> Last I checked, the tractor and plow are doing a lot more work than 3 farmers, yet we've got more jobs and grow more food.
Not sure when you checked.
In the US more food is grown for sure. For example just since 2007 it has grown from $342B to $417B, adjusted for inflation[1].
But employment has shrunk massively, from 14M in 1910 to around 3M now[2] - and 1910 was well after the introduction of tractors (plows not so much... they have been around since antiquity - are mentioned extensively in the old testament Bible for example).
That's his point. Drastically reducing agricultural employment didn't keep us from getting fed (and led to a significantly richer population overall -- there's a reason people left the villages for the industrial cities)
I'm not sure that's what they meant. Read like this:
> the tractor and plow are doing a lot more work than 3 farmers, yet we've got more jobs and grow more food.
it sounds to me like they mean "more job and grow more food" in the same context as "the tractor and plow [that] are doing a lot more work than 3 farmers"
But you could be right in which case I agree with them.
But where will office workers displaced by AI leave? Industrialization brought demand for factory work (and later grew service sector), but I can't see what new opportunities AI is creating. There are only so many service people AI billionaires need to employ.
No, you cannot ignore every argument by claiming someone else made it before. Make an actual response.
What new opportunities does the LLM create for the workers it may displace? What new opportunities did neural machine translation create for the workers it displaced?
In what way is a text-generation machine that dominates all computer use alike with the steam engine?
The steam engine powered new factories workers could slave away in, demanded coal that created mining towns. The LLM gives you a data centre. How many people does a data centre employ?
"the post-training performance gains in Qwen3.5 primarily stem from our extensive scaling of virtually all RL tasks and environments we could conceive."
I don't think anyone is surprised by this, but I think it's interesting that you still see people who claim the training objective of LLMs is next token prediction.
The "Average Ranking vs Environment Scaling" graph below that is pretty confusing though! Took me a while to realize the Qwen points near the Y-axis were for Qwen 3, not Qwen 3.5.
In my short testing on a different MoE model, it does not perform well. I tried running Kimi-K2-Thinking-GGUF with the smallest unsloth quantization (UD-TQ1_0, 247 GB), and it ran at 0.1 tps. According to its guide, you should expect 5 tps if the whole model can fit into RAM+VRAM, but if mmap has to be used, then expect less than 1 tps which matches my test. This was on a Ryzen AI Max+ 395 using ~100 GB VRAM.
Running a 247GB model reliably on 100GB VRAM total is a very impressive outcome no matter what the performance. That size of model is one where sensible people will recommend at least 4x the VRAM amount compared to what you were testing with - at that point, the total bandwidth to your storage becomes the bottleneck. Try running models that are just slightly bigger than the amount of VRAM you're using and these tricks become quite essential, for a significantly more manageable hit on performance.
No real fixed benchmarks AIUI since performance will then depend on how much extra RAM you have (which in turn depends on what queries you're making, how much context you're using etc.) and how high-performance your storage is. Given enough RAM, you aren't really losing any performance because the OS is caching everything for you.
(But then even placing inactive experts in system RAM is controversial: you're leaving perf on the table compared to having them all in VRAM!)
I'd like a reference for it being rug pulling. What happened with OpenCode certainly wasn't rug pulling, unless Anthropic asked them to support using a Claude subscription with it.
My comment reports only facts and a few of my personal opinions on professional conduct in journalism.
I think you and I have a fundamental divergence on the definition of the term “hit comment”. Mine does not remotely qualify.
Telling the truth about someone isn’t a “hit” unless you are intentionally misrepresenting the state of affairs. I’m simply reposting accurate and direct information that is already public and already highlighted by TFA.
Ars obviously agrees with this assessment to some degree, as they didn’t issue a correction or retraction but completely deleted the original article - it now 404s. This, to me, is an implicit acknowledgment of the fact that someone fucked up bigtime.
A journalist getting fired because they didn’t do the basic thing that journalists are supposed to do each and every time they publish isn’t that big of a consequence. This wasn’t a casual “oopsie”, this was a basic dereliction of their core job function.
> I’m simply reposting accurate and direct information that is already public and already highlighted by TFA.
No you aren't. To quote:
> There ought to be social consequences for using machines to mindlessly and recklessly libel people.
Ars didn't libel anyone. They misquoted with manufactured quotes, but the quotes weren't libelous in anyway because they weren't harmful to his reputation.
Indeed, you are closer to libel than they are.
For example, if these quotes were added during some automated editing processes by Ars rather than the authors themselves then your statement is both harmful to their reputation and false.
> These people should never publish for a professional outlet like Ars ever again. Publishing entirely hallucinated quotes without fact checking is a fireable offense in my book.
That's going perilously close to calling for them to be sacked over something which I think everyone would acknowledge is a mistake.
One could argue that failing to catch errors in AI generated code is a basic dereliction of an engineer's core job function. I would argue this. That is to say, I agree with you, they used AI as a crutch and they should be held accountable for failing to critically evaluate its output. I would also say that precisely nobody is scrutinizing engineers who use AI equally irresponsibly. That's a shame.
> Given that LLMs are just lossless compression machines, I do sometimes wonder how much better they are at compressing plain text compared to zstd or similar. Should be easy to calculate...
> The right script, with the right prompts can be tailored to create a loop, allowing the premium model to continually be invoked unlimited times for no additional cost beyond that of the initial message.
(For those unaware, AWS doesn't have a VM monopoly, and the market dynamics seem similar)
reply