txhwind

If we assume resources are infinite, any effort would look ridiculous. To achieve great goal, we have to limit ourselves deliberately on some aspects. It reminds me that "compression is intelligence". By limiting the parameter space, the LLM could extract deep representation from training data.

Competitors will try to touch everyone on the list.

Why is a plan "mode" needed? I really hate reading plans in conversation UI. One line in AGENTS.md is always better, and you can customize it to fit your document convention. From the communication perspective, the planning process is most about creating clarity and alignment, on things like scope, constraints and decisions, among people (and agents now), often requiring multiple rounds. The planning tool built in Codex or Claude Code is not persisted, version-controlled, nor well accessible. In practice, My AGENTS.md contains an instruction about writing a document before starting implementation. Usually I won't read that document, because I don't want to micromanage agents. That document mostly serves a historical purpose like ADR, helping me find out what agents missed, made mistakes on or misaligned on, if needed.

every tax can be expaned to a infinite series of rent.

However, the bank is happy to extend the loan infinitely for people with enough assets. It's questionable that whether such loans are cash flow neutral.

Most back and shoulder pains are caused by bad habits. Cutting something from your body doesn't help a lot. Go to gym or physical therapist first.

A free-to-read PDF host site is valuable enough. Most academic publishers have a pay wall.

I guess they want to use the email CC concept showing the agent is a side assistant.

It looks like a Transformer encoder post-trained on classification and regression tasks. The encoder-only model is less noticed in recent years, but this product finds a nice application for it.

In imperative languages, it's a leaky abstraction not reducing the cognition overhead, compared with the plain loop.

If the contraint is not in the signature, and cannot trigger a test failure with typical implementation, it doesn't exist.

other questions:

- can it be used to relax timing order requirement in pipeline parallelism? Each node update lambda on communication, and optimize weights at other time. - given that BP is using SGD, can batches and T share the same timeline in optimization, while keeping the descent direction?

Backprop can also be viewed L local signal passing, but it requires global timing coordination. The Lagrangian enables out-of-order update.

Nice introduction to a simple but useful idea! The Lagrangian works like a time-smoothed optimizing direction state, but it can be placed on any wire, even at non-differentiable boundary! Can it be better than existing training methods for discrete components like argmax, MoE or VQ-VAE? Maybe networks can be composed by a lot of learnable discrete components, or even bits and gates finally.