Revanche1367

Time complexity is hardly pure math, there are tons of practical uses in addition to the "purer" parts.

This has basically happened already in all but exact form: https://statmodeling.stat.columbia.edu/2026/08/27/258/. I bet somebody out there is working on a textbook right now as I type this, or maybe they're already about to click the button to publish (maybe even ask the LLM to click it for him/her). The case talked about at the link is particularly egregious, it seems like some otherwise intelligent and mostly grounded people, even tenured academics in STEM fields, are completely unaware of the dangers of relying on LLMs for accurate knowledge. AI is either inducing some kind of psychosis in them or people with pre-existing problems with psychosis are just using AI as a tool, it's getting harder to differentiate between these scenarios.

>gender is very clearly a social construct and we shouldn't get hung up on colors or names

Why? Is there some reason we shouldn't get "hung up" about social constructs? Can you argue a reason for this without reference to other social constructs?

The link doesn't support your claim. There is no evidence of a reversal, and the original researcher most subsequent authors cited apparently never suggested there was, only that it was inconsistently applied.

I would think people who find it hard to accept their loved one’s body decomposing would not like the idea of it burning to ashes either.

I know there is some controversy, I would say that the field of study also “causes” controversy in a sense if the disagreement has to do with the subject itself, that is true for any field. For example, set theory “caused” significant controversy in the early years of its development, yes it was people in disagreement and insulting one another, but the fundamental questions were from the subject matter itself.

Regarding widespread perception of math as useless, I guess I was thinking mainly of academic circles rather than the general public in that statement. Even while I hated math, I knew it wasn’t useless because my late father (who was a civil engineer) was very good at it, and he was adamant on teaching it to us. It didn’t work on me back then though lol.

As I was getting more interested in math while getting my degree in computer science, I realized much of my own issue with math earlier in life was because of the same thing just about everyone else that struggles with it has: the lack of any practical relevance in my own life because of how it’s generally taught. And the fact that many math teachers themselves don’t know the foundational/philosophical aspects of the subject, which would go a long way I think towards explaining that relevance if introduced in some pieces earlier on, rather than treating it like a rote memorization subject and being so focused on continual practice to develop heuristics with exercises that some students couldn’t care less about. Of course, that works for some people clearly, but it didn’t for me at the time and I’m of the opinion high schools should focus more on discrete math like set theory, and formal logic rather than calculus, which is of little relevance at that stage unless you’re also invested in studying physics. Sets and logic on the other hand imo are general enough right at the beginning to be relevant to just about everything people of just about any age are concerned with.

The quote reads to me like “as long as you can analyze the solution after the machine solves it and make it suggest new ideas based on that, you’re almost the same as a discoverer yourself.” I’m deliberately avoiding “mathematician” here because I think Tao probably didn’t intend to take it that far. But, for many people that difference will be indistinguishable, especially because there will very likely be “math bros” just like “ai bros” and “crypto bros” before them who want to pretend they’re mathematicians just as we see people pretending to be engineers and financial experts based on how fast they got on the hype train.

I’m no mathematician, though I do have an undergrad degree in math, and unless I had a graduate degree or made substantial contributions at that level, I would never consider myself anything but a “student of mathematics,” and that too not just because I have a degree in the subject, but because I still have a great interest in certain branches of math and I keep learning in my spare time from books or from adjacent applied fields I continue to formally study. As a kid, I grew up hating math but I became interested in it later in life because of coding, computer science and formal logic, pushing myself a lot through my college years to get some foundational understanding of the subject at a deeper level than the general population, just to get my foot in the door, so to speak.

Even if I’m an outsider from the circle of mathematicians, this whole AI takeover of mathematics leaves such a bad taste in my mouth because I’ve long felt that math was one of those things that everyone could agree is a useful and beneficial thing for humanity in general, and it was one of the least divisive fields, where people from different parts of world, from different stations in life could participate, and very few points, especially today, caused any controversy (unlike just about every other subject). It feels like a great loss for humanity to lose a brunt of our engagement with it to the machines, like a whole wing of a gigantic library starting to be closed off from us indefinitely because some people decided that we don’t _need_ to access it anymore, except on an even grander scale possibly than my imaginary library metaphor.

You're not contradicting anything the poster said. The reason for a lot of Chinese students and workers having high IQs wasn't the point of discussion, the sheer quantity was, and most people can see they have a lot of such people.

I had been on the pro plan for a few months and recently upgraded to the $200.00 AI Ultra plan because I've gotten good mileage out of 3.8 flash while keeping in mind that it's far from perfect. Argon is better in my experience in terms of reliability (especially in not claiming to have done things it clearly didn't) but I have to agree with the hallucination part, for a "SOTA" model (before Carbon came out at least), it hallucinates significantly more even on moderately complex tasks compared to Astra or Opus while being much slower than 3.8 flash (as expected of course). Another irritating thing is that like the 3.6+ models (but thankfully less often), it has a nasty habit of sometimes recognizing unstaged changes in the branch as not being its own work (whether done manually or with another agent) and deciding to undo them or even delete files. That's something I don't expect any SOTA model to do without at least asking the user.

Gemini so far has also had terrible computer-use capability or tool-use capability compared to ChatGPT (haven't used enough of Claude for these tasks lately to really compare). I tell ChatGPT to make some formatting changes to a LaTeX file I gave it and generate a PDF, it get's the PDF generated most of the time; occasionally the download link doesn't work, but that seems to be a server or link-generation issue, the actual PDF is generated and it quickly fixes the broken link when told. With Gemini, it's a long struggle and I use LLMs for formatting LaTeX a lot, so sometimes if ChatGPT is doing some long-running task for me, I have to begrudgingly give the document formatting task to Gemini, and it's dreadful every single time.

Before Argon was a thing, it once generated a PDF file, then started claiming PDF generation was not part of its functionality and the link that was previously working stopped doing so. I think I had inadvertently jail-broken a functionality because it refused to generate PDFs again and said the first time was a mistake. With Argon, it doesn't deny having the capability, but it takes much longer than ChatGPT, and messes up more formatting than it fixes. Its PDF viewer is also trash and seems like it gets built from scratch every time you want to view a PDF.

Plagiarism is on the part of the companies making them, the machine cannot really intend things and thus are not the things that are "guilty" of plagiarism. It also doesn't have to be the thing that decides anything, you can still decide anything of importance while acting as a guide.

I and countless others have been using them quite effectively for complex coding tasks since at least 2023. This seems to be a skill issue in itself or you haven't used any frontier models in about 3 years. Living in denial about new technology isn't very beneficial generally; look around you, the industry has already changed significantly, people like Linus Torvalds who are tough on even good human coders are willing to use LLMs nowadays.

If this was in another country that is not predominantly inhabited by people of western/central European heritage, US mainstream media would've been calling these "undercover" agents the "secret police" or "stasi" or some other term heavily connotated with tyranny and oppression.

Yeah...some people complain about OpenAI and Anthropic's model naming schemes but at least it's fairly obvious from the names how they scale, how is ordering for Argon and Carbon supposed to be interpreted? And there's the 3.x and 2.x models which I don't think have elemental names (at least not public ones). The obvious atomic number or standard weight approach doesn't work.

Maybe it's a Russian bioweapons company masquerading as an Internet traffic control company pretending to be in bioweapons manufacturing so any snoopers think it's just anti-LLM obfuscation.

Except in this case the monkeys are pretty good at coding as long as you have enough bananas (tokens). They can follow directions pretty accurately and work independently for at least short periods and incrementally test the work even if you’re not present onsite monkey-handling all the time. I don’t understand this hard anti-ai stance from open-source groups except as a submission volume issue. If and when nearly the entire professional industry changes to LLM-based code in most cases, what are they going to do to find those relatively few people who are still willing to write everything manually and spend their spare time maintaining these projects?

And warfare of course, the stuff they won’t openly talk about but we might find out about months or years down the line when it’s used to destroy another school or hospital with people inside.

How would we know if we can't identify them and they don't tell us? The ads are among us...

>many people were only familiar with LLMs as the generic AI tool for everything

That’s exactly how the biggest vendors have been marketing them, it’s no wonder non-experts think of them that way. I’m not an expert myself but have a passing familiarity with theory as a SWE and grad student who did some research for machine translation and currently am in a computing-adjacent field. I use LLMs heavily for work and have never come across any coder-oriented tutorial, guide, etc. encouraging use of LLMs as decision models specifically, even if they are being used that way “beneath the hood.” Jev’s announcement seemed sort of new functionality to me but even in the first post about it on HN (where I heard about it), iirc people were pointing out that LLMs do the same thing, at least to some extent.

It says at the bottom of the main page some articles at least are written with AI assistance in the linked article for “November Singularity” and mentions the author cannot guarantee accuracy.

While the people behind this project seem to be real and apparently have credentials, this article is giving me serious whatever the textual version of uncanny valley is called. It hits so many of what reads like assorted buzzwords meant for marketing but guised in programmer-speak. One moment, they’re talking about the Curry-Howard correspondence and then next moment they’re describing what a data type is to people who are presumably more than casually familiar with LLMs. And the worst offense imo is linking to that “November Singularity” bs clearly written by AI to sell hype to people without much of a technical background (blog writer even admits some articles are written with AI “assistance”).

Finally! As a SWE, I was getting worried scientists would be left out of the increasingly-worried-about-getting-replaced club that SWEs and mathematicians are currently enjoying being members of.

I wasn’t talking about altogether different fields, perhaps you meant branches of the same field in your last part?

While this is somewhat true, people also overstate how narrow a Phd student’s general understanding of the field really become as they specialize. For pretty much any serious field, specialization is inescapable without spending decades minimally studying the field via many branches, but a Phd isn’t just preparing you to have a specialized understanding of a specific branch within the field, because very often contributing something mostly new and worthy to that branch requires a mix of depth and breadth, the proportions of which depend on the exact field, branch and specialized topic. Most Phd curriculums I’ve seen still require students to take several advanced courses outside of their main topic area and take exams in multiple areas within the field. Somebody working in only one area would need to have at least graduate level understanding of multiple branches to get through these. I have also seen Phd students who are not broadly interested or comfortable in certain core branches of their field so there are all types and different requirements for different programs.

"Endangering" is an exaggeration based on what actually occurred it seems. What the car did was technically in violation of traffic regulations but it was careful to not hit anybody or cause significant damage to people's property as it did so. This is essentially no different than a careful driver ignoring (intentionally or otherwise) a sign restricting them from entering an area, but still driving through the area successfully without doing any harm. Technically, traffic rules were broken, by mistake (an algorithmic one most likely) in this case, but there were no lasting effects.

People make mistakes like that all the time, and sometimes they still get ticketed for that kind of thing, and other times they get a simple warning or don't get any form of rebuke at all. How many of us have never accidentally gone into a one-way traffic lane from the opposite side and managed to still drive through and get back on track without killing or maiming anybody? It's unlikely anybody would get charged with a crime for this, nor should they over a simple mistaken rule violation that did no harm.

Replying to correct my phrasing in the edit, meant to say “so that humans can reduce the need to think about things _like_ survival and comfort,” but ended up saying the exact opposite, but the end kind of suggests the earlier part was a mistake so hopefully it wasn’t misunderstood.

(I change my phrasing multiple times as I’m writing and sometimes leave earlier changes in without realizing and mistakes like these happen, quite a lot in fact. I blame it on ADHD and anxiety, both of which I was professionally diagnosed with at the age of 33.)

Yeah, wouldn’t they rather be doing real impactful engineering work, like figuring out means to display even more ads on personal user devices, maximizing the number of hours kids spend doom-scrolling per day, or making robots that are deliberately designed to take jobs away from as many workers as possible?

But, imagine instead being an engineer of any kind today and you’re only getting paid to work on boring things like advanced automotive technology and building cars that can maneuver themselves in midair and might possibly fly in the future? Ridiculous.

(This message should not be taken as support for either Musk or any of his views or business ventures.)

Your last sentence reminded me of a quote by Alfred North Whitehead which I read many years ago in one of my CS textbooks: “Civilization advances by extending the number of important operations which we can perform without thinking about them.” I think about the essential idea of this statement quite often but didn’t remember the exact phrasing. As I went to look that up, I was thinking to myself something like “bet he would’ve changed his mind if he knew about LLMs!” but in finding the quote with its wider context, it appears he would’ve been thrilled:

“It is a profoundly erroneous truism, repeated by all copy-books and by eminent people when they are making speeches, that we should cultivate the habit of thinking of what we are doing. The precise opposite is the case. Civilization advances by extending the number of important operations which we can perform without thinking about them. Operations of thought are like cavalry charges in a battle — they are strictly limited in number, they require fresh horses, and must only be made at decisive moments.”

Apparently, part of his philosophical thought was that the ultimate goal of mathematics is to reduce (maybe even eliminate) how much thinking we do. Not exactly what I thought he was going for all these years lol. I wouldn’t read much more into it but it’s not hard to see how the statement could be perceived as this renowned early 20th century mathematician, logician and philosopher implying that he would’ve have loved the idea of machines reducing the need for research and development in engineering and mathematics (even if only to allow researchers time to focus on more important things).

Edit: As I’m not familiar with the breadth of his works and know of them primarily through secondary sources, I can also see how what he said meaning that the purpose of mathematics is automation in a very general sense, so that humans can reduce the need to think about things other than survival and comfort, and being a mathematician, he likely saw mathematics, philosophy and knowledge in general as one of those “leisure” activities we should be thinking about instead.

This is one of the biggest benefits. I'm not a plugin user so I don't remember how they're different from skills. Not speaking about the tool in OP specifically but regarding skills in general: you can have essentially the entire task or some of its subtasks (which you'd ordinarily direct to the agent using the skill, the main benefits of which are clearer natural language directions) handled with assistive tools written in any language your environment supports. You can additionally supply the skill with its own config files, data files, etc., building essentially a small repo within particular skills and making these resources however closely tied to the skill as you prefer.

This is particularly useful when the agent would otherwise be doing a task that is programmable, or the agent is inconsistent or unable to do it in an efficient, performant or other preferred manner. The programmatic tool assistance means at least some parts of the task (possibly its entirety) could be handled in a well-tested and deterministic manner. How reliable the determinism, optimality, etc. are depends on how well you or the agent writes and verifies the code, tests, spec, etc. in the first place of course, but that only needs to be done correctly once.

The other great benefit is reduced token use, potentially a massive reduction. Imagine a task Claude takes 10 minutes to prepare to do because it's reading and analyzing files for the 3rd time in the last 4 minutes, then it's asking for permissions several times in between to do this and that even after you put it in auto mode, then it's "thinking" again for the thousandth time about how to approach the task even though you had it repeat the task a hundred times already...by the time it actually starts doing the task, the most recent models are likely to have churned through tens of thousands of tokens (and that's a conservative estimate for frontier models) in the first 10 minutes. But, with tool assistance, you can have Claude directly handle some parts of that same task which, for example, needs unpredictable analysis or decision-making each time, but also have it use a set of highly optimized programs to handle the bottleneck parts or the parts you want done "your way." You could potentially cut down those 10 minutes of prep time + whatever minutes to actually do the task down to seconds, with barely a couple of thousand tokens used up! Just an example of course, but I've benefited from this strategy repeatedly across various tasks, simple ones and very complex ones.

Another great thing is that the tools don't have to be tied to the skill at all, they can live anywhere in your codebase. You could even manually run the scripts like any other to handle certain tasks in their entirety or do the prep work before setting an agent loose on the remainder. There are many different approaches with different advantages. It's like actually using the agent as your...well, agent! But now you're supplying him with the latest gadgets like M/Q does for 007.

No, it was this one: https://news.ycombinator.com/item?id=47112892. Last posted on HN 8 months ago, that's when I saw it. I misremembered the Windows version, it was 3.11. But, the site looks to be significantly different now and has different Windows versions, XP being the default emulation.

Tried it today on some relatively complex mid-sized real world tasks with several thousand lines of code generated by Argon. My impression is not good. Maybe it's due to first-day high user counts, but it seems to be doing worse than 3.8 flash in some coding tasks. What I noticed was that it often forgot significant parts of detailed, well-structured and precise instructions (which is somewhat rare with newer Claude and ChatGPT models) and not just for small details. It also seems to spend too much time thinking even when the task was a quick and simple one.

Without going into details, for a small part of a larger task, I asked it to classify different content types of different files and formulate a structured mapping (e.g., in JSON) between content types and another set of files (deliberately left vague here but not to Argon) from my codebase. I even offered it a multi-part draft of a classification system to start with and told it that if it can find other classification systems from international standards, academia or industry more suitable to the task than my rough draft, we can discuss replacing my suggestion. There were other parts to the instructions too, written out like an article with bulleted lists, precise program-like wording, etc. Then, I sent the prompt and left for a while.

When I came back, it was nearly done with the implementation but the classification system it came up with was a mess. It even admitted to making up "bogus" classification standards when I pointed out some errors. It also kept leaving gaps multiple times when I referenced errors to fix in subsequent instructions (and not once apologized the entire time lol). So my first impression was subpar, not really comparable IMO with Opus >=5 or Sol >=6, and seemed to leave larger gaps than 3.8 flash typically does. The most irritating part to me was not only the fact that it was making up things, but that it ignored my instructions to discuss things with me before deciding to shift away from my draft; I don't recall any major LLM ignoring an instruction like that in at least a year.