HN user
I don't understand this point of view. They buy the book, they use it for training. When you read a book, you use it to train your brain. Is that copyright infringement?
Downloading from Pirate Bay is indeed theft, someone put the scan of the book online, you download it without paying, the author and publisher of the book do not get any royalties. Very different situation.
There's nothing there that isn't immediately replaceable.
Indeed. But that was not my point. My point is that we still have this old impression that training cost is dominated by compute and it is hugely expensive, and the Chinese labs can short circuit that by distilling the American frontier models. I don't think the training compute cost is a big factor anymore for the American frontier models, because of the reasons I gave. If the Chinese models can get the training compute cost down by a factor of 100, that's not going to make them 100 times cheaper, and not even cheaper by a factor of 2. Maybe 10% cheaper or so.
People who claim that the Chinese open weight models have some type of manifest advantage don't realize that the close weight models have a huge advantage as well: the researchers from OpenAI, Anthropic, Google, xAI, Meta are not dumb, they can read the white papers written by DeepSeek, Moonshot, etc, and they can inspect all those architectures and they can pick and choose the best tricks there are out there, and of course, they have access to their own in-house secret sauces.
Sure, any model that is not at the frontier can use the frontier model to generate synthetic high quality training data, so this can reduce significantly the training costs.
But at the scale of OpenAI, Anthropic and Google, it is quite likely that the (raw) training cost is very high anymore. Here's a few heuristics:
1. All the hyperscalers see a huge demand for inference. They can't deploy datacenters quickly enough to satiate all the demand they see. But, it's is impossible for the inference demand to be constant throughout a day or a week. If you use the times when the demand is lower than the peak demand (which is almost all the time) to dedicate the spare compute capacity to training, then your the cost of training compute is zero.
2. It is likely that increasingly a higher cost of the "training" is actually setting the guardrails, which is essentially post-training. As we've seen, without proper guardrails, the US Government won't allow you to serve inference. Anthropic was hit directly, but OpenAI delayed their 5.6 release as well to make sure the US Government is ok. This part of the training cost can't be reduced easily by using synthetic data generated by other models.
3. The frontier labs are also investing more and more in building an ecosystem around their models.
I am not a frontier lab insider, but take a look at the jobs posted on the Anthropic career page [1]. There are 74 jobs in "AI Research and Engineering" and by my count at most 15-20 are related to pure model training (of pre-training or RL type), and the rest are post-training, safety and security, alignment, interpretability, productivity and lots and lots of other things.
I think it's the opposite.
Kimi K3 has 2.8 trillion parameters. We don't know the number of parameters of ChatGPT 5.6 or Opus 4.8, but it's probably in the same region. Fable/Mythos are rumored to be around 10 trillion.
So, K3 is directly comparable with ChatGPT 5.6 and Opus 4.8, and the price is not so much lower:
K3: $3/$15 per 1 Mtok input/output ChatGPT 5.6 Sol: $5/$30 Opus 4.8: $5/$25
This is not a watershed moment. It's a competitor converging to the same capability and trying to undercut your prices, but not by a lot.
As for the open weights? For now, Kimi K3's weights are closed, and I don't expect the situation would change.
I didn’t assume that. The author makes the opposite assumption: that because nobody has ever proposed a path to technological feasibility, that thing will forever be technologically unfeasible. I simply find this argument to be very weak.
That is true. What should we conclude from this statement?
Shrinking that into a 500 kg seed — or even Freitas’ original 100-ton seed — is not an engineering detail. It may be the entire problem.
How many AI tells can you count there?
But honestly (see what I did there?) the AI slop is reasonably cleaned up in this piece.
However, the essence of the argument has two deep flaws. One is that the time to complete an interstellar voyage is extremely long and you need some exergy, yada, yada, yada. We could start with sending self-replicating probes to the asteroid belt. There is zero chance that we'll attempt to send self-replicating probes to a different star system before we send them inside our own solar system. And the second error is this:
Bootstrapping this loop [...] is a chicken-and-egg problem that no study I am aware of has worked through at the level of actual process flowsheets.
The fact that the current technology is not adequate, and nobody even attempted to solve such a problem is a weak argument. Three hundred years ago nobody had "worked through the process flowsheets" of making an injection molding machine, or a 3D printer, or a power drill, yet they are all available now.
That's an interesting idea, but I don't think you can use a mass driver to shoot something from an higher orbit to a lower orbit.
Spirits are not capital intensive. Building a rocket is very capital intensive. You'd like to be able to reuse it. But if one single mission takes 30 years, then you can reuse this rocket once, at most twice. Let's say you reuse it twice, you amortize the capital cost over a period of 90 years. Now, let's say someone builds the exact same rockets, but they do missions between the asteroid belt and Mars. Each trip takes about 2 years. Everything you can source on Titan, you can find an asteroid to source it from. By the time you reuse a Titan-bound rocket once, you reuse an asteroid-bound rocket many times.
The Hohmann transfer time between Saturn and Mars/Earth is around 16 years. So, all the ships used for such a supply mission would reach their destination at least 32 years after they’ve been built, assuming we build them on Earth.
A given model has a shelf-life, which these days is measured in months, not years.
Not all new models are trained from scratch. ChatGPT 5.3 to 5.4 (and likely 5.5) was basically the same model, but probably trained a bit more, not a new model from scratch.
The "someday" when frontier model providers can enjoy their current high inference margins without the burden of significant training costs is never going to arrive.
That is debatable. I believe the moat for the frontier model providers is the compute. At the level of 10 trillion parameters (that Fable/Mythos are rumored to have), you need serious compute to serve inference, and you also need serious compute to train. Will DeepSeek, Qwen, Kimi, GLM come up with a 10T new model anytime soon? I doubt that. People keep saying that the Chinese labs are catching up to the US big 3, and measured in months the gap is now only 4-6 months. I doubt a Chinese version of Fable/Mythos will be released in the next 12 months.
People think mathematics is about proving theorems.
I think that's just an accident of history.
When we write software, we very seldom write proofs that our algorithms are correct. We just write tests, and we also run the algorithms and when they fail we know we have a bug and then we proceed to debug, fix, and add new tests (if we are disciplined, but most of us are). In time, by usage and testing, we gain confidence that our battle tested software is correct, mostly. And we tell people that we will never be 100% confident that any software is bug free. But that's a slight lie: if we wanted such confidence, we would start using provers, and create bug-free software. That possibility exists, but it's just extraordinarily expensive.
Well, in math that's the only possibility, and we use it. And it is indeed extraordinarily expensive, but it's also the cheapest among the alternatives. The alternatives are 2: be rigorous and do these proofs, or be sloppy and allow bugs to creep in, and allow an entire school of math to collapse like the Italian school of algebraic geometry [1].
There is one more alternative. If a particular math theorem has some applicability, then you write a program and use it in real life. In time you eliminate the bugs as much as you can, and you get to the steady state of "virtually bug free". At that point you don't have a solid proof that the theorem is correct, but in general you don't really care. Because you feel that a formal proof is just a thing one would pursue for getting academic satisfaction only.
[1] https://en.wikipedia.org/wiki/Italian_school_of_algebraic_ge...
The performance, if we trust the benchmarks, put it at Sonnet 4.6.
I don't trust these benchmarks. I used a number of times Kimi K2.7 and I was disappointed. It would run in circles for things that Claude would do in one shot. However, my usage was via Ollama cloud, and I have no idea if they serve the actual model or a quantized version, and it was the quantization that degraded the performance.
The great news, in my opinion, is the precedent. If Microsoft is now serving Kimi K2.7, then very soon they might start serving GLM 5.2, and that is indeed a very competitive model.
What I meant is: for a human who has the context, like someone who works in the field, or took some courses, or read some materials about this. A human who just has a vague idea about fissile isotopes and needs to google the properties of U-235 or Pu-241 would do no better; they would probably do much worse.
I just tested it with a slightly tricky question
> If you could run a nuclear reactor with U-235 as fuel or Pu-241 (both mixed with 95% U-238), which one would you choose and why?
In any case, LongCat-2.0. gave a very well reason but incorrect answer that Pu-241 is preferable.
I then tested on Qwen 3.7 Plus, and it correctly answered that U-235 is preferable because of its much higher delayed neutron fraction. I then went to Gemini Flash, which answered the same, with much more confidence, and with much stronger arguments, and the speed of the answer was much higher.
Overall I rate Gemini Flash the best, Qwen 3.7 Plus an acceptable second, and LongCat-2.0 an ok'ish third, if you have nothing better.
Because it's pretty likely some of these have negative long term effects. And some of those long-term effects are probably pretty serious.
Many people taking these kind of drugs are gambling with their well being. No doubt that in many cases that gamble is worth it.
You use the word "gamble" and it is correct, but the connotation is a negative one.
We take a risk in everything we do. We seldom know the long term effects of anything. Right now I drink tea, and I drink tea quite often. Is this healthy? Does it have long term negative effects?
With GLP-1 medications, deciding to not take one, if you are overweight, is also a gamble. A pretty huge one. Obesity very often leads to type-2 diabetes, and the first step is prediabetes. Should someone with high BMI and high A1C avoid GLP-1 because we don't yet know the long term effects? We do know the long term effects of being overweight.
I don't think open weight model will get as good as Mythos/Fable or GPT 5.6. Not in the next few years, and perhaps never.
The Chinese Communist Party is not any happier than the US Government to see millions/billions of people being able to use incredibly dangerous models.
Here's an unpopular opinion: this might be the only way to deploy advanced models. A lot of people compare advanced AI with nuclear weapons. Creating white lists of users that are allowed to use advanced AI feels wrong. It feels against everything that the Constitution stands for. That men are created equal and they are free to pursue their happiness. Now they are free to pursue that happiness only if the US Government signs off on that. It hurts to only think of that. But I'm afraid there's no other way. These models, in the wrong hands, can result in unfathomable devastation.
How do I know? Dario Amodei said that when he explained why Anthropic has to limit the US Government's usage of its models [1]:
> Some uses are also simply outside the bounds of what today’s technology can safely and reliably do. Two such use cases [...]: Mass domestic surveillance. Fully autonomous weapons.
[1] https://www.anthropic.com/news/statement-department-of-war
Because in reality, in most cases it works. I worked in many places that had large offshore teams that I worked with closely (India mostly, but also Hungary, Poland, Argentina, Morocco), and people were mostly happy with the arrangement.
There are some cases where the outcome is bad (like the case at Ford now), and lots of people point out to that and say "I told you so". But those are the exceptions, not the rule.
Proliferation is irrelevant
I would not be so cavalier about that.
It was not a Candu reactor what India have used
It was a heavy water reactor, called CIRUS [1], based on the NRX reactor that later evolved into the CANDU reactor [2].
But there are easier and cheaper pathways - uranium enrichment distributed to be stealthy
Iran tried that. It is not easy to keep centrifuges stealthy, you need tens of thousands of them, and they need to be organized in cascades, you can't have one here, one there. You need a few hundred here, a few hundred there, and you need conversion and deconversion facilities (of uranium oxide into/from uranium hexafluride). It is virtually impossible to keep such a program secret, the more distributed it is, the more people need to be involved, how do you keep a program secret if there are thousands of people working in those numerous facilities?
[1] https://en.wikipedia.org/wiki/CIRUS_reactor
[2] https://en.wikipedia.org/wiki/CANDU_reactor#Early_efforts
Global reserves are around 6 million tons, with estimates putting potential reserves around double that
In the mining industry reserves are a technical term. They can be proven, probable, likely, etc. qualifying a deposit as a reserve of a certain grade costs money. Reserves are used as colateral for secured financing, so in some cases the cost is justified. But if the sum of reserves is about 100 years of current consumption (our case here), mining companies will not spend one dollar more to certify new reserves.
For all practical purposes, uranium is an inexhaustible fuel, even if we never develop fast reactors.
I love nuclear reactors and CANDU are quite cool. But I don't think that today we have any reasons to build CANDU reactors, except possibly that Canada can demonstrate they can build them for cheaper than others can build light water reactors. The ability to build is something that has little to do with the technical merits of a nuclear reactor design. But all things being equal, a PWR or a BWR should cost less per GW than a CANDU reactor and have other advantages too:
- main problem with CANDU: proliferation. India was able to build nuclear weapons after using a Canadian built heavy water reactor (basically a CANDU reactor) [1]. There is no guarantee that another country will not try something similar in the future, the design has no built in proliferation resistance. An operator can remove irradiated fuel at any time, and if the IAEA discovers they engage in plutonium manufacturing and they get on a black list, they can manufacture their own fuel quite easily, because CANDU uses non-enriched uranium. With light water reactors, you need enriched fuel, so if you are flagged as a proliferator no fuel manufacturer will be allowed to sell you fuel, and it's going to be much harder for you to manufacture your own fuel, since you can't enrich. If you can enrich uranium, you might as well try to build a uranium bomb (like Iran is trying to do). Also, with light water reactors, you refuel only at discrete times, generally about 18 months apart, so it is much more difficult to extract lightly irradiated fuel without being caught by the IAEA.
Now some less important problems:
- because CANDU uses non-enriched uranium, it produces much more nuclear waste per GWh compared to light water reactors. Nuclear waste is not the boogeyman nuclear anti-advocates make it to be, but still, if you generate 5-10 times more nuclear waste than the mainstream alternatives, it is less than ideal.
- there is one positive reactivity feedback loop in a CANDU design. Because of that CANDU designs are not licenseable in the US. The Canadian nuclear regulator is comfortable that the design is stable [2], but if you can choose between a design with one positive feedback loop and one without any positive feedback loop, why would you choose the first?
- heavy water is a worse moderator than light water (by a large factor). It [1] https://en.wikipedia.org/wiki/India_and_weapons_of_mass_dest...
[2] https://www.cnsc-ccsn.gc.ca/eng/resources/research/safety-an...
Unfortunately its just a small boiling water reactor.
It is not just a small boiling water reactor. It is a 300 MW-electric boiling water reactor, and if successful, it will be followed by 3 more of the same type for a total of 1.2 GW-electric. That is more than an AP-1000 reactor, and much less risky.
I dispute that. Enshitification happens when the user does not pay for the product. The provider of the service has to find alternative ways of monetization. As a user, this feels shitty, but you get what you pay for. With AI, if you pay for usage, you can demand that the service provided not be shitty. You can even litigate, but the most important leverage is the threat to take your business elsewhere. Enough people do that and the provided service stops being shitty in a hurry.
But they are using wind:
A majority of the generation will come from large GE Vernova (NYSE: GEV) turbines.
It sounds like a huge amount of money and a clear sign of defeat. The US (or maybe Europe, who knows?) is paying Iran some form of tribute.
But, there's another way to see this: if you pay someone money, you end up with a lot of say in what they can or can't do. That money does not come with any guarantee; the party offering the money will keep the money flowing based on their own perception of good behavior of the party receiving the money. The post-Trump administrations will find themselves in the possession of a tremendous amount of leverage with Iran.
The US could make it clear that if earth moving equipment is spotted within half a mile of the Isfahan, Fordow or Natanz previously bombed facilities, they will be bombed again [1].
[1] https://en.wikipedia.org/wiki/Nuclear_program_of_Iran#Main_f...
There might be. But, from the US point of view, the tolls are not a strategic problem. A conspiracy theorist might even argue that tolls in the Strait of Hormuz make the US oil exports more competitive, and therefore more profitable. The strategic problem for the US (and the rest of the world, Israel obviously included) is for Iran to acquire nuclear weapons. One might say that the Memorandum of Understanding is quite vague about that. But the natural conclusion from last Summer and from the last few months is that the US will not hesitate to bomb Iran if it tries to resume its pursuit of the bomb. This does not need to be codified in any agreement, the threat is there implicitly. The MoU is more focused on the carrots than the stick, but the stick is there, in the background, and Iran knows it.
Maybe, but in the end if the cost of launches is less than 10% the cost of the GPUs, then why is the project non-sense?
Yes, I am for real. The Colossus datacenter that xAI built cost $7 billion.
[1] https://www.globaldatacenterhub.com/p/can-elon-musks-7b-colo...