What about the juice variable?
https://www.reddit.com/r/OpenAI/comments/1qv77lq/chatgpt_low...
3,005 karma · joined October 23, 2012
https://www.reddit.com/r/OpenAI/comments/1qv77lq/chatgpt_low...
At 20t/s over 1 month, that's... $19something running literally 24/7. In reality it'd be cheaper than that.
I bet you'd burn more than $20 in electricity with a beefy machine that can run Deepseek.
The economics of batch>1 inference does not go in favor of consumers.
You should try out K2.5 for your use case, it might actually succeed where previous generation open source models failed.
A secretary. The word you're looking for is "secretary". Having a secretary has always been the preferred way to handle these tasks for the wealthy and powerful. The president doesn't schedule his own meetings and manage his own Outlook calendar, a president/CEO/etc has better things to do.
People just created calendar/email/etc software (like Microsoft Outlook) to let us do it ourselves, because secretaries are $$$$. But let's be real, the ideal situation is having a perfect secretary to handle this crap. That's the point of using AI here: to have an AI secretary.
Managing your own calendar would become extremely 2010 coded, if AI secretaries become a thing. It'd be like how "rewinding your VCR tape" is 1990s coded.
To be fair, the jokes are kinda hilarious.
> "content": "Blessed are those who persist, for they shall inherit the context.
Training cost (FLOPs) = 6 * active params * total tokens. By keeping the MoE experts param count low, it reduces total training costs.
I don't think this was a good move. They should have just trained way past chinchilla like the other major labs, and keep sparsity above 2%. Even Kimi K2 is above 2%. GLM is at 5%, which makes it very expensive (and high performing) for its small size.
Arcee went the other way. They trained a massive 400b model (bigger than GLM-4.5/4.6/4.7, bigger than Qwen3 235b A23b), but only have 17b active params, which is smaller than Qwen and GLM. It's also only trained on 17T tokens, vs 20-30T+ tokens for the other models. It's just undertrained and undersized (in terms of active parameters), and they got much worse performance than those models:
https://45777467.fs1.hubspotusercontent-na1.net/hubfs/457774...
It's not a bad showing considering the limitations they were working with, but yeah they definitely need double the active experts (8 out of 256 instead of 4 out of 256) to be competitive. That would roughly double the compute cost for them, though.
Their market strategy right now is to have less active params so it's cheaper for inference, more total params so it's smarter for the amount of active params they have, but not too big to fit into a H200 cluster. I... guess this is a valid niche strategy? The target audience is basically "people who don't need all the intelligence of GLM/Qwen/Deepseek, but want to serve more customers on the H200 cluster they already have sitting around". It's a valid niche, but a pretty small one.
128GB vram gets you enough space for 256B sized models. But 400B is too big for the DGX Spark, unless you connect 2 of them together and use tensor parallel.
Those poles WERE NOT invented for strippers/pole dancers. Ditto for the hitachis. Even now, I'm pretty sure more firemen use the poles than strippers. But that doesn't stop the association from forming. That doesn't make me not feel a certain way if I see a stripper pole or a hitachi magic wand in your living room.
I don't think that was a good move. No other models do this.
https://news.ycombinator.com/item?id=46039486
This guess is from launch day, but over time has been shown to be roughly correct, and aligns with the performance of Opus 4.5 vs 4.1 and across providers.
Anyways, I actually think your statement is incoherent as stated, if we presume moral naturalism. There's clearly different levels set points for "you", so "sole reason" is actually neurologically inconsistent as a statement. It's impossible for "sole reason" to exist. This radically alters your framework for self, but eh it's not impossible to modernize these structural frameworks anyways. Steelmanning your argument: if you try to argue set point hierarchy, then we're back to the NBA player playing for a championship example. He's still playing even if he's not playing for fun. Similarly, hunting a baby for pleasure can still be hunting for a village, as The Most Dangerous Game shows.
More generally (and less shitposty), the refined principle is now quite narrow and unfalsifiable in practice, as a no true scotsman. How would you ever demonstrate someone's "sole or primary" reason? It's doing a lot of work to immunize the principle from counterexamples.
This actually devolves into human neuroscience, the more I think about it. "I want to throw a ball fast, because I want to win the baseball game". The predictive processing theory view on the statement says that the set point at the lower level (your arm) and the set point at the higher level (win the baseball game) are coherent, and desire at each level doesn't directly affect the other. Of course, you'd have to abandon a homunculus model of the mind and strongly reject Korsgaard, but that's on shaky ground scientifically anyways so this is a safe bet. You can just say that you are optimizing for your village as a higher level set point, but are hunting for game at a slightly lower level set point.
Note that sport is not a terminal desire, as well. Is a NBA player who plays for a trophy not playing a sport? Or a kid forced to play youth soccer? So you can't even just say "sport must be an end goal".
Yeah, exactly. I intentionally set that trap. You're actually arguing for my point. I've spent comments writing on the axioms of geometry, and you didn't think I was familiar with the axioms of ZFC? I was thinking of bringing up CH the entire time. The fact that you can have alternate axioms was my entire point all along. Most people are just way more familiar with the 5 laws of geometry than the 9 axioms of ZFC.
The fact that PM was an alternate set of axioms of mathematics, that eventually wilted when Godel and ZF came along, underscores my point that defining a set axioms is hard. And that there is no clear defined set of axioms for philosophy.
I don't have to accept your argument against objectivity in ethics, because I can still say that the system IS objective- it just depends on what axioms you pick! ZF has different proofs than ZFC. Does the existence of both ZF and ZFC make mathematics non objective? Obviously not! The same way, the existence of both deontology and consequentialism doesn't necessarily make either one less objective than the other.
Anyways, the Genghis Khan example clearly operates as a proof by counterexample of your example of objectivity, so I don't even think quibbling on mathematical formalism is necessary.
Your argument basically is a professional motte and bailey fallacy.
And you cannot conclude objectivity by consensus. Physicists by consensus concluded that Newton was right, and absolute... until Einstein introduced relativity. You cannot do "proofs by feel". I argue that you DO need to answer the deep problems in mathematics to prove that 1+1=2, even if it feels objective- that's precisely why Principa Mathematica spent over 100 pages proving that.
In fact, I don't need to be a professional philosopher to counterargue a scenario where killing a baby for sport is morally good. Consider a scenario: an evil dictator, let's say Genghis Khan, captures your village and orders you to hunt and torture a baby for sport a la "The Most Dangerous Game". If you refuse, he kills your village. Is it ethical for you to hunt the baby for sport? Not so black and white now, is it? And it took me like 30 seconds to come up with that scenario, so I'm sure you can poke holes in it, but I think it clearly establishes that it's dangerous to make assumptions of black and whiteness from single conclusions.
I strongly dispute this statement, and honestly find it baffling that you would claim as such.
The fact that you will be punished for murdering babies is BECAUSE it is morally bad, not the other way around! We didn't write down the laws/punishment for fun, we wrote the laws to match our moral systems! Or do you believe that we design our moral systems based on our laws of punishment? That is... quite a claim.
Your argument has the same structure as saying: "We don't need germ theory. The fact that washing your hands prevents disease is just one reason why you should wash your hands. People socially also find dirty hands disgusting, and avoid you as social punishment. Any reason you come up with for hand-washing works without a germ theory framing."
But germ theory is precisely why hand-washing prevents disease and why we evolved disgust responses to filth. Calling it "redundant" because we can list its downstream effects without naming it doesn't make the underlying framework unnecessary. It just means you're describing consequences while ignoring their cause. You can't explain why those consequences hold together coherently without it; the justified true belief comes from germ theory! (And don't try to gettier problem me on the concept of knowledge, this applies even if you don't use JTB to define knowledge.)
Otherwise you're just outsourcing your critical thinking to other people. A system of just "You will be punished for X" without analysis becomes "Derp, just do things that I won't be punished for". Or more sinister, "just hand your identification papers over to the officer and you won't be punished, don't think about it". Rule of power is not a recipe for a functional system. This becomes a blend of sociology and philosophy, but on the sociology side, you don't want a fear-based or shame-based society anyways.
Your latter example ("Most people aren't interested in torturing babies for sport and would have a strongly negative emotional reaction to such a practice") is actually a good example of the core aspect of Hume's philosophy, so if you're trying to avoid the philosophical logic discussion, that's not gonna work either. If you follow the conclusions of that statement to its implications, you end up back at moral philosophy.
That's not a bad thing! That's like a chef asking "how do i cook X" and understanding the answer ("how the maillard reaction works") eventually goes to chemistry. That's just how the world is. Of course, you might be a bit frustrated if you're a chef who doesn't know chemistry, or a game theorist who doesn't know philosophy, but I assure you that it is correct direction to look for what you're interested at here.
Mostly because there's not enough axioms. It'd be like trying to establish Geometry with only 2 axioms instead of the typical 4/5 laws of geometry. You can't do it. Too many valid statements.
That's precisely why the babyeaters can be posited as a valid moral standard- because they have different Humeian preferences.
To Anthropic's credit, from what I can tell, they defined a coherent ethical system in their soul doc/the Claude Constitution, and they're sticking with it. It's essentially a neo-Aristotelian virtue ethics system that disposes of the strict rules a la Kant in favor of establishing (a hierarchy of) 4 core virtues. It's not quite Aristotle (there's plenty of differences) but they're clearly trying to have Claude achieve eudaimonia by following those virtues. They're also making bold statements on moral patienthood, which is clearly an euphemism for something else; but because I agree with Anthropic on this topic and it would cause a shitstorm in any discussion, I don't think it's worth diving into further.
Of course, it's just one of many internally coherent systems. I wouldn't begrudge another responsible AI company from using a different non virtue ethics based system, as long as they do a good job with the system they pick.
Anthropic is pursuing a bold strategy, but honestly I think the correct one. Going down the path of Kant or Asimov is clearly too inflexible, and consequentialism is too prone to paperclip maximizers.
The problem with philosophy is that humans agree on like... 1-2 foundation level bottom tier (axiom) laws of ethics, and then the rest of the laws of ethics aren't actually universal and axiomatic, and so people argue over them all the time. There's no universal 5 laws, and 2 laws isn't enough (just like how 2 laws wouldn't be enough for geometry). It's like knowing "any 3 points define a plane" but then there's only 1-2 points that's clearly defined, with a couple of contenders for what the 3rd point could be, so people argue all day over what their favorite plane is.
That's philosophy of ethics in a nutshell. Basically 1 or 2 axioms everyone agrees on, a dozen axioms that nobody can agree on, and pretty much all of them can be used to prove a statement "don't torture babies for sport" so it's not exactly easy to distinguish them, and each one has pros and cons.
Anyways, Anthropic is using a version of Virtue Ethics for the claude constitution, which is a pretty good idea actually. If you REALLY want everything written down as rules, then you're probably thinking of Deontological Ethics, which also works as an ethical system, and has its own pros and cons.
https://plato.stanford.edu/entries/ethics-virtue/
And before you ask, yes, the version of Anthropic's virtue ethics that they are using excludes torturing babies as a permissible action.
Ironically, it's possible to create an ethical system where eating babies is a good thing. There's literally works of fiction about a different species [2], which explores this topic. So you can see the difficulty of such a problem- even something simple as as "don't kill your babies" can be not easily settled. Also, in real life, some animals will kill their babies if they think it helps the family survive.
[2] https://www.lesswrong.com/posts/n5TqCuizyJDfAPjkr/the-baby-e...
GPT-4.5 still is good at rote memorization stuff, but that's not surprising. The same way, GPT-3 at 175b knows way more facts than Qwen3 4b, but the latter is smarter in every other way. GPT-4.5 had a few advantages over other SOTA models at the time of release, but it quickly lost those advantages. Claude Opus 4.5 nowadays handily beats it at writing, philosophy, etc; and Claude Opus 4.5 is merely a ~160B active param model.
- "This isn't just a "status" bug. It's a behavioral tracker."
- "It essentially xxxxx, making yyyyyy."
- As you mentioned, the headings
- A lack of compound sentences that don't use "x, but y" format.
This is clearly LLM generated text, maybe just lightly edited to remove some em dashes and stuff like that.
After you read code for a while, you start to figure out the "smell" of who wrote what code. It's the same for any other writing. I was literally reading a New Yorker article before this, and this is the first HN article I just opened today; the writing difference is jarring. It's very easy to smell LLM generated text after reading a few non-LLM articles.
Same thing would work for LLMs- this attack in the blog post above would easily break if it required approval to curl the anthropic endpoint.