gruturo

> It isn't generally a problem to hoard unused room. It is a problem to hoard unused room in an environment where room is scarce

So... San Francisco?

You, madam, sir or other, win the thread today. Chapeau.

By the same line of thinking, his SF home has 6 bedrooms, it would be more efficient to transform it in a community housing, why is he hoarding all the bedrooms? He can't even physically be in more than one at a time.

And that's just one of 3 he owns in SF alone. Also he has 2 kidneys and 2 lungs, while he could get by with just 1 of each, and there are people badly needing them.

I will renounce my hardware when he renounces all homes and bedrooms. Until then he's a hypocrite.

Quite a few core people in Raspberry Pi are ex- (or even current) Broadcom engineers, that's what they're familiar with.

You may want to reconsider your choice as I was just blocked from a residential IP in Germany, Firefox on Linux.

Yes LLMs fail and fail often. I use a council of 3 for my health related conversations. They have recommended stuff which proved useless, I even wasted a little money in supplements which later iterations of the same LLM explained are not useful at all. I failed to explain to the LLM that some of what I was taking (baby aspirins) was not a doctor's prescription but my own initiatives, so it didn't dissuade me from continuing them until it came up much later - something my own doctor also agreed should be dropped - she had also assumed it was something I had been prescribed in the past!

Regarding Gell-Mann amnesia, I literally mention it in a separate post in this same thread :)

And I'm sorry if your experience is that most medical professionals would react like that. This isn't absolutely the case for me. I show up with things printed out, don't come across as a lunatic who googled a symptom and just wants to be agreed with, and I explain that my LLM research is so I ask fewer stupid questions in the limited time the doc has for me.

Two things can be true at the same time. And whataboutism isn't changing any of that.

>Why do we expect that for other professions it works differently?

Gell-Mann amnesia most likely applies here.

This, 100%, and I'll add (apologies in advance for bringing AI into this): In an era of extremely accessible and competent language models, you have no excuse not to use one to verify your doctor's (or any hired professional's) work. AI won't replace a professional but can spot mistakes, suggest solutions, educate you on a topic, and, most importantly!, has all the time in the world for you, something your doctor doesn't.

Case in point - just last month: To bring my LDL further down, I was prescribed a doubling of my existing statin dosage. I checked with an LLM and it turns out it's a crap idea, as doubling only weakly increases the effect, while the unwanted impacts on the body increase. Supplementing the existing dosage with bempedoic acid or Ezetimbe would be a way better strategy. I brought this up with the doctor, who immediately agreed it's a better idea and changed my prescription.

It's your body, your health, your future, your loved ones, your home, your career, your investment, your savings. If you care, check the work, the tools exist and are now ridiculously accessible.

I really dislike that term. It's stupid and supports a distorted narrative.

They arbitrarily decide 1M API tokens cost X, and any subscription where usage costs less than that is "subsidized". No it isn't. They may be profiting on subscriptions as well, or losing on full API costs as well, or anything in between, and that X is anyway more reflective of competition than their costs. For all we know they did a rand()%1000 to determine it in the first place.

Honestly it may just be that lots of people - especially those with deep pockets - are going to buy the Duo instead. Not a mention in the article, maybe someone is short on Apple and could use a dip, so conveniently forgets a logical explanation.

Very cool!! Thanks, appreciated!

It's frankly pointless. Pretty much nothing produces them natively, so it's usually a generation loss from JPEG, not enough compression gains without annoyingly visible loss of quality, still has format many format and functional limits so it solved nobody's problem at any time, pretty much.

And much (though not all) of its improved compression (before it damages the image too badly) was anyway matched by the better JPEG encoders which were released and improved over time. It sure benched very well against a 30+ year old JPEG encoder implementation though.

JXL has a pixel-accurate JPEG-input mode with no generation loss which already beats WebP, a native mode which is way way better even, it addresses many more use cases and is not Google's pet project (saying this while thanking them for all the foundational work from which modern image/video codecs benefit - JXL itself first and foremost!) so it DOES have better than a snowball's chance in hell of becoming the defacto standard of the next 10-30 years and we'll soon see it in silicon.

Seconding this. Flash-next (and let's not forget, it's a PREVIEW of the 4 architecture - with the "real" 4 rumored coming later this month) is the first model I can run on reasonable hardware (2 thoroughly obsolete P100s off ebay at ~$100 each plus the RAM I could scavenge from other PCs at home) at a reasonable speed (22, with GPUs in layer-split due to llama-cpp's limitation on qwen4-exp arch, and no MTP. Strata could double these numbers).

It's.... the real thing, for the first time. If you cut me off cloud models today, I would get plenty of utility out of this thing.

(Others may have had the same feeling from GLM5.3 or Deepseek 4.1 flash but I never had a chance of running those.)

While this is generally true, it's _a little_ less true the larger the model is.

Also, quantization techniques have improved - the I in IQ3 stands for imatrix - Importance Matrix - it is a bit more surgical in what it cuts. The result is a model where the most important weights are even Q6 or above, the least important Q2 or even below, overall it takes the space of a Q3 but with better results.

Assuming Valve funds this work, just buy their products. If possible, avoiding credit cards as the CC companies are trying to dictate policies about which games are acceptable and which aren't. I buy the occasional steam card with cash. I encourage everyone to shift a bit of their spending away from CC's in order to reduce the power of these companies to tell us how to live our own lives.

I wouldn't know where to start from.

Most intersections are atrociously designed. Protected left turns everywhere, with their own lane and their own traffic light causing slowdowns and triggering also when nobody needs them.

Many pedestrian crossings could be moved 10m further from an intersection, allowing up to a couple of cars (wishing to turn right) to wait on the pedestrians without blocking all the people who were just going to proceed straight ahead. On the other hand, this problem would often not even occur if we hadn't wasted a lane for a protected left - there now would be a dedicated lane for those going right.

Right-on-red is legally permitted (there's a dedicated street sign for it) but almost never implemented. It's so frustrating.

An overwhelming majority of the traffic lights I encounter is completely non-adaptive. I wait an average of 40 seconds every morning at a specific one which is, in aggregate, causing entire human lifetimes of wasted time on a monthly if not weekly basis.

Zebra crossings are often outfitted with a pedestrian traffic light which is red (rather than off) by default. If the road is completely clear, you could just cross, but no, you must press, wait up to half a minute, and finally it turns to green for you. Now you also stopped a car or 2 which need to wait for your green to finish. If it defaulted to OFF, you could have crossed 20 seconds ago without interrupting anybody's life, and the traffic light would still still be available for elderly or disabled (or whomever wishes to press and wait, while not having it as a default).

...I could continue.

Better not, it would just get me even more pissed off at how crude and incompetent German traffic engineering is. Pretty sure some frustrated worker is making it worse on purpose - you don't reach certain peaks of crappiness accidentally, it takes _work_.

I have an English/Italian bilingual keyboard set up as my default.

Every single f**n time I begin a sentence with "È" (Italian for "she/he/it is") - which is an extremely deliberate gesture - you have to keep the E tapped until the accents pop up - it will immediately correct it back to an unaccented "E". It's nuts.

There is a metric ton of intent behind my selecting, keeping tapped, waiting for the accents, picking one - it's so obviously NOT accidental, and this utterly cretinous piece of junk changes it .

> You seem to be a little confused about how oxygen toxicity works

Thanks for the condescending tone. Am I?

At 1.4 bar of O2 partial pressure (therefore, with a 21% breathing gas mixture, at around 66m depth), oxygen toxicity causes convulsions and unconsciousness. If someone doesn't grab you and take you up, this likely results in death.

Btw I had to google the precise values and depths to be accurate, but I'm not talking out my ass. I'm a certified CMAS and AOW diver with 80ish logged dives. Hardly a veteran but this was in my certification course and it's a very real thing - enlighten my confused state with your better knowledge please.

The raise is even worse if you look further back than 2021. Netflix was 7.99 for ages, while the "floor" in 2021 is 13.99. Against the original baseline, the new 19.99 price is not a +43% but a +150%.

Most importantly 13.99 already in 2021 bought LESS than 7.99 or 8.99 bought until a while before. Thanks to the fragmentation, the catalogue was way larger before, and as of 2022 ads were introduced to further insult the customer.

High seas it is.

A weekend? How about a full week at the absolute bare minimum, if not more.

3 days is a complete joke. Competitors have battery lives measured in weeks rather than days (a famous, and still 100% appropriate, quote by Garmin. Now 5 years old and counting, and the situation is still as ridiculously bad).

Seriously what is it even doing to justify going through the battery so fast? Are they just bad at programming?

You go significantly deeper? You're aware that shortly past 50mt the partial pressure of oxygen causes neurotoxicity, leading to convulsions and death? Anything happening past those depths involves extremely expensive hardware, custom trimix breathing gases and most importantly a hell of a good reason to even bother doing it - that is NOT the environment for an Apple Watch.

> Or licensing costs.

Which don't exist. BMW famously tried to blame Apple for trying to charge buyers an annual fee for Carplay and Apple _publicly_ called their shit out. They confirmed they don't charge anything - neither one-off nor recurring. Manufacturers may incur certification costs, there's nothing after that.

Just did a couple quick searches and nothing's changed since then AFAIK

Shaka, when the walls fell.

I hate the banners and the ugliness too but they are designed precisely to do that, and adtech maneuvers to ensure the hate is directed at the wrong source - the lawmakers instead of the people doing all the spying.

GDPR 1.1 should address all that - no legitimate interest exclusion of any kind, ONE SINGLE CLICK to reject all, no witholding service at ANY degree unless consent is granted, a 3rd option (I offer to pay to not be tracked), and a mandatory disclaimer on the cookie banner saying in clear terms: "Tracking is spying. If we were not tracking you and invading your privacy, this banner would not be necessary at all". Maybe even revive the "do not track" header by mandating that websites react to it accordingly, obey it 100%, not even show a banner if the header already tells them what to do, and ask users if they want this set or not, without a default value which would give an excuse for complaint from the people spying on you.

This is a fantastic idea and I'd like to add my voice to the choir asking you for more information :) Don't let us rebuild everything from scratch if we don't have to

> What's next?

Killer robots which can't be outrun and which can give chase through urban landscapes

> What a time to be alive

Indeed.

Yes, and if you have the PCIe lanes (say, an x16 lane - actually delivering 16 lanes! - to each GPU) it's also quite performant - it's called a tensor split in llama-server.

If your motherboard/cpu doesn't actually have those (few do outside some xeons, epycs and threadrippers) you can still do it - it's called a layer split and will work even with 1 lane per GPU. Each GPU will work at its maximum speed, but only 1 will be active at any given instant - imagine a relay race.

(Didn't mention which PCIe generation - obviously the higher the better. At v4 and up, even 8 lanes per GPU would be enough for a performant tensor 4-way split)

Edit: If you have more than 1 user at a time, the GPU can actually all be working all the time, if there are enough parallel requests to serve. But you need enough KV cache for all the sessions you're running in parallel.

The reason for running those insanely low quants is to fit in extremely limited memory budgets. The first thing you sacrifice is speed, then context and accuracy (up to you in which order). IQ2_XXS and below is desperate/proof of concept territory. If you have a spare half gig for the MTP drafter, run a larger quant instead, it will be less incoherent, and damn the speed, it won't be garbage at least. Only around Q4 I'd allocate the comparative luxury of more memory for a speed increase. At least on a dense model. MTP makes a lot more sense (but helps statistically a bit less) on an MoE.

Qwaiting for that 3.8-35B-A3B

Let's try a thought experiment. It's not subsidized. It's not "worth" the tens of times it costs if accessed through API. Let's drop the subsidy word, they're selling you a product, and they're trying to sell other people a similar product at 30x the price with some excuses. If either (or both) turns out to be unsustainable, they haven't been subsidizing you, they just had a crap business model (or a perfectly good one, if the goal was an inflated valuation, an IPO and then a crash once the losses are socialized).