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

Seriously people need to publish intermediate results. This is like sitting on a product idea for 12 years and then someone ships a worse version and gets all the credit.

That is also why they don't publish. They fear their peers/competitors would take the intermediate result which may be the critical step to solve it quicker than them and take the main credit/limelight.

They publish the intermediate when they admit to themselves they there are not going to make any progress in their lives and that intermediate result is their limit.

It is not how all mathematicians think, but the ones who spend 5-10years or more working on a problem do.


That just seems a doom spiral of making the field obsolete

In tech, people publish interesting, non game changing half-ideas all the time. There's incentive to doing that (edification, employment, cool points).


it’s crazy for you, someone who i assume has no personal stakes or experience in the industry, to use a word like “need” so boldly.

But in the industry I do have stakes in, it's a common practice to be rewarded for sharing early and often, to build in public, and gain credibility not just because you claim the prize for the end solution.

I think a lot of outsiders feel frustrated with public investment in academia, when the structures seem to reward someone spending 12 years on a solution with no clear signs of intermediate progress


No shade for the guy who worked on this for 12 years, because he didn't make the incentives that cause this, but all I get from this is that humans could have solved this a decade ago if the incentives were towards collaboration, rather than keeping progress secret in order to eventually get all the credit.

Do human mathematicians need to publish more incrementally? Seems many are hoping they’ll save up their findings for one big bang finding?

IMO academia could benefit from lower key, more tech blog style discussions of their work, and get away from this attempt to only publish the final result.


If you publish incrementally you don't get to write your name on discoveries in blood, win prizes, etc.

Someone else scoops up and runs ahead with your work.

In an ideal world, where there is no ego and clout chasing, your proposed way is of course much better.

Agent swarms doing research for example don't have these ego issues.


That sounds like a broken system not actually helping the field advance.

People should feel free to write about half baked ideas and whimsical hypothesis. That should be rewarded.

This focus on being first on a problem is about to become very outmoded.


It is, I find it liberating in some sense. You will have all of these ego people grasping for straws, trying to desperately prove their value just to be sniped and outdone by a random model for 0.06$

At some point when they have come to terms with that perhaps their value was their place in society regardless of their output, we can start building something meaningful.


This is basically the way the world works, not just mathematics or academia. Someone always captures most of the value and it's not usually the one that did most of the work.

I don't think the tech space works this way. People get a lot of credit for half-baked ideas / moving fast breaking things. For better or worse.

Funny you mention "moving fast, breaking things" as that comes from Facebook. Zuckerberg captured most of the value and the Winklevoss twins were left will relatively little in comparison.

Maybe now that LLMs are claiming credit for mathematical advances, humans can be free from this status-seeking game and actually start doing maths

Trouble is, nobody's going to call this hypothetical approach the Softwaredoug Half Baked Whimsy.

This is an issue across academia. Unlike the technology sector which is flush with 'risk capital' academics rely on grants. In order for grants to be dispensed, the funders need to have some way of quantifying a researcher's productivity. Papers and citations are the primary way to do this.

There may be opportunities to strategically 'thin slice' your research into a dozen groundbreaking papers, but generally it does not serve you well to publish half of an insight if you are giving an opportunity for other labs to scoop you and funnel all of the scarce funding to their institutions instead.


Some are cautious about hinting their competitors. Incremental pushes are hints for everyone else and they get scooped.

That seems like an outdated model and not really in the spirit of scientific collaboration

Scientific collaboration happens through publishing, not giving away the ideas/credit before publishing. Sure, there's collaboration, but that's not with the entire community, especially not in math which for big results is an all-or-nothing endeavor.

Also outdated since when? We've just seen a scandal about OpenAI scooping Levent on Navier Stokes https://en.wikipedia.org/wiki/Navier–Stokes_priority_controv...


For a proof like this, I assume many of the components are interesting enough to be published on their own. The authors just don't.

As I understand how credit goes for mathematical discovery, if you do 99.9% of the work and then I come in and put the last couple steps that make it actually work as a proof instead of a bunch of useful lemmas, the proof has my name on it.

You gotta eat.

Agreed. I know nothing about the field, but I always assumed that mathematics was the gold standard of an open, collaborative scientific process (especially considering how mathematicians are known to hold most other fields in contempt for not being as "pure" as them). If one mathematician sitting on a work-in-progress proof in total radio silence, for 12 years, is not uncommon, then clearly I thought wrong.

>Do human mathematicians need to publish more incrementally?

Yes, if you want a job in academia. Publications begets funding which begets more publications etc. Ideally you just solve the problem you want to solve and not worry about putting out partial results. But there is a time pressure to prove your productivity - you are not going to have much to show for 3 years of work if you aim to solve only the hardest problems in math. Interestingly, it's not clear the author held an academic position after his PhD.


Yes, they need to use Agile/Scrum. They need to create small JIRA stories that can be completed in a single math sprint. 12 years on a single story makes it hard to manage the expectations for the math customers.

Then it would be a game of 9 ball - last "potter" wins.

low key tech blog content can be automatically generated now

Something I don’t quite know, maybe someone closer to math can answer this question

It seems we’re making tremendous progress on the back of existing abstractions and concepts with the brute force ability to generate language and verify proofs.

For example, Tao says about the Riemann proof:

> appears to use the full power of modern developments on the Langlands program and arithmetic geometry

What I wonder: how do we create new abstractions and concepts? Like whatever “Langlands program” is? That then feedback into the system, and allow for even more sophisticated mathematics.

Are we assuming AI will do this? Or does this still need to be something at least agreed-upon by a mathematics community - in the same way programmers talk about a “linked list” without having an hour long conversation to define it axiomatically? And without that social aspect, without breaking down long proofs into Lego block common abstractions, math may plateau? In part because we lose the ability to ask meaningful questions?


This has been ambiguous so far. Most people guessed "the AI can't build new theory yet, but can do insane brute-force, which turns out to be surprisingly powerful, and probably the theory building will get there eventually". I think that in the next 14 days, as this gets processed and understood, there's a good chance we finally settle this.

Compare Elliot Glazer a month-ish ago:

> I occasionally talk to experts on all the MP's. Their rough summaries on what is needed to solve them usually go:

> Hodge: not a clue

> BSD: not a clue

> RH: not a clue

> P vs NP: not a clue

> Yang-Mills: "Well..."

https://news.ycombinator.com/item?id=50252673

So if we're in "not a clue" territory, maybe that's that.

I fear that it gets muddled again with e.g. the concurrent human announcement and that in retrospect RH turns out to be easier than expected, like many other problems before.

Math is hard, but how hard math is is also hard...


> I fear that it gets muddled again with e.g. the concurrent human announcement and that in retrospect RH turns out to be easier than expected, like many other problems before.

The proof in the human announcement is until unverified. I don't expect it to hold water, but to be sure we still have to wait like a month or so to get a few AI assisted analysis. (There are a few pure AI vibe-refutation, but I want to wait until a human checks the AI objections.)

>> the AI can't build new theory yet, but can do insane brute-force, which turns out to be surprisingly powerful, and probably the theory building will get there eventually

A lot of cutting edge math is close to brute-force search, like to throwing stuff to the wall to see what sticks. But at the high level instead of throwing bolts and nuts, you throw BMW and Ferrari.

For example, it's very hard to know how hard a problem is when you already know the solution. It's much easier to verify that the solution is correct than imagining how hard is guessing the tricks to solve it.

Usually after a big proof, parts are reused by the community to prove similar stuff and followups. After a while, the tricks get a name, perhaps lemas are upgraded to theorems, and they may get their own book or course. It's similar to refactoring in programming.

From Andrew Wiles:

> Perhaps I could best describe my experience of doing mathematics in terms of entering a dark mansion. You go into the first room and it's dark, completely dark. You stumble around, bumping into the furniture. Gradually, you learn where each piece of furniture is. And finally, after six months or so, you find the light switch and turn it on. Suddenly, it's all illuminated and you can see exactly where you were. Then you enter the next dark room...


CAVEMAN SUMMARY: AI no need theory, only people need. Fable understand OpenAI math. Fable explain well to human. People the problem, not AI.

I was reading a mathematician's take on a recent OpenAI proof (their area of expertise). They argued that models don't lack theory building, they just work at such a high level of abstraction that they don't need it.

Theory is for us, so we can understand what the models are doing. They gave the example of Fable, which understood and explained everything perfectly right away (written by OpenAI's unreleased model in terse mathematical notiations).

This was totally intuitive for the AI (Fable), but it took this person two weeks just to get a high-level grasp of it. And this person is one of the few people on earth who can (he is an expert in this niche of mathematics).


I imagine more semantically comprehensible proofs benefit LLMs as much as they benefit humans, from a time/energy/iq perspective.

Ergo, there are likely numerous deeper insights that will emerge when smarter models have the chance to go over brute forced, millions of line Lean proofs. Likely things humans have not conceived, but I believe will truly advance understanding. Post-train the existing models with that knowledge, and see them resolve these same problems with far less effort.

Just a hunch. I'm neither an AI researcher nor a mathematician.


Humans experience reality, and as they do so they gradually build maps—models—that allow them to reason about it with less overhead. Those maps are hopefully useful for some purpose, but also necessarily wrong/incomplete.

LLMs only operate on our pre-existing maps in their training dataset. (The first “L” is the key: language is itself a map.) They may be useful at refining those models, but they do not experience and therefore cannot get us to the next peak over.


I love how you wrote these half assed ambiguous/incoherent theories right under a thread where ai's made some of the most immense mathematical developments in perhaps a century.

It's not incoherent. It's describing an abstraction ladder.

It's how you graduate from 1+1+1 to 3*1. And so on.


Certainly, I am fully in support of more forms of data being added as the native mode of thought. Visual and tactile and audio based data would be necessary next step for physical robots.


> I love how you wrote these half assed ambiguous theories right under a thread where ai's made some of the most immense mathematical developments in perhaps a century.

Rest assured, not nearly as much as I love how someone on such a respected forum runs out of arguments in their first reply to me.


Because you made up incoherent undefined nonsense, and in your previous argument chain literally had to ignore reality to salavge your points. You would have had to define what you mean by experience and explain why sense organs that accept text are not sense organs while sense organs that experience visual and force and audio data are senses. Modela are multimodal in any case these days. And yes I am certainly rooting for ai to debe developed that also traina in visual, audio and force data natively along with language. Wouls be fun to see the next wave of copes and excuses aftet that.

Just for the benefit of others, are you referring to this?

https://web.maths.unsw.edu.au/~davidharvey/research/subnlogn...


Yes! Thank you for linking it here. I couldn't find it for the life of me.

> This was totally intuitive for the AI (Fable)

I'm regularly getting replies from a Fable-related model about our code base and problem space which would suggest that it totally inuitively groks it when it confidently explains and asserts things, and more often than I'd like it turns out that it's wrong, the thing is actually different and/or more complicated than what it made it out to be, and when I point that out it apologizes and agrees with me, rinse and repeat.

I'd be very careful with such claims. LLMs have been around for long enough for us to still fall for its confidence.


This is weird, so you think letting agent loose with "solve cancer, and giving them unlimited compute, with no further guidance, will yield results??

People seems to have lost their sense of judgment. It was the same with Deep Blue and AlphaX, the world's problem did not end. If we are to make progress in many field, it is still people who will have to build new theories, etc.


I thought theory building WAS the process of creating powerful abstractions?

Open question - are these proofs brute forcing, or are they generating new theory at a level of abstraction and complexity we can't understand?

Controversial statement: I think we've always done math by generalising from concrete experience. So concepts like numbers, spaces, shapes, distances, and curvature are based on physical experiences we use as anchors before expanding them far beyond physicality and looking for patterns in the abstractions.

If AI is generalising from patterns in latent space, that's a different kind of abstraction, and one that's much more alien to us, because it's going to be much, much bigger - too big to fit inside even the smartest human mind.

So AI theories and abstractions may be valid in true/false terms, but we will never be able to comprehend them in their raw form - just as we can't experience latent space directly, and will only experience tiny slices of it as outputs to our queries.


> If AI is generalising from patterns in latent space, that's a different kind of abstraction

How's latent space conceptually different from neuronal activations in our brains?

> just as we can't experience latent space directly,

Your LLM also doesn't "experience" their activations in the latent space either when generating tokens.

I'm not saying LLMs are not conscious/sentient/etc. Just saying if you ask your LLM which activation was stronger that led to a particular solution, they won't be able to answer without using external tools to inspect it.

It may just be that LLMs have such a huge context window (short-term memory) compared to humans, that they can process a lot more raw data without having to compress them into theories yet.


Human short-term memory and LLM context windows are very different things.

More generally, your mental model of how the human brain works seems to be very much aligned with how LLMs work and I would very much caution against that.


Then It's just a case of unambiguously defining the question, and getting the model to explain the answer.

I swear, if it says 42, I'll be laughing.


I feel it is the same in computer security, it will be interesting to observe for both, if among the proofs remaining to make or new, if a human could make them whereas models weren't able to do so, and in sec what new bugs are found that models couldn't. Also it would be interesting to see if new classes of proofs or bugs resist, despite their attempted learning to the models once known. Could something resist to their learning? Is there something human could durably do that models or future models can't?

The recent progress in AI maths I think has made it a lot more obvious and visible what maths is at its core: The study of interesting patterns that emerge from the consistent application of simple rules.

It's quite similar to the Game of Life and the community around it that tries to find interesting ever fancier gliders in that way.

The thing about maths though is that what counts as big and interesting is socially determined and chosen by the community over decades or centuries.

It remains to be seen whether human mathematicians will be content with also having the AI decide what's interesting.

One possibility I suppose is to find an objective definition of what's interesting, e.g. some sort of centrality score to a theorem within the tree of theorems it enables.


> find an objective definition of what's interesting

Say that any louder and the lecture hall will be lousy with economists!


A lot of people have raised the possibility of math may plateau because humans no longer have the insights for new abstractions, but the truth is nobody knows.

Langlands is interesting because it is most conjectural (and not completely clear what the precise statements are -- hence "program" and not "conjecture"), but proving special cases turn out to be tremendously powerful. One special case implied Fermat's Last Theorem, for example.


Maybe they'll do what I do with programming: tell it to suggest new abstractions I can use. I tell the computer all the time to think of ways to create new classes or refactor the codebase in a way that makes more sense.

If you ask for abstractions to use in your code you don’t get anything remotely “new”. I do it but I get well known, proven ideas (which is great)

The Langlands program is as old as 1967, it translates hard problems in number theory (like primes and polynomials) into problems in geometry and harmonic analysis (like symmetries and waves), where we have other tools to solve them.

The Langlands program just makes sense. The best most succinct definition of the set of real numbers can be put as "a complete ordered field." We usually experience them as decimals and infinitesimals, etc but when you think about them in the context of those three meaningful, powerful words you realize harmonics and geometry make so much sense. And also how primes etc are connected to quantum


Chat is this woo?

No. It's about connecting two areas/domains of mathematics.

An example would be how you can specify complex numbers in multiple forms:

- as a (real, imaginary) pair using the definitions of operations like addition and multiplication;

- as a 2x2 matrix, with 1 and i having matrix representations;

- as a 2D vector (angle, magnitude).

The idea is that if you prove something in one area you can apply that proof to the other domain using the equivalence mappings without having to prove all of those in the target domain. The basic construction is:

- have the problem in domain A (e.g. a complex number in (re, im) form);

- apply mappings converting domain A to domain B (e.g. representing the complex number as a 2x2 matrix);

- apply the proof from domain B;

- apply mappings converting domain B to domain A;

- QED

In a way, the Langlands program is like a subset of Concept Theory.


Given what we've seen, I see no reason to think that AI couldn't produce meaningful abstractions like perfectoid spaces or rough paths, and even give them memorable names. The pertinent issue to me seems to be the opposite of your concern. Right now, a scenario of 'runaway math' seems conceivable, where machines produce ever more complex results about ever more complex abstractions. At a rate exceeding the capacity of the human mathematical community to comprehend (the capacity limit might also be just physical - even if every single result could still be understood by humans, there might not be enough humans who care with enough time to go through all of it). The more real constraint seems to be economical, as in who would dedicate the hardware and energy to producing the math, and why would they make the results public.

In this case, deepening our math knowledge becomes an empirical process like a physical science. We have reams and reams of 'truth' in the physical universe, but nobody knows what its or how it connects to our current understanding. Now, mathematicians exist to empirically probe and test their assumptions against this massive 'physical' edifice of endless proven, interconnected maths like trying to understand the chemistry of a human cell.

It's there, we can observe it, but we still need to study it / take it apart to be able to explain it.


even the basics of LLM, we dont have full understanding of math behind LLM.

would be a great idea for mathematicians to build the "Standard Model" of AI/LLM from mathematics perspective. Currently its resembles alchemy and brute-force more than anything


100% what a great reply!

It's highly probable that each AI proof will basically produce its own abstractions, since it doesn't need to establish a "research program" that should be "understood" by the "community".

Maybe we'll use AI to read through all these AI proofs and write them in a common language as "lectures" for the 3-4 people who will still be "interested" in learning new "math" (I count myself among them).

(Sorry for the abundance of scare quotes.)


> It's highly probable that each AI proof will basically produce its own abstractions

I am not looking forward to the further silo-nation of math, where even the abstractions live only in one small area.

I think your later comment clarifies that you see this as a negative.

We build abstractions to transport understanding across fields. It is plausible to me that in some huge proofs (like FLT), there could be a role for proof-specific abstractions. In general, though, a lack of abstractions that cross boundaries makes me think the abstractions are not sufficiently powerful/useful.


What? Virtually everyone who writes a math paper "produces abstractions" in the course of doing so. It's a more routine and mundane activity than perhaps you are imagining.

The scare quotes are indeed doing a lot of work here -- for instance, I read the ones wrapping "community" as implying there won't be a mathematics community that will need to understand new research, which is a fairly bold claim deserving some justification.

You should try being a bit more explicit in general. Particularly so, given that this is a domain and community that you don't seem especially familiar with. By that I mean "be explicit about what you mean, clarify your analogies -- better yet, avoid them until you understand the topic deeply -- and define important terms"


> It's a more routine and mundane activity than perhaps you are imagining.

I know. I am a mathematician. I mean that the abstractions will only be done in this mundane way. (I produced some myself.)


Exactly like in commercial work - dark forest, where swarms of hungry AI equipped copycats are swarming to steal and exploit any original idea anyone has, thus demotivating anyone from coming with ideas.

Nobody knows the answer to that question, we are discovering in real-time what frontier mathematicians and frontier AI are able to do to solve previously unsolved problems.

That being said, its probably going something akin to how experimental and theoretical physics have an interplay with each other.

Can we develop more and more all encompassing abstractions (theories) that explain and verify experimental results; translates to, can we come up with more and more theorems/abstractions that explain formal proofs (and perhaps provide explanations for other known proofs)? Probably yes, and likely will also be accelerated by AI.

How this fits into the discipline of mathematics is something we are currently finding out.


It reminds me of the old movie "Pi" (a good movie). A crazy mathematician finds a new transcendental number and people are chasing him to try to get it because it's so powerful.

He tells them "I could give it you, but you wouldn't know what to do with it."

A new proof doesn't really do anything on its own if we don't know how to make use of it. It's like handing a set of Vernier Calipers to a Chimpanzee. The growth and innovation from developing the proof is where the real value is.


Langlands program[0] was proposed by the Canadian mathematician Robert Langlands (1967, 1970) seeking to relate the structure of Galois groups in algebraic number theory to automorphic forms and, more generally...

[0] https://en.wikipedia.org/wiki/Langlands_program


cutting that quote immediately before "representation theory" is very funny in a matmul-heavy zeitgeist; representation theory is representing algebraic structures as matrices where you can bring more familiar tools to bear.

> What I wonder: how do we create new abstractions and concepts?

By keeping mathematicians on the payroll and letting them do their jobs. It's always been their primary objective and it will continue to be.

> Are we assuming AI will do this?

It can't because it is fundamentally about asking questions about what we as humans find interesting.The machine can't know that. It could suggest questions and it can help answering them. But what is "interesting" is fundamentally subjective.


> What I wonder: how do we create new abstractions and concepts?

Mathematicians are already afraind they might not undestand all of the applications of existing concepts. When completely novel things without names start poping up they will be completely freaked out, like go players over the move 37.


My take on this is that new abstractions and concepts are sort of efficiencies you may use to tackle a problem. The more abstractions you have, the less energy you need to come up with a proof. So its like a spectrum.

One of the things I wanted to understand or simulate was, can we remove some abstractions and concepts and come up with a proof of some conjecture like Fermat's Last Theorem? I assume this proof involved a lot of modern abstractions. But if we can ask Claude to use only rudimentary concepts that a high school kid knows, could we come up with a proof that, while possibly huge, is correct?


LLMs are great at ‘seems’. the big question is, is it real? I’m automatically suspicious when the proof is 250k pages long.

Put these new proofs that are being made into the dataset for training the next model. Just use the bitter lesson and increase the amount of proof data.

Of Outsourcing the brute forcing the parts one struggles at leaves more space for exploring what’s left.

> Of Outsourcing the brute forcing the parts one struggles at leaves more space for exploring what’s left.

Except the tasks are probably not so neatly divided.

* Exploration and insight might be a side effect of a human doing "brute forcing."

* AI may vacuum up all the intermediate successes and quick wins that keep people motivated, making it practically impossible for a person to "create new abstractions and concepts" at the frontier.

* etc.


>Are we assuming AI will do this?

Yes.


So what is a greater accomplishment, formulating the hypothesis, or coming up with the proof?

I think the proof.

Although I assume in the future AI will be doing all of it, and humans are going to be less and less capable of comprehending the outcome.


My experience on teams is there are the high agency AI adopters. Management likes this group. Then there’s what management would characterize as those not adapting (or adapting enough). This is the group getting laid off / not advancing in the environment. And this group is particularly bitter about AI disrupting an otherwise stable career, with wishful thinking they’ll be able to go back.

I've seen the first group vibe up garbage and dump maintenance on the second group. It may be unfair to blame AI for that when management is ultimately responsible, but that's what a lot of people are doing. In their defense, management is being mislead by AI hype men who are saying AI is so intelligent it shouldn't be making dumb mistakes.

Agreed, its more complicated than GP describes. There is clearly lots of value, but there are also lots of poorly led organizations creating a mountain of problems for themselves. We are all still figuring out what AI really means just for software development, which is the most useful economic activity it can do by an enormous margin.

Human minds are still needed to understand many things related to developing correct software, especially in business domains. There is still no AI running a profitable business, because that is not a verifiable task that can be scaled in RL like math or coding. So extrapolating from a math proof that no one can understand to a economically useful agents taking over most jobs is quite suspicious.

What is changing now is RSI truly beginning to happen. Much of ML is a verifiable task that can be scaled, and a lot transfers from general coding as well. Leading researchers are predicting their own obsolescence as early as next year, and that is something I take seriously, though I cannot judge the likelihood myself.

As improvement accelerates, we may see generalization happen at a scale that means we don't need RLVR for every profession. We've been surprised by it before.


> because that is not a verifiable task that can be scaled in RL like math or coding

Is it not, or are the feedback loops just longer? I could imagine being able to build a pretty good simulator of business operation for various kinds of businesses. (Some decent ones already exist as games.) From there, if you can build a good enough simulation, I can't see why you couldn't RL it.

Of course it's easier said than done to generalize it to the real world, where the possibility space is large. But I don't necessarily see an impediment to getting there eventually.


The feedback loops are quarters and years, so no you cannot scale it like you can scale math or coding. This is the reason we are seeing the most progress in those domains. Even ML is harder to scale because test runs are expensive and consume significant calendar time, so research "taste" is still important.

A full simulation would be, but I suspect even a partial simulation of evolving a software project though multiple "checkpoints" over a quarter would improve model performance

"Running a business" includes lots of real-world interactions that would be hard to model. An AI trained on an accurate simulation would probably get torn to shreds by scammers in the first six months. The models that exist today have enormous blind spots.

Honestly, the second group kind of brought this upon themselves. You cannot survive in tech without having an open mind and moving fast.

Are we losing valuable traffic? The person who wants the quick answer seems like a less valuable customer than the one that knows they want to read a full article.

In my area, also a college town, most neighborhood cafes I go to still spark conversation and act as a place where people can meet. Yes some people work there. Or read books.

But I actually think cafe culture is pretty healthy in walkable places.

One problem is zoning has made it so “neighborhood cafe” is hard to build in the US. Commercial in residential isn’t often allowed.


Plenty of cafés around Paris, which is very walkable. What I see, even outside usual tourist hotspots (I live in a regular residential area with no tourist attraction around), is mostly groups of friends or people working that don't mingle.

The zoning situation in the US keeps coming back when subjects like these are discussed on hn. Now, I've never lived there, so I can't talk from experience. But even though I can see why that situation would definitely not help, I think that solving that is far from sufficient. Something else seems to be happening in Western urban societies.


I met a lot of people in my local cafe when I used to study. A few things I think that works:

Be there all the time. Cafe is like your second living room. Eventually you will notice the other regulars, and they will notice you. It also means that the coffee and food need to be reasonably priced.

Cafe cannot be too busy. It needs to be a chill atmosphere. Chat with hosts first. This often impossible if the cafe is too busy.


I think capitalism pushes a cafe to stay as busy as possible, charge as much as possible, and cycle guests as quickly as possible. In my experience "good" cafes are destined to go out of business.

In the West rents are just too high, so there is only room for the most savvy business people. When I was in Vietnam many people ran cafes from their own home, so they only needed very few customers to survive. Some even had a cafe just for fun. This also ties back to the zoning laws a few comments up.

Also that its intertwined with asking the corrupt administration to relax export regulations

Or do the build out to get to economies of scale with it faster.

And that helps us how...?

Scaling out causes us to climb an efficiency learning curve faster than bespoke, one-off weird data centers that require a jet engine to run

Again, how does that help us in regards to GP's point? Surely you understand that efficiency does nothing for us if we emit even more on the way there?

It increases the share of electricity compared to other types of energy. That makes it more worth it to build new production capacity.

This approach feels like we're driving toward a cliff, and rather than slamming on the brakes, we'd be choosing to hit the gas. Because if we just go fast enough, we might fly!

Even if everyone globally went carbon-neutral today, we'd already be pretty darn fucked, because is literally no buffer left! Temperatures would still increase for a few years, before settling at levels that will already cause catastrophic famines and escalating wars. And now we must consider that just looking at the raw numbers, we've never been further away from being carbon-neutral (plus/minus 1-2 years).

(let's please skip any discussions who should receive how much blame, since that's not relevant to this larger point)

Every bit of additional emissions today will kill more people and animals in the coming decades. And since Jevons paradox means that scaling will not cause the removal of our existing carbon infrastructure until it becomes fully cost-prohibitive, which is not purely a question of market forces, since politicians are already artificially keeping fossil power alive. They can just keep subsidizing the fossil industry!

So you're proposing that we put all our cards on the outcome that has a tiny chance of working out. Meanwhile, the current AI buildout is already strengthening the fossil industry, since suddenly there's huge demand for gas power!


Jevons Paradox

I have a Stackoverflow mode for my agents (essentially a tutor) and sometimes I spend time hand coding but heavily guided by an agent. It helps A LOT to make careful tradeoff when succinct and line by line care matters.

The experience still matters.

I have had experiences out of my technical depth where I’ve instructed an agent to do something technically nonsensical. The agent tried to do it, creating a mess. Nothing works, I get frustrated, only to realize I have a fundamental misunderstanding of something technical.

A similar thing happens when you ask for something, but the complexity for the perfect solution is 10 x the complexity of a simpler, 90% solution. And sometimes you only realize that looking at code (or feeling lots of brittleness)


Fwiw, that's not what the parent was talking about.

You're trying to middle road it. He was talking about actually treating it as a black box, so just give the model your expectations and goals it needs to meet, not tell it a rough solution for it to apply in detail.

The expectation/goal approach in a ai edited project works better then the middle road since opus5.5 ime. Depending on your project structure, it may not be viable however without large refactorings


For me it's really hard to find a balance with LLMs. Even trying one to add .NET IR codegen to my compiler that already had a working C++, JS, and C# backend for it to use as a reference. It just kept doing really odd stuff and when we got to generics it had done such a bad job with the prior work it struggled to collect the information it needed from .NET to do generic calls correctly, and mid session decided to just make everything reflection based with no notice to me. It had a doc of the AST with notes on what the nodes mean and how different constructs are represented, it's a pretty typical ML family language. Plus a working C# backend to reference.

Other times it will one shot things near perfectly that I was expecting to just use as a reference. Demo pages for work I've done its so fast and gets good enough right out the door

Every time I've tried to use it to write something small in my compiler? :"( pain. It's the perfect intersection of looks good and covers 90% of cases fine but explodes far away because some subtle invariant is wrong. I can't always think of the invariants up front without writing the code myself


When did you try that? The rate of improvement of models has been hard to keep up with.

I recently asked Gemini 4 Argon a question about a gcloud api, and to answer the question it decompiled the Google Java client library (without me asking) and figured out how it worked.

I do wonder if models struggle with .NET IR because there may be less info about it in its training data. When you’re working with open source systems it has much more to go on.

I’ve used models to generate ONNX IR (executable format for deep learning models), and that worked well. But the ONNX IR is much simpler.


This was about 2 months ago using 5.6 Sol and 6.0 Astra (I used Astra for the initial steps and setup and discussion and used separate prompts with Sol in the Codex desktop app afterwards).

The issue that it ran into is that it needed to keep track of .NET assembly and type information to do do generic calls (you need a reference to the actual type/method information, which you would need to collect and keep for lookup when compiling your AST into .NET). I didn't ask it to do generics until a handful of prompts in, and the codebase was not terrible but sketchy enough that once the LLM realized it needed type information it hadn't been tracking, it tried, flopped a couple times, and reached for reflection to avoid the problem

This is the kind of thing I might have considered to flag to the LLM ahead of time if I had written a .NET IR generator before, but I'm just a long time user/typically focus on native code or transpilation. If I was writing it myself I would have noticed early on that generics should be filed under "weird shit to verify" given it requires me to track info that is not in my AST, but when I'm prompting an LLM my mind is not on that, I'm focused on the higher level constraints + semantics (ie how suspend/resume works or rules around product types etc), and not details like that

I think it could been affected by a lack of training data, I don't know how many feature complete compilers for typed languages exist out there for .NET. C# and F# of course, and there are compilers for SML, Koka, and maybe some others I cant think of, but most were for dynamic languages written many years ago.

.NET IR was standardized before .NET core even IIRC and everything I use in the .NET stack has been MIT for a long time now


I wanted to reply to this point separately instead of one massive reply

> I recently asked Gemini 4 Argon a question about a gcloud api, and to answer the question it decompiled the Google Java client library (without me asking) and figured out how it worked.

I've noticed the same sorts of things happen and its kind of funny, because at times it is the right move, and others I have to interrupt the model and remind it that there is a uh, documentation, it can just skim real quick

It is super handy but also a little funny how quick they are to "let me just reverse engineer this real quick" before considering a short search


Expectations and goals can be technically nonsensical. Agents have gotten really good, but they can’t magically solve intractable problems. If you’re out of your technical depth it can be easy to set unrealistic expectations without realizing it

Then you're still middle-roading it from my perspective. Treating it as a black box means you're only specifying what comes out, essentially.

But to preempt a "I can't do that in my brownfield distributed monolith I'm maintaining at $dayjob": yes, that's why I said it may not be possible without refactoring.


I think we are talking past each other a bit? I’m curious about GP’s perspective here

My point is that you can ask for unrealistic outputs even if you’re treating the system as a black box. This is the classic case of higher level managers asking for infeasible results from their engineering team because they haven’t considered they’re trying to solve the wrong problem. Depends what you’re doing but I don’t think agents tend to push back on badly specified goals unless you’re proactive about it


Yeah I think this is the main issue at play. Bad engineers are now 10x bad engineers with the ability to hide behind the mask cause they can output something that looks right, but they don't understand it. You can "make" an agent do ANYTHING. There is no rigor, there is no taste, there is no risk. Anecdotally I have seen this play out IRL. I worked with a 3-month boot-camper that somehow stumbled into a software developer role at a small company. The amount of slop that he was creating was unprecedented. No formal training, no idea of abstraction, letting/making the agent reinvent the wheel for everything. The LLM was doing the whole "you're absolutely right" so he was also under the impression that he was a real engineer doing the "right thing". Concretely he convinced our manager that we need to re-write our web-app in Go because it was so much faster(0 users btw) and safer than python. He had an agent build out this test suite and there was a metric like Go vs Python DB operations, Go being 107x better. Seemed sus, looked into it and he(his agent) removed connection pooling for the python side but for the Go side, 100 concurrent connections, 10 idle and a connection lifespan of 1 hour. His reason for that was that the python code kept crashing every time his agent adding pooling so he had the agent removed it. This sheet got printed as is and distributed to our C-suite to much praise. I know this is a very specific example but I fear this is happening at some level everywhere. Where in the past a bad engineer to this degree would stick out like a sore thumb if they were surrounded by competent people that have rigorous standards. An LLM/AI doesn't have to worry about what happens if things don't work and when it's confirming that YOU are a genius every step of the way. It's going to be a disaster when the reckoning comes. I am not a AI luddite but software engineering is just as much as a culture as it is art and science. Without that backstop I just think that we will see a huge degradation in quality because there is no incentive for LLMs/AI to write "good software".

Sounds like that colleague is on track for a Distinguished AI Builder title and promotion

Hopefully this is sarcasm cause if so, haha. Otherwise seek help :P

Sure theoretically. practically this isn't a real problem for me or anyone else i know

In the black-box approach, how do you know what is a reasonable task to give it? Just ask the agent about that also?

You don't give it tasks. You give it what you want. It decides the tasks, and then executes. Then you give it feedback. "black-box approach" here is just a fancy word for vibe coding.

I want to be rich, healthy, happy, young, smart, beautiful. Same for all my friends and family please. How many tokens?

PS: make no mistakes and tell me in advance if I forgot anything that I could regret later.

Or on a more technical side: I would like all problems that any app/company/product could solve to be solved in the best possible way, so that nobody has to even think about what kind of problems even could be solved. Go iterate on that but do not make any mistake, do not ask any questions, be done by tomorrow, do not interpret what I said in any bad way.


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

Search: