I think this is due to the fact that the language game that we have collectively decided to play, with AI, is not yet mature enough. In the language game of “giving a gift”, for instance, there is a distinction between giving a gift that you made yourself, and one that you bought at the store, but it is not that great, nobody cares a lot anymore. Right now, the language game of publishing words on the internet is strongly associated with the notion that these are “my words”, and the fact that they might have been written by a machine instead, introduces some discomfort. But it’s possible to imagine a future where this language game will have evolved enough that the notion of authorship by a particular machine entity, will not in itself be such an important issue, anymore.
LawZero, Yoshua Bengio’s startup, also proposes to automate scientific research and experimentation, from the perspective of safety, by being explicitly “non-agentic”:
For those interested in this type of climate data visualization apps, I have worked on this one in the past, which is actively maintained with a lot of love, and very nice:
For "something that is published" (which includes a comment like this) I clearly dislike it too, but for chatting / texting, I realize that I often use it more than my interlocutors, and I'm not sure why. There's a part of lazyness I guess, but also a vague sense of "conveying the impression of a never ending stream of communication", which is closer in my mind to the essence of the chat medium. In French, there is also the additional layer of "using the accents or not".
Very interesting project! I cannot resist mentioning an old project of mine that was made in a very similar spirit, but way before any LLM: wrapping a classic Lone Wolf gamebook around a very crude text parser: https://projectaon.org/staff/christian/gamebook.js
I had written an entire "framework" for it, in JS (so in theory more books could be supported), but it never went anywhere: https://github.com/cjauvin/gamebook.js
Author here: my learning objectives, in order of importance, are: (1) vocal understanding, (2) speaking, (3) reading, and (4) typing and writing (far further). As I explain, I'm mostly bypassing the problem of typing by using screenshots (ChatGPT's OCR capabilities are very good, and Anki works very well with it too).
I've been using ChatGPT to learn Persian (as a third language) for more than a year now (along with a heavy use of Anki), and it's incredibly useful and surprisingly good, for about everything: romanization, OCR from screenshots, deep explanations of complex and subtle stuff, etc.
Counterpoint: I actually really enjoy his style of writing, which I find clear, patient (you must appreciate visual examples and exploration though) and very often challenging and stimulating (recent examples: the posts about the bigger brains, Conway GoL engineering, and biology / evolution). I find he regularly introduces intriguing and useful ideas, like the distinction between "brain-like" computers (which includes neural networks) and more general, Turing-like mechanisms, and I find his overarching concept of computational irreducibility (even though he didn't invent it) quite profound in its implications. I would add that his posts read like an ambitious research program in progress (like a book written one chapter at a time) and that is why I think certain concepts (like ruliology) may appear obscure at first, if you didn't read a lot of stuff that comes before. One tiny nitpick I have: certain language tics, like the constant use of the "And, yes" pattern (he really uses this a lot I wish someone somehow told him).
If you consider this from the angle of Wittgenstein's "language games", you could say that the problem would be "simply" to distinguish between these two, quite different, language games, and act accordingly.
> Emacs is a powerful tool, but it also demands a lot from its user. Eventually I got tired of dealing with the host of plugins and customizations that I needed to keep my system running the way I wanted. I'm at a point in my life where I would rather spend my spare time on hobbies, hanging out with family and friends, and otherwise not messing around with a patchwork of ELisp code snippets that I've cobbled together from various sources. I gradually stopped using Emacs in favor of more modern tools that are less flexible but also less of a hassle.
I don't know how many times I've read a variation on this. It took me a very long time, but now I pretty much made my peace with that: I use Emacs (for certain things), I use VS Code (with Emacs bindings), I use Apple Notes.. I don't find that it's possible or reasonable anymore, the desire to be "pure" and use only ONE tool to rule them all. The same for messaging apps, chatbots, etc.. I now embrace extreme diversity.
What I find interesting about spaced repetition is the underlying thesis that raw memorization, in certain contexts, is playing a more important role for learning than what some modern education ideas would make you assume. In mathematics or programming, for instance, there is this idea that understanding a concept is better than memorizing algorithms or recipes (derivation methods for instance). But spaced repetition challenges that, in a sense.
There's a surprising gulf between word recognition and overall sentence comprehension. I'm learning Farsi with a combination of Anki and Youtube videos and sometimes I find myself in the weird state where I recognize every word in a sentence, but yet cannot assemble its overall meaning.
For a complete solution requiring many traditional high-level components like templating, forms, etc, then yes, clearly Django. But for something looking more like a REST API, with auto-generated documentation, I would nowadays seriously consider FastAPI, which, when used with its typed Pydantic integration, provides a very powerful solution with very little code.
> Is there a good primer/community consensus on how to configure shell so all history gets saved in sensible manner (including continuously, so it survives a crash)?
I find that zsh is quite good at addressing some of the issues you mention.
Your answer somehow suggests that solving such problems is merely "recognizing which algorithms to apply".. which is another way of saying that they are more "pattern matchers" than "true reasoners". I would think, on the contrary, that these problems (at least the tougher ones that are coming in about two weeks) require more than pattern matching, but I'm not sure exactly what are my thoughts on that.
Because we can assume that these problems necessarily cannot be in the training set of any current LLM (and are hard enough, i.e. they should be more than mere variations on existing problems), I'm wondering if they can be a good benchmark to get a better sense of how good they really are at problem solving?
For building a serious business CRUD app, to this day, I've never found a JS UI toolkit that matches the power and scope of the 4.X era of Ext JS: https://docs.sencha.com/extjs/4.1.3/
You say AI, but even more precisely: I believe that his ideas form the core basis of modern LLMs actually (their probabilistic word sequence modeling underpinning), which is indeed extremely impressive and deep.
Could this be the answer I needed to run an SQL Server image that refused to run on my M3 MBP? I was about to, sadly, try Docker Desktop, because of that.
I absolutely agree. Recently I had to work on a complex client app, and I could simply not believe the amount of trouble you have to go through when you want to increment the version number of things like React, MUI, webpack, TS by.. one!