> tokens. 2. Talk. Take the hidden state output from each rationale, and pass through a "mixing head" before passing to the language modeling head to get logits for next token prediction. 3. Learn. Optimize parameters for rationale generation to increase the likelihood of rationales that make correct future text most probable. Some interesting notes: First, they only tried this with a strong pre-trained language model (Mistral-7B). I'm not sure how or if this would work when training a model from scratch, because the model would not be able to produce strong, valid rationales. They acknowledge, and I agree, that this approach would probably improve the performance of much larger models, just as CoT improves disproportionately with larger models. Second, because the base model does not recognize the new meta-tokens, IOT force valid thoughts from the get go, they initialize the token embedding for <|startofthought|> and <|endofthought|> to equal the token embedding for the emdash `---` because that generally denotes thoughts in writing. This was my favorite part of the whole paper, and I actually think you see some of the consequences of this choice in their examples. The examples almost read like they belong between emdashes. They also note that their loss function which mixes the normal language modeling loss with a loss relative to the generated rationales is equivalent in theory to language modeling loss for optimal language models. Empirically though the result is a more capable model. They use REINFORCE to optimize a model to produce valid rationales. My understanding here is that they reward rationales which make correct continuations more probable. There are stability challenges with this approach. One of the most important architectural decisions is how to incorporate thoughts with talk. They do this with their mixing head, but experimented with some other methods as well. I'm pretty sure this is a somewhat common thread in every reasoning-based paper I've seen, I guess because it is literally the central problem in reasoning. I've slowly been consuming some of the papers out there on building reasoning into language models. It's interesting to see how connected all of these ideas are, and it definitely feels like there's something "in the air" w.r.t. reasoning breakthroughs.

I knew the new Twitter logo looked familiar