AGI Is Still 30 Years Away – Ege Erdil and Tamay Besiroglu

I just used o3 to design a distributed scheduler that scales to 1M+ sxchedules a day. It was perfect, and did better than two weeks of thought around the best way to build this.

You just asked it to design or implement?

If o3 can design it, that means it’s using open source schedulers as reference. Did you think about opening up a few open source projects to see how they were doing things in those two weeks you were designing?

yeah unless you have very specific requirements I think the baseline here is not building/designing it yourself but setting up an off-the-shelf commercial or OSS solution, which I doubt would take two weeks...

While impressive, I'm not convinced that improved performance on tasks of this nature are indicative of progress toward AGI. Building a scheduler is a well studied problem space. Something like the ARC benchmark is much more indicative of progress toward true AGI, but probably still insufficient.

Designing a distributed scheduler is a solved problem, of course an LLM was able to spit out a solution.

Dunno, in work we wanted to implement a task runner that we could use to periodically queue tasks through a web UI - it would then spin up resources on AWS and track the progress and archive the results.

We looked at the existing solutions, and concluded that customizing them to meet all our requirements would be a giant effort.

Meanwhile I fed the requirement doc into Claude Sonnet, and with about 3 days of prompting and debugging we had a bespoke solution that did exactly what we needed.

why would I do that kind of research if it can identify the problem I am trying to solve, and spit out the exact solution. also, it was a rough implementation adapted to my exact tech stack

the other models failed at this miserably. There were also specific technical requirements I gave it related to my tech stack

as noted elsewhere, all other frontier models failed miserably at this

the future is more custom software designed by ai, not less. alot of frameworks will disappear once you can build sophisticated systems yourself. people are missing this

“It does something well” ≠ “it will become AGI”.

Your anodectical example isn't more convincing than “This machine cracked Enigma's messages in less time than an army of cryptanalysts over a month, surely we're gonna reach AGI by the end of the decade” would have.

The point is that AGI is the wrong bar to be aiming for. LLMs are sufficiently useful at their current state that even if it does take us 30 years to get to AGI, even just incremental improvements from now until then, they'll still be useful enough to provide value to users/customers for some companies to win big. VC funding will run out and some companies won't make it, but some of them will, to the delight of their investors. AGI when? is an interesting question, but might just be academic. we have self driving cars, weight loss drugs that work, reusable rockets, and useful computer AI. We're living in the future, man, and robot maids are just around the corner.

That's a future with a _lot_ more bugs.

So you could stick your own copyright notice on the result, for one thing.

I'm not sure what is your point in context of AGI topic.

youre assuming humans built it. also, a ton of complexity in software engineering is really due to having to fit a business domain into a string of interfaces in different libraries and technical infrastructure

im a tenured engineer, spent a long time at faang. was casually beat this morning by a far superior design from an llm.

Because that path lies skill atrophy.

AI research has a thing called "the bitter lesson" - which is that the only thing that works is search and learning. Domain-specific knowledge inserted by the researcher tends to look good in benchmarks but compromise the performance of the system[0].

The bitter-er lesson is that this also applies to humans. The reason why humans still outperform AI on lots of intelligence tasks is because humans are doing lots and lots of search and learning, repeatedly, across billions of people. And have been doing so for thousands of years. The only uses of AI that benefit humans are ones that allow you to do more search or more learning.

The human equivalent of "inserting domain-specific knowledge into an AI system" is cultural knowledge, cliches, cargo-cult science, and cheating. Copying other people's work only helps you, long-term, if you're able to build off of that into something new; and lots of discoveries have come about from someone just taking a second look at what had been considered to be generally "known". If you are just "taking shortcuts", then you learn nothing.

[0] I would also argue that the current LLM training regime is still domain-specific knowledge, we've just widened the domain to "the entire Internet".

is this because the LLM actually reasoned on a better design or because it found a better design in its "database" scoured from another tenured engineer.

That doesn't mean the one what manages to spit it out of its latent space is close to AGI. I wonder how consistently that specific model could. If you tried 10 LLMs maybe all 10 of them could have spit out the answer 1 out of 10 times. Correct problem retrieval by one LLM and failure by the others isn't a great argument for near-AGI. But LLMs will be useful in limited domains for a long time.

What's the point holding copyright on a new technical solution, to a problem that can be solved by anyone asking an existing AI, trained on last year's internet, independently of your new copyright?

who cares?

Does it matter if the thing a submarine does counts as "swimming"?

We get paid to solve problems, sometimes the solution is to know an existing pattern or open source implementation and use it. Aguably it usually is: we seldom have to invent new architectures, DSLs, protocols, or OSes from scratch, but even those are patterns one level up.

Whatever the AI is inside, doesn't matter: this was it solving a problem.

What else is going to build it? Lions?

The only real complexity in software is describing it. There is no evidence that the tools are going to ever help with that. Maybe some kind of device attached directly to the brain that can sidestep the parts that get in the way, but that is assuming some part of the brain is more efficient than it seems through the pathways we experience it through. It could also be that the brain is just fatally flawed.

Here on HN you frequently see technologists using words like savant, genius, magical, etc, to describe the current generation of AI. Now we have vibe coding, etc. To me this is just a continuation of StackOverflow copy/paste where people barely know what they are doing and just hammer the keyboard/mouse until it works. Nothing has really changed at the fundamental level.

So I find your assessment pretty accurate, if only depressing.

I was pointing out that if you spent 2 weeks trying to find the solution but AI solved it within a day (you don’t specify how long the final solution by AI took), it sounds like those two weeks were not spent very well.

I would be interested in knowing what in those two weeks you couldn’t figure out, but AI could.

Because as far as you know, the "rough implementation" only works in the happy path and there are really bad edge cases that you won't catch until they bite you, and then you won't even know where to look.

An open source project wouldn't have those issues (someone at least understands all the code, and most edge cases have likely been ironed out) plus then you get maintenance updates for free.

> Because that path lies skill atrophy.

Maybe, but I'm not completely convinced by this.

Prior to ChatGPT, there would be times where I would like to build a project (e.g. implement Raft or Paxos), I write a bit, find a point where I get stuck, decide that this project isn't that interesting and I give up and don't learn anything.

What ChatGPT gives me, if nothing else, is a slightly competent rubber duck. It can give me a hint to why something isn't working like it should, and it's the slight push I need to power through the project, and since I actually finish the project, I almost certain learn more than I would have before.

I've done this a bunch of times now, especially when I am trying to directly implement something directly from a paper, which I personally find can be pretty difficult.

It also makes these things more fun. Even when I know the correct way to do something, there can be lots of tedious stuff that I don't want to type, like really long if/else chains (when I can't easily avoid them).

It is depressing but equally this presents even more opportunities for people that don't take shortcuts. I use Claude/Gemini day to day and outside of the most average and boring stuff they're not very capable. I'm glad I started my career well before these things were created.

ive got ten years at faang in distributed systems, I know a good solution when i see one. and o3 is bang on

it was two weeks tossing around ideas in my head

If you thought about it for two weeks beforehand and came up with nothing, I have trouble lending much credence to that.

All sorts of stuff containing no original ideas is copyrighted. It legally belongs to someone and they can license it to others, etc.

E.g. pop songs with no original chord progressions or melodies, and hackneyed lyrics are still copyrighted.

Plagiarized and uncopyrightable code is radioactive; it can't be pulled into FOSS or commercial codebases alike.

Someone raised the point in another recent HN LLM thread that the primary productivity benefit of LLMs in programing is the copyright laundering.

The argument went that the main reason the now-ancient push for code reuse failed to deliver anything close to its hypothetical maximum benefit was because copyright got in the way. Result: tons and tons of wheel-reinvention, like, to the point that most of what programmers do day to day is reinvent wheels.

LLMs essentially provide fine-grained contextual search of existing code, while also stripping copyright from whatever they find. Ta-da! Problem solved.

Ignoring the copyright issues, credit issues, and any ethical concerns... this approach doesn't work for anything not in the "database", it's not AGI and the tangential experience is barely relevant to the article.

It is unsurprising that some lossily-compressed-database search programs might be worse for some tasks than other lossily-compressed-database search programs.

idk why people here are laser focusing on "wow 2 weeks", I totally understand lightly thinking about an idea, motivations, feasibility, implementation, for a week or two

I find now I quickly bucket people in to "have not/have barely used the latest AI models" or "trolls" when they express a belief current LLMs aren't intelligent.

I agree. AI has made even mundane coding fun again, at least for a while. AI does a lot of the tedious work, but finding ways to make it maximally do it is challenging in a new way. New landscape of possibilities, innovation, tools, processes.

>Because that path lies skill atrophy.

I wonder how many programmers have assembly code skill atrophy?

Few people will weep the death of the necessity to use abstract logical syntax to communicate with a computer. Just like few people weep the death of having to type out individual register manipulations.

Call me back when ChatGPT isn't hallucinating half the outputs it gives me.

So 10 years at a FANG company, then it’s 15 years in backend at FANG, then 10 years in distributed systems, and then running interviews at some company for 5 years and rising capital as founder in NYC. Cool. Can you share that chat from o3?

How are those mutually exclusive statements? You can't imagine someone working on backend (focused on distributed systems) for 10-15 years at a FANG company. And also being in a position to interview new candidates?

Yeah that's the thing.

Personal projects are fun for the same reason that they're easy to abandon: there are no stakes to them. No one yells at you for doing something wrong, you're not trying to satisfy a stakeholder, you can develop into any direction you want. This is good, but that also means it's easy to stop the moment you get to a part that isn't fun.

Using ChatGPT to help unblock myself makes it easier for me to not abandon a project when I get frustrated. Even when ChatGPT's suggestions aren't helpful (which is often), it can still help me understand the problem by trying to describe it to the bot.

Who knows but have you read what OP wrote?

"I just used o3 to design a distributed scheduler that scales to 1M+ sxchedules a day. It was perfect, and did better than two weeks of thought around the best way to build this."

Anyone with 10 years in distributed systems at FAANG doesn’t need two weeks to design a distributed scheduler handling 1M+ schedules per day, that’s a solved problem in 2025 and basically a joke at that scale. That alone makes this person’s story questionable, and his comment history only adds to the doubt.

Most programmers don't need to develop that skill unless they need more performance or are modifying other people's binaries[0]. You can still do plenty of search-and-learning using higher-level languages, and what you learn at one particular level can generalize to the other.

Even if LLMs make "plain English" programming viable, programmers still need to write, test, and debug lists of instructions. "Vibe coding" is different; you're telling the AI to write the instructions and acting more like a product manager, except without any of the actual communications skills that a good manager has to develop. And without any of the search and learning that I mentioned before.

For that matter, a lot of chatbots don't do learning either. Chatbots can sort of search a problem space, but they only remember the last 20-100k tokens. We don't have a way to encode tokens that fall out of that context window into some longer-term weights. Most of their knowledge comes from the information they learned from training data - again, cheated from humans, just like humans can now cheat off the AI. This is a recipe for intellectual stagnation.

[0] e.g. for malware analysis or videogame modding

You can put me in that bucket then. It's not true, I've been working with AI almost daily for 18 months, and I KNOW it's no where close to being intelligent, but it doesn't look like your buckets are based on truth but appeal. I disagree with your assessment so you think I don't know what I'm talking about. I hope you can understand that other people who know just as much as you (or even more) can disagree without being wrong or uninformed. LLMs are amazing, but they're nowhere close to intelligent.

Wow, 12 per second on average.

Who hired you and why are they paying you money?

I don't want to be a hater, but holy moley, that sounds like the absolute laziest possible way to solve things. Do you have training, skills, knowledge?

This is an HN comment thread and all, but you're doing yourself no favors. Software professionals should offer their employers some due diligence and deliver working solutions that at least they understand.

the commenter never said they came up with nothing, they said o3 came up with something better.

I would say there's a big difference with AI though.

Assembly is just programming. It's a particularly obtuse form of programming in the modern era, but ultimately it's the same fundamental concepts as you use when writing JavaScript.

Do you learn more about what the hardware is doing when using assembly vs JavaScript? Yes. Does that matter for the creation and maintenance of most software? Absolutely not.

AI changes that, you don't need to know any computer science concepts to produce certain classes of program with AI now, and if you can keep prompting it until you get what you want, you may never need to exercise the conceptual parts of programming at all.

That's all well and good until you suddenly do need to do some actual programming, but it's been months/years since you last did that and you now suck at it.

There is one very specific risk worth mentioning: AI code is a potentially existential crisis for Open Source.

An ecosystem that depends on copyright can't exist if its codebase is overrun by un-copyrightable code.

That's a future paid for by the effort of creating current frameworks, and it's a stagnant future where every "sophisticated system" is just re-hashing the last human frameworks ever created.

It's not an existential crisis. You just don't merge radioactive contributions.

If it sneaks in under your watchful radar, the damage control won't be fun though.

true and with AI I can look into far more subjects more quickly because the skill that was necessary was mostly just endless amounts of sifting through documentation and trying to find out why some error happens or how to configure something correctly. But this goes even further, it also applies to subjects where I couldn't intellectually understand something but there was noone to really ask for help. So I'm learning knowledge now that I simply couldn't have figured out on my own. It's a pure multiplier and humans have failed to solve the issue of documentation and support for one another. Until now of course.

I also think that once robots are around it will be yet another huge multiplier but this time in the real world. Sure the robot won't be as perfect as the human initially but so what. You can utilize it to do so much more. Maybe I'll bother actually buying a rundown house and renovating myself. If I know that I can just tell the robot to paint all the walls and possibly even do it 3 times with different paint then I feel far more confident that it won't be an untenable risk and bother.

> and his comment history only adds to the doubt

for others following along: the comment history is mostly talking about how software engineering is dead because AI is real this time with a few diversions to fixate on how overpriced university pedigrees are.

I’ve had similar things over the last couple days with o3. It was one-shotting whole features into my Rust codebase. Very impressive.

I remember before ChatGPT, smart people would come on podcasts and say we were 100 or 300 years away from AGI.

Then we saw GPT shock them. The reality is these people have no idea, it’s just catchy to talk this way.

With the amount of money going into the problem and the linear increases we see over time, it’s much more likely we see AGI sooner than later.

its not dead, its democratized

Write me when humans will achieve hallucination levels smaller than ChatGPT.

Bingo. LLMs are consuming data. They cannot generate new information, they can only give back what already exists or mangle it.

It is inevitable that they will degrade the total sum of information.