If I understand correctly companies like OpenAI could run LLMs without having access to the users new inputs. It seems to me new users data are really useful for further training of the models.
Can they still train the models over encrypted data? If this new data is not usable, why would the companies still want it?
Let's assume they can train the LLMs over encrypted data, what if a large number of users inject some crappy data (like it has been seen with the Tay chatbot story). How can the companies still keep a way to clean the data?
What is so special about IUT? They say the theory is "out of this world" but in what sense exactly? Did Mochizuki found a new interesting way to look at some ideas?
It's extremely complicated. The original document he wrote up was 500 pages of maths introducing effectively an entirely new theory. I studied number theory in uni, tried to read it, and understood barely anything of it.
Which obviously leads to the epistemological problem that the article points out. You had extremely good mathematicians like Scholze look at it and thought he found a flaw, then one guy from Arizona disagreeing that it is a fatal flaw and claiming to have fixed it, which Scholze doesn't agree with.
So what do you really make of it if only a handful of mathematicians can engage with it, and they can't even agree with each other. Probably the biggest value of IUT is that it puts to the test what even counts as a proof.
It kind of introduces a fun thought experiment, of a super high-level, complex equivalent of the Monty Hall Problem (which is so counterintuitive that even very intelligent and mathematically literate people will outright refuse to accept the established truth). How would we ever establish truth on something so monstrously complicated that only ~10-100 people in the world could possibly understand and at the same time so divisive that there cannot be a strong consensus?
I'm not a mathematician (but I've seen exact sequences and commutative diagrams) and to me the stuff out of his IUT papers[0] looks borderline LLM-generated. I can only imagine what the LaTeX source looks like.
You can express `a + b` or `a * b` in their regular algebraic notation or you can express them as a lambda expressions
ADD = λab.(a S)n
MUL = λxyz.x(yz)
Manipulating these expressions instead of algebra, you can suddenly compute things such as "+ * +" (Plus times plus). That will yield you another expression for sure, but we don't even know what that means.
So maybe an analogy would be, it's like you developed a field where, from that mess, you could derive important insights and even turn them back into proofs
And there's debate on whether all invariants truly are maintained throughout the entire process
>Would managing the git repo include CI/CD?
Not really, we already have some people who manage our Jenkins instance.
>Do you have a CS degree?
I don't, I have a scientific background and then I did a coding bootcamp. I don't really plan to get a CS degree since I already have hobbies that take all my free time.
Do you have any insights on what an analyst is doing in a company? Is it different from what I do now?
>I'm not so sure about the spaced repetition side of things.
From what I could read, many studies have shown this is an effective technique. From my personal experience it only works if you understand what you are trying to remember. For example, trying to memorize some complicated formulas that you don't understand is doomed to fail.
I think you misunderstand. They aren't doubting SRS. They're doubting that MathAcademy is doing it. They logged in after 2 weeks and had no reviews waiting.
Just a heads up that there's an explanation for this; I responded to the original comment (https://news.ycombinator.com/item?id=41556644) but I'll also paste here in case anyone misses it:
1. New tasks are selected only as you complete existing tasks (so if you come back after 2 weeks, you need to complete some existing tasks from 2 weeks ago to get new tasks selected based on your knowledge profile right now).
2. We are often able to implicitly knock out due reviews with new lessons. We're not just doing plain vanilla spaced repetition. We're doing a highly efficient novel version of it that we call Fractional Implicit Repetition (FIRe). I have a writeup on this that gained some traction on HN recently: https://news.ycombinator.com/item?id=40954571
I recognize myself a lot when he talks about the feeling of losing knowledge acquired years ago. I would be very interested with MathAcademy if it can help me fight this!
I am also wondering if it can help me learn new knowledge, especially new topics in maths I have never learned before?
Also what is the difference between MathAcademy and Brilliant?
I have read many skeptical opinions on Brilliant, so it would be interesting to have a comparison.
Brilliant doesn't have spaced repetition, it has much less coverage and doesn't have enough exercises to let you really master subject, but it's an easy to use mobile app with nice visualizations that can give a conceptual understanding of many things. If MathAcademy and Brilliant covers the same topic, Brilliant may be more visually appealing and contain animations. Brilliant is nice for getting overview of many things, but doesn't really go that deep.
MathAcademy is great for learning new topics that you haven't studied before, provided they have a course in that subject.
I've tried both and am currently using MathAcademy to learn new knowledge. It's very, very good at incrementally building out concepts and gradually adding complexity. I had no luck at all with Brilliant. MathAcademy practically forces you to bust out a pencil and paper and dive into problems, but Brilliant was more like watching a kind of neat YouTube video on a topic.
I reckon it can and I am about to give it a shot, myself. It sounds like a Duolingo sort of structure, which can be boiled down to "use it or lose it" which is not hard to support with compelling evidence. Say what one will about Duolingo, but my Spanish and German are passable because of it (no, I am not fluent, but I can get by in common situations). Ideally, MathAcademy will cultivate the same results.
I used to use them.
I really loved the fact that you can ask questions really freely and get quite raw answers (I don't like the so called "reddit style").
Then I stopped browsing them since it felt like doomscrolling even though these platforms were not designed to encourage that.
What?
/s