Is there a polite, constructive way to tell him no word has been put into the "original author's mouth"?
What he's reading there is just the following sentence in the book, separated with a comma...
>
TextBlob does a pretty good job at the translation: "C'est une vérité universellement reconnue, qu'un homme célibataire en possession d'une bonne fortune doit avoir besoin d'une femme!".
It can be argued that TextBlob's translation is far more exact, in fact, than the 1932 French translation of the book by V. Leconte and Ch. Pressoir:
"C'est une vérité universelle qu'un célibataire pourvu d'une belle fortune doit avoir envie de se marier, et, si peu que l'on sache de son sentiment à cet egard, lorsqu'il arrive dans une nouvelle résidence, cette idée est si bien fixée dans l'esprit de ses voisins qu'ils le considèrent sur-le-champ comme la propriété légitime de l'une ou l'autre de leurs filles."
In this case, the translation informed by ML does a better job than the human translator who is unnecessarily putting words in the original author's mouth for 'clarity'.
Maybe it's me reading too deep into it, but that article sounds like the author had strongly held predetermination that simplistic ML translations that generate malignant deformations of English are "obviously objectively better", and could not accept the evaluations from actual first language speakers that they don't even make sense. IOW the fight had already happened and the author won by power. Two points that stuck up to me:
> [x] Try some more sentences. Which is better, ML or human translation? In which cases?
Any English second language speaker just knows this: There are basically zero cases where ML translation is in any way better or more accurate. That includes DeepL and ChatGPT w/GPT-4 proper. It's just tap water at restaurants. Which is a fantastic choice to pair with actual drinks. It's weird this sentence appears here at all.
> [...]and why is TextBlob so good at translation? Well, behind the scenes, it's using Google translate, a sophisticated AI able to parse millions of[...]
And this part. Maybe it's just me, but I think it might be showing that the author first tried to hand-roll translation, as seemingly needlessly lengthily elaborated up to this part. It could be that they then either faced technical challenges or failed validation by certified Frenchmen, and had to rewrite the section as a guide to use Google Translate API.
There seems to be an endemic misconception, mostly seen among but not limited to American people, that American English is a perfect language that is also completely disconnected from internal thoughts and intents and monologues, that simplifies language translation problem to a simple matter of "convert[ing] the formal grammar rules for one language, such as English, into a non-language dependent structure, and then translate it by converting back to another language" as the author claims, while in reality even the possibility of "non-language dependent structure" is still under debate. This kinds of attitude always existed, but now it's borderline beyond annoying.
Unless something I can’t predict changes, using LLMs for EVERYTHING smacks of bloat. They’re just not very efficient at many tasks. I don’t know why in a world where people complain relentlessly about bloat in web development, how it could possibly be right for everybody to just use something as heavyweight as LLMs.
Because while you and I rightfully complain about web development bloat, normal people who use the products made with those tools don't care, as long as it does what they need the product to do
How about using an LLM to help you write a MuZero-like model designed for a specific task? (Also MuZero took like 12 hours to train on old hardware, so my MacBook might be good enough here) obviously we’re not here yet, but it doesn’t seem far away. Hell, you could train a small LLM specifically just to do this.
I don’t know, it’s not my field, but the applicability of an LLM in a lot of current AI fields such as motion control, vision, planning, etc. That is classic AI stuff, I don’t think LLMs are appropriate yet.
Too big, too slow, too much resources etc. And it’s not even clear to me (mind, who is ignorant) that the LLM is some generic model suitable for all AI like tasks.
Making a big splash right now to be sure, but seems to me there’s still room for the core concepts folks have been working on for a long time.
Yes. You need traditional ML to detect, predict, cluster, etc. Then you can layer on LLMs for reasoning on this (provided you have existing documented reasoning on such predictions).
Edit: At least that’s what I’m doing. I could be wrong though.
i dont really understand what an LLM is going to do for you in probably 99% of all problems that have an ML solution except maybe write a block of code using sklearn or whatever to get you started. There are several reasons for this:
1. there is no reason to pay the api costs for an LLM to ingest data for you and do something with it when basically all it will be doing is writing the python codes for you that you will eventually be using
2. the LLM doesn't represent some sort of conceptual understanding of whatever you are trying to do to solve your ML problem, so you can't rely on it to be clever and answer questions or brain storm new ideas
3. even if you have a reason to use an LLM in some data processing pipeline it will only be one stop on the information super highway you are trying to create. you probably are going to use it to do something, but you probably also are going to be doing other things (e.g., image segmentation, time series analysis, etc.).
LLMs are great. but they are really just like, one more tool to have, they aren't the only tool.
I just have a hobbyist curiosity in the area, I think learning ml basics removed alot of the magic of ai hype, I at least like to think I'm less susceptible to hype and bs. I feel like I have a good idea what is going on and what the limits are of models and how to use them. There is a bigger world of ai besides llm
Depends on your goals though, using llm just seems like using any other api to me.
Definitely read up on ML basics if you are going to build anything. And by ML basics I mean the concepts of supervised/unsupervised /reinforcement learning - and in particular model evaluation. It is absolutely critical for any ML-based system to have a structured way to judge how well a system works: whether it works well enough (according to some target), or which of two alternatives is better, and that the nature of the errors are not critical or maybe some of them can be mitigated etc.
My personal approach, as someone not in the field but who has tangential interest (perhaps intellectual and not industrial interest) in it, my approach is to start from scratch, implementing the basic ideas in my favorite languages with no supporting libraries. I personally like this approach, but I doubt it gets anyone a job. So it just depends on your goals.
I think it is more relevant to learn how the technology that will take over most junior tasks works, before it takes over everything and the only job left is being LLM Architect, like in those offshoring projects, the difference being those offshore folks aren't there any longer.
The answer to your question depends on whether you think LLMs will remain the state-of-the art solution in the domains you are interested over the next few years/decades/timescale you care about.
Relatedly, I recommend Andrej Karpathy’s “neural net: zero to hero” YouTube lectures. He breaks things down to the fundamentals in this really nice way, and you follow along in your own jupyter notebook.
Looks pretty nice (high-level skimming of content .. I work in the field but more DL focused. Mathy ML is my nemesis).
Two questions related to learning in general:
1) I feel I have a good undergrad level grasp of ML but not at a grad level grasp. The math is a bit overwhelming. I am not a fan of conferences like Neurips. Any one else try to conquer this challenge and have a story to share?
2) A bit off topic but the XR course they have linked on the page is also cute. It lists both Unity and MRTK. I have a Hololens2 and am curious about spatial mapping and awareness (just to learn at the moment). Any suggestions on what is a good stack for this area? I have very little 3D graphics background .. unity seems a bit too high for serious work and an engine seems too level :(
It's just a big ad for Azure. There are probably way better things out there that really teach you machine learning, instead of just the product palette of microsoft.
Maybe you are right but "In this curriculum, you will learn about what is sometimes called classic machine learning, using primarily Scikit-learn as a library" should be useful outside of Azure
What he's reading there is just the following sentence in the book, separated with a comma...
> TextBlob does a pretty good job at the translation: "C'est une vérité universellement reconnue, qu'un homme célibataire en possession d'une bonne fortune doit avoir besoin d'une femme!".
It can be argued that TextBlob's translation is far more exact, in fact, than the 1932 French translation of the book by V. Leconte and Ch. Pressoir:
"C'est une vérité universelle qu'un célibataire pourvu d'une belle fortune doit avoir envie de se marier, et, si peu que l'on sache de son sentiment à cet egard, lorsqu'il arrive dans une nouvelle résidence, cette idée est si bien fixée dans l'esprit de ses voisins qu'ils le considèrent sur-le-champ comme la propriété légitime de l'une ou l'autre de leurs filles."
In this case, the translation informed by ML does a better job than the human translator who is unnecessarily putting words in the original author's mouth for 'clarity'.
https://github.com/microsoft/ML-For-Beginners/blob/main/6-NL...