The ones in cars need to be heated up quite a bit in order to work, and you still need reference air. Otherwise, I'm pretty sure that CO2 isn't a problem but rather an indication of a lack of oxygen in the first place, so it technically could work... just not if you're measuring the environment itself.
Same with ds4-flash. Low power is 13t/s vs 26t/s, power usage is ~30w vs ~100-120w. I still use high power because my m5 max is basically a server with built-in UPS and screen (and an off-work media machine, blender cycles toy, etc - growing fond of it, actually), but even 10tok/s is usable with a relatively smart model. Caveat is that the code that I touch barely has any boilerplate and I like keeping it lean.
Modding communities are still going. Kids, afaik, are still playing around with hosting minecraft servers or whatever is en vogue/cool/meta/etc nowadays. DIY 8-bit computers are gaining popularity.
IMO the fact that something's become very mainstream doesn't necessarily mean it's been watered down for everybody. There will always be people with various levels of curiosity and enthusiasm.
DS4-Flash definitely stands its ground when I'm obviously wrong (i.e. me reading ifneq as ifeq for several minutes straight), and I've seen at least once a "thinking" trace that was almost verbatim "the user has changed this". That's local, so thinking traces are raw. Pretty sure the more powerful models (500+GB weights, closed SOTA, etc) are even better at this - haven't had GPT5.5 with codex sugar coat things for me.
The apple silicon chips basically beat everything in bandwidth. Highest amount of memory controllers (i.e. channels) for a given capacity. That's the main party trick.
My 2nd hand ~$200 (minus a 256gb SSD upgrade) T400 was the best laptop I've ever had. Comfy everything, best laptop keyboard I've ever had, not worrying about dropping it on concrete from 2 meters (on the big extended battery, no less). Coil whine when switching p-states, no IPS, that's about it.
...and then you have hypervisor-based cheats, hardware cheats and whatnot. I'd say that AI flagging of suspicious cases + additional targeted scrutiny is the way forward - for competitive platforms, that is. That, and trust factor - I practically never get bad games when I play alone in cs:go/cs2 (~20k mmr eu, lem/smfc prior to that) - both in terms of somebody cheating and in terms of people that are full of themselves in one way or another. I'd say that combining these techniques should be very effective.
I'm just (occasionally) vocal - i.e. overall a minority. Pretty sure there's way, way, _way_ more people that just quietly do this. I'd even say that the current market makes you appreciate such companies even more.
p.s. a bit of a windows fanboy as well - used to do drivers for it, kind of like the internals / driver model / etc... but I really dislike the path they've taken, and there's nothing else like it.
Finally, I have an old projector setup with an x360/x1x on it right now (hc4000 + diy frame w/ dark energy abyss + 758 v3 + lsr305 + some subs - rag-tag), so I have a good excuse as well :P
I get to support people that are very involved in making sure that a long list of x86 win32 software that I want to be able to run plays well with linux and osx (not-quite-directly, but the crossover folks are on it) - regardless of whether it's on steam or not. Plus general linux desktop work in the "make games play well" department.
+100 on this. In addition, if you don't outsource your thinking and you're willing to go through all this, you absolutely don't need the top tier models.
mbp16 m5 max 128gb, antirez/ds4, deepseekv4-flash. Works well for relatively dense (let's say <20k LoC per project) C codebases that are essentially a bunch of custom specialized stores, http servers, network infra, media transformers, etc.
Runs through Pi with a custom prompt (basically "don't speculate blindly, isolate things, make them traceable and measurable, then verify") and behind a pretty restrictive bwrap setup - RO bind everything other than ~/.pi, cdw and a separate tmpfs, unshare almost everything other than the network - for which I use a network namespace that only allows tcp connections to a specific ip and port (i.e the inference mac) - i.e. netns exec into bwrap.
Can't compare it to SOTA or higher-requirements models on what I work on - policy. That said, on a bunch of test pieces - it obviously isn't gpt-5.5, it definitely lags behind k2.6/glm/ds4-pro, but it absolutely is usable. Of course, on such codebases, forget about one-shotting or trusting it blindly or anything of the sort - you ask it, guide it, restart the context from time to time to have a "fresh dice roll" and to keep the context small and clean, etc. Compared to anything smaller (incl. all the usual local qwen models) - on a test piece, it figured out that memfd and mmap were used for setting up a ring buffer with natural wraparound handling (double mapping the first page at the end) and didn't tell me "this is for sharing memory between processes" or some other BS.
Performance as described in the tables in the readme here:
https://github.com/antirez/ds4
...with a bit less than half that at "low power" (30w). Both are usable.
At low effort (i.e. active recovery can-do-this-all-day-every-day) levels, it's quite efficient. That assumes that you're fine with whatever let's say ~120w continuous can get you, which is what the m5 max 16'' mbp peaks at (in my case) during inference (ds4-flash)... so I guess it's quite usable, actually :)
Far from everybody is using it for massive code generation - code size is still very much a liability. I currently use it (ds4-flash, local, code isn't open to non-local) to modify and trim a certain internal project for a different use case. Tons of review passes, fuzz test generation and running, "give me your thoughts on this", etc, etc - basically iterating over and over and over. Funnily enough, my intrinsic knowledge of the codebase improves massively like this. It is involved, but the acceleration is real - and I don't have to worry about wrist nerves as much.
The only people I know that have LoC/token use/etc metrics imposed on them work for big corps where such things are (or used to be) en vogue.
Given that you can run quantized flash on 128g ram, and there's a heavy focus around it (DS4)... I'd say that it's pretty feasible for a decent amount of devs. Never thought I'd buy an MBP but here we are.
n.b. I can't use nonlocal models for a big chunk of my work, so there's that as well.
Ehh... can't really hit "chatbot" limits on the $20 plan. Pretty sure the limits are not token based for that in the first place, and if it spews out a ton of stuff, it takes me longer to go through it and I end up asking it follow-up questions in a way where it replies... _relatively_ concisely. Still, gimme robot back. On a good note, it almost managed to call me stupid.
Codex has also been fine, but I'm guessing they know better than to tweak it like that, given their target users.