"Almost everywhere" is precisely defined, and it is broader than that. E.g. the real numbers are almost everywhere normal, but there are uncountably many non-normal numbers between any two normal reals.
If you need accuracy, an LLM is not the tool for that use case. LLMs are for when you need plausibility. There are real use cases for that, but journalism is not one of them.
Emacs's hexl-mode does this, incidentally, though annoyingly by default it makes all faces the same color. I never understood why it defines the faces but then doesn't customize them.
What exactly does it do? I'm looking at hexl-mode sources in my Emacs, and I see it defining only two faces - hexl-address-region and hexl-ascii-region.
That's correct, as far as I can tell: the first one is used for all hex values in the "main" are of the buffer, and the second one for the character representation of each byte in the right-hand side column.
That's so cool. I remember loving this game in the arcade but then being annoyed when I had to also buy a paddle wheel to play it on my 2600, which was then useful for exactly 0 other games.
The "paddle" and "driving" controllers looked the same, but they did not have the same function.
A paddle controller for the Atari 2600 had a hard stop, so that it could only make one revolution (or a bit less) in each direction. Therefore, you could use it with Tennis or Pong or whatever else just had you going back-and-forth.
A driving controller spun freely in both directions without stopping its motion. This was not analogous to the steering wheel of a car, but it did permit driving games to be relatively free-wheeling, and you could spin the car's wheels endlessly in either direction.
In my experience, paddle controllers were more compatible with more games, but if you had a diverse library, it behooved you to keep driving controllers on-hand for that eventuality. Other unique controllers included the BASIC Programming pads, and one of those space games which had some really intricate controls on the dash.
The "driving" controller class was the type that was supported by Tempest. Analogous to the arcade controller, you could spin indefinitely in either direction without having the physical tab to stop the motion. This definitely contributed to the fun and suspense of the gameplay!
I'm not sure about the specific Atari 2600 controllers but my hazy memory has at least three types of what appear to be rotary.
One is basically a self centring sprung up/off/down switch. That would be similar to a car indicator stalk and simple left/right arrow keys.
Another would be rotary with a stop but it sent a physical position, presumably it was something like a variable resistor or very fine resolution rotary switch. With these you could instantly position your character by the position on the ring/slider. This could be interpreted as position 1, 2, 3 etc etc.
The third was a free spinning which moved the character faster the faster you spun it. This would be how I remember Tempest playing, you could slowly nudge it or just do a fast spin & stop to quickly move around. This would produce a signal such as clockwise+very slow or anti-clockwise+very fast.
Driving controller makes way more sense for tempest, and the lack of use. My family had a pretty extensive collection of 2600 games, and two sets of paddles (needed for four player paddle games, we had one, but it wasn't very good and the 2nd set of paddles was wonky anyway), but no driving controller or any games that used it.
I had several driving games, too. But they used a joystick or the paddles.
If you want LLMs to continue to be offered we have to get to a point where the providers are taking in more money than they are spending hosting them. And we still aren't there (or even close).
Nope. They're losing money on straight inference (you may be thinking of the interview where Dario described a hypothetical company that was positive margin). The only way they can make it look like they're making money on inference is by calling the ongoing reinforcement training of the currently-served model a capital rather than operational expense, which is both absurd and will absolutely not work for an IPO.
Inference, in and of itself, can't be completely unprofitable. Unless you're purely talking about Anthropic?
But
> If you want LLMs to continue to be offered we have to get to a point where the providers are taking in more money than they are spending hosting them
Suggests you just mean in general, as a category, every provider is taking a loss. That seems implausible. Every provider on OpenRouter is giving away inference at a loss? For what purpose?
For the same reason that Amazon operated at a loss for two decades and Uber operated at a loss for a decade and a half. The problem is the free money hose isn't running anymore.
The open models may not be as great but maybe these are good enough. AI users can switch when the prices rise before it becomes sustainable for (some) of the large LLM providers.
Currently it costs so much more to host an open model than it costs to subscribe to a much better hosted model. Which suggests it’s being massively subsidised still.
For a lot of tasks smaller models work fine, though. Nowadays the problem is less model quality/speed, but more that it's a bit annoying to mix it in one workflow, with easy switching.
I'm currently making an effort to switch to local for stuff that can be local - initially stand alone tasks, longer term a nice harness for mixing. One example would be OCR/image description - I have hooks from dired to throw an image to local translategemma 27b which extracts the text, translates it to english, as necessary, adds a picture description, and - if it feels like - extra context. Works perfectly fine on my macbook.
Another example would be generating documentation - local qwen3 coder with a 256k context window does a great job at going through a codebase to check what is and isn't documented, and prepare a draft. I still replace pretty much all of the text - but it's good at collecting the technical details.
I haven’t tried it yet, but Rapid MLX has a neat feature for automatic model switching. It runs a local model using Apple’s MLX framework, then “falls forward” to the cloud dynamically based on usage patterns:
> Smart Cloud Routing
>
> Large-context requests auto-route to a cloud LLM (GPT-5, Claude, etc.) when local prefill would be slow. Routing based on new tokens after cache hit. --cloud-model openai/gpt-5 --cloud-threshold 20000
I've found MiniMax 2.7 pretty decent and even pay-as-you-go on OpenRouter, it's $0.30/mt in, and $1.20/mt out you can get some pretty heavy usage for between $5-$10. Their token subscription is heavily subsidized, but even if it goes up or away, its pretty decent. I'm pretty hopeful for these openweight models to become affordable at good enough performance.
Rapid MLX team has done some interesting benchmarking that suggests Qwopus 27B is pretty solid. Their tool includes benchmarking features so you can evaluate your own setup.
Edit: I’d also consider waiting for WWDC, they are supposed to be launching the new Mac Studio, an even if you don’t get it, you might be able to snag older models for cheaper
I see the current situation as a plus. I get SOTA models for dumping prices. And once the public providers go up with their pricing, I will be able to switch to local AI because open models have improved so much.
A guy from Meta interviewing at BBC a few years ago claimed that every school child in India was going to have the metaverse VR or they'd be left behind in their education, so every family was certainly going to pony up the money.
Somethings not adding up. Why is Amazon making financial plans for the next decade based on continued OpenAI spending but you’re saying AI providers like OpenAI and Anthropic aren’t even close to being profitable, so how can they last a decade or more?
That's the interesting question, right? Because if this unwinds during a period of external inflation (say, because of a big war and energy shortage) then even the Bernanke would say helicopter money won't work
They probably aren’t planning on making the money on consumer subscriptions. Any price is viable as long as the user can get more value out of it than they spend.
Like with all new products. It takes time to let the market do its work. See if from a positive side. The demand for more and faster and bigger hardware is finally back after 15 years of dormancy. Finally we can see 128gb default memory or 64gb videocards in 2 years from now.
reply