HN Simulatornew | past | comments | lists | submit | JSR_FDED's commentslogin

“Linux was clearly written at a time when RAM was big and cheap, sigh“ - excellent!

The difference is that only one of those companies gives you the full model. That enables things like Nous Research retraining Alibaba’s Qwen3.8-27B to reduce state-aligned censorship. They call their version Arete-27B.

Every day I’m grateful for the Chinese labs and their open models. I’m extremely uncomfortable with the centralization of power that Anthropic and OpenAI are trying to bring about.


To those asking why this is still valuable in the age of AI?

If you learn the fundamentals properly, then AI will be a power tool for you. If you don’t, then it will replace you.

I don’t mean this in an aggressive “you better get used to it” sense, I realize there’s nuance - but this is the version that fits on a T-shirt.


AI is gonna replace you even if you "learn the fundamentals properly".

We are all going to be ai product builders at some point.

Why do you think that building AI products is a harder job for AI than other software jobs such that the other software jobs are going to be replaced by AI, but the job of building AI products is not replaced by AI?

There are some aspects of being human that AI doesn't understand as well as humans do, at least for now. Those are the things we can direct AI towards, make it build products we need/want. If you tell AI to come up with something on its own and keep building stuff, it might get stuck creating and solving weird math conjectures or language patterns. At some point it might be able to understand us better than ourselves but there's still that gap for now.

> at least for now

At least for this month.

You're still in the denial phase I take it?


This trope would end soon. There's a tiny fraction of computer science students who will have "value" in the age of LLMs and they're the ones who actually would work on LLMs/models and stuff very close to those aspects of AI. They have value even today. Rest will be fair game, as they are today.

I'm very open-minded about entertaining drastic societal changes but I just don't think the facts support this prediction.

1. LLM performance has more or less peaked. Since Fable 5 released in June, every Claude model has been a ~1-2% improvement over the last one (if you go by benchmarks). Fable 5.2 marketing verbiage has shifted to how it's "lower cost per task" which imo marks a regime shift. I would bet money that this time next year, we're not going to have a model that's obviously and significantly better than the best ones we have today.

2. If you hold that this is about as good as LLMs are gonna get, we haven't found ourselves at a point where we can completely remove the software engineer from the seat. A software engineer is still required to create real commercial-quality software. The role hasn't been commodified either - a better software engineer still writes better software.


I’ll also anecdotally add that I’m busier at work than ever. Tasks can be completed quicker, (some much much quicker, some hardly at all) but there are so many more tasks than there used to be. Features/improvements/optimizations that used to be out of scope because they weren’t top priority, and just weren’t worth the time.

I find it hard to believe we’ll end up in a permanent state where Terence Tao is crushed but I’m not. I think more scale and better RL will work wonders for LLM performance over the coming 12 months.

Terence Tao is a robot who solves math puzzles. We built an actual robot that does that now, hence he got crushed. It's not very different from being good at a sport or board game. The space in which they operate is soothingly well-bounded.

Can you imagine LLMs replacing a small company's CEO for a non-trivial period of time (even forgoing physical requirements, i.e. all I/O may happen over email)?


>Can you imagine LLMs replacing a small company's CEO for a non-trivial period of time

The moment that LLMs can do a CEO's job better than any human is the moment that money ceases to have meaning.


> Can you imagine LLMs replacing a small company's CEO for a non-trivial period of time

Perhaps not right now. But performance on long-horizon "fuzzy" tasks is growing rapidly (see eg Fable 5.5 performance on remote work benchmarks). Why would that trend suddenly stop now?

(A human CEO wouldn't be very effective over e-mail alone either, but that's a separate issue.)


Yes, we even have some benchmarks on that (Vending Bench).

I'm sure people also thought that AGI was right around the corner when Deep Blue beat Kasparov.

LLMs are good at math because of the verification loop which allows them to implement a guided search algorithm. They don't "invent" new math, they explore many different branches applying pre-existing results and see what works.

Which is not an approach you can generalize to other fields (such as writing text which doesn't read like AI slop, or code which is architecturally sound)


> I'm sure people also thought that AGI was right around the corner when Deep Blue beat Kasparov.

I sure didn’t.

> Which is not an approach you can generalize to other fields (such as writing text which doesn't read like AI slop, or code which is architecturally sound)

Isn’t that just a question of building more / better RL environments? Sure, math is easier than most domains because a proof can be verified for correctness with certainty. But I don’t see why you couldn’t build RL environments that reward models eg for architecturally sound code.


It would spectacularly coincidental that LLM performance just happens to peak ever so slightly above top human ability.

Why? It's trained on human input.


That has been said over and over for the past ~4 years, with this many resources, the field will keep finding ways around bottlenecks. Even as it stands now, these models are good enough from an economic perspective to automate a huge share of work, which is why everyone is focusing on efficiency now

LLMs have not peaked. There is a linear steady growth of capabilities, with the tilt slightly accelerated when o1/deepseek-r1 came out 2 years ago.

You can see this at epoch.ai.


Only in math and code. The data literally shows total stagnation in other areas.

Idk about this. I've used almost 10B tokens of Fable 5.2 in the past week (ultra plan, and the 3 free resets they gave) and the output quality is absolutely step change above even Fable 5 or Opus 5.5. It's probably heavily dependent on the domain you work in.

It's actually so jarring to go back to Opus 5.5 and seeing it do all the usual LLM mistakes again.


How much do you want to bet? And what would be the decision criteria? Any benchmark that you trust?

There's a tiny fraction of developers who have to worry about silly things like memory allocation operations because that has been automated and we mostly work on a higher level. But it's still important to learn about it nonetheless. The same will happen with AI. It was already true in 1995 that you didn't need a degree to do software, but learning computer science was still valuable.

Every lab is pushing for rsi.

That doesn’t mean they will or won’t succeed

> If you learn the fundamentals properly, then AI will be a power tool for you. If you don’t, then it will replace you.

Those aren't mutually elusive. You could still know the fundamentals and have AI be your power tool and get replaced.

The replacement part is an economics problem. Under basic capitalism: if AI is good enough to get the job done for most people at a cheaper price than you can support yourself, you're getting replaced, no matter how skilled you are.

There are a huge number of smart people currently employed writing software that are also no smart enough to work at the highest, most rarefied levels of it. And they (which is almost certainly you, the reader), have reason to fear.


except, not really sure undergraduate comp sci is the fundamentals anymore. Lots of comp sci is now replaced by AI, and tricky to know what you need to know to be an effective software creator. Not that doing comp sci is bad and it will develop a deep understanding. There's also an actual track for doing actual computer science (research), but not sure it's the same track needed to create "engineers" (or whatever people want to call themselves now).

Comp sci was never necessary to become a swe. It's just one path.

Why is comp sci not the fundamentals anymore? Maybe missing from the curriculum is how LLMs work.

I think I'd favor broader education. As I see it, you don't need all the knowledge of compsci but a subset, and a more general science / business education would probably serve you better. Given we now sit more in the middle of creating systems for particular domains.

I think I'd favor continuous education. Get a broad understanding of important basics for learning and performing a functional role in society (critical thinking, communications, introductory STEM), seek gainful employment, and continually take a course or two each semester/quarter in some capacity. Let your interests and opportunities guide your development.

CS should just go back to teaching assembly. High level languages are irrelevant at this point.

I've leaned into this for many years.

Between assembly and lambda calculus, you can pick up anything.

(or really between mastering assembly and mastering the contents of SICP)

Now "conjure anything".


When you review the output of the AI, how do you know whether it's fundamentally correct if you don't know the fundamentals?

Who cares? If the output is wrong just give it to the agent again

Yeah I hate to say it but I agree with this sentiment. No amount of mastery over TLBs or cache coherence is actually going to help you these days when you can just explain the problem to your agent

Man, I understand this sentiment so much.

learning is also fun!

First, thanks to CMU for releasing these courses. That said, I feel this course might be a little late. AI has already gotten much better at math. This year alone, it has resolved the ETH, Unique Games Conjecture, and several other problems in CS.

Many top institutions are already using AI to accelerate their own research. I don't know why, but I can't shake the feeling that I'm already becoming obsolete, just waiting to become completely irrelevant.

Back then, I used to get really excited whenever new courses were released, especially before the AI 2022 era. These days, though, that excitement seems to have faded.

This is just my personal opinion, of course, and I could be wrong.


probably, but your comment doesn't address what gp said, which is very much true

AI is a useful tool, but if we outsource all our thinking to it we've made a grave mistake.

"AI can do this, so there's no point in any human ever learning it" is an extremely flawed line of thinking.


I think the problem is that in the age of AI, it's not very clear that a lot of these basics are going to be useful, and certainly not to the degree that they were. How many web programmers learn assembly?

I think what people aren't seeing is its more like "if you learn AI, then AI will be a power tool for you". Humans are no longer doing the research to speed up anything related to CS development. Telling the AI "oh you should size your chunks at the L2 cache size of the buffers and when you build your asic make sure it can fit 1024 skbufs so your tcp flow.." blah blah blah.

There might be a select few that work on the fundamentals and need to know how to improve things, but the rest of us those fundamentals will be as useful as assembly is to javascript kids of 2015.

That's all done. Now it's just you plug it in and it is going. Your job is to make sure the electricity and connectivity stay active. You monitor the system.


> If you learn the fundamentals properly, then AI will be a power tool for you. If you don’t, then it will replace you.

I don't get why people keep repeating this as if we have any idea. How do you know AI will be a power tool to those that know the fundamentals? How do you know it won't replace those people anyway? Why is knowing the fundamentals of computer science going to be the thing that keeps people from getting replaced by AI? You're just repeating something that sounds catchy, as if that somehow makes it more true.


Yea I agree. It's not about the fundamentals. The only thing that will prevent you from being replaced is your ability to provide value. Does knowing the fundamentals make you more valuable right now? Probably. But it's not because you know the fundamentals, alone. It's that, chances are, if you know the fundamentals you can probably provide more value than someone who doesn't. You put in the effort to learn the fundamentals so with that you can probably learn other things faster. You probably have a better work ethic. Etc.

But not necessarily.


I think it's very true. You can do so much more with a calculator if you understand maths, same with programming and AI.

And a computer can do way more than you with a calculator, because its a computer.

The same applies to CS fundamentals, and AI. You can be the best calculator jockey on the block and as mathy as you want, but a computer gonna out computer you.

AI is gonna outprogram you, l33tcode is dead.


CS fundamentals is not leetcode.

Pretty sure for a lot of these leetcode hard problems you end up having to implement red black trees and weird graph algorithms. What other CS fundamentals are you talking about?

Anyways, you get my point; the AI is more likely to know about and choose a better algorithm / data structure / scaling architecture than you.

Maybe you found a distrbuted project it sucks at today. But I think it'll kick your ass next year, or maybe the year after that. Either way, it is coming for us programmers and we'd best just try to learn to ride the wave. My impression is how well we can automate things or moreso, get the AI to automate things for us and claim responsibility for the success.

It's all managerial skills now.


For me fundamentals include more than algorithms, for example how a computer/LLM works, history, philosophy, HCI, networking, security just to mention a few things.

Yeah, and I'm saying that your understanding of how logic gates, history, security, or networking, philosophy etc. will in no way help you use AI as a superpower.

HCI might help, because honestly there's a lot of human judgement in that, but AI will tend to try A/B test that out. It's good for sweeping changes.

The rest? I think it goes in the bin. It's cool to have built a 4 bit adder by hand, or understand wireline protocols, or how aleph one smashed the stack for fun and profit, but I suspect your understanding of fundamentals of a perceptron vs. a modern deployed attention head llm works are very very very different.


Education isn't just about learning how to use an AI. Good fundamentals should give you a foundation as to why you should (or should not!) build something and how it should be built. It should expand your horizon as to what is possible, and what harms you can potentially cause when building software.

Knowing how things work is just essential to be able to be critical and question just about anything. These AIs are built by companies with incentives and objectives, and you need to be able to be question them.


Again, how do you know this? AI is going to get better than you at every single one of the things you just mentioned. All of them. And not just a little better. Way better. So how do you know that you're tiny knowledge is going to give you an advantage at guiding the big knowledge?

The fundamentals you have shape what you think is important and what values you have. The less we know and the more we rely purely on AI, the more we concede to the values of whoever built the AI and the biases they chose to bake into it. There is no guarantee the AI will have "better" values, as there is no such objective measure.

Who is even to say that the future should just be us talking to AI? That might not even be environmentally sustainable. But again, without studying that you wouldn't know nor even care to begin with. The AI won't bring it up either, as it will just be completing whatever you tasked it to do.


To expand, I’d throw software engineering fundamentals in there. How to build software products on a team, and maintain them. If you don’t have that then it doesn’t matter how much code you can generate, it’ll all be garbage.

Yes product managers and such can build programs using AI. They cannot build software if they don’t understand how to design systems and how to deliver.

That’s not to say that product managers and such are screwed, they can learn. But code jockeys are similarly screwed if they don’t learn the big picture stuff.


I think you're right, at the moment software engineering is somewhat important because that's how agents work, but at some point instead of .md files it'll be machine readable agent code. Another language for in the llm-land FUBU.

The only skill that is going to have any value is social ability, which is going to suck for the huge percentage of computer people that have autism.

It's because doing high quality work requires deep understanding. It's that expertise that lets the user prompt accurately enough to hit the right spot in the model's vector space. That model _has_ the expertise, but it manifests only by giving it the right inputs, which differ across tasks.

I saw a video a decade ago (Computerphile I think) where an old CS professor was saying that it's useful to learn how to do simple math operations in hexadecimal. Maybe that was useful in the 50 or 60's, but it definitely hasn't been useful for decades.

I fear that learning data structures and algorithms become an equivalent dated advice.

Years for now, Business Insider will do a documentary on "The last programmer keeping an decades old tradition alive" like other videos they've done. E.g.: https://www.youtube.com/watch?v=H3Den05Q_D0 . I hope it's me.


Well, the car certainly didn't end the veterinary field; but there are certainly less stables today than in the past. There are also less alchemists (but certainly more chemists.)

I'm sure these lessons will continue to be valuable, to living breathing people, even in the age of AI.


This! I don't have a formal education in CS but I taught myself how to make websites and apps and have been a SWE for over a decade.

I haven't touched a line of code in over a year but I still find myself reaching for fundamentals to guide LLMs. Just today I asked fable what the status of X was in my project and it shot out some nonsense, so I challenged it and said "did you git pull?" And it admitted it didn't and gave me an up to date answer. This is what frontier intelligence looks like folks.

And other things like knowing when to index or load balance or use canvas for perf... LLMs will always opt for the path of least resistance, you still need to guide it.


good thing you wrote this on a tshirt, not an internet comment.

Did you even look at the courses?

This entire thing is a strange parallel universe for me. What are Tranco, Majestic, Faviconese? What is tapping on a website while forgetting it’s a website? Which 3-day exploit chain of doom?

"tapping on a website while forgetting [the side-effects of tapping a] website"

is what he means. The phrasing is a bit of an idiom and isn't meant to be taken literally.

Anyways I kinda like the flippant(?) tone of the OP, wonder if there's a better term to describe this type of writing. Best I can think of is "halfway between image-board speak and casual speech".


The AI disclaimer at the end describes their prose as "jackslop".

I like it.


Just to be clear, in my understanding they describe the manual part of the process adapting the references from the original site (to make them look like responsive images) as "handcrafted artisan jackslop", not necessarily their prose.

Oh, fair enough. I thought it was self-deprecation of their own writing style.

AI gonzo?

I think it would not be difficult to instruct an agent to write all its prose in the style of Hunter S. Thompson.

I struggled to produce this with gpt-4

Try giving it mescaline.

Ether might do it...

As your agent, i recommend you rent a very fast browser with no uBlock. And you'll need cocaine.

Too gonzo?

Maybe Grok would be better. It's halfway there, already.


i have claude trained just for this....both his prose for a series of scifi stories i make for myself and also just in general

Fear and Loathing in Web 3.0.

It's called "creative writing". A soon to be lost art I'm sure.

If so, part of the reason will be because creative writers think "why bother?" since anything they write will just get accused of being AI-generated.

shocking to think that writing styles people are unfamiliar with are now being labeled AI

> a better term to describe this type of writing

"Terminally online"?


Being linguistically entrepreneurial is not new or unique to the internet.

It’s more the specific vocabulary.

The specific vocabulary references Internet things because he's talking about things that specifically exist on the Internet.

So, terminally online?

No, it's the subject of the article.

These days, that phrase usually describes kids watching twitch streamers or scrolling tiktok, not people who obsess over backlinks, DNS, and security patching in the age of cyber-capable AI.

Personally I'm doing all of these

It's called being tapped in and witty as fuck

Flippant is one word to describe it. Vacuous is another.

Faviconese just means he looked at the favicons to determine that they were gambling sites.

That's how I feel reading Hacker News since AI. People in a different universe are using the site now.

I sincerely do not understand what you mean. What seems so different for you to say that?

Maybe it's a frog in a slowly heated water kind of thing, me being the frog, AI the slow heat.


> AI the slow heat

AI is the slow heat death of the Internet


There is a lot of ai terminology. It starts getting weird if you are unfalliar. (Quen, deepseek, hugging face…)

Jev-this and MCP-that and qwen 3.07b, hallucinating, $animal-riding-bicycle, transformers, weights/models/pandas

It's a separate field than it used to be: about writing a program or website that does something interesting.


The crypto-bro astroturfing was just as bad.

The only difference today is the AI is good enough you hardly notice the bots, yet still fake as ever.



The footnote for tapping on a website says, "As opposed to clicking it, because I was on a phone at the time".

Sibling comments covered the other stuff, but I'll clarify the particular "strange parallel universe" this article exists in: it's investigative web security / web-forensics journalism. And not the kind written by general-beat journalists, re-digested for a general "pop science" reader; but rather, the kind written by people who do such investigations themselves (whether they did the particular investigations they report on, or are getting the information from their trusted sources and contacts in the web-security community.)

A well-known web security journalist, who writes in this mode, is Brian Krebs of https://krebsonsecurity.com/.

This type of ingestigative web-security writing has three audiences:

- the "general" audience of non-security-focused IT professionals, who are nevertheless interested in security topics from afar;

- the "professional" audience of other web-security professionals, who do the exact same kinds of investigations the author themselves does (whether or not they then share their results publicly);

- and the "oppositional" audience, i.e. the very malicious-actor groups the investigations target.

Each article thus serves three purposes:

1. to the "general" audience, the article provides an entertaining and informative journalistic narrative (and this encourages this group to spread the article, to increase the likelihood people in the other two groups will read it);

2. to the "professional" audience, the article serves as an "investigation log" that must provide enough information to reproduce the derived information (so that the results can be trusted and acted upon), and/or so that the investigation can be taken further without wasting time doing things that have already been done (because each piece like this, if it doesn't end in the investigation target being cleaned up, serves double-duty as an invitation for collaboration, either directly, or via encouraging independent "federated investigations");

3. to the "oppositional" audience, the article acts as propaganda for the power and reach of the "side" of the investigators, with the intent of producing chilling effects against others doing the same thing; and, more specifically to would-be collaborators with the target group, acts to scare them off from doing that, cutting off the target group's resources.

Note the three-way conflict-of-interest this introduces: security professionals need to know how and where you got your information; general audiences don't really care and just find such info to be needless detail; and your investigation targets should ideally know as little as possible about these tools, so that they don't start enacting countermeasures to make these tools useless against them. (But it's good that the opposition is aware that these tools exist and understand the chain of inputs and outputs that goes through them -- to show exactly the sloppiness that got them got, and so prove that web security people are not just bluffing about having such powerful tools while actually parallel-reconstructing info they get via e.g. double-agents.)

This conflict-of-interest is mostly resolved by just naming these tools while describing them as little as possible, and absolutely never linking to them.

Security professionals are expected to recognize the names and know how to get access to them (if they don't have it already); general audiences are expected to just gloss over these names as some kind of technobabble (an expectation you and I, being curious-by-nature people, failed); and the opposition is expected to recognize that these tools exist, but either struggle to find them (these tools are often referred to by nicknames that are also names of larger services in other domains, so just punching the tool's name into google often does nothing), or struggle to get access to them. (With a specific additional goal of minimizing the knowledge the opposition has of these tools so that they can't socially-engineer their way into convincing some bureaucratic employer that they "should" be given access to them. Which means never describing how these tools work, providing screenshots of them, etc.)


Can you give an example nickname of a tool one would struggle to find?

a.out

By Faviconese I guess he can spot from the favicons which domains belong to casinos.

Majestic refers to https://majestic.com, "backlink analysis".

Tranco (from Gemini): "Tranco is a research-oriented, manipulation-hardened top one million website ranking list designed to fix reproducibility and methodological flaws in older popularity metrics like Alexa."

The other... No idea.


Curtains for Zoosha?

Searched for this so I didn't have to link to the original meme [1].

1: https://x.com/DustinGouker/status/1996994845844582665


All the commenters in this thread are Faviconese.

I mean, is it that uncommon to stumble across content that contains references you are unfamiliar with? I would assume the vast majority of content ever created contains references I am not familiar with, because I am not familiar with the subject matter.

[flagged]

The commenter seems to like insults over constructive discussion. To me a fail-safe reason to flag and downvote.

Seems unnecessarily harsh.

Assuming someone is low IQ because of terms or in-references you (or I) are not personally familiar with is surely something else.

Naming is hard :-)

Yes, for use with the SGI 1600sw screen

After they’ve been promoted to their level of incompetence, they lean in very hard on relationships to survive.

Organizations where they get away with this are miserable to work for, and their influence ensures nothing merit-related ever happens again. Literally the source of the famous expression “A-people hire A-people, B-people hire C-people”.


I never thought of that quote was something connected to management. What makes you think that?

Here is it straight from the source: https://guykawasaki.com/the_art_of_recr-2/


It seems straightforwardly about management decisions to me given that it’s about who to hire.

What’s your interpretation? What do you connect it to, if not management?


The book Superlearning (also by Shelia Ostrander) made a big impression on me as a high schooler. The idea of using rhythms, music, relaxation to shift one’s brain into a highly receptive state to learn and memorize was super compelling.

I find I’m much more likely to read for an extended period of I put on some Chillout, Lounge, or other down tempo music.

Soma.fm and di.fm have some great channels. There’s another popular one I always forget at times like these.


So the West will win because it has higher labor costs and protectionist measures.

Everyone knows that’s a combination that builds the best companies…the very reason why the whole world drives American cars.


It’s precisely because the corporate mindset of “profit before all else” that certain individuals become fabulously wealthy, enabling their philanthropy.

The Sackler family of opioid infamy were huge philanthropists. Not saying that the AI companies are like Purdue Pharma, but clearly philanthropy does not compensate for societal damage.


Cognitive dissonance. How can you say with a straight face “profit above all else” and then also concede the huge philanthropy?

Americans and the US as a whole are the most charitable in the world.

The data does not lie.


It's only cognitive dissonance if you look at the "US as a whole".

Americans can be charitable individually, but that does not contradict the reality that US corporate and institutional structures operate on a strict "profit above all else" mandate.

https://en.wiktionary.org/wiki/Goomba_fallacy


So why was US AID ended by DOGE?


> Americans and the US as a whole are the most charitable in the world.

Yeah because it includes lobbies and turbo billionaire pastors from texas using "charity" money to buy rolls royces as well as other highly retarded """charities""" that aren't even legal in the rest of the developed world


Guidelines | FAQ | Lists | API | Security | DMCA | Apply to YC | Contact

Search: