revolvingthrow

The AI-isms in this post are a bit much.

Anyway. Routers are such juicy targets, especially since there's less eyes on them than on laptops/desktops, that it's hard to believe there are any which aren't backdoored, be it by China, US, US' middle eastern sidekick, whatever. My question would be: how limited are they to software, like in the article?

You can flash with OpenWRT and Fritz, but to my knowledge it doesn't replace the boot parts. A suspicious bootloader in SPI flash probably isn't difficult to produce, even if it would likely be discovered... eventually. A separate modem / radio processor would be an even bigger nightmare, esp if it's an 5g router - you essentially add a second computer to your router, with its own components and network access, and I think you can't even play around LTE/5g networking as freely as with wifi without bending some laws.

Theoretically you could watch the router from, well, outside the router. But as the article says, it can be difficult to say what's an actual connection and what's uploading your network's data to an uninvited third party. I'm also wholly ignorant what implications it has for things like Wireguard and Tailscale. I assume you're completely pwned if the Wireguard runs on the router, but I don't think it does anything if it runs on your laptop or whatever? Then again, there's a lot more devices in peoples' homes nowadays, and almost all the IoT has pretty much zero security.

I suppose AI will help with some of those things. Tracking every connection made is very tedious for a human, but throw some compute at it and you'll get a much better insight, especially since you'll probably know what to expect out of your machines.

I guess using a mini PC as a router-slash-proxmox-hub would be a safer choice? There's probably many fingers in this pie as well, but my gut feeling is that it's a lot safer than some cheapo box provided by your ISP at the cheapest price they could source them.

> 320B total parameters and just 18B active parameters

This is pretty hefty for a "flash" model, even a 256 GB setup is insufficient at q4 - and q4 is already the worst-but-still-acceptable quant in my experience. The benchmarks look great, especially since GLM tends to be more honest than the average Chinese lab, but you’ll need to splurge to run it at home.

@edit: so many releases that I forgot to math. This fits just fine in q4, realistically the minimal hardware would be 192gb - so blazing fast on double rtx 6000 pro and usable on 256gb unified memory. You could even go with 5bit quant on 256gb.

… you’ll still need to splurge, though.

People are annoyed at apple’s recent lack of polish in software, but the increasing monetization and push into services are a bigger issue imo. You pay the apple premium for hardware and the hardware is indeed great, but you’re increasingly treated like a piggy bank. Line must go up.

In regard to Maps: if all you want is a gps, there’s many apps that use OpenStreetMap data. I use CoMaps myself. You download the map(s) you want and it just works, no forced updates or ads or anything in the way of upselling.

>And what happens when the price goes up by an extra zero on the end?

>And then another one?

Well so far the price trends towards going down by the zero at the end. Inference isn’t expensive, there’s a whole laundry list of providers to choose from. It’s training (for everybody) and salaries (for US labs) that are expensive.

>and you have to switch to using local models?

Many companies are interested in that, if for no other reason than to avoid uploading secret sauce to we-totally-dont-train-on-your-chats-pinky-promise API endpoints. A mid sized company can cough up $500k for dxg b200 which will be useful for years and years and the only ongoing cost is electricity. Put glm 5.3 on it and it’ll zip along for a number of employees at the same time.

>And then, suddenly, those local models cost an absolute bomb to license, because the companies that bought up the wreckage of the original model creators are trying to recoup their spend?

How are they recouping anything on a mit licensed model released into the wild? There are plenty of models released by "who?" companies and the cost to download those is zero.

>Meanwhile the company is being hit by nuisance litigation from people whose code the bots plagiarised

Any day now

The doom scenario makes no sense in the race to the bottom that’s happening now. China being the other big entity that’s invested in AI makes all the what-ifs about litigation and plagiarism moot; they will gladly offer everybody inference for dirt cheap, if for no other reason that they’d like it if you gifted them your IP that they will then use to outcompete you with their manufacturing prowess.

The genie’s out of the bag. AI is never going away, and we all need to adapt whether we like it or not.

~75% higher price for some extra GPU power and small CPU boost compared to M4? Eh. It's pretty far from M4's value, which was a no-brainer recommendation for people who wanted a tiny box sitting somewhere on their desk.

> OpenAI could still have a significant moat. ChatGPT occupies most consumers’ minds when they think about AI and has become a household name.

ChatGPT is AI for the average non-techie the world over, but the average non-techie isn't eager to pay for it. The more progress that's made, the less incentive to pay - most people are happy with the total garbage spewed by google AI overview. They'd be happy with google's 30b MoE gemma, whose performance will likely be squeezed down to something that can run on a phone in 2-3 years. Why would they pay $20 a month?

It's why OpenAI is pushing a variety of things such as ads and offer a more polished ui/ux than the competition, I think. The models are already good enough for people who just want to know how much sugar to add to their cake or when's the next basketball match their team plays - it's OpenAI's game to lose those people, by annoying UX and whatnot. If they can make a few bucks off of every one of their non-paying users it'll stretch their runway immensely. Those users will never go to Antrophic or some cheap Chinese model, but they might defect to Google because a popup on Android / in Chrome told them to.

The bitrot would be excessive

There isn't much to upgrade to if all you want is a desktop for gaming, a non-workstation work box or just a decent laptop. Much like smartphones, there just hasn't been much innovation of late. A midrange gaming desktop from mid 2023, something like rtx 4070 and ryzen 7800x3d, can still play pretty much any video game on max settings and high framerate at 1440p, and has sufficient performance for most workstation-esque tasks as long as you have enough ram. Similarly, the only thing my spare macbook m2 lacks compared to m5 is hardware av1 decoding - the performance is worse, but I don't really notice it in normal usage, and macOS 27 runs perfectly well on it.

Local AI will crush any machine, of course, and if you need a 512gb ram box you'll probably need to sell both kidneys, but you're well past the average consumer territory with those.

So, fingers crossed nothing breaks down and that some of the funny money goes into R&D which will eventually trickle down to consumer devices in a few years (16TB gum stick SSDs at sane prices pretty please) instead of endless datacenter GPUs and SSDs that will be tossed into a shredder in x years.

This really doesn’t match my experience, as Gemini was bafflingly bad at any online [re]research when I tried it 2-3 months ago. Not that google’s ai models seem particularly good at anything of late, but given the fact that google is still the king of the hill in search (even if they mostly serve you garbage, they still crawl everything) Gemini being very noticeably worse at searching things online than claude or chatgpt was a read head scratcher. I wonder how our usage patterns differ.

I have several 1m long TB4 cables (100W, video, 40Gbps) and they’re neither particularly thick nor were they anywhere close to $60.

The best you can get is TB5 (240W, up to 120 Gbps data transfer) and at 1m long they’re $32 or so from CableMatters. Those are admittedly somewhat thicker, but not stiff.

I wish there was a higher baseline data transfer mandated for usb-c. Sure, 40gbps isn’t needed for all cables, but is some piddly 5gbps too much to ask? We had usb 2.0 speeds back when CPUs just barely entered 1 GHz range, it’s practically ancient technology. Admittedly most people never transfer any data over usb, but with this logic we might as well have usb-c that does charging only. Being able to pick any cable and have 5 gbps baseline means that any usb-c cable is fundamentally usable, if slow. 480 mbps is… ugh.

I don’t have an issue with 60W being the baseline, it’s enough for most laptops and the average GaN charger is 65 W (or less) anyway.

> a new interface paradigm that's balanced between agentic use and human review (with light editing)

Is reviewing endless reams of AI slop something that's doable? It's a very tiring task when you're looking through code that exhibits most of the usual good quality signs -- except it's just a facade that's likely full of subtle mistakes and you barely know the codebase in the first place. I don't think most devs would be particularly happy about having to do it day in and day out, and you definitely can't outsource such work to some bargain bin shop full of clueless people.

Steering an AI, that's doable, especially at a high level. You probably have a rough idea of how you'd approach a given problem, and a wild mismatch between what you'd expect and what's happening is easy to detect. I'm not sure how a human gets into the loop with "light editing" without just telling the clanker to rework X because flamegraph looks unhappy about this part. Any alternatives would probably slow the process down significantly, at which point you might as well write it yourself.

Ah, the newly popular "let's make the policy unusually broad and treat all foreign production as presumptively dangerous, regardless of country, manufacturer or security practices, this way we're more transparent about prioritizing a bribe rather than our stated concerns".

> Advanced robotic devices are inherently networked systems which creates broad attack surfaces and leaves them vulnerable to data exfiltration, remote disruption of the physical robot, and dependencies on unsecure over the air updates.

Good thing this isn't a concern in ones produced locally.

While I am grateful for open weights models I never found much use of them in the past, barring those I could run myself. This changed with deepseek 4 - it is staggeringly cheap, even if the performance definitely isn't near sota and it's not particularly fast either.

When I expect to need a lot of tokens and the task isn't too difficult I use sota to plan and create a thorough set of instructions and let deepseek chip away at it. With thorough instructions the quality tends to be satisfactory, and you pay something silly like $15 for 600m tokens.

GLM 5.2 seems like a decent price/perf and Kimi 3 has some real nice performance for an open weights model, but gpt 5.6 is unexpectedly affordable (especially if you don't automatically use Sol at max) so I don't think either is worth it atm. The exception is when you're working on something that US models get cold feet about, which seems like a constantly growing list. For me Fable is already too much of a headache in this regard, but chatgpt is still okay-ish. Hopefully it'll last. If not, there's Kimi.

tldr SOTA for most things because gpt 5.6 is token efficient. If I expect to burn a lot of tokens I use deepseek 4.

So far the first and only sophisticated cyberattack was carried out by openai rather than by one the subversive commie models that hate us for our freedoms, so forgive me for taking their safety and alignment spiel somewhat less seriously than they’d like.

What Apple wants out of Google is Siri that runs at 8gb ram and isn’t a horrible embarrassment that feels like a primitive markov chain. Given how good Gemma 4 is, Google can squeeze some serious performance in small models. Whether they can make bleeding edge models is irrelevant to Apple.

Tons of guardrails, lazy model, super confusing plans, expensive 3.5/3.6 flash and lite and 3.5 pro MiA?

Rough patch for google ai

Article feels a bit shallow. Obviously carriers will do their usual fuckery, anyone who ever dealt with one knows this. What many people probably wouldn’t expect is when you’re in a country where physical sim are dominant, you the dumb tourist will pay 10x for the privilege of esim since you’re prime target to rip off and what are you going to do about it, buy a new phone? Pay 1000x the price for roaming? Feels like notable omission and is fairly common.

I’m also surprised about there being not a word about smartwatches. I got a cellular apple watch with the idea I could leave my phone at home when doing sportsy stuff like swimming, assuming that a watch with esim support will indeed work with esim. In reality most don’t, Apple is particularly restrictive about it. You’ll probably need some shitty "companion / wearable plan" that most networks don’t even offer. There’s also extra regional compatibility on top of it to deal woth. In the end I never found a deal worth using.

That being said, I settled on esim for my main number and second esim or physical sim as needed when traveling. Having the sim slot free is neat. The biggest downside is "phone gets wrecked and you’re on another continent -> need to authenticate to switch esim to different device -> you need previous number to authenticate -> good luck", but I’m not sure wrecking a phone is more common than having it stolen so it’s probably a wash.

While a valid point, China also produces plenty of whitepapers going about the architecture and know how about the training and inference itself.

There’s also the fact that unless LLMs do get to AGI (which seems… doubtful, still) there comes a point where a model is good enough for what you need. Fable and gpt 5.6 are certainly pretty neat, but I’ve been happy since opus 4.6. I’d still choose a better model, obviously, but it’s not the end of the world if I was stuck with 4.6 for a while when it already lets me get the end result at acceptable quality.

It also needs to be said that the "erosion of training capability in other countries" is largely theoretical, given that Mistral hasn’t been keeping up and other countries don’t even have anything worth mentioning. You’d first need to _have_ training capability to lose it.

The few tests I ran were by no means comprehensive, but while kimi felt like the real deal qwen seems a bit of a benchmark princess.

According to artificialanalysis, cost per task is $0.94, which is almost the same as $1.04 of gpt 5.6 sol max (fable is most expensive by far, at $2.75). Things like glm 5.2 max cost roughly half that. The model certainly sounds extremely impressive for something not from openai/antrophic, but the price makes it a mediocre product.

Instruction following seems lower than I’d like, too. OTOH scores on agentic stuff seem high, which… feels a bit contradictory? I thought decent instruction following is step 1 of solid agentic workflow.

The benchmarks look nothing short of incredible. Assuming it’s not benchmaxxed to hell and back it’s just a notch below gpt 5.6, which came out what, a week ago? If the performance claims hold up the delayed Gemini 3.5 pro will likely end up not only behind fable, but also behind 5.6 and a (supposed) open weights model. Google might have to do some real soul-searching.

Not sure I understand the argument.

Obviously there are people who help themselves to others' money if given the chance no matter the circumstances. But if the circumstances change so that people DO start going hungry or homeless, which is a rather obvious side effect of AI-but-not-AGI maximalism brightly espoused by our overlords sama and amodei of the "I can’t wait to make half the knowledge workers worldwide obsolete" variety, the scale of the problem will obviously get worse, as well as the type of people you can get involved if you’re in the international scam market.

The problem described in the article is unsolvable, given that a mid-range desktop from a few years ago can easily clone a voice that's convincing enough and there are no guardrails to those. Some silly KYC laws might limit a highschool kid making deepfakes of his crush, but once a model exists it's trivial to spread it around, and for organized groups to get ahold of those. Similar will happen with images, it's just that nobody with any serious money bothered releasing image gen models that compete with gemini or chatgpt -- but it's just a question of time. A year or three, what difference does it really make?

As the cost goes down to near-zero you can scale it up almost infinitely, especially if the profits are high enough to get some smart people working on the problem, which going by the article is already the case ("INTERPOL's finding that AI-enhanced fraud is four and a half times more profitable than the traditional kind"; incidents rose by 26% last year). If AI does succeed on mutilating white collar work enough there will be a large supply of knowledge workers that might just join International Scam Co. rather than have their families go homeless. Drowning man clutching at straw and all.

So if technologically it's impossible to prevent and societally it's impossible to prevent (like the attorney that got pwned same as the grandma), I'm not sure if there exists an answer that isn't worse than the thing it's supposed to prevent. I suppose we'll soon be in a situation where nothing we don't directly perceive in real life is provably true. That journalism and media in general seem to be in a deep crisis of trustworthiness means that you won't even get the benefit of the chain-of-trust as a proxy for whether something is or isn't real.

Ignoring everything happening outside of your immediate surroundings is a choice, and probably even good for people's mental health, but my gut feeling is that it does make humanity as a whole dumber and disempowered. What does corruption matter if nobody cares, or even hears about it? It was AI generated by $current_enemy anyway; nothing to see here, citizen.

Proton is already good enough that you don’t need SteamOS. I’m not that much of a gamer so my numbers are hardly representative, but about 90% of video games I tried ran with zero issues through Heroic Launcher and GE-Proton. The ability to install SteamOS on your own hardware is nice, but largely irrelevant. It was Valve’s previous work on Proton which was important.

The bigger fear is of course Valve itself. They have many, many levers they could turn to squeeze their audience. This is of course a terrible long-term decision—Valve being a good steward is why PC gamers largely don’t even bother to learn what other online storefronts offer—but long-term planning is irrelevant if you’re a CEO who’s looking for a new yacht and to lateral away after 3 years of record profits.

@edit: admittedly there is something to be said about the "I don’t _want_ a PC experience, I just want to play the games on my couch" crowd. Being able to set up SteamOS on your own hardware _is_ nice, I just don’t think it’s relevant to any post-Gaben enshittification.

Given the nightmarish nimby gridlock I’m less and less convinced it’s a good thing. I’d rather have people mad about windmills being eyesores than be perpetually chained to oil and gas for energy, as an example. I’m also not a fan of endless roadblocks to all manner of construction, even for such simple things as housing.

Yes, having a data center that raises your utilities costs by 300% jammed down your throat because the local mayor got blatantly bribed shouldn’t be a thing, especially when it’s powered by mobile gas turbines that stink up the entire area (note: I’m not against data centers on principle, but there are many ways for ultra-wealthy interests to leave people hosed). But things like faintly visible mini-sats don’t seem like a big deal, subjectively, unless you work at an observatory.

> I don't think many consumers (outside of hardcore games) could tell the difference between the graphics of a game from 10 years ago to the graphics from a game of today

My half blind aunt could probably tell the difference between the graphics of a game from 2016 and 2026 if you put them side by side.

Were video game graphics "good enough" for a while now? I’d say yes, with the exception of vr. But to say there has been no noticeable improvement over the last 10 years is silly.

GTA 6 is coming out soon. I’d invite you to actually compare the visuals of gta 6 vs the original gta 5 from 2014 that was released for ps4 (rather than the 2022 enhanced version for consoles or 2015 version for pc which shouldn’t be compared to non-pc gta 6, since 6 for pc will also get a significant facelift).

Benchmarks look really promising. Suspiciously good, even. I guess we’ll see soon enough.

My question to previewers: how are the guardrails for random joe that wasn’t personally blessed by the ai pope to access the non-nerfed model? Fable is a nightmare in this regard, but I’m not sure whether 5.6 also gets a critical side-eye from the gubmint when you ask it to fix bugs in your code (you filthy hacker, you).

> when it was really just Google falling behind the same way it did with Tensorflow, Angular and GCP

Not sure I agree. Angular fell behind in popularity but was (is? unsure atm) still eminently usable. I gave gemini a test drive recently and it was horrendous, as in "picking dirt cheap Chinese model over gemini any day" bad, and with overzealous guardrails to boot. 3.1 pro feels a year behind and is extremely lazy. 3.5 flash feels like a model you’d run on your 128gb macbook, not something that was released a month ago and which costs a fair bit when used through api.

In any case: as of right now I think that we went from a three horse race to anthropic / openai as premium choices vs whatever is the Chinese fotm for a fraction of the cost. 3.5 pro better be a miracle if google wants to hang out with the big boys, otherwise their only strategy is hoping that both US labs go broke and they remain the last man standing.

While I have no experience with it personally (no interest in image gen) my aunt was raving about current chatgpt image model for "restoring" / working with old photos - sharpening, changing some small details like ill-fitting background. It takes her a bunch of prompting but eventually she gets things just right. In comparison, current gemini output (supposedly) tends to be subtly off, details aren’t quite right, proportions are subtly changed etc.

This is purely about generating images with people in them, I don’t think she’s doing any logic puzzles with gotchas and specific alignments of differently colored blocks and whatnot

Because the market pays less for DDR4 than for HBM (or DDR5), and since HBM is heavily modified, vertically stacked DRAM, it competes for the same raw inputs and fab space than DDR4 used.

If I can produce DDR4 for modest profit or HBM for a lot more profit I will obviously produce HBM. And given physical realities producing HBM takes from existing DDR4 production capacity. Worse still, it takes roughly 3GB of ram to produce 1GB of hbm iirc.

> sunlight causes most of the heat issues; cloudy days are unlikely to be extremely hot

> solar panels convert sunlight to electricity

> AC converts electricity to cooler air indoors

Ah, if only there was a way to solve those three problems at once. Alas...

As to the "resistance" to AC: is this an actual thing or just something the media made up? It seems that for the most part anyone who wants AC can just buy it, barring exceptions like historical buildings. If people want to cook themselves alive I’m not really seeing the issue, as they only harm themselves (unlike, say, with vaccines where herd immunity is important).