Hacker Newsnew | past | comments | ask | show | jobs | submit | program_whiz's commentslogin

I know I gave a long rambling response, but I think dwelling on how the successful were already winning can be depressing. It can be helpful to become more stoic about it by realizing its kind of just the way things are (resignation and acceptance). Do I like realizing that there is probably nothing I can do to create the next FAANG? No. But its also freeing / helpful to realize that there is a range of what one can do in life based on the system, and to thrive / succeed with the range one is given, taking what opportunities are available.

People are sad the RH was solved because it was a machine / privileged person. This is how humanity works, countless human lives filter through a giant genetic optimization algorithm, and a select few randomly hit on meaningful progress for the species. So hopefully we can just accept the blessing that progress / technology is being built by someone, and everyone (mostly) gets to have some of the fruit falling from the canopy.


This is true, and unfair, and can make one angry, but its also kind of inescapable. Imagine a system where anyone could easily do such things, starting a large business or scientific lab, or proving the RH, for example. In such a system, if any person can do those things, they have little value / are not noteworthy. If someone has access to more resources, training, etc, then they will also be able to do those things, but bring more resources to bear (people, money, time), and therefore their creation will have more "value". As a result, whatever accrues to these endeavors as a result of "value" will go more towards the person with those extra resources, and cause a rich-get-richer spiral.

Its kind of an inescapable feature of free markets and trade, regardless of how high the floor gets, differentiated inputs create differentiated outputs. Even if the input is "one person has a far superior mind", it might create a difference, even if everyone is free to do something.

The HN crowd seems to be biased towards thinking that someone should receive outsized rewards due to their inherent intelligence (a matter of luck and genetics), but not other lucky factors such as looks, connections, family name, inheritance, timing, etc.

Is it lucky that these people had access to unlimited compute and cutting edge models no one else did, and time/resources to work on Riemann? Yes. Does that mean probably tens of thousands / millions of people, if given the same chance, could do it? Yes. Is that fair? No. Is there a way around that? No (for the reasons I gave above).

The counter-argument would be "everyone should have access to models with all these powers", in which case "everyone can solve the RH", in which case "solving the RH isn't impressive / hard", and only solving something most people can't solve is noteworthy, and hence anyone with even more compute would be seen as "winning unfairly".

To take a more neutral example: music. Most anyone can make amazing music today with VSTs and DAWs. So baseline music has little value (whereas hundreds of years ago just hearing/making quality instruments was a rarity). But some people bring more musical skill due to training, background, genetics, exposure, mentorship, practice, etc. All of those things require resources, luck, and circumstances. They also need time and resources to work on their compositions. Some, like Hans Zimmer, were lucky enough to land on some big movies. Are there thousands of musicians equal to or more talented than Zimmer? Yes. But like I stated above, its an escapable feature of a system where something like "can bring more value because got lucky movie deal" is a thing, if "can make better music because talent" is a thing.


It is probably sufficient. If we had an infinite army of willing robot slaves smarter than most people who would work for free, we could probably live in utter paradise. Instead we will fight over who controls them and use for warfare, but the bar to achieve a kind of utopia with resource abundance is probably only a few years away.

We're already living in resource abundance! Yet don't feel we're in paradise. Famine is hell, but living in an all-you-can-eat buffet is no utopia. It's an abomination of sorts.

To be fair, most people do not have time abundance and safety abundance. The typical person in the United States has a level of access to material goods that would make him the envy of almost anyone from 200 years ago, but 1) he has to work and so has limited free time, and 2) he cannot be sure that the material abundance won't be taken away from him either by poverty or by ill health. He can go to the all-you-can-eat buffet and sit there, tired from work, worrying about tomorrow.

If people had time abundance and safety abundance, they might be much happier. There would still be potential problems about loss of meaning, or about feeling a loss of status because AIs are smarter than humans or whatever, so I'm not sure that humans would actually feel like they are living in paradise. Probably not, but at least some of the problems that the typical person has wouldn't be present.


I don't know with how many rich people you've interacted with, but my personal experience is that they're full of themselves and think money = iq and they aren't gonna like your average joe experiencing life like they do. Guess who owns these machines?

They’re not going to be employed either lol and will be caught up in some AI honeypot or whatever too.

I’ve met a lot of rich people. They don’t think about you.


They don't need employement, they just need ownership.

They will when they're replaced by the board with a robot lol

Or maybe the bitter lesson is a flawed over simplification. Computer vision is only possible because we used intuition about significant features in vision to make it tractable for GPUs with custom designed kernels. Tossing a generic model with more compute is pointless, it's the invariants and assumptions that make the models effective.

Same for sequence modeling with attention, highly generic RNNs which can in theory model any sequence are just incapable, regardless of compute thrown at it. And a ton of engineered features like tokens are required to make LLMs work (just using the pure data is not scalable).

Bitter lesson was only true back when the engineered features were hand made vs statistically derived, and then it stuck around making people think a bigger model is all we need to get to AGI, when the reality is the exact opposite.

Technically an N-layer nnet can model "any function" but tossing a 10T plain nnet will not even approach a 10M param model with proper features, structure, and design. I call this the savory lesson (intuition and design matter)


Hillariously everyone in here is imagining the choice is between the overworked and grumpy human workers and a benevolent, gentle, dedicated robot (who looks like a model).

If robots take over, the same forces that have pushed the human workers to the brink will happen there too. Shuffled into your cell where the robots tend to you as quickly as possible, then move along. You think the nursing home is paying robots to sit around and chat? These things are depreciating assets!

Hah, all these posters think they are going to be tended by a gentle supermodel who comforts them and listens to all their problems. The reality is probably: you get to use TV / internet for your "social", a machine that can automatically empty your bedpan / waste, and a gantry system that can move stuff around, rotate you, and dispenses your medication so a human never has to see you.


What you describe sounds great. If we get automated way to bath, go to toilet and eat. Then it solves like most of the issues with taking care.

In addition if technology would somehow allow to type and may be control computers using some neural interfaces, then it would be really great for all people who are paralyzed.


Yeah, and there are a bunch of these sort of technologies already in use.

Eg. many hospitals already use something like the gantry system you mentioned, and also have zoned air mattresses that can sufficiently vary pressure to avoid the need to turn bedridden patients.

So I think you're right - once we have enough of this basic assistive tech available cheaply enough, it'll become universal in care homes and staffing levels will be slashed. And, at that point, where does the cost pressure come from to provide an incentive to build these putative humanoid do-it-all robots?

More likely we'll find ourselves on a plateau with chatbots and single-purpose bedpan robots for a very long time. Maybe supermodel carebots will eventually become available for billionaires, but they'll be the equivalent of travelling by private jet while the rest of us make do with bargain basement Ryanair flights.


Follow the chain back, pretty soon the prompt just becomes "make a good product", or "fulfill this market's needs" but why prompt that? The agent can prompt itself with such high-level phrases. If your only job is to babysit such a process, there can be a running loop that just says something like "scan all products in market, find a niche, research customer fit, propose products, brainstorm and critique, then implement and sell."

If humans need to oversee those even higher level agents, then the prompts become abstract like "make money" or "pursue human flourishing". That leads to the alignment problem (AI controls the paperclip factory). But ultimately, if aligned, agents can prompt themselves with such basics as well.

The final prompt will eventually be "run everything", and it will ship with the agent which runs on a machine in the cloud. Ultimately there is no need for human prompting at all. We can finally let the machine run itself completely and just do something else.

Everyone saying they are becoming "project managers" or "CEOs" are just at t+1 on the curve of elimination. The timeline of cope:

Agents are just auto-complete, I still manage the code.

Ok fine, you don't read the code, but you guide the design and marketing.

Alright, the design and coding are solved, but you still need to choose market and customers.

Ok sure the agent can do better market research/fit, but you at least still need to give basic strategy and concepts. You still need a CEO!

Ok the agent can figure out basic strategy, and run a company more efficiently, but still, you need to guide it fit human interests more broadly.

Ok it turns out that was just adding "don't be evil" to the system prompt, and its running autonomously 24/7. You still need humans to consume to have a viable economy.

Ok so the oligarchs didn't care, the economy was redefined to just be exchange of work between autonomous machines through financial derivatives traded by black boxes to make numbers move.

Now the earth can finally rest in the matrix, all that remains is a black box moving information around to create economic value. No one will be around to observe, but AI will create massive returns for virtual shareholders, and the GDP will eventually reach 1 quintillion dollars per second, a number humans never could have achieved. All nations (now just virtual analogs without citizens) will move to a global currency which maintains perfect stability through automated monetary policy. The output of all processes is efficient and maximized, the final perfection is achieved -- the mathematically optimal conversion of resources and energy into large numbers which increase the reward function.


It would be funny if AI ended up achieving communism. Maybe this was the missing piece of technology for Stafford Beer's Cybersyn project.

It certainly is a good possibility.

A clarifying way to think about this for sudoku:

For every clue, mark the probability of sampling its number in the row/col to zero. Also mark the probability of selecting a "clue cell" for modification to 0.

Pick a random cell for modification, based on its "energy" (higher energy get modified first):

    1. Every conflicting cell gets energy +1.0 (try to resolve conflicts before new empty cells).  E.g. two 7s on the same row both get +1 energy.

    2. Empty cells have small energy 0.1

    3. Filled cells with no conflicts have energy 0
For the selected cell, choose a digit. Look at numbers in its row/column, every occurrence of a number in the row/col drop prob of sampling by 90%. Example:

    | No. of 7s | chance of 7 |
    | --- | --- |
    | 0 | 10% (equal to other digits) |
    | 1 | 1% |
    | 2 | 0.1% |
Obviously this would get more complex with more numbers on the line (this is just checking 7 on otherwise empty line).

This is essentially what these models are doing. It provides a kind of "wave function collapse" search with random exploration based on "energy", but energy is just defined as conflict / unsolved. For something like a maze, energy is disconnected path segments or walls, etc.


See my comment below, I explain what it means. Energy is just a number that comes out of matrix which learns the relationship between different states of the system being searched. For example, suppose we're talking about sudoku. You would have a matrix that has W_ij (weights between spaces). For each of those, there are 9x9 numbers (energy between two of the digits). A combination like W[0, 1](4, 4) is high because you would have a 4 in the first space, and a 4 in the 2nd space (which breaks the rules). Most of the values are low energy (e.g. 4 and 2 is energy 0, no conflict). However, as different numbers are placed, they are checking for the lowest energy number across the whole board, and the "clues" are locked in.

So if you had a "clue" like 2 in a space, then the energy for 2 on that row / column would be high, but other numbers low. You just sum this up and pick the lowest energy number. As you add numbers, you can hit conflicts (e.g. two 7s in the row), changing one of them will be the next lowest energy move. The probability of changing a number is inversely proportional to its energy. So changing the conflicting numbers is the usual most likely move.

So in the best case it works like this: For each empty cell, pick a number at random, but its weighted by the other numbers that affect it (e.g. if 7 is in the row, the chance of 7 is 0.1%). Continue doing this on empty cells, unless there is a conflict in solved cells (higher prob of mutation), in which case change that number first.

So its a slightly fuzzified version of a very fast concrete algo. "Energy" is just a number from training, its defining by the task at hand (in this case, "energy" is how much conflict numbers have on the board, 0 meaning no conflict and solved if every cell is filled in).


Nice explanation. Very intuitive use of the word “energy” by the researchers, imo, because of its relationship to entropy and “low energy states” (good / desirable / stable) and “high energy states” (bad / undesirable / unstable).

Big caveat if anyone thinks this could replace or strongly augment "reasoning" models (LLMs, RNNs) -- its more like an energy-based search for a discrete problem (graph coloring, solving mazes, sudoku).

It works by training a weight matrix (N-to-N) on different solutions. For example, in sudoku you're training a weight matrix for every space to every other space (81 c 2), with 9x9 numbers (~240K weights). So then the "energy" for every number in a space is W_ij(a, b) where ij is a pair of spaces, and (a, b) is a pair numbers you want to swap (e.g. 4->2 in the space). Pick the lowest energy move each time (argmin of this whole space), and iterate.

This has a few limitations. First, its discrete and has a limited space definition, getting this to work for arbitrary reasoning seems hard (what size/shape graph would you use for arbitrary reasoning). Second, the ruleset is defined apriori (e.g. you need to know you have a 9x9 grid and 9 numbers, the rules of the game are implicit in the checker). Third, the actual search algorithm is somewhat bespoke to task (less of a factor, maybe use generic graph coloring as the basis of any problem), but for something like agentic reasoning its not clear how this would be adapted. For example with mazes, we need to know what constitutes a valid path and a solved maze, but for language/code rules like that are hard to formulate.

Fourth, this doesn't admit of an easy way to use "fuzzy reasoning" that would be required for integration with any kind of human language or image-based system, where continuous vectors are useful (the graph edges and colors are on/off or discrete).

There are ways to adapt, but this approach fits pretty solidly in the "deterministic, fixed problem size" realm.

In addition, for training you need correct solutions and counter-examples (strongly labeled). For something like "fuzzy logic" training or "natural language" coming up with clear examples and counter-examples would be challenging to fit into this model. LLMs are only possible by leveraging vast quantities of unlabeled data. Not saying impossible, but challenging. For example, if training on human code, what constitutes a counter-example? You basically already have to have a method of getting correct and sufficiently incorrect code just to train this type of system, and its unclear how to represent those changes.

For all of these problems, given the constraints and amount of bespoke logic required, solving with dynamic programming is far faster (including arbitrary graph coloring problems). I only say this to mean that if we could solve all the problems I mentioned above and get arbitrary reasoning into the form of graph coloring, it would still probably be more efficient just to solve that with good old dynamic programming / recursion than with a model like this that needs O(N^2) params versus the graph and needs millions of training examples for different domains. They brag about solving sudoku, but a CPU-based approach to solving it the good old fashioned way on is probably 10-100x faster. If we were using an H200 like they did, we could easily solve the whole sudoku dataset in less than a second.

But it _is_ a very interesting approach, partly because the energy strategy encodes good / likely guesses, so that is kind of interesting vs dynamic programming which is kind of a brute-force approach, so maybe for very large spaces it would be more efficient (but then you have the issue of the parameter space blowing up).


I'd love to wire it into a model and do some training. Can't see any way to obtain it though.

Yeah any method I'm thinking is better solved another way. If you need continuous graphs, then this devolves into GNN or transformers. If you truly have discrete graphs, just use standard programming? I dunno, maybe this energy approach is better on some problems (more efficient), but getting training examples where you automatically extract the graph representation, and the energy rules seems like a big leap. If you already have the energy rules in code, then the "just program it" approach is stronger. I left a comment above with an example how to sample this just using ordinary code that would be a lot more efficient imo (but keeps the random weighted search), and is probably better than brute force.

Reading the friendly article:

It is worth noting that of the 290 OpenAI theorems which come with Lean code, around 38 of them have the property that the Lean code does not even state, let alone prove, the main theorem of the paper.

So even if LEAN is 100% bug-free, its still on the programmer to correctly express the logic and proofs, or they could be proving completely unrelated facts.


38 of them is a bit egregious wtf

It also speaks to why I remain skeptical of relying too much on coding agents for code review. They are still pretty iffy on implementing the right thing, and even worse at detecting when a PR didn’t implement the right thing. Ignoring that reality is how you do like Anthropic did and increase your rate of merging changes from 2500/day to 30,000/day without a corresponding increase in the rate at which you deliver useful work product.

OpenAI did a data dump of work in progress, leaving it to the mathematics community to make sense of it. I don’t think they claimed that there were Lean proofs for everything? Did they warn that some Lean proofs were incomplete?

Why did they stop in the middle anyway? Was there a justification given for that?

The cynical answer is that their primary goal was never to meaningfully contribute to mathematics. It was to pick some low-hanging fruit for the marketing department. So they did the 10% of the work that generates 90% of the buzz.

It’s more likely that their goal was internal research, to test a new LLM before releasing it. But, then they had something that mathematicians would be interested in, so they threw a data dump over the wall. It’s a rather rude way to do it, but this isn’t a company that particularly cares about the things that mathematicians care about.

OpenAI no longer needs to do anything to get publicity. There are some organizations that are so famous that everything they do is newsworthy and anything they do will be covered whether they like it or not. Remaining low-profile is no longer an option.

The people claiming that everything is marketing haven’t come to grips with the inequality of fame.


This would be plausible except for the fact that there was a press release.

There was a blog post. It's the bare minimum when announcing something to the public, not a sign that they're going to use it to promote themselves in a significant way.

There's no "low profile" for a company like OpenAI. Either you let the public (including the press) know or you don't, and keeping the results secret would be bad.

If they just told a few people, it would leak eventually and then you'd have articles based on rumors, along with accusations of a coverup. For an example of the difficulty, look at what happens when people try to do coordinated disclosure of a security bug.


The more likely answer is because that is what the agent decided to do given inconsistent instructions. No one read what it said about what it was doing before releasing the presser.

They could be proving completely unrelated stuff. Or even entirely trivial things like 1 + 1 = 2 over multiple millions of lines of lean code.

Look, don't want to derail the conversation but proving that 1 + 1 = 2 is not trivial. Try it out. Why does 1 + 1 = 2?

succ(0) + succ(0) ?= succ(succ(0))

succ(0) + succ(0) [by `+` def]=> succ(succ(0) + 0) [by `+` def] => succ(succ(0)) QED


I'm sorry, what is that and what does it have to do with 1 + 1 = 2? I don't understand, can you explain?

If it is a Socratic question, I don't understand what hidden complexity you are pointing to.

Lean axiomatization of natural numbers goes roughly like: 0 is a natural number, for each natural number a, succ(a) is a natural number (plus some additional properties of succ that don't matter for this proof: succ(a) != a and so on). succ corresponds to the intuitive notion of a next counting number (which number of things you get if you add another thing to the collection). Mathematical object designated as 0 corresponds the intuitive notion of nothing. Therefore succ(0) is the number of things that you get if you add a thing to nothing, that is 1. succ(succ(0)) is what you get after 1, that is 2.

Lean axiomatization of operation `+` is, basically, a set of rewriting rules that I apply in the proof. And it's constructed in such a way as to correspond to the intuitive notion of counting two sets of things together.


>> If it is a Socratic question, I don't understand what hidden complexity you are pointing to.

All of it! What is a "natural number"? Why is 0 a natural number? What does it mean for a natural number to be the successor of another? What are those additional properties of succ that don't matter for this proof? Incidentally, what is a proof? What is a "Lean axiomatization (OMG)? What is an axiomatization (OMGOMG)?

And so on. Notoriously, the proof of 1 + 1 = 2 is on page 362 of Volume 1 of Principia Mathematica. That's because if you really, really want to axiomatize arithmetic you have to do a hell of a lot of legwork beforehand and it's not "trivial" at all to do so in a self-contained and consistent way.

For example, consider trying to get a five-year old to understand your proof. Would that make any sense to them? They know that 1 + 1 = 2 because they learned that at school. Can they understand the proof you're trying to give here?

And btw, you didn't really give a proof, it's only a sketch proof and you didn't even use the words "by induction". I'm pretty sure you have to, at some point, say "by induction" and oh boy, try to explain that to a five-year old.

Edit: btw, lest this is thought of as pedantry, the point of automated theorem proves is that they work in this extremely anal way, i.e.t hey have to take all those "trivial" steps which human mathematicians normally don't bother with. And that's where they draw their power from and why we can rely on their proofs and trust them in a different way than we trust human proofs.


The actual proof is

theorem one_add_one : 1 + 1 = 2 := by

  rfl
> That's because if you really, really want to axiomatize arithmetic you have to do a hell of a lot of legwork beforehand and it's not "trivial" at all to do so in a self-contained and consistent way.

That's because when you develop natural numbers from logic you need to jump thru hoops. Rewriting rules a-la Peano aren't less rigorous.


What is rf1

A proof strategy that, basically, computes both sides of equality and succeeds if they are the same expression. We have no free variables here, so it's a straightforward application of the definition of addition.

Correction: "We have no quantified variables here"...

A proof strategy in Lean? And I'm supposed to assume a correct Lean kernel that for what the OP calls a "trivial" statement? See what I mean? You and the OP just assume too much.

Isn't that pretty much by definition? 2 is defined as the number that is 1 more than 1.

I'm sorry. Where is that defined?

dictionary.com, among others.

two: a cardinal number, 1 plus 1.

(Yes, I am aware that dictionary.com is not a math textbook. Still, I think it is not mistaken. We define two to be one more than one.)


That would make it an axiom [0], which it often is. But you don't _prove_ axioms.

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


Look, if you don't know that 1 + 1 = 2, I don't know if AI proving theorems is going to be interesting.

Arithmetic is one of the oldest proofs we even have, and quite an underpinning to mathematics.

If AI can't construct either the inductive proof or the axioms required, then it can't do any of the rest.

Which is most of the point - it probably can. And there is a proof in Lean for it. But if you don't understand the difficulty involved, then you don't have the background to know when the proof generated and handed to you, has holes in it.


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

Search: