rocho

I've been using Fastmail with my own domain for a few years and never had a problem. Web and mobile apps are great.

I've set up IMAP syncing with my main Gmail so that it also acts as a backup. Eventually I'll ditch my Gmail for good, but sadly I'm not there yet.

Gemini models have consistently been the best, by far, at writing and explaining.

Indeed, that's what I was getting at in my last paragraph. Businesses invariably determine that speed of delivery is a superior metric to optimize for, rather than learning and maintainability. In academia the situation is not so clear cut, but there are still powerful incentives to deliver faster. We'll see what happens.

As a software engineer, I found that a codebase that is LLM-assisted or entirely produced by agents gets harder and harder to understand intuitively over time. By writing code manually, you build a very precise mental model of how it works. The time it took was the price for that. Now, the time cost has been largely removed but the payoff also disappeared: it's essentially impossible to maintain the same mental model with the speed and volume of AI contributions. Reviewing code is not a substitute for writing it, and new contributions to the codebase are essentially bound to LLM usage. So now we're in a position where the codebase and its capabilities grow fast, but are not backed by the full understanding of the engineers. AI tools have become our crutch.

Math is very similar, and requires an even higher degree of mental understanding to produce advancements. The effort spent trying to prove theorems and connect theories pays off in deep intuition; removing that struggle significantly impairs our understanding.

So no, you can't "be done with that, and go do other great work" without paying the understanding tax. And over time that's just not sustainable.

A simpler analogy would be gym training: the effort expenditure and the strength gains are inextricably linked. You can't go in there and start lifting weights with a forklift while expecting strength improvements.

In the case of software development, businesses have decided that speed gains are more important than maintainability, understanding, and independence. Will the same happen in math? The future in which mathematicians become LLM assistants (as opposed to LLM-assisted) and spend their days "turning the crank" just to review what comes out is quite bleak.

Strawman. There's very little pressure, if any, to drink all the time. On the other hand, businesses are pressing employees to constantly use AI tools. The sad fact is that in many cases, they do result in faster work, but at the same time they carry so many negative effects for the individuals.

Why do you consider bonds usury?

I absolutely agree, although even that doesn't solve the root problem. The underlying LLM architecture is fundamentally insecure as it doesn't separate between instructions and pure content to read/operate on.

I wonder if it'd be possible to train an LLM with such architecture: one input for the instructions/conversation and one "data-only" input. Training would ensure that the latter isn't interpreted as instructions, although I'm not knowledgeable enough to understand if that's even theoretically possible: even if the inputs are initially separate, they eventually mix in the neural network. However, I imagine that training could be done with massive amounts of prompt injections in the "data-only" input to penalize execution of those instructions.

I find Gemini is outstanding at reasoning (all topics) and architecture (software/system design). On the other hand, Gemini CLI sucks and so I end up using Claude Code and Codex CLI for agentic work.

However, I heavily use Gemini in my daily work and I think it has its own place. Ultimately, I don't see the point of choosing the one "best" model for everything, but I'd rather use what's best for any given task.

That's not DeepSeek, it's a Qwen or Llama model distilled from DeepSeek. Not the same thing at all.

You don't need to propagate it, you just need to show the gradient of the current position alongside with the classical evaluation, to give more context to the viewers.

That's correct. I saw a paper recently that showed how LLMs performance collapses when they are trained on synthetic data.

Why would they be? Cursor took an existing editor and added some AI features on top of it. Features that are enabled by a third party API with some good prompts, something easily replicable by any editor company. Current LLMs are a commodity.

To you maybe. But if Claude or any other competitor with similar features and performance keeps a lower price, most users will migrate there.