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

Funny, that website ( https://qwen.ai/blog?id=qwen3.8-omni-flash ) downloads automatically 441 MiB of mp4 files... (using Firefox on Linux PC - noticed it because my network chart in Gkrellm spiked for several seconds).

Exact same test executed on my RTX 3060 (12 GiB VRAM, PCIe 3.0 4x slot):

  [ Prompt: 95.0 t/s | Generation: 26.5 t/s ]
(the test's prompt is very short but with longer ones the I get ~200 prompt processing rate, but I was hoping for a better token generation rate...)

Am I understanding correctly that no draft model exists (will never exist or just currently does not exist yet)?

There is no draft file in Huggingface's repository ( https://huggingface.co/prism-ml/Ternary-Bonsai-2-27B-gguf/tr... ) and in the file "scripts/download_models.sh" of the demo repository ( https://github.com/PrismML-Eng/Bonsai-demo/blob/main/scripts... ) I see this remark:

  if [ "$_family" = "bonsai2" ]; then
    the projector ships in the same repo; Bonsai 2 has no dspark drafter

> 192 GB GPU memory

How did I miss that? I was still sleeping I guess

Hehe - yeah, I was pretty sure that you just overlooked it - pointing that out was irresistible :o)

I know nothing about "vLLM" -> anything better compared to "llama.cpp"?

I started with "Ollama" (precompiled version) and it worked and was good enough to understand the very basics.

Then I downloaded the sourcecode of "llama.cpp", compiled it with specific compilation options for my GPUs (CUDA/nVidia using proprietary module on Gentoo Linux) & CPU (AMD), and the same model ran twice as fast -> since then I stuck with "llama.cpp" (and "ik_llama.cpp" in very few cases).

I honestly don't know what made "Ollama" (precompiled) so much slower than "llama.cpp" (compiled locally) at that time and I'm too lazy to doublecheck now, in any case I now absolutely love all the knobs that "llama.cpp" has to tune your hardware setup & your workload, which is the reason why I recommend it.


The parameter "reasoning_effort" is something new, or am I wrong?

  Qwen3.8 comes with official support for reasoning_effort, which can be used to adjust reasoning depth and control cost:
    - xhigh (default): for complex tasks demanding thorough analysis
    - medium: balancing accuracy and speed
    - low: efficient reasoning optimizing for speed and cost
  In addition, preserve_thinking is enabled by default for all workloads for the best out-of-the-box experience.
Asking because in my case (OCR of scanned historical "National Geographic" magazines) the LLM trying to merge text split into separate columns was running in circles from time to time and needed a lot of prompt tuning when using Qwen 3.0/3.5/3.6 (still needs from time to time).


I'm using Qwen 3.5 for OCR, and reasoning_effort is supported there too. I found that it can be loop-prone (though somewhat less so) even if you set reasoning_effort to low.


Let's add as well "SpaceX’s Terafab will rely on natural gas power plants, not Tesla solar panels" to the mix :o(

https://techcrunch.com/2026/08/07/spacexs-terafab-will-rely-...


Also an important part that keeps getting left out for some reason: they're doing so illegally. They are running unpermitted turbines. They claimed the turbines were off and journalists had to use thermal cameras mounted on drones to prove that they are indeed running. Then a non-profit had to sue them to get them to agree to turn off the turbines by 2027...

That's how bad our system is at giving any consequences to white collar crime.


and they (and many other AI data centers) run on some of the environmental/health hazard worst kinds of gas turbines available (i.e. cheap portable ones)

A well build gas power plane can be one of the cleanest fossile fuel power plants with neither huge issues of air nor sound pollution (compared to other fossile fuel, it still isn't perfect by far)

Same btw. for data center cooling, my office was for 3 years above a data center of a super computer in the mittle of a city and there where no sound issues at all (through yes it was a special case in many ways).

I mean yes, it costs more and many well established companies for high quality gas turbines are booked out.

Still using the kind of gas turbines which can bring back literal acid rain shouldn't be an option at all.

Lets also not speak about the environmental/health damage the production of the natural gas through fracking does.

Which bring us to a common nonsense argument, people going on and on about hoe economically expensive/harmful climate protection is, but totally missing that many oft he circumvented laws are not for climate protection but are for environmental protection for the sake of humans health.

To make thinks worse most of this data centers seems to be more about making money, then being in any way relevant for winning any AI race...


They probably take portable ones so they can claim it's not a permanent installation and they can skip a lot of planning permission that way.


Problem is all the (relatively) efficient and (relatively) quiet turbines have long since been snapped up, so the AI sloppers have to go with inefficient, noisy ones for the foreseeable future. And given that no-one knows when the AI bubble will finally burst, no manufacturer wants to sink hundreds of millions into new plant to produce turbines in a year or two when the sloppers may be barely alive on taxpayer cash injections.


No, there are extreme consequences only when toes are stepped on. For example SBF... If it was run by Musk or Trump you can be sure there would be no sentence. I'm not arguing what SBF did isn't a crime. I'm saying if that deserves 20 years, then everything else should also be equally punished.


White collar crime is the only crime that should be eligible for the death penalty.


I don't support the death penalty at all but I have to admit I think society would be better off than what we currently have


Don't understand why. AI training can be run during the sunny days. Could be a good PR opportunity for Musk, I'm surprised he doesn't take it.

Especially if he buys from his own company.


When you spend many billions buying GPUs, you want to maximize utilization, and not turn them off during the night.


When you spend many millions buying a government, you want to maximize your return on that investment, not just play by the rules like everyone else.


You could argue that Musk spent 44 billions buying that.


It’s an industrial supply bottleneck and rollout timeline incompatibility rather than a Dr Doom Evil Genius Complex that prevents everyone from deploying solar. The race dynamics make the discount rate on GPU deployments very high.

Solar is preferable when you have a ton of unproductive land and longer timelines to set it up - it’s certainly going to be cheaper than gas when everyone goes sets up NG plants as option 1 for new energy capacity.

The other confound is that China is the primary supplier of battery and solar tech, I’m sure that supply chain risk becomes part of the calculus when setting up so much infrastructure.

For scale, the current DC is about 8k acres, and the solar to run 7.6GW is almost 40k acres.


Natural gas generators are also a supply bottleneck. https://www.naturalgasintel.com/news/turbine-shortage-threat...


_decent, clean_ gas generators have a bottleneck

cheap, dirty, loud "tbh. shouldn't exist" kind fo gas generators are far more available and what is widely used

but also huge part of the solar supply bottleneck is artificial due to import regulations from china and both the US and EU having failed to secure meaningful domestic solar supply chains.


America produces a ton of LNG. Its cheap and corporations will ALWAYS choose cheap unless the government steps in.


»It’s an industrial supply bottleneck and rollout timeline incompatibility rather than a Dr Doom Evil Genius Complex that prevents everyone from deploying solar.«

No, it's just physics.


no, it's not physics but politics which are the problem here


The reasons are:

- Running billions of dollars worth of GPUs only 30-50% of the time insanely wasteful. Everyone else is running them at 99% duty cycle or close to it.

- Solar is big and generally has to be built far from where the data centres are. Transmission is already a bottleneck in many states. Adding tens of gigawatts requires non-trivial large scale infrastructure investment by… the utility companies, not SpaceX. Not to mention regulatory hurdles, etc.

- Battery production is already mostly maxed out, so obtaining tens of gigawatt hours of capacity on short notice is a decent percentage of global production.

Gas and other fossil fuel power plants can be deployed adjacent to the data centres for effectively zero transmission loss and no long powerlines that upset the NIMBY types.

They can work 24/7 and are largely available right now, but as others have pointed out the turbine blades are a limiting factor with these too.


> Solar is big and generally has to be built far from where the data centres are.

these data centers dont care that much about network throughput - you can sneakernet the training set into place and inference on the text isnt that taxing on the network

there's a reason the ai datacenters are being built wherever the power is cheapest and theres the lowest regulations. there's no real physical location requirement, so you could actually put them close to the solar power generation


Solar works as well as it does only because we’ve spread it out and connected it to a grid. When it is overcast in one spot it is sunny elsewhere.

That “elsewhere” has to be far enough away to get different weather.

This means a grid capable of transferring 100% of the needed power.

So even if you plopped a new DC in the middle of a field of panels you’d still need a ~1 GA grid connection… or a battery array many times the size as you’d otherwise need just to cover night time usage.


All those companies are sitting on literally hundreds of billions of dollars in cash. If they need this so badly, let them build the connection and other infrastructure.


Cash doesn't buy time machines.

They want to build the data centres right now, but the power companies and other energy suppliers are booked solid for 5+ years out.

Money or not, you can't just click your fingers and instantly get regulatory approval, or spin up a facility to build parts that only a handful of places have ever figured out how to mass-manufacture.

Gas turbine blades are especially difficult, but batteries and panels are not far behind. There are a surprisingly small number of shops capable of making power switching gear and transformers too.


> They want to build the data centres right now,

And I want them to behave in a decent way, but universe does not ensure we get what we want.

They should be forced to build a grid in a way that does not speeds up global warming. What they want should be irrelevant.


I hope you get paid for apologizing for the worst people in the world.


Idk. about any specific data center but many of this data centers aren't for AI training / making advancements on the AI race. They have sufficient resources for that already AFIK.

They are to run AI, so that the 100th app can force an AI agent down your throat weather your want or not /s. (Or in more optimistic words, so that the AI adoption can spread as fast as their is demand.)


The problem is that the grid can’t get built out fast enough. You’ll probably see most AI data centers switch to grid power in the coming years using mostly wind and solar because it’s simply cheaper as well


They could (and should be required to) build out their own solar/wind/battery infra as part of standing up a new datacenter. So much money is being wasted on poisoning our air and further enriching the richest companies in the history of the world.


Remember when they said they couldnzt electrify the grid for ev's so they opted to just not do anything with ev's. Now datacenters are getting rushed in so fast and breaking the grids so hard they can illegally running generators


EVs aren't popular in the US due to the Republicans being beholden to oil and gas interests and the Democrats to unions. The grid was just an excuse.


Both sides! We need to make up democrats against EV, because conservatives are!


One side being worse isn't an excuse for the other being bad.

Regarding EVs, the US started to fall behind in 2019, and the gap has only continued to widen, regardless of administration:

https://ourworldindata.org/grapher/electric-car-sales-share?...

Electric buses should be a no-brainer, yet New York comes last among US states in a recent market survey:

https://theicct.org/publication/zero-emission-bus-and-truck-...


"Fast enough" for what?


I would assume that training pales in comparison to inference when it comes to energy use, taking into account the number of inference users


I wouldn't assume that. I would wish to see numbers on that first.


»Don't understand why.«

Because you cannot turn a large datacenter with thousands of aggregates, power supplies and so on off and on like a toaster.


You totally can. I've seen cases a rat gnawed a circuit breaker and blacked out a DC. Fires too. It's not an industrial manufacturing facility. Computers can just be shut off and on. There's also nothing needed to bootstrap the other parts like cooling etc.


A rat taking out a DC resulting in a black start event isn't the same thing as being able to turn a DC on and off at will. The thousands of computers in an DC running various bits of software don't just power on like your Windows computer and result in a working data center after a couple of minutes automatically.


Datacenters literally have UPS power supplies to let them gracefully shutdown and start up.

A 4 hour grid battery and a weather forecast would easily let a training run follow energy availability.


Are the UPSes actual batteries or just backup generators? If they're generators then that defeats the point. If they're batteries then it's kinda just regular solar then?


But that's the point. It's not like a solar DC would suddenly lose power because of cloud cover. Large scale batteries can provide amble ride through inclement weather to allow a graceful shutdown, and weather forecasts at that scale are accurate enough to predictively model your energy availability.

AI training is unique because it's not interactive workload serving.


when the GPU capex overshadows everything else your idle hardware is burning enormous amounts of money


I totally can. And if it's powered by solar (low inertia) there's less obligations to supply side, solar panels will just get warmer a tiny bit.


Batteries exist, and are getting better all the time.


I don't see why this isn't obvious. Terafab is a FAB not a datacenter.


Inference almost certainly dominates training workloads at this point and that can only be expected to continue to become more lopsided


and would be cheaper in the long run? or maybe they could even do solar + storage.

i guess even at this stage “gas is easier” is still winning…


Elon has aligned with the hard-right. He'd rather own the libs and hurt the global south than do that.


You’re talking about one of the most hated person on hearth with one of the worst reputation. The literal Nazi salute guy. I don’t think he cares much about positive PR like that


Hospitals can also just run on the sunny days.

Renewable energy only makes sense in a society that is agrarian. The second you have to make anything industrial you need baseload. That baseload can either be coal or nuclear.


In modern times you combine solar energy with wind power and batteries and you already have 90% of the load covered. Then combine that with cheap, easily adjustable gas power plants to complete the picture for cloudy, windless winter days.

There's really no need for coal or nuclear power plants anymore.


Building gas power plants that run only a few weeks per year is very expensive. Just compare prices in Germany (gas) with China (coal) or Finland (nuclear).

A more economical way would be to build nuclear power plants that power electrolyzers most of the time and provide power to the grid when necessary.

In the US, gas power plants make more sense due to ridiculously cheap gas, but even then the capacity factor of those power plants is much higher than 10%.


If you run a washing machine sure.

If you need to run an industrial process, not so much.


Comparing 24/7 AI training to operating hospitals 24/7 is a bit of a false equivalency, don't you think?

Besides, hydroelectric is commonly used as baseload where available. Additionally, there are big strides happening in power storage options to smooth the effective baseload of solar and wind.

I would argue that coal energy only makes sense in an anachronistic society doomed to fail.


Where available is limited to Norway. Building enough mountains per capita in the rest of the world is somewhat difficult.


You've been confidently wrong a couple times now - or maybe intentionally misrepresenting the facts?

In the US there are several states where hydro is the majority source of baseload electricity: Washington, Idaho, Vermont, Oregon, etc.

Outside of the US there are several countries that source more than 90% of their electricity from hydro - at higher percentages than Norway even: Paraguay, Albania, Democratic Republic of Congo, Nepal, Namibia, Zambia, Tajikistan.

You also don't have to build mountains to utilize hydro. There are areas that are utilizing pumped hydro to store the kinetic energy to smooth the baseline.


Hydro is the primary source of electricity in Canada, around 60%.

Quebec produces far more electricity via hydro than any single US state, some of which gets exported to the northeastern US.


Clearly there is no industry in Canada then! /s


I've worked in energy and ran the numbers. Renewables only make sense for a pre industrial or deindustrializing society. When you need power to make shit they are useless. Which is why the industrial revolution moved from water to coal in the first place.


Oh you ran the numbers!? If only you had led with that.

Even China has seen a stark reduction in the percentage of electricity generated from coal plants over the past decade. Are they deindustrializing?

Everyone knows coal is a dead end. Maybe you should run the numbers again.


https://ourworldindata.org/profile/energy/china#what-sources...

That's a funny statistic to focus on when the actual use of coal for electricity has increased by 50% over the same period.

It's almost like they are using it to cover baseload or something.


You realize that link shows the exact opposite of what you're saying it shows, right?

It shows that the percentage share of electricity generated from coal in China peaked at ~80% in 2007 and was at ~54% last year.


You do realize that Chinese electricity production quadrupled in that time?


Yes - what's your point? It seems like coal wasn't the golden ticket you're making it out to be so they built alternatives.


Do you understand what baseload means?


Yep and China is leading the world in recent and active construction of traditional hydro (dams) and pumped hydro storage to reduce their reliance on coal for baseload. I'm done with your bad-faith arguments.


Their 30 year plan is to move to nuclear for base load, and they are on track to manage it. The only people who think China is moving to a pre 2020 like western grid are people who don't know the first thing about China.


The above listed Washington, Idaho, Vermont, and Oregon are all industrialized societies.


Unplug a ventilator for five minutes and a patient dies.

If unplugging a data center for five minutes can have a similar effect, then every AWS or cloudflare outage would have been a mass extinction event.


Or we can turn off the net negative data centers in order to keep our Hospitals running smoothly, rather than taking the entire grid down so some psychopaths in Silicon Valley can LARP at creating their silicon God.


Gentoorian here as well, cheers.

I use it since ~20 years on most of my notebooks/PCs/root servers mainly because of its flexibility and I often end up wanting to use the very latest versions of software (e.g. ZFS, kernels, etc...) without struggling as most things are available from the official central package repository. I admit that updating/upgrading the OS is a pain from the time perspective (because of the long compilation times). As a programmer/developer the experience is as well fantastic as most of the libraries, compilers, interpreters, etc are very well integrated or easily installable.

I did try to use a few times Linux Arch, but I never fully liked its AUR.

For VMs I basically use always either Debian or Linux Mint with Xfce.

For gaming, mediacenter-PC and webradio-miniPC I use Linux Mint with Xfce.

It's funny that whenever I need assistance with some task (fixing a problem, doing something weird, ...) and I mention to Google Gemini something like "I'm using Gentoo Linux blahblah" then often Gemini starts its reply with something like "Ah, a picky Gentoorian, I see - then let's do this the complicated way..." :o)


Infos: https://oracleoftime.com/how-to-buy-new-swatch-x-omega-moons...

ESTA = "Electronic Swatch Timepiece Application", as pun/mockery of the US "Electronic System for Travel Authorization".


For the ones that don't know "nmon", have a look at it as well! (press "h" to see the list of available monitors - press it again to make it go away, press "q" to quit)

https://nmon.sourceforge.io/pmwiki.php

Especially disk throughput and I/O (keys "d" & "D") can be very useful.


I know it from some AIX machines that we have at job, also topas comes to mind


very useful tool. I install it on every machine where I have control. I appreciate the "Wide" CPU usage graph, that can handle huge core counts easily


> I have trouble getting my computers to copy data between each other anywhere near the theoretical limit for a 1 Gbps NIC, despite being connected through only a single switch.

Maybe your LAN cables have deteriorated?

Reaching 100 MiB/s should be easy peasy for a PC; in my case (2 10Gbps switches) by using nvme SSDs I reach almost 950 MiB/s and with 4x raidz1 HDDs I reach 450-650 MiB/s.

Until a few years ago I was smoking in my flat and often I had to change lan cables (whatever's in the smoke was sticking to the connectors)


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

Search: