But everyone learns by example! How is this any different from a person just reading the outputs of Fable, learning, then producing output. Surely reading outputs, gaining knowledge, then producing work isn't illegal, or all art/writing would be illegal.
Funny how that argument seems so vacuous in this situation, yet others find it compelling when justifying the mass theft of art and writing for model creation. In this case the model is "just learning priors" before it "creates its output which is novel", nothing problematic.
Does this expend less carbon than you captured? Skimming, gathering, transport, drying, compression, burying. I assume to reach "gigatons captured per year" this would require heavy machinery and industrial buildout. Not that its impossible, I'm just not sure its a net-negative in CO2.
This article shows that we can reliably cause phytoplankton blooms, but doesn't link it with actual "carbon removal". I suppose CO2 is temporarily removed in the form of the bloom, but I believe historically these blooms have a low efficiency for "permanent capture" because the plankton decay or are consumed. The only way to effectively capture would be to somehow create the bloom, then sink the phytoplankton en masse deep within the ocean where the would be neither eaten or decayed. Maybe possible with the right setup, but hard to scale.
yep, all those coders paying $200-$500 per day to use claude once subsidy ends will be seriously rethinking how much they really want to vibe-code "rewriting X in rust". Helping people write word docs, recipes, and emails isn't going to justify $15K per month subscriptions either.
This is a sight-reading test. If a musician practices a piece for thousands of hours, it would no longer be an effective sight reading / creativity test. The purpose of the test was to see how models would compose something novel requiring the ability to compose orthogonal, normally unrelated, components into a coherent image.
the nature of the test was to see if the models can effectively compose an image of a novel concept outside the training set. If they are trained on it, it ceases to be an interesting test to some extent.
I think your response might be sarcasm, but actually this very situation demonstrates the truth of the adage.
In this case, competition and information-sharing is driving intelligence to become a commodity, with ever shrinking margins above compute+hardware. If this is the case, the incumbents can't recoup the many billions they have borrowed.
Cornering a market makes a winner, a winner who can charge large margins on a product you don't have an alternative to.
However, the point you might be thinking of is "competition is good for consumers" which is true. Thiel's sentiment was "competition makes companies into losers", as they become low-margin commodity factories, which is also true.
Yep, time to pay back those campaign contributions in the form of "chinese models embargo". USG has already shown more than willing to dabble in picking what AI companies can do what.
Actually its interesting, I wonder if the recent freezes on Fable/Mythos and GPT 5.6 where actually prepping so that a "chinese model not allowed" play would be more pallatable / excusable. But then again, that's attributing 4d chess to an admin that has been making rookie mistakes.
Not sure you're wanting an explanation, but it comes down more to equivalent algorithms than rigid categories. For example there is a P algo to sort a list of numbers, but not to solve a sudoku (NP). However there is a polynomial algo to check sudoku (spaces ^ 2 if you check every space against every other space for rule violation).
However, the reason all NP algos are part of the same category is because you can solve any problem in NP by switching the problem into another problem in the same category and solving that. For example, you can turn sudoku into a graph coloring problem, which is also NP. You can turn sorting (P) into something like balancing a tree, which is also P.
The major question is "is there any algorithm that would allow us to change some NP problems into P problems, solve it, then use it for the original problem". E.g. could we take graph coloring and turn it into sorting a list of numbers?
So basically, if there is any way to bridge the two, then it might mean every NP problem is actually solvable by a P algorithm, under some transformation. This would be immense because it would completely change the way we solve those algorithms and greatly reduce compute costs.
While this seems far-fetched, realize that there are some problems that seem extremely expensive if done the naiive way, but are actually solvable in P. For example, you _could_ write an exponential sorting algo (try every element in every position), but clever people found a way to make it efficient (P). So its possible we just need the right algo to completely change the landscape of computing.
However, as you say, its almost self-evidently true that P != NP, but has never been proven so (to do so, we need to prove that no such algorithm can exist). But clearly, solving an exponentially complex problem using a O(log n) algo would be remarkable.
To take a concrete example, currently the best algos to exhaustively check a board game like chess or go are exponential (NP). Its easy to verify the winner, but its exponential to enumerate every possible move (e.g. 80^turns states). If we found a polynomial way to solve this (even by converting to something simplified), then it would mean we could exhaustively search chess polynomial to the number of moves (e.g. turns^100). This changes it from "cannot be done in the lifespan of universe" to "its possible with a powerful computer in measurable time". We already use heuristics and estimates to explore the exponential space in efficient time, so if we had a polynomial algo chess, markets, optimization, and other NP problems would be extremely efficient to solve.
Isn't the question at hand whether its bad for mathematics if "prove the circle has area 3" devolves into 500 lines of inscrutable lean code, vs just having `pi r^2 == 3`? Sure they both "proved" it true/false, but knowing an answer isn't as useful as knowing why its an answer. Knowing an answer does have some value, its just not as valuable. If I can't work it out myself, I just trust the oracle.
Now if you ask "does the area of unit circle equal 4?", I don't really know, but we can go back to the oracle and ask again (we haven't learned the general pattern).
Also, I'm not sure that assuming this 'area of circle' question was cutting edge math, that the oracle wouldn't say 'yes, to a certain level of tolerance'. Can't count how many times I've seen agent decide a test needs to be loosened or deleted because its an "edge case" or "blocking". If you don't understand the proof you might get back 'yes' for some versions of 'is the area of unit circle 3' (depending on complexity of that ask).
Its an inevitable state of any market system (especially if regulated). Eventually it will become more profitable to use the gains of capital to buy protection from regulators, instead of investing in ever smaller business improvements. Using money to defeat competition is ultimately inefficient and generates lower returns for investment. Unless regulators are totally insulated from the market, they will have a stake in selecting winners and will eventually construct moats that others can't cross, as that provides the highest returns to capital, and the greatest rewards to regulators.
Monopoly is the most market-efficient vehicle to deliver returns to capital, and the most natural state of the market; one player using advantages and gains eventually destroys all opponents. Smaller players can never gain a foothold due to the incumbents being so efficient and far ahead, and it makes more sense to merge with the front-runner, allies, or be destroyed (hence the competitor pool keeps shrinking).
These are features of the system that naturally emerge without counterveiling forces.
For example, AI companies will shortly find its cheaper to just get the government to constrain their competition. The alternative is many companies spending trillions to eek out profits, a poor state to be in. Regulators want money and power, so its in their interest to create this protected state, as the "free and open market" isn't buying elections or vacation homes. And of course, any unprotected competitors left behind will die, consolidate, or sell to the victors; so we will eventually have a "winner takes all" system where one or two big players dominate. Any startups will either be quickly destroyed as people ask "why use a worse product", or will sell to the monopoly when they realize they can't afford to spend $1T training models and building data centers, and complying with all the regulations.
My point is that protectionism (in any form) isn't something to bring down over time to encourage competition -- the system can't naturally function that way, as it would require each player to go against their own interests. Instead, protectionism is a natural ever-increasing good that will be cultivated for the controlling capital and regulators in the system. We only see the "free market" operations during a time before market / regulator capture, as that's the time when there aren't yet dominant players who can guarantee power and money to the regulators, and there isn't enough consolidation of capital to immediately destroy all competition, but its an unstable market state.
"approaching" is doing some work there. $30K today will get you 90-144GB usable VRAM with solid system RAM and disk and CPU. A single B200 chip at 180GB is $40K. Unfortunately that is nowhere close to being able to run a 750B param model. For something like that, we're getting closer to 1TB VRAM (8+ H200/B200), and then 1M context KV cache is many more GBs on top of that.
That's a $500K-$1M+ rig as of now. That's a lot of $200 subscriptions to break even, but reasonable if you are paying Anthropic $25/M tokens. Then of course there's the power, cooling, and maintenance to consider...
But yeah, I can see if the prices come down 10x in a few years, or crater after the bubble, $30-40k might get you a decent machine.
What's crazy is the prompt must be something like "pro-AI but still believable and measured", since its "fixed my iOS app albeit with back and forth". Interesting, they know the HN crowd for sure.
hey that's perfectly understandable, and yeah, definitely there is a group that wants you to "do the work for them". I was reacting more to a kind of question/chat shaming I've seen -- where engineers act like they're optimizing 99% of their precious time and mental energy and can't be bothered to make a humanized workplace experience for others.
honestly a sad state. Obviously there is a reasonable threshold, but trying not to speak to anyone until you've done a ton of work / research when they know the answer is just sad. Like what's wrong with asking a question? We've entered this anti-human hellscape where asking a question in slack (async) is somehow a crime, like posting an opinion without a double-blind study to back it up (burn him!).
And the same people who are complaining about time wasting of having to ask/answer a question from a coworker which might create a modicum of civility and connection in this bitter cruel world are the same shit-posting on social media and doom watching youtube all day. "My flow can't be interrupted, I need all my energy to refactor this column from VARCHAR to TEXT, and to update this button from onClick inline to using a named closure".
Please, the reality is that we sold human connection for an illusion of productivity and the bitter pill of isolation where we all now feel guilt and shame for wanting to talk with other people (through an albeit disconnected and disembodied asynchronous channel).
If anyone responds with "I don't have time to respond I'm so busy", please realize you are proving my point. You are literally doom scrolling YC for no reason and alienating / pushing away coworkers to argue with internet strangers, sad.
I think he meant "show me a true linked list / node graph in rust that isn't unsafe". The reason being its not possible using c-style pointer following (or without just putting everything auto-pointers). What you've shown is exactly the tradeoff they were referring to. In rust, the answer is: make sure lifetime of all memory is explicitly managed, then use integers for the 'links' between nodes.
His point was that for his programming, he wants to be able to make real pointers and real linked lists with memory unsafe, which Rust makes difficult or opaque. For example with linked list, you could simulate (to avoid unsafe), by either boxing everything (so all refs are actually smart pointers), or you can use a container with scoped memory lifetime, and have integers in an array that are the "next" pointer. In addition to extra complexity, the "integers as edges" doesn't actually solve the complexity, it just means you can't get a bad memory error (you can still have 'pointers' that point to the wrong index if you're rolling your own).
Same with your graph code. Using a COO representation for a graph does in theory make it "memory safe" (albeit more clumsy to use if you are doing pointer-following logic), and it also introduces other subtle bugs if your logic is wrong (e.g. you have edge 100 but actually those nodes were removed, so now you're pointing at the wrong node).
I think the point (which I agree with for things like linked list, graph, compiler) is that depending on your usecase, the "safety" guarantees of rust are just making it harder to write the simplest most understandable code. Now instead of: `Node* next` I have lifetimes, integer references, two collections (nodes and edges) to keep in sync, smart pointers, etc. Previously my complexity was to make sure `next != null`, now its a ton of boilerplate and abstractions, performance hits, or more subtle bugs (like 'next' indices getting out of sync with the array of 'nodes').
If there was a way to explicitly track the lifetime of an arbitrary graph/tree of pointers at compile time, we wouldn't need garbage collection -- its not solvable at compile time, and the complexity has to live somewhere.
Yeah the usual mott and bailey. Monday -- AI is taking over the world, tremble in fear! Tuesday -- sure it did a boneheaded thing, its just a tool, no better than an intern, actually its _your_ fault, all the data in the entire world isn't enough to train this system not to delete prod!
But everyone learns by example! How is this any different from a person just reading the outputs of Fable, learning, then producing output. Surely reading outputs, gaining knowledge, then producing work isn't illegal, or all art/writing would be illegal.
Funny how that argument seems so vacuous in this situation, yet others find it compelling when justifying the mass theft of art and writing for model creation. In this case the model is "just learning priors" before it "creates its output which is novel", nothing problematic.