I asked Claude to come up with 7 problems to solve that were not part of any benchmarks. Besides speed, I asked it to put together a scoring that would reflect models that perform well on the 7 problems and in a reasonable time-frame (RWS, the definition is in the artifact). I valued models that produced good answers in a reasonable time better than models that had very high throughput but kept spitting useless stuff. Overall, IIRC most of them still had a >30% chance of producing wrong/non-working replies.
I then tested Strata ( https://github.com/niko1221/strata ) and the results are not even close. You can basically ignore everything in that table and use Strata instead, you can use bigger/less-quantized models and have way better results. Basically any otherwise-decent model degrades noticeably at Q < 5, and KV cache at Q4 also contributes a lot.
Still, a 20$/mo Claude subscription is way cheaper/more efficient if you code 2-3h in the evening than running stuff locally, unless you really have hardware than can hold models that get most things right at the first try. You can iterate way faster and cover a lot more ground in 2-3h using a Claude subscription, or OpenRouter if you prefer to change models. You get faster replies that are 99% correct at the first try, whereas using the local model you'd be spending 30% of your time fixing things that don't work and prompting it to do better.
Nowadays I use the local model to delegate tasks that would burn a lot of tokens, like sifting through big data archives for indexing and such.
QFN Q4 K XL Unsloth is basically identical to the bf16 model. It’s KL divergence is very very small. It is competitive with the full model. Also I benchmarked iq3 xxs on Strata for 3 days against deepswe and it scored better than the published benchnarks for the model. If the gsq rco quant prettain tasks fit well it is a great model quant. My data shows it takes 20% more tokens basically than q4 xl. A4 xl is within noise on all my repeated task benchmarks over the public model serving and my own runs of q8. I published the deeswe results. You have to be very specific about model choice and know these models and their quirks.
This is not about if a frontier AI provider is cheaper. It is about having control over your data and use of AI.
> QFN Q4 K XL Unsloth is basically identical to the bf16 model.
Yes, but like original poster I have 16GB of VRAM and 64GB of RAM, I was referring to non-Strata benchmarks :)
I did quickly run a couple things on Strata but haven't gotten around yet to benchmark bigger models/customize them with Strata, but will do rather soon. So far Strata/Swift IQ2_XS performed really well compared to non-Strata/Qwen3.8 quantized variants.
> This is not about if a frontier AI provider is cheaper. It is about having control over your data and use of AI.
Yep, totally get it, and to be honest getting to run ablated models sometimes can also be useful.
> Also YMMV in “faster” — I have less than 5K into my home AI lab shenanigans and it’s pretty fast :)
My rig is... ~1.5k, maybe 3k if we count the CPU and GPU upgrade since 2019 :-D I really wanted to grab one of the latest mini-PCs (Halo, Spark, and other minor assemblers) for local setup, but to be honest I can't justify a 6k+ purchase for me and my wife alone. Perhaps in 3 years these mini-PCs will get to the ~2k ballpark (perhaps used), and although older by then, I might be able to run things 24/7 locally as well.
With 32 GB more of RAM you can run ISTA GSQ RCO IQ3 XXS fast. You can make it work now but it is tight. You can stream experts from disk with builds now. You will get an okay token rate. Try it out if you have some disk space. It is a very good local model. I benched it at that quant. Sonnet 5 class at coding benchmarks.
Get a couple or V100S GPU. Figure out cooling. They will run it /fast/ with the right engine. :)
Social/IM platforms are literally where the violent ones organize. Nobody has issues with the peaceful protests, but in the past couple weeks it's become very clear that violent rioters use the excuse of the peaceful organized protest to infiltrate and wreak havoc, using the peaceful demonstrators as essentially meat shields. That needs to stop. And without readily available social/IM platforms, there's no way the violent ones can coordinate in the hundreds/thousands on a couple hours notice.
I have no idea what happens in France now, but in the past of many other countries the main place where the violent protests were really organized was not on any Social/IM platforms, even in the cases where it appeared to be so, but at the police/secret police/government HQs, where covert agents were prepared for triggering and/or leading violent actions during the protests, in order to discredit the peaceful demonstrators and enable direct action of the police against them.
Nowadays whenever you see "violent protests" and "destructions of private property by vandals during protests" you can never be certain whether they are organized by people who are against the government or by the government itself.
There are multiple videos from France showing undercover police instigating and breaking things.
Then, the government is pushing cities to denounce the most popular Left party as the instigator of the riots.
They are preparing to ban them next elections for this. It's pretty clear it's all a political maneuver to ensure it's once again Macron's party versus Marine's far right party, hoping it ensures their win against her
How do you figure that? Why would their members have called to vote for the left parties in the areas where they were losing the last legislative elections if they were preferring Le Pen?
I'm not saying Macron is happy with the "extreme left", but it seems to me he's much more against the "extreme right". And not just him: pretty much every politician and their dog is trying to distance themselves from the Le Pen's party.
They normally have to ask for dates, permits, are normally driven by groups that have been around for long, they have websites, they have a list of asks, etc etc.
Compare that with violent ones that see the opportunity coming with 3-4d heads-up (because peaceful protests are normally covered with a little advance - 3-4d - by mainstream press), create an Instagram/TikTok/Whatsapp/Snapchat group, spread it aggressively thanks to zero cost of social/IM, and their only objective is to smash everything in their path (and if they manage, steal some loot to resell).
IM cannot be stopped. When a government broadcasts they will censor them, the very next step is for protesters to install something like Bitchat, which does Bluetooth-based, mesh/P2P IM-ing, and is specifically designed to resist even full-on Internet shutdowns.
The best way for governments to deal with protests is to have policies in place that fulfill what We The People want. Everything else is a band-aid atop an ulcer: it hides the problem for a little while, but the problem itself keeps growing regardless.
If there are massive protests, it's clear the reconciliation hasn't reconciled things properly, and a better reconciliation must be done, preferably one that doesn't result in more massive protests from either the same or other huge groups.
In fact, remarkable long-term lack of massive protests would be a very good metric for determining whether ongoing reconciliatory efforts are moving in the right direction. If such protests start happening, correct course sooner rather than later, and measure success by a newfound lack of further massive protests.
> violent rioters use the excuse of the peaceful organized protest to infiltrate and wreak havoc
This is not true, and very much the kind of propaganda spewed by the (far-)right, please do not trust it.
The entire crowd chants very explicit chants against the government and police, gathers around fires in the street, shouts at police when they attack the demonstrators or fire gas canisters and flash-bang grenades. The crowds that get out are not getting smaller, despite mainstream media and government equating the protests with violent looters, so it seems everyone marching feels a need to. There are looters, but they're a tiny, insignificant minority.
The protests are violent because the repression is violent, and because they are marching against real suffering.
You cannot throw out the current neoliberal political regimes based on peaceful protests alone, I thought that by now that would have been obvious for most of us adults in the room who went through Occupy and the peaceful protests of the early 2010s. Tahrir wasn’t peaceful, throwing out Assad wasn’t peaceful, why do people still have the expectation that throwing out a regime that puts an investment banker as its leader is supposed to be peaceful?
Presumably, because the people who want to change things do have in-system ways to do so: joining into a new Party, preaching their views, getting elected, growing, until they convince enough people to give them the power they need to change things.
In practice there may be many roadblocks in that, so this isn't absolute. But if that's the case, it's still on the people who notice the roadblocks to convince most everyone these are real. Without that any protest will lack widespread support and be suppressed regardless.
Empirical history contradicts your claim. In reality people protest when life feels bad and when life feels bad for many many people, there are really big protests. If it gets to this point, the system has already failed, because life already feels bad.
Arguing that the system has escape hatches is just an excuse. The escape hatches have also failed, clearly, since they either haven't been used (which is their fault) or their use did not fix the system.
political movements that pose a threat to the regime's interests are suppressed much more aggressively and in much earlier stages, debanking etc.
there are many NGOs aligned with the current protests, and they receive substantial government funding, because they are in symbiosis. the regime doesn't provide funding for ideologies that actually threaten its interests.
public opinion has now reached a tipping point, any further vandalism/arson/violence will not provide a net benefit to the regime. that's why now they finally try to stop it.
I once worked in L1 support, and because we had to take calls @07:00 sharp (we were the only number a company of 40k people could call) and Windows updates were still a multi-hour endeavor, I stuck to rebooting daily (or shutdown+startup) all devices I use heavily, including MacOS.
Over the past 8 years that has basically solved every bug all my colleagues ran into with MacOS. At some point my boss wasn't getting calendar notifications anymore for a couple weeks, "did you try and reboot?" fixed it in 5m. He later mentioned he hadn't rebooted in 7 weeks.
I don't understand why people treat Apple devices like unbreakable analog AM radios, they're very complex computers like any Windows or Linux machines, with a huge state machine beneath it that is easier to reset daily/weekly than to keep working properly for months.
After its commercial span is over, over 90% solar and battery materials can be recycled to make new panels and batteries. Compare that with nuclear where we need to stow away basically-forever-radioactive spent material.
We have a giant fusion machine in the sky delivering energy in great excess what we use, and we only capture a tiny fraction of it at the moment. We can make each household practically independent. Fission and (hopefully soon) fusion will be good for meeting grid demand when renewables can't, but a distributed grid is simply more resilient to everything, from local events to country and nation-wide disruption of central generation points.
Apparently it's still mostly about Aluminum and Glass as that's 95% of the weight of panels. The silicon is often recycled for precious-metals recovery and not circular. ROSI does try to do high-purity recovery. It seems volumes are still on the smaller side, probably because most of the volume has been deployed in the last 5-10y and they have 20-30y life expectancy. Still a lot of recycling.
This, I repasted my wife's Zephyrus laptop which had liquid metal for the Ryzen APU (6900HS), regular thermal paste for the GPU (3080) and thermal putty for VRMs.
Redid APU and GPU with PTM and VRMs with paste. -10C on average on the APU with way lower idle temps, and GPU went from being thermally throttled at 60W to being able to push until 110W. Fan is basically silent for 3 months now, whereas previously I could hear it from the next room.
I've always used source-based packages, but chromium was the one I simply dropped in favour of Firefox.
On a 16c32t / 64G machine, compiling Firefox takes ~1h with LTO and PGO, while Chromium already took 5+h in 2019 without PGO. Then they removed jumbo-build because there were some issues and kept adding code and toolchain deps (jumbo-build is a build step that merges a few hundred source files before compiling to do the header processing once, shrinking the build from 5h+ to ~1-2h). Last time I checked the build had 70k+ build steps, with the final link step easily taking 10m+ with LLD. Answering some other comments: ccache doesn't help, because every version will have some minor header changes somewhere, so basically the entire cache gets invalidated even within the same major (e.g. 151.44.22.11 to 151.44.23.28).
And I have no idea what kind of driving he's doing to spend 2-3k$/y . I got Front Upper Control Arms replaced under warranty, exactly once, in the lifetime of my 2019 M3 (also Fremont-built). I do a yearly inspection just to catch issues before they have a chance to become really problematic, but for the past 4 years I've always been told that "the car's underside is just perfect, there's nothing wrong with it".
> speed limits are generally well below what is safe
I used to think so, but no, not in a school zone or cities at least. They might be safe for the car, but not for pedestrians and cyclists. Those are road users, and speed limits take them into account as well. That's also why you don't see 25mph limits on highways: pedestrians are not allowed there, and the lanes are designed to support high speeds, so the risk for all allowed users is lower and limits can be higher.
Couple that with the fact that everyone wants to ride the fine-limit (and not the road limit) and a 50mph road becomes effectively a 55-60mph road, because you know, shaving 1m off a 1:30h commute is such an achievement.
Some people here complain about 30mph limits, but in EU many city/school zones are becoming 18mph or even 12mph. Two reasons:
1. An impact at 12-18mph might break your leg, but you otherwise have a very good 100% recovery chance. At 30mph, major injuries become more common, and you risk crippling another human being for life. And again, 12-18mph for many people really means +5-10% or more depending on the fine, so 15-25mph in reality.
2. (not accounting reaction time in the following) At 12-18mph, an SUV can break in 13-30ft on wet pavement, which means you might bump someone and they might break a bone or two, but will otherwise probably be okay. At 30mph that's 115-150ft, which means with the little space you might have, there's no way you're going to stop in time to not run them over and cause them major damage.
The other thing is the squared taper-off of speed when braking and I used to not think about this until someone showed the math. If you're going 30km/h and a kid jumps out 10 meters in front of you, you can just barely come to a stop. If you're going 50 km/h, you'll run them over with 30km/h remaining speed and come to a stop about 10m further down.
> Are they dumb or is it intentional to keep their cities small and exclusive?
Madrid has 3.4Mln people and is the second-largest city in the EU. How big do you want it to become?
I disagree on "just build more". There was a time where salaries moved more or less line with cost of living, and especially rents, but with AirBnB and estate funds sucking up any and all places they find, there's an enormous pressure that drives up rents for common people, which already struggle as-is to make ends meet. "Just build more" would simply keep pushing those towards the edges of town over time, while the city becomes a glorified luxury hotel that only rich people can afford. I welcome new regulation to combat this.
I just hope this event shatters any hope for the ultra-right parties that voted against it, and that people remember this. Ultimately, if anyone will build more, it'll still be the large real estate companies and funds that keep increasing rents on their tenants.
> Madrid has 3.4Mln people and is the second-largest city in the EU. How big do you want it to become?
Bigger, obviously. Because the number of people who want to live in Madrid or visit it is higher than that. This can create a lot of economic surplus. Part of that surplus can be used to pay off locals (e.g. with housing in the newly built stock) so they win too.
The real problem, as far as I can tell, is that there's no socially accepted system for paying off locals. So locals end up resisting growth, because there's nothing in it for them except higher rents, and everyone loses.
Last time they touched the housing laws one of the brilliant things they did was to force the owner of the flat to pay the expenses for signing the leasing contract of their tenants, before that change this was paid by the tenant. What happened is the landlords switched from long term rentals to short term rentals because they did not modify that clause for those type of rentals, so the tenants got the worse outcome: rental contracts of 11 months instead of 5 years, which also opens the door to massive rent increases every 11 months or eviction instead of just the inflation rate of the year under the other contracts.
They also keep increasing the difficulty of evicting non paying tenants, taking a lot of rentals out of the market because owners are simply afraid of not being able to evict a non paying tenant, or making prices increase to account for that risk.
> was to force the owner of the flat to pay the expenses for signing the leasing contract of their tenants
The reality is, anyone renting in Spain can readily tell you things don't work this way, if you query this, they go with someone else. Tenants paying the contract fee has an enormous momentum and it's way to easy for landlords to ignore it, in our hot housing market.
Nothing wrong with estate funds owning homes as long as you keep building. They will invest at slight profits and that is why to give flexibility that renters need.
Just need to prevent them from hoarding land or delaying building.
Tokyo has almost 15 million and works very well as a city. China is building many new cities close to the big cities to allow housing. The young population needs to live in the big cities because where the jobs are. Freezing prices will not solve the problem for the vast majority of the people besides the ones that were lucky to born in the place before the freeze.
You keep building until everyone has a house. There is no such thing as "too big".
The Dutch government designed entire cities and reclaimed land from the sea in the 1970s and it's already full. Human civilization is endless growth or death.
I also reject "just building more", at least in already too big cities. I think a more viable long-term approach would be to foster the economy in such way that creates more jobs in smaller cities. Spain needs cities like "Düsseldorf" or "Austin".
I am sceptical of this kind of central planning. Fighting against organic growth and artificially trying to increase jobs in other cities seems like it would cost a lot more as a whole than trying to solve the problem where it actually exists.
You are right to be sceptical and should be. However in this particular case, what we see today, is exactly the outcome of central planning, from the previous regime that ended in the 70s, which effectively decimated industry on the countryside and forced everyone to live in large cities.
> Probably because pop_os and Cosmic are so niche and their market share so insignificant
Consider that it's packaged for many well-known distros, so pop_os install base alone doesn't tell the whole story: https://system76.com/cosmic/download
I asked Claude to come up with 7 problems to solve that were not part of any benchmarks. Besides speed, I asked it to put together a scoring that would reflect models that perform well on the 7 problems and in a reasonable time-frame (RWS, the definition is in the artifact). I valued models that produced good answers in a reasonable time better than models that had very high throughput but kept spitting useless stuff. Overall, IIRC most of them still had a >30% chance of producing wrong/non-working replies.
I then tested Strata ( https://github.com/niko1221/strata ) and the results are not even close. You can basically ignore everything in that table and use Strata instead, you can use bigger/less-quantized models and have way better results. Basically any otherwise-decent model degrades noticeably at Q < 5, and KV cache at Q4 also contributes a lot.
Still, a 20$/mo Claude subscription is way cheaper/more efficient if you code 2-3h in the evening than running stuff locally, unless you really have hardware than can hold models that get most things right at the first try. You can iterate way faster and cover a lot more ground in 2-3h using a Claude subscription, or OpenRouter if you prefer to change models. You get faster replies that are 99% correct at the first try, whereas using the local model you'd be spending 30% of your time fixing things that don't work and prompting it to do better.
Nowadays I use the local model to delegate tasks that would burn a lot of tokens, like sifting through big data archives for indexing and such.
reply