You can try the latest GLM 4.6 https://z.ai/ . Their coding plan is $6 a month and performs on par to Sonnet 4 for my personal task. Sonnet 4.5 still has an edge though. All of ZLM’s models are also open sourced so you can run it locally if you want
I am mostly retired but I am thinking of restarting a solo products mini-company next year. I have been looking at much less expensive options like Alibaba Cloud, GLM, Kimi K2, etc. There is a recent Stanford study showing most US startups are using less expensive Chinese models, but I think usually hosted in the US.
For now I am happy enough with Gemini and GPT-5 because my usage is so lite that anything is cheap. For many engineering use cases, Gemini-2.5-flash-lite works well enough.
How do you use GLM? With codex —oss? Or, just ‘raw’ with no agent-wrapping coding environment?
I use it directly with Claude code [1]. Honestly, it just makes sense IMO to host your own model when you have your own company. You can try something like openrouter for now and then setup your own hardware. Since most of these models are MoE, you dont have to load everything in VRAM. A mixture of a 5090 + EPYC CPU + 256GB of DDR5 RAM can go a very long way. You can unload most of the expert layers onto CPU and leave the rest on GPU. As usual Unsloth has a great page about it [2]
Idk if this is the reference but it’s in the same direction:
„ These days, when entrepreneurs pitch at Andreessen Horowitz (a16z), a major Silicon Valley venture-capital firm, there’s a high chance their startups are running on Chinese models.
“I’d say there’s an 80% chance they’re using a Chinese open-source model,” notes Martin Casado, a partner at a16z.“ —- https://ixbroker.com/blog/china-is-quietly-overtaking-americ...
I recently got an “older” generation robot vacuum for this reason. I wish the dreame debug board was more accessible. The designer and creator purposely made it this way so people would learn how to solder. I could solder but I no longer have the time or patience to source my own PCB and parts. Thankfully, someone sells one premade on Tindie.
The overall "I want you to have a project, not a product" vibe put me off using this when deciding what robot to buy. I want a clean floor, not another hobby that turns into a chore at the least opportune moment.
Luckily, someone explained to me that in practice once you've set up, it usually just works, and it runs completely on the robot (i.e. no second device/server/homelab that you'd have to maintain) and since updates are optional, you shouldn't be required to deal with it unless you want to.
(The offputting statements are at https://valetudo.cloud/pages/general/why-not-valetudo.html: "it is very much not [a product]. [...] Instead, it is highly idealistic, anti-consumerism, anti-hypergrowth and anti just-continuing-what-we-do-now. [...] these aspects are baked into its design. There is no way of using it without being constantly confronted with them. If you’re not willing to reflect, introspect, grow and most importantly stop, you will not be happy with Valetudo.", plus this "not selling PCBs out of principle so people have to source and solder themselves" - I can solder, but I prefer to delegate boring, efficiently automatable tasks to robots in a factory and would much rather pay someone 10 bucks for a finished board than pay more for the shipping of individual parts and end up with 4 extra unpopulated PCBs that I have zero use for.)
I wholly agree with the anti-closed vibe. I even run Valetudo on my roborock. But lol I just want to control my vacuum cleaner locally and that's it. I haven't updated mine for years exactly because I just want shit to work and not have to be constantly fixing things and "be confronted".
CarPlay is probably the top of my list of features when I buy a car. I can careless about performance specs after a certain threshold. But not having CarPlay would straight up go into my do not buy list (yes, I wouldn’t buy a Rivian or tesla either.)
Manufacturers ditching Apple Carplay/Android Auto support will, if not immediately, inevitably pursue rent-seeking behavior in the form of paid subscriptions for services people could otherwise just have for free (and likely better) via phone.
It is a crazy step for GM in particular because it is not like they have any distinct advantages or differentiators as an automaker. Rivian and Tesla have specific niche appeal. I am not sure anyone is clamoring for a Chevy in the same way. Lack of Apple car play is a feature your average user will be aware of and consider when purchasing a new car.
GMs top end vehicles have been plagued with exploding engines recently so to remove CarPlay is a crazy move. I wonder how much additional revenue they expect from pushing ads and their shitty infotainment experience
I used to feel this way, but nowadays I've come to realize that Carplay's UX is inferior to my iPhone's UX on a mount. As long as I have Bluetooth or Aux-in, I'm fine.
(That isn't to say that I think GM will somehow produce anything other than a captured rent extraction tool)
I only care about bluetooth for music and handling phone calls, don't get the point why having the actual car dashboard show phone stuff that relevant.
You might as well be making the “I’m okay with my 2004 Honda accord with a tape adapter” argument.
There is no cost reason to exclude the option. Even if I don’t use it, if I’m buying a $30-50k new vehicle it better have it, even if that’s for the sake of resale or future family members I might pass the car down to. My 2016 has it, why am I tolerating the removing of such a feature?
If you want unscientific evidence you’ll notice that the Honda Prologue (has CarPlay) outsells the Equinox EV it’s based on.
> You might as well be making the “I’m okay with my 2004 Honda accord with a tape adapter” argument.
I love how authoritatively you say this, as though this is meant to be just so patently absurd it doesn't require any further argument. This is an attempted reductio ad absurdum, but the absurdum in question fails to actually be abusrd.
Not everyone wants to drop $50K every few years on some liquid-glass-operated subscription-based monstrosity. Some people much prefer cars operated by knobs and dials, that are easy to repair, and dirt-cheap to replace. It's simply superior tech.
I think it's hilarious that I can upgrade my 2004 Honda Accord and add CarPlay into it with a myriad of screen sizes and options, features, cameras. There's an entire world of car stereo / infotainment gear out there that is cheaper and better than ever. Meanwhile, everybody with their new cars are just stuck with the terrible built in systems, or can only upgrade with a $1000 worth of audio processing gear and multi-channel amplifiers.
I think this is a negative recency bias, nobody makes aftermarket stuff for brand new cars because nobody with a brand new car is trying to augment it.
Your link proves my exact point. You need some specific one-off gear from some no-name android tablet manufacturer that doesn't fit any other car. LOL@ this thing running Android 12, which was out of support in March of this year... Much aftermarket support wow.
The 2004 Accord has a DIN slot in it. Which has a myriad of options available for it, despite the dwindling aftermarket.
Have you seen how to upgrade a 2023 Honda Accord infotainment system? You can't without major work. You can use the factory head unit, and feed all the amplified signals into a $1000+ sound processor, with a bunch of other modules specifically built for the car, then run the speaker outputs out to some crazy-ass 8 channel amplifier (because more speakers means the stereo is better for some reason), then feed that back to your speakers. And then at that point, all you've upgraded is the audio, not even the head unit itself. And why does it even have a center channel again?
Also, people obviously care about what infotainment is in their cars, as there is a huge amount of people saying they won't buy a car without CarPlay. Sounds like people DO want to augment their infotainment systems. Nobody is trying to do it, because you can't do it. Imagine if you could buy the car you wanted, and install your own accessories in it.
Look, I wasn’t even getting into proprietary or not. Obviously generic DIN was better than the present automotive industry situation.
You made a claim that you can’t upgrade these systems without $1000 of gear and I proved that idea wrong with a single link. It doesn’t really matter that it’s Android 12 - nobody really cares because all it has to do is run CarPlay and Android Auto. As long as the system can do that it’s infinitely upgradable from a software perspective. Nobody’s actually using the base Android system.
I also have a friend that spent under $500 to add CarPlay to a Chevy Sonic with a similar system and they’re very happy with it.
And that’s why I want a car to have CarPlay and Android auto, because it negates any need to upgrade the system down the line. The upgrades happen on your phone.
Imagine if you could install your own accessories in it…like the one I linked? I mean, again, I get it, it’s not a simple DIN setup but for the 1% of people who are interested in upgrading their car system this this is a real product you can buy for basically any car model. I owned an Alpine head unit for my 2005 Volkswagen and I wouldn’t really describe it as not janky compared to the OEM head unit, but the thing had Bluetooth and that’s all I needed.
My "real lived experience" with a CarPlay device that I installed into my 2016 Toyota Sienna is that it was slow, annoying, and I actually removed it. You can have it for free if you are in the Calgary area, my email is in my profile.
The UX difference compared to my phone, and the change in speed / etc.. was infuriating. The experience with my phone was ALWAYS better.
An argument based on the desires of fiscally illiterate people with 19.99% APR loans on oversized cars would make more sense if I worked for an auto company, thankfully I don't.
Yeah that’s a Toyota problem, not a typical experience of old wired CarPlay systems at all.
And again I’m not making some kind of pro-consumerism buy a new car right now argument, I’m just saying that in 2025 CarPlay and Android Auto are high demand features that a lot of people insist upon.
I’m not making some kind of profound statement on the state of the car industry or whether infotainment is too deeply integrated into vehicles.
I’m just saying if it was time for me to buy a new car I’m thinking twice about buying something that’s not giving me phone mirroring, just like I want my car to have FM radio even if I rarely or never use it.
It’s almost same price as Halo11, but seems to be far better product and fit! During first look of it, it felt expensive, but after looking at Halo11, it seems affordable :-)
I think this may be the dilemma GM and others are in. Given that it's a must-have, how much can Google and Apple charge the manufacturers for a license? $2000 per car? 5000? More? I dont think you can necessarily criticize GM's decision (and making a public announcement that all other car manufacturers will read) without knowing the upstream costs.
There aren't really any licensing fees for Android Auto. Apple charged a few dozen dollars for the right to produce a car with the integration, rather than a license as such. There's integration costs and hardware, so it's not free, but generally it's cheaper than the alternative.
Android automotive, the system GM is discussing here, is more expensive in every way than Android Auto. The reason they're switching is that Android Auto/carplay don't give GM enough additional monetization options for customers.
Is it possible Apple/Google initially charged a nominal fee when the technology was brand new, and now that it has widespread acceptance and is mandatory for many car purchasers, they turned the screws and are now charging an exorbitant price?
It's possible, but it wouldn't lead to this kind of action. Automakers already have a way to deal with expensive features that are critical to a large segment of the customer base. They limit it to higher trims or select models, depending on the cost constraints. That's what GM does with ADAS for example.
They're eliminating phone projection entirely here, which means they think the feature is incompatible with their business model.
What is the economic incentive for Google and Apple not to do this? They've convinced much of the public that it's a necessary feature (moreso than SiriumXM, for example, which also starts free then costs a lot), and better than what the manufacturers can develop internally, so why allow the manufacturers to integrate it for such a low price?
I don’t know but my best guess is in the low hundreds given the Vline sells for $679. Just installed one in a 2008 Lexus with a resistive touch screen and ancient interface, now I’ve got CarPlay.
It's basically the feature to a huge number of consumers. Various numbers I've seen place not having it as a dealbreaker for somewhere between 15% to 80% of buyers, especially the younger ones. It's more important than virtually any other feature you can name.
Wild how different people’s expectations can be. I’ve never even owned a car with any of these systems, yet they’re dealbreakers for others. Not saying they’re good or bad, just remarking on how something one person has never even thought about could be an absolute requirement for someone else!
It's surprisingly common. A lot of standard features (reclining seats, AC, large fuel tanks, back seats, large trunks, fuel efficiency) aren't important to all buyers, but they're important to some buyers. Over time they creep into the standard feature set as manufacturers design new models because the economies of scale from additional customers and a simpler production line outweigh the additional costs.
Once you get used to CarPlay it is hard to go back. Your apps are right there and synced without effort. Whatever music/podcasting app is there. Your choice of navigation/maps. It is just incredibly convenient and frictionless. The interface is simple and easy to navigate and when implemented correctly is incredibly responsive.
It’s definitely about expectations. I suspect that a lot of people, once they’ve used CarPlay or (presumably) android auto, would find it very difficult to go back to the experience that the typical car manufacturer is shipping.
I’ve never used a car that I was impressed with the built-in navigation or anything else that I might otherwise do on the phone. Aside from Google and Apple just shipping more polished products in general than the teams that seem to be building infotainment systems, the specific phone integration also makes a lot of stuff easier. Yes, I can use voice to ask Siri to do something. But I can also just use my phone to type in an address or whatever else. I cannot do that with Chevy’s built-in system. I also frankly don’t want to pay Chevy or any other car manufacturer a subscription fee for up-to-date maps or traffic that my phone has already.
The number I’ve seen is that CarPlay specifically is in the top ~2 prios for 79% of new car shoppers in the US. I think there’s probably enough mass here, and GM doesn’t have the niche EV/SW cachet that Rivian and Tesla have despite lacking CarPlay so I don’t think this will go well for them.
No longer in that industry, but I worked on one of the first generation of semiconductor equipment for production when GAN first started picking up. Took about a decade before we saw it prevalent in consumer electronics. While this is interesting, I don’t see why DLC process won’t do something similar to this paper?
Ehh, is it cool and time savings that it figured it out? Yes. But the solution was to get a “better” version prebuilt wheel package of PyTorch. This is a relatively “easy” problem to solve (figuring out this was the problem does take time). But it’s (probably, I can’t afford one) going to be painful when you want to upgrade the cuda version or specify a specific version. Unlike a typical PC, you’re going to need to build a new image and flash it. I would be more impressed when a LLM can do this end to end for you.
Pytorch + CUDA is a headache I've seen a lot of people have at my uni, and one I've never had to deal with thanks to uv. Good tooling really does go a long way in these things.
Although, I must say that for certain docker pass through cases, the debugging logs just aren't as detailed
Yup. The beauty of it is that the underlying ai accelerator/hardware is completely abstracted away. There’s a CoreML ONNX execution provider, though I haven’t used it.
No more fighting with hardcoded cuda:0 everywhere.
The only pain point is that you’ll often have to manually convert a PyTorch model from huggingface to onnx unless it’s very popular.
I’m a software engineer now but I was a design mechanical engineer for a decade. In America, mechanical engineers always felt like the bottom of barrel. The pay, benefits, and authority was always worse, even at the big tech companies. Culturally, America wants manufacturing back but doesn’t give it the respect it deserves.
I think it's partly the economic conditions. You have some great innovation in AI and it's worth billions. You have some in cars and it just holds off the decline a bit while manufacturing moves to Asia.
I know several mechanical engineers who ended up in semiconductor processing and they make bank. Maybe not senior google coder/engineer money but probably in top 10% of senior engineers easily
Culturally, America wants manufacturing back in order to increase the prestige of non-college men. Some amount of spillover to engineers might be tolerated, but if the pay, benefits, and authority are seen to mostly accrue to high-SAT-scoring engineers (just like in Silicon Valley) then a manufacturing renaissance will be considered an abject failure.
As for mechanical engineer many things that are puzzling to average engineer are easy to understand / recreate by mechanical engineer, especially in 3d printing era
I do a lot of small mechanical projects for fun. But designing something for one or two pieces is very different from designing something that is made in the hundreds or even thousands.
I self host a lot of things, pihole and adguard is one thing I no longer self host for about five years now. $20/year for NextDNS for the whole family is worth every penny and most importantly spouse approved. My spouse doesn’t mind what we self host as long as the friction to use it is not too high.
NextDNS support now doesn't even bother to respond when you face a problem. If you are lucky a fellow user will comment which most probably won't solve the problem and it would rather be a "same here" comment. I had to stop using before even my first year's subscription finished.
I have two pi-holes running concurrently, mainly so it doesn’t ruin the internet for my wife if one goes down. In 4-5 years of running pi-hole I’ve had I think 3 complete failures, 2 were due to cheap SD card corruption and one due to a failed upgrade to pihole v6.
I also excluded most of her devices from any filtering by the pihole because she wants to be able to click the sponsored links and ads on Google. Whatever.
That’s why nextDNS is nice, there’s a “allow affiliated link” setting. So it blocks the ads but allows your wife to click on sponsored ad links. How’s do you manage Adblock when you’re not on your network? That’s the main draw of NextDNS for me. Works more or less anywhere
Huh interesting feature, I'll have to check it out today to see if there's enough improvements over pihole to warrant a switch.
I'm effectively always on my network because I use Wireguard to VPN back in to home, so I can easily access my server and RPi dashboards. Though at this point I've whitelisted a few dozen domains that were giving my wife or I issues, and excluded most of her devices because she doesn't want to be on it, so it's pretty hands-off. The only time I have to disable the pihole nowadays is when I'm unsubscribing from an email list and the link is a tracking link. And that's with over 3M domains blocked.
I've never specifically noticed WireGuard anywhere in the top battery consumers on either Android or iOS. Friday I was out of the house all day, and Wireguard running on cellular all day used 1% of my iPhone battery.
Yes but that’s the case for any DNS (which is why your ISP loves it when you use their DNS, for “marketing” data they sell to others). However, similar to pihole and adguard, you can turn off logging if you want.
reply