torginus

The only argument they have here is that they use GB300 GPU's

I don't follow events closely, but the US has constantly flipflopped on what sort of GPUs the Chinese are allowed to have, not in small part because much of the AI boom's valuation is based on demand for US-made hardware, for which the Chinese have inexhaustible and well-financed demand.

So even this feels a bit hypocritical to me, but my understanding is that Chinese native AI hardware is getting good enough that labs dont feel a huge disadvantage by being forced to buy at home, even if they'd have preferred to buy US chips.

Which is a situation that was manufactured by the constant thread of having their access to advanced GPUs revoked.

But our VMs start fast, so fast that even if we created the DNS entries before creating the VM, our users still had to sit around waiting for DNS to propagate, which occasionally took minutes, not seconds.

To my (limited) understanding, this is not a good idea, and is an unfixable problem from the server side. Companies, VPNs or ISPs or routers, often use their own DNS servers, and those can have caching logic, which means it doesn't matter how fast your own DNS implementation is, as the users lookup request wont hit your DNS server, it'll hit an intermediate cache.

I've always found it weird how ludicrously poweful personal hardware has gotten. The fact that you can't buy a midrange CPU with less than 16 cores, or that a high-end gaming GPU is 100+ TFLOPS just blows my mind. The fastest supercomputer in 2004 was 70 TFLOPS. Absolute crazypants level of power, and companies were fall over each other to get people to buy it.

Yes, but how much would be a static Sonnet 3.5 be worth today? Its just about 2 years old. I'm not even going to ask about 3yo models like GPT4.

Upper bound of AI progress - recursive self improvement. In this case AI will be responsible for building better models, making people who own datacenters the winners. Anthropic/OAI is cooked.

Lower bound of AI progress - plateau. Progess is slowing, focus is on serving a meaningful peak capability at the lowest possible price. There's been news today that Google is building a Gemini chip with weights baked into silicon. Considering a chip's lifetime of 2-3 years at minimum, and that a 2-3 year model today would be useless today, they're expecting they wont make a similar amount of progress in the next 3. Game is about selling at the lowest margin. Anthropic/OAI is cooked.

So their survival rests on the presumption that AI progress will fall between these two extremes.

Eh, I take what I can get. A lawyer doesn't argue that his client is innocent (though that might be tha case), because of some innate sense of justice, but because he gets to paid so.

Just because a guy wears a fitted suit doesn't mean he'll be good at handling your money. Likewise if you buy tens of thousands of euros worth of incredibly overbuilt teutonic industrial equipment in matching colors and snap-in cabinets doesn't mean the guy who builds the PLC program in ladder logic will be any good at it - or that the next guy will have an easier time troubleshooting it than if the whole thing ran on an Arduino scavenged out of an intro to robotics kit.

Experts know there is stuff worth spending on, and there's stuff like the ESP32 where no amount of money will buy you something better, and no amount of money will save you, if the guy building the whole thing is an idiot.

The problem with super smart people is that they often succeed without having to or bothering to understand even the most basic common sense techinques.

I had a friend who told me a story, that when he was a kid, he found a copy of the Borland Pascal DOS compiler on his dad's machine. It came with a demo demonstrating the graphics library, but that was the extent of his introduction to Pascal, most of the syntax he figured out by trial and error.

He went ahead and build a quite elaborate tank game, with things like enemy AI, multiple weapons and destructible terrain.

The code was a fever dream, but the game was pretty impressive, and enjoyable.

One of the things he didn't intuit, was that you could not only use built in functions, but define your own. So when he needed a code block a second time, he just copied it, so you can imagine how pleasant that code looked.

I cannot make sense of a single word of this essay. What sort of insight am I missing? There's not a peep here about write barriers, futexes, OS schedulers, critical paths and/or how they relate to Zen or whatever. This article can be easily recycled as the Zen of compiler design or Zen of audio engineering or whatever, without substantially having to chance much of the words, thats how generic it is.

I'm genuinely considering that this is just a foreword and I missed the clickthrough link to the meat of the article.

Yeah, it sounds more like they found 19 regressions after the migration so far.

4. One probable outcome of an open-weight-model-dominant world is full AI communism, which is precisely what China proposes: rather than a market product,

https://www.theregister.com/software/2000/07/31/ms-ballmer-l...

lol

I think it's safe to assume the Chinese models will try to steal your corporate IP and business know-how.

It's also safe to assume US models will do that too.

I've been trying Chinese models, like GLM 5.2, as substitutes for Claude or GPT5.X, and my experience is that they underperformed them in real world metrics like prompt adherence, hallucination or code quality, even though they benchmark better.

Afaik, for a MoE model, total size doesn't matter that much for how heavy it is to run, the size and number of active experts at any given time does. Of course, you still have to store the whole model in fast memory, so there's that, but the reason these things are getting this large is because it doesn't really affect runtime that much.

I mean, lots of things have been developed in European countries, and said countries either never commercialized the technology well enough, or lost leadership. I'm sure that prior to WW2, Brits would've treated the idea that they could be out-engineered by the Germans with ridicule.

From what I can tell, China has a capitalism problem, not an industrial wealth problem - they could house every one of their citizens in a decent apartment and give them gadgets and a car to drive, yet half the country is third-world poor still.

But that's because in a market economy, you can't just give away things, and those poor people don't produce much of value by capitalist standards, so the can't pay for those things.

I'm sure the CCP is working on the problem of how to give these people those things without crashing the economy.

Sorry, but I have a thought that's off topic: If Kimi is good enough to improve itself, what's preventing someone who owns a big datacenter and nothing else, to just run Kimi to do AI research, thereby rendering the frontier labs more or less unnecessary?

Opus/Sol/Fable are valuable because of their reasoning and coding ability, not their bedside manners.

While still not okay, I suspect the latter is what gets stolen by Chinese distillation (and some evidence suggest this happens the other way round with US models talking in Chinese)

This is good and interesting research, and I think it proves Kimi was trained on Claude's outputs, but only for English conversation and essays. Their style is similar, which is tbh somewhat expected for a Chinese company - I don't think most American companies (especially small labs making LLMs) could figure out a good way of talking to Chinese customers, without copying some of the existing Chinese models.

This doesn't imply that Kimi's reasoning capabilities or coding ability has anything to do with Anthropic (which is the most valuable part), or the model's strength comes from distilling Anthropic.

Considering by this benchmark, Opus sounds extremely similar to Fable, this isn't really evidence of a ton of Fable output being trained on (and even then, most likely for superficial stuff, like style)

I also have a suspicion that in original Chinese, these models don't sound anything like Anthropic's ones, though I have no proof of that.

This. A decade ago, during my summer internship in the first year of college, I got put on a hardware project, basically designing a remote monitoring device for a telco.

I wrote the software, and my friend designed the PCB for the digital side. Basically I had to be a decent C programmer, and read the datasheet for the MCU we were using (which was way more work than using something like an ESP, but still). My friend had to lay stuff out on the PCB and place secondary components according to various application notes.

Overall not simple, but not rocket science.

On the other hand, the guys doing the analog daughterboard were industry veterans with decades of experience. They had a million concerns, from signal integrity, to voltage protection, back-EMF, etc. I was not even near the ballpark of that skill level.

Later on, I worked on more complex HW projects, stuff with FPGAs, DRAM integration and 8 layer PCBs, where you had to run full electromagnetic simulations because when you run at 1+ GHz, suddenly signal propagation is not instant, and traces behave like antennas.

I also did software there (a smaller slice of it this time), but there was an army of engineers working on the whole project which took a year, and not a second of it was spent in vain.

There has recently been a post about the Cyberpunk genre here on HN. I find the parallels between that and the AGI hype noteworthy. Authors in the 80s imagined a dark and dystopian future, driven mainly by the growth of computer technology.

But if we compare what what was described in Neuromancer, vs how personal computing and the internet affected the world, we see basically very little overlap.

AGI Doomerism likewise has no basis in reality considering what LLMs have been shown to be capable of. Not saying they wont be transformative in some way or isnt already, but essentially the way these AGI prophets predicted things will go down will be completely inaccurate save for the vaguest tems (something bad will happen at some point, and AI will be involved)

'Open weight' is roughly analogous to 'freeware' - software that's free to use, but closed source. We haven't seen a third category - 'paid weight' - software where you pay for the weights, but you keep them and have to run them on your own machines - analogously, updates could so if you don't want your LLMs cutoff to stay at a point forever, you have to keep buying new weights. These are standard practices in software pricing.

Can't you do infinity read replicas? I don't think the write load on lobster.rs is severe - maybe 1 comment/sec, with an acceptable replication delay of 60 sec?

If you can do that, you can have infinity instances.

A lot of the ways in which the zig compiler works doesn't use pointers, it uses indices. This stuff is easier to write as safe code, not less easy.

I respectfully disagree. This is only true if you view malloc as qualitatively different from a piece of code that gives you an index for a free object in an object pool.

Provided you don't ask/give back memory from/to the OS, what malloc is doing is giving you an index (pointer) into a pool of bytes, while manipulating an internal bookkeeping structure.

Use after free is just you using an index after said bookeeping structure has marked that piece of memory as available for something else (and perhaps claimed already).

If you have an array of Node structs to represent a graph (like the AST in Zig), and use indices to represent references, you have essentially zero protection in Rust that helps you with finding Node-s that have been used for something else etc.

The 'asking memory from the OS' aspect for malloc doesn't really change the safety of your language compared to this where it matters - if you do use-after-free on a page claimed by the OS, you get a segfault, which immediately tells you there's a problem, which is much better than silent corruption.

At least with malloc, you get debug allocators, or other features that can help you in this case. If you are careless with indices in an object pool, and overwrite stuff, essentially, it's up to you to figure out what went wrong and you have no tools to help you.

First, this is legal:

       if err := do_something() ;err != nil {
         return nil, err
       }

Like, if you go by function: ReadData(has err)-> ParseData(has err) -> ProcessData(no err)

Instead of having functions that can fail dynamically all over your codebase.

Considering the Chinese are one of the major contributors to AI, I would think this was a solved problem by now, at least in some other CLI based coding agent.

If you look at the list of contributors, you see Claude and Copilot, but on the other hand, You can see a huge, functional codebase going back to as far as 2024, so its safe to say its not vibecoded.

https://github.com/DavidObando/gsharp/tree/9579dd6626801f089...

I don't find Go especially verbose - in fact, due to the economy of syntax, Go code without the usual syntax sugar features ends up similar in length to other curly brace languages. I tested this with C# and Typescript, where I migrated projects from these languages to Go - the overall volume of code stayed roughly the same.

I think people who want to project a 'cracked' (god I hate that word) agentic engineer vibe. But my experience with basically everyone in my immediate vicinity, is that people have no respect or awe for the 'tell the robot to do the thing' workflow.

They could've at least made something custom, and claim it was designed with help from GPT 5.6

The price for that HW basically implies it either has sizeable margins or is made with artisan methods.