Hacker Newsnew | past | comments | ask | show | jobs | submit | stymaar's commentslogin

Why do you think NVDA is the most valuable company on planet Earth?

Because they're engaging in Enron-esque accounting tricks but because they admit what they're doing it's not illegal?

1. “The X is not Y, it's Z” 2. 3 months old account

Bad bot.

(I still wonder what these not operators have to gain in that process, but they sure want HN karma).


Interestingly, the presumed bot uses unicode apostrophes (’) instead of ', yet misspells "its".

Tamping down on verbosity, removing em-dashes and adding misspellings are the obvious fixes to apply when your first attempts at spamming get flagged for being too on-the-nose: https://news.ycombinator.com/threads?id=coder-pm&next=487167...

There are "humanizer" scripts that these people like to use. This one is not the typical style I see in my online moderator adventures, which is usually "lowercase everything and remove all punctuation except periods", as if we are repelled by em-dash use and not the tortured prose.

I like the removal of a random definite article, to simulate an ESL writer and garner sympathy for using LLMs as a "tool to clean up posts". Heavens, they would never even think of using their Claude subscription to be a slop factory, that would cross a line.


gotta be careful. inevitably people will start emulating the LLM writing style since they're so overexposed to it

n-gram per-layer embeddings[1][2] might be it.

[1] https://sebastianraschka.com/llm-architecture-gallery/per-la...

[2]: See DS 4.1-Flash and Qwen-3.8-Next.


this is to offload VRAM to DRAM (for GP comment), and makes no difference for URAM

You can definitely offload n-gram embeddings to storage; they're very sparsely used (only a few KB fetched per token) so this is quite effective. Loading to DRAM only becomes necessary if they are a bottleneck to overall performance (which might happen if you're doing very wide batches and everything else uses super fast VRAM/HBM).

I was looking at the qwen-next-flash, and the weights would fill my OEM Spark on their own, before the n-gram. I'm unclear if offloading to disk can work here, is that what you are implying is possible?!

Check out eugr’s TP=1 sparkrun recipe :)

It’s an NVFP4 quant, but it fits, and is surprisingly capable.


do you have a HF link? HF search is not uncovering it for me

(or is it somewhere else)


https://github.com/spark-arena/eugr-recipes/blob/main/recipe...

This one!

I'd recommend pointing your agent at it (after installing sparkrun), and asking it to research the absolute latest in TP=1 Flash-Next - mine grabbed particular vLLM nightlies and mods to improve performance, and it was well worth it.


I have a quirky vLLM on k8s on 2x OEM sparks setup with 9 models available to me. I'm not keen to run nightly vLLM, too many issues with it in the past. Going the qwen-next path means displacing things I use daily :/

I have a watchful eye on the diffusion ~ Jev/Kev PR

https://github.com/vllm-project/vllm/pull/57250


For what it's worth, Flash Next outperforms every other model that is available to us on the GB10 in all of my testing; though if you have two sparks then the TP=2 version is even better and easier (I don't think you'll need the nightly for that at all, just use the recipe)

I'm so tempted to buy a second one...


prices have gone up quite a bit...

I'm running embedding, reranking, and policy tuned models too, and a Jev/Kev when that's landed. Flash Next is not a substitute for those

I have OpenCode/Fireworks to access big models



Nah, I’m streaming ngrams off NVMe on my Spark-alike right now. Works surprisingly well (except for when I accidentally bottlenecked it through my NAS)

What kind of throughput do you see on what models?

GB10 boxes have way more compute than they have memory bandwidth, which nicely fits medium sized MoE models with speculative execution (MTP, DSpark/DFlash, etc)

Qwen 3.8 Flash Next (what I'm running basically entirely now) sees 30 / 35.0 / 45 tk/s for prose, analysis and code respectively for actual use (not short context benchmarking) with Pi. Thinking blocks are ~35tk/s or so.

The GB10 having so much compute is great for prefill too, 2000-3000/s for 14k to 64k token prompts (cold cache too) in the quick benchmark I did. 3500tk/s for warm cache which is nice :)

When I accidentally streamed my ngrams over the 2.5Gb/s network, it cut all the throughput down in half basically. Especially notable for the time-to-first-token, which is what clued me in that I'd messed up somehow!

For Qwen 3.8 27B, I got it up to a consistent 20tk-25tk/s but 27B thinks so much that it was honestly too painful: Flash Next is as smart, as useful, but much faster for real agentic dev usage IMO

Laguna S 2.1 saw similar numbers to Flash Next if I remember right, but their latest updates means it doesn't quite fit a GB10 128GB anymore at full context which is a shame.

Note: these are all NVFP4 quants (usually a dynamic one where some tensor layers are left at full precision though)


I personally stopped caring as much about the tok/s as the agents are largely in the background, and so have also moved preference from MoE to dense

I want to see about fine-tuning these models a bit on the GB10 to tame that over thinking and some other behaviors (like using tools I don't use)

qwen 3.8 seems to have been trained with some `rkt` that messes with tool outputs to "save tokens"


check out the spark arena website, its the raison d'etre

interesting, peer comment seems to indicate this is a possibility as well, will have to take a deeper look

n-grams can be kept on SSD, no need to hold them in any kind of RAM (at least w/o batching)

Am I missing a joke? WTF is URAM?

unified memory, not sure if anyone uses URAM, I human hallucinated it

Flash[1]: 309B total / 15B activated parameters

Pro [2]:, 1.02T total / 42B activated parameters

[1]: https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Flash-RL

[2]: https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Pro-RL



Those this mean they've fine-tuned this Qwen 3.5 9B on output from the V2.6 model?

It is a 9B agentic model developed by Xiaomi MiMo through supervised fine-tuning of Qwen3.5-9B on MiMo-generated data

curious why the HF pill (on the right) always has inaccurate values

Packed 4-bit weights are often identified as u8 byte arrays in the safetensors metadata, so the HF UI counts them as half a weight each.

I would think there is sufficient information in the various config files to work this out, based what I've seen in my own quant artifacts.

Yeah, I agree it's probably fixable, but I think a naive interpretation of the metadata is probably the source of this bug (which HF has had for as long as I can remember)

I believe it's because this model is natively fp8 (for the most part), and that display struggles native quants.

I noticed the same, and I wonder as well.

I suspect they are calculating something in the weights or config, I see it pretty consistently with quants

more like 500B in FP8

> Okay, but reading also requires substantial thinking.

I don't think anyone would argue reading a book takes as much thinking as writing one…


Perhaps not 1:1, but, personally, I can read for much longer than I can write. And so I'd argue that this duration increase matches any deeper thought that comes with shorter writing bursts.

That should tell you something about how much effort (thinking) is required for the activity. The fact you writing is t something you can do as much as reading means you use your brain a bit more. I would claim that consuming information always uses less thought then generating information. For a human brain at least.

Most people[1] interacting with chatbots don't have a paid subscription and they do interact with the free-tier LLMs that are Luna and Haiku, so I still think it's relevant.

[1]: not on HN obviously, but IRL, and probably among FT's readership as well.


A free claude account with no subscription gets you access to sonnet and I believe uses it by default over haiku

> The villagers are tired of bartering

Come on… It's been known for more than a century that bartering has never been the default mean of exchange among humans, can we kill this myth eventually?!


This is obviously not an attempt at anything remotely resembling documentary history. It’s more of a just so story explaining what problem each step solved.

Unless of course one thinks that smooth gray stones from a riverbed two miles away reflects some historically significant point in money’s evolution.


I'm not going to explain David Graeber's Debt to you, but yes, this exact 18th century fantasy of the emergence of money in caveman times contains the same kind of misunderstandings that also leads to most people having an incredibly misguided view of how the contemporary global financial system operates.

I certainly don't think anything like this actually happened. And the linked article doesn't really do so either. I do think it can be a useful illustration of what problem each technology solves.

In another thread you talk about how the internet really is a series of tubes. We all know that "tubes" can be a useful analogy for understanding certain things about circuits and things built upon them. But we also don't think that tubes are more than a very basic analogy. Tubes don't really do packet switching.


By leading with the improvement-order-barter story, the article announces a bias that borders on propaganda. It's maybe worth objecting to even if the goal isn't to be historically accurate.

The problem that was actually being solved when money was invented was how to get recently conquered villagers to support the soldiers that recently conquered them. (Demand taxes in the coinage that you pay your solders with, then they can live off the local economy and you don't have to feed them directly, which might be hard if they're very far away--tax collectors can take years to show up, but the demand for the new currency is instant, so your soldiers get fed right away).

The system of interpersonal debt that predated money likely outperformed money in a variety of ways from the perspective of the users. It was only for the conquerors that the invention of money solved any problem at all. It's hard to keep reading the article when it starts with the story that is used to distract from this.


There's a central problem, a blind spot, that plagues economics: because the fundamental credo of the discipline is that it's all about free people engaing in mutually beneficial trade to maximize their respective well-being, it completely neglects how important armed force and violence are in shaping the economic system we know.

Agreed. We should acknowledge that it has historically been about ensuring that conquerer grandpa and conquered grandpa end up with different qualities of life even after they're too old to swing a sword. I.e. extending violence begotten structure into places where violence is impractical.

The problem is the story is wrong, and that leads to misunderstandings of what money is and how economic institutions arose in fact vs mythology.

For starters, debt records came before money basically everywhere we have evidence for the invention of money.

i would very strongly suggest reading David Graeber. It completely revolutionized my understanding of how human social and economic institutions emerge and evolve in actual fact vs imagined myth.

We can just tell the real story, it's far more interesting and useful than the myth.


You are arguing a point I didn’t make.

No, what I and the other commenter are saying is you don't understand why the article is misleading because you don't know the actual scholarly knowledge here.

In your other comment you use wording along the lines of "still useful to show what problem each technology solved" but the part you're missing is the explanations are largely wrong, not just in the sense of not telling actual history but you will come alway with mistaken conclusions because the imagined/example history is fundamentally wrong.

So, consider if you're getting this from multiple people, perhaps you're missing something on the topic.


There are two commenting critics. So "multiple" is true, I suppose, but only quite weakly so. And there are more than 4x that many upvotes on the top-level comment. So I don't think it is as simple as "everyone is saying you are wrong."

Not only bartering has never been the default mean of exchange, but also the idea that money were created to optimize the trading process is also wrong.

For an item to become currency in a society, it must already be something people value independently of its use as money, and people must expect others to accept it as well. Say the villagers live near the sea and find some beautiful shiny stones that everyone wants. Those stones can then become currency.

A good example of this is tobacco in colonial America. It came to be used as currency because it was already widely traded and valued.


Salt was similarly used as a currency. It didn't expire, could be weighed, had inherent value in it's role for food preservation, etc

So a systems using money is just barter system where one commodity is overwhelmingly involved?

It's not “just” that, because monetary systems don't work like that at all.

The point of GP is that while the myth says “in the beginning there was barter, and barter is inefficient so we created money to allow trading more efficiently” the reality is more “In the beginning there was trade, and some traded goods became money because they had the correct properties”.


Isn't that what I just said?

Is that Marcan's vtuber persona (Asahi Lina) that changed name since Marcan doesn't work on Asahi Linux anymore?

Just curious: do you know the expected XP gain over the first 90 minutes for a human player (either a medium-level one or a speedrunner)?

How is that irritating? I find it funny that the article gives you a way to distinguish the way the author pronounces it just from their writing.

It always feels a bit like hitting a speedbump mid-flow!

Mostly breaks concentration/focus etc.


Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: