revolvingthrow

When the highest offices of the land are packed to the gills with liars and grifters and anyone with a brain can observe there’s no downside to such behavior, "just be honest" rings rather hollow

This is a real head scratcher. Unless this is a very short term action it seems to have only downsides for everybody:

- people pay much more for US models than Chinese models because right now they're the best. Once they're no longer the best (since you don't get access to them) why would anyone pay several times as much for the same result?

- once you get a high amount of tokens flowing into China instead of US companies, they will train on those chats and their rate of improvement will only accelerate, making US models even less attractive over time

- the sky-high IPO are dead in the water, since their story of "we will replace a good chunk of all knowledge work in the world, capturing a few % of total global spend relating to it" turns into "we will make a bunch of money out of a few dozen S&P 500 paying for the best, and some pocket money out of whoever uses our overpriced models that are as good as Chinese models" - far less money overall. Losing access to untold billions of investor money certainly won't improve performance for the US labs

- all the non-US people start asking themselves why they're funneling money to US corporations who barely share any of the secret sauce compared to Chinese corporations who share plenty when it comes to LLM, including the models themselves (at least for now)

- Chinese models have significantly less guardrails, making for better end-user experience

- there is a small but non-zero chance Euros get off their asses and invest into AI, making something halfway decent and further fracturing the market which cuts into US profits

So what's the benefit here? I thought the Mythos situation was the current admin taking revenge on Antrophic for not kissing the ring, or simply looking for a bribe, but no matter which way I look at it it's a self-own. The only way this would make any sense is if AGI is imminent, which I don't think even the boosters are arguing at this point.

Theoretically US could outlaw Chinese models, but I'm not sure what it's supposed to accomplish as the rest of the world certainly won't, especially as long as they release open weights models that you can run without phoning home.

>what's the timeline of the RAM shortage ending?

Barring unusual market forces like Taiwan invasion the timeline to ending the acute shortage seems to be mid 2028. The AI still has plenty of money to burn and is the biggest driver, but we’re also shortly before gaming consoles ought to release a new gen (although who knows whether they won’t get delayed for a while). There was even going to be a small upgrade cycle for nerds waiting for 2nm fabbed devices, same as pre-ai datacenters looking for power efficiency. Plenty of pent-up demand, too, as many people simply make do with what they have but will upgrade once the silliness stops.

If you’re looking for ssd/ram prices to go back to the low of 2024/early 2025 it probably won’t happen before China catches up, which will be a while yet. There is some build up of new capcity happening from current manufacturers but it’s significantly less than what the demand increased by.

Oof, that’s a ~20% increase across the entire lineup. Ram and storage are particularly expensive, as can be expected: mbp m5 pro $1700 -> $2000, m3 ultra $4000 -> $5300. To be expected, there’s only so much margin apple is willing to lose and everybody else already increased prices.

I’m surprised that iphones didn’t get a price raise while neo did. Neo seems like a clear market share attempt so that they can upsell on services, I would’ve expected either both of those or neither to get dinged.

People using google’s models: am I holding it wrong or are the guardrails really overtuned?

I had the dubious pleasure of testing gemini of late and I kept running into refusals. How do I transfer a sim number from one provider to another? No. What should I consider when making backups on ntfs less prone to data loss and more bitrot resistant? No. Evaluate this piece of code? No.

I’m not sure if it’s cold feet from the mythos situation or what, but it reminds me of the dark days where you couldn’t use ai for much of anything. But then I go to chatgpt 5.5 and it does mostly everything I want outside of the usual cybersecurity boogeyman that you run into now and then.

The funniest thing about this post is not the fact that some people took it as anything but satire, but that it’s likely very close to what the true believers at Antrophic actually think.

Ah, those wacky terrorists and their non-aligned models, trained on copyrighted data to boot. Remember, the only thing that stops a guy with an evil god-in-a-box is a guy with a benevolent god-in-a-box, and only Antrophic can lead us to the second one – but only if we act together as a nation and ban those subversive open weights models!

I subscribed (with some distaste) to the $20 tier to check out fable. I got two refusals in a row on innocuous tasks, then ran out of quota halfway through the third one.

Truly, the future is here.

After saying for weeks of how Mythos is in a league all of its own you’d think it was a bit more than the usual iterative few % on the benchmarks (and even more guardrails as a bonus).

IPO gonna IPO, I suppose.

Why learn to read and write when everyone carries a computer capable of TTS? Why learn anything when your pocket computer has access to AI doing the thinking much better than the average highschooler and has 100x the knowledge?

I’m not meeting with friends to work on our speeches, I’m meeting friends to do friends stuff. Go join toastmasters if you want to do work stuff as a fun pasttime.

Amusing that just when the big three AI providers from US raise prices significantly, even for the mini models, you’ve got a Chinese model slashing their already-cheap offer by 75%. Not to mention you can run this model on your own hardware, although admittedly even the flash stretches the meaning of local for individual people.

Semi-related: has the rate of published exploits picked up as if late, or is it simply the fact that there’s hype around ai as security tool (offense or defense) so it’s simply in the news more often?

Feels like there’s something new every other day - linux, windows, mobile, various commonplace tools used by everybody, the list goes on

A local Answer Machine is the dream, especially when the internet is decaying and generally on its last legs, but the hardware requirements seem like a huge mountain to climb. Things are progressing tremendously - deepseek v4 flash is very good for what it is - but even that goes beyond any reasonable local setup, which imo is 128 GB ram + 16 GB vram. 4 ram slots on a consumer board craters ram speed, 256 gb macs are too expensive, and even then the inference is ungodly slow.

On the other hand… v4 flash model is actual magic compared to what was available 2 years ago. If the rate of improvement stays as is, we’ll get a similar performance in a ~120B model in a year, which is viable (if expensive) for everyman hardware. Possibly you’ll be able to run its equivalent on a ~$1200 laptop by 2028, which for me-in-2020 would sound straight out of a scifi movie. A good harness that lets the model fetch data from other sources like a local wikipedia copy from kiwix could do a lot for factual knowledge, too; there’s only so much you can encode in the model itself, but even a cheapish (pre-curent prices) 2TB drive can hold an immense amount of LLM-accessible data.

Big caveat: I don’t see local models for programming or generally demanding agentic tasks being worth it anytime soon. You likely want bleeding edge models for it, and speed is far more important. Chat at 20tok/s is fine; working on even a small codebase at 20tok/s, especially on a noticeably weaker model, is just a waste of time. Maybe it’s a PEBKAC but I have no idea how people make any meaningful use out of qwen 3.6.

Is it possible to dual-boot on android? It sounds defeatist but I no longer believe it’s possible to change course - the increasingly authoritarian governments, google and most moneyed interests are all on the same side, so it’s just a matter of when.

Being on the palantir-approved google ranch for the few Apps You Need + graphene (or some other alt OS) for everything else would be quite inconvenient, but still better than carrying two phones, which nobody wants to do.

Regional pricing makes sense for products that don’t have ongoing costs or where most of the input cost can be offset by local labor. You’re not buying server racks nor electricity at 1/3 of the price to serve poorer markets

Given the topic and the fact llm providers charge global rates, the absolute take-home money is much more relevant. Even if you live like a king on $1000/mo, 5.5 pro is still $200.

Boomers love slop. Even when they know it’s ai (and it can be increasingly difficult to tell even for people who don’t struggle to send an email) they love it almost as much as they love political ragebait, and they love political ragebait more than their own families. My ancient grandpa is on some facebook groups and they share bottom of the barrel ai videos and images all day. If it can be consumed with zero effort, it’s great. They couldn’t care less about whether any care of effort was involved in its creation, whether it has any value whatsoever, whether it’s made up bullshit. Not a whit.

They also have money and can vote, so there will be an endless avalanche of slop being generated every single day, enough to bury organic content ten times over.

Damn, those li-ion sure seem like a bad long-term solution. Very convenient that you use lifepo4 for at-home battery storage, and either lifepo4 or possibly sodium for grid scale.

Inverters aren’t a problem. China produces roughly half of them worldwide iirc. They’re dominant but you can source from elsewhere without an issue.

LiFePO4 is almost purely China, but those will last you 20 years, which is roughly 365 times as long as if you’re cut off from oil.

If openai / anthropic / google were the only game in town then yea, we’d already be paying 5x as much as we do. But local models are so close to sota that it just isn’t going to happen. If I’m a lawyer getting billed $500k/yr on $600k profit I’d rather buy a chonky server and run a model that’s 90% as good and get my money back in 2 years, then pay $5k electricity on $600k profit.

Nobody will successfully lobby for banning local models either, it just isn’t going to happen when the rest of the world will happily avoid paying 80% of their profits to some US bigco for the privilege of existing.

Eh. Those results would be noteworthy if it was a a MoE. A 120B dense? Firmly in meh territory.

I used (and will continue to use) most of those. Quick rules of thumb:

- markdown is .txt with just a tiny bit of syntactic sugar/syntax highlighting, and you can export it to pdf or html

- quarto is markdown-but-I-want-to-execute-code-blocks-inside

- typst is latex but modern, with 90% less cruft and 10% less functionality (academia, hating everything modern, will also hate you if you use typst)

- pandoc is how you export to pdf/html/whatever

By and large, it’s obvious which tool is needed when. There’s of course more, like asciidoc, but I struggle to think what isn’t being covered by the markdown/quarto/typst combo. Some wysiwyg editor maybe?

> pricing "Pro" $3.48 / 1M output tokens vs $4.40

I’d like somebody to explain to me how the endless comments of "bleeding edge labs are subsidizing the inference at an insane rate" make sense in light of a humongous model like v4 pro being $4 per 1M. I’d bet even the subscriptions are profitable, much less the API prices.

edit: $1.74/M input $3.48/M output on OpenRouter

For flash? 4 bit quant, 2x 96GB gpu (fast and expensive) or 1x 96GB gpu + 128GB ram (still expensive but probably usable, if you’re patient).

A mac with 256 GB memory would run it but be very slow, and so would be a 256GB ram + cheapo GPU desktop, unless you leave it running overnight.

The big model? Forget it, not this decade. You can theoretically load from SSD but waiting for the reply will be a religious experience.

Realistically the biggest models you can run on local-as-in-worth-buying-as-a-person hardware are between 120B and 200B, depending on how far you’re willing to go on quantization. Even this is fairly expensive, and that’s before RAM went to the moon.

Anthropic’s $20 plan gives you such a pittance of tokens that it’s borderline unusable for anything more than a few scripts or a toy app. If $20 is all you have you’d do _much_ better going with chatgpt

I really wish the benchmarks were even slightly trustworthy for AI models. ~120B are the largest models I can run locally. Naturally I grabbed the 122B Qwen3.5, which had great benchmarks and… frankly, the model is garbage, worse than glm air 4.5 IMO. But then, qwen famously benchmaxxes.

And here we have another release. The benchmarks are just a tiny bit worse than qwen3.5 (for far less tokens). Am I to take it that the model is worse? Or does qwen’s benchmaxxing mean that slightly worse result of non-qwen models means a better model? I’d rather not spend hours testing things myself for every noteworthy release.

Ah well. Mistral has been fairly decent so worth taking a look. Obviously they’re behind the big 3, but in my experience their small models are probably the best you can get for several months after each release. I’m not sure how it works as a sales funnel for their paid models, same as with chinese models - people likely just go for google/openai/anthropic in this case - but I’m thankful for their existence.

Eh, I found several interesting things from various tracking tools. Take a nap? Sleep is destroyed this night. Exercise in the evening? Same. Not something I’d pay attention to without noticing the chart afterwards.

There’s also the motivation factor. I’m not sure of the total %, but I certainly did some exercising just to fill the daily goal. Nothing life-changing, but for the price of a cheapo apple watch se once every 5 years or so, more than worth it.

It’s not unlike simplistic time tracking on my iphone. I spent a lot of time on bullshit websites. Obviously I knew it was happening, but the sheer magnitude was surprising. It’s akin to acute pain letting you know there’s a health problem vs something brewing in the background that you are vaguely aware of, but have no motivation to truly care about - one is far more noticeable than the other

It is difficult to put into words how much I dislike macos 26. I held out on upgrading for a long time since there were so many horror stories, but to my surprise both iOS and ipadOS 26 aren’t really any different than 18. Maybe because you don’t really do any proper work on it? The graphical differences aren’t anything major when the apps fill the whole viewport anyway.

But macOS? Good lord. I can only hope 27 will unfuck things somewhat, there are so many small annoyances and all of them add to a constant sense of unhappiness throughout the day. I’m really tempted to downgrade back to Sequoia. At least the M4 will be good enough for years if this truly is the new path Apple will take.

I usually end my rides at 24-28 kmh average and outside of cyclists and e-bikes the amount of people that ride faster than me is pretty much zero. 70 is completely absurd, especially for "youth".

I don’t really care if a bunch of reckless kids want to gamble with their lives, but the place to do it is clearly on the street, not on the sidewalk. Those ebikes are pretty much dirt bikes, and nobody sane argues those should slalom between pedestrians, just like cars or motorcycles don’t drive on sidewalks. I don’t want to constantly be on the lookout for a 60 kmh vehicle careening right into me whenever I’m outside, which increasingly happens with food delivery "bikes" as well. There’s no place for them on crowded beaches either.

Wanna go as fast as cars? Cool, do it on the streets, it’s what they’re built for. Helmet (or even clothing) optional, I suppose, it really isn’t my problem.

It doesn’t. I’m not sure it outperforms chatgpt 3

I don't understand the target audience of ipad air.

The base ipad is "really big iphone, with a few laptop-esque features". It's reasonably cheap for what it offers, especially if you want a highly mobile media consumption device and handwritten input.

Then there's ipad pro, which is wildly overpriced for its specs -- m4 pro has half!! the ram that the cheaper m4 macbook air has, which is laughable for a 'pro' anything, especially if you have apple intelligence enabled - you get what, 3GB of usable ram once you take OS and apple intelligence into account? Yet, aside from the crazy sticker price, the hardware is a lot better - the 120 Hz OLED display looks amazing and is way brighter, the speakers are quite an upgrage, full blown thunderbolt port for external display and so on. The OS is still toy-like, and ram is pitiful, but there is place for an ipad pro.

And then there's air which is... base ipad with an M-series chip and pretty much nothing else? The display is barely any better than base ipad, the storage and ram are pitiful, the speakers are from the baseline ipad and so on. Just about the only saving grace of the M4 one announced here is 12GB ram, which is the absolute lowest those really ought to have, and really puts into perspective how utterly miserly Apple was about ram pre-AI. I don't understand the value proposition - you want the baseline you buy a much cheaper base model, you want more you get the pro, right?

To be fair the asking price is far less than pro but the upgrades over base model seem so minuscule that I just don't know.