I agree with you in general, it's not "Open Source". But in many cases you're allowed to fork them, and create and distribute derivative works (finetunes) from open models, and the training and inference code is also permissively licensed. That's not shareware.
And working with the binary weights of a huge pretained model is much easier and cheaper than doing it based on the entirety of its humongous source datasets.
In case someone wants to see for themselves how "instantly" it really was. Skip to ~9:50 when it gets real sluggish.
Edit: if you don't wear nostalgia glasses, you can watch the bits of screen being redrawn piece by piece after something moves or a menu closes. Everything is flickering. Opening and closing a menu takes up to one second. The text being typed in the editor lags. This is not more responsive than a modern system, far from it. Obviously.
To give a corresponding example, try moving the mouse while loading from floppy disk on Windows 95, ten years later. The mouse pointer judders!
This does not happen on a 1985 Amiga, even with a fraction of the CPU power, which it needs all of it to redraw UI damage as seen.
The reason is that the Amiga system prioritised input processing above normal tasks, it had preemptive multitasking where hardware timers triggered a CPU interrupt to let the kernel switch to a new timeslice, and of course reading a floppy disk was done with hardware interrupts and DMA transfers because the OS knew exactly what hardware it was dealing with, while PC compatibles couldn't guarantee those transfer modes (https://wiki.osdev.org/Floppy_Disk_Controller) so Windows 95 would inevitably handle floppy drives with polling.
Windows 95 was hamstrung by DOS compatibility. If you installed OS/2, you could have butter-smooth mouse cursor movement even while formatting a floppy disk and running a DOS (or even Windows!) VM in another window. And if you had a XGA card (which no one did) you had a hardware mouse cursor sprite, just like the Amiga (except only one boring sprite and no cool audio/video tricks)
There were SCSI floppy drives. Rare, though. However, with the right controller, not necessarily Adaptek, and drivers for it, they enabled all of that, too.
If the Amiga's mouse-pointer sprite was removed and the pointer was rendered onto a bitmap, it would still be 50Hz. There is more than enough time to do that.
The flawless refresh rate comes from the fact the update is executed as part of a level 3 priority CPU interrupt triggered by the vblank refresh. Interrupt code reads the hardware register of a dedicated chip that has been measuring the mouse's potentiometers this whole time, to see how much they've moved. The mouse itself is a passive device.
It then sends a input event to Intuition, whose priority 20 input.device task (the highest priority task in any normal Amiga system) processes the event and updates the hardware sprite coordinates, but could equally write it onto the bitmap. It has the time to do that because everything else in the system waits for it.
Compared to a modern PC, which also uses a hardware cursor... the mouse is an active device and sends its own potentiometer readings as USB input events, which arrive to be processed by an independent USB subsystem in the kernel and is bundled with all the other USB traffic from devices attached on the same bus. A vblank interrupt couldn't get the mouse coordinates even if it wanted to.
I should correct myself here and say it's not potentiometers, but that the mouse has vertical/horizontal wheels that let infrared light through to detectors, and the hardware is counting a quadrature-encoded pair of pulse trains, so it can tell how many steps back/forward each axis has moved.
It also estimates that, if the hardware counters are sampled every vblank, it's accurate (no missed wraparound) for the mouse moving at up to 3km/h! But it's still possible to whip the mouse around faster than that, so action games might like to sample the mouse more often than once per frame.
Apple loved doing things in software, ever since Woz wrote the Apple II disk routines in software so their floppy drives didn't need their own controller (the Commodore machines had a whole second CPU running an operating system in theirs). This is also why the one thing that WOULD lag the cursor on a classic Mac was formatting a floppy disk, since it required interrupt-level timing.
Even to this day Apple loves running things in software - the speakers on modern Mac laptops are driven in software - macOS has a daemon that runs and adjusts the speaker volume, tracking the sound pressure sent to them to keep them from being blown out. When Asahi Linux added speaker support, they had to reimplement this from scratch.
I had a sidecar with my Amiga 1000, which allowed you to use an XT MFM hard drive + controller card for the Amiga side. Once kickstart + the software loaded, it was incredibly fast. It was weird having a full IBM PC + Amiga in one conjoined machine.
But I agree that predictable slowness is a thing. One only need to discuss traffic jams vs a 20 minute car trip without them.
Hence the caveat "Aside from the loading times," :)
The "instantly" refers to user input: clicking a button, typing, or moving the mouse was as fast as it could possibly get because of how close the OS was to the bare metal, with no fluff bogging down the experience. Simply put, there were no wasted frames, as there were none to waste. Nowadays, at the very least, you get a couple millisecond of latency. That was not the case here.
Also, this is a machine from NINETEEN EIGHTY F//G FIVE, and it's still probably faster than your average PC running Windows 11.
The GUI was fast because it was fixed layout, MUI on the Amiga was slow as well. We tend to go that way to get UI that is easy to create. You see the same in all GUI libraries the more CPU you have the more you use it.
MUI was slow because it was written in C and went all-in on BOOPSI, as well as supporting fully dynamic layout. It could likely have been faster, but the sort of people who paid for it were also the sort of people who bought CPU accelerators.
Also gp talks about input responsiveness (e.g. key and mouse input), and that has indeed much less latency on those old-school home computers than on modern systems where each input event goes through layers and layers of abstraction bloat before causing a visible change on the display.
Floppy based Amigas were incredibly clunky. Each file shown in the Workbench GUI has its own icon file (".info" file) which is loaded separately, resulting in additional time for each file in a "drawer" (Amiga folder.)
All of our posts have been well received by an insanely high percentage of people who have interacted on here -- most people clearly find what we're doing interesting and relevant to the HN community (AI evaluations). A flag seems pretty aggressive! Especially when the top comment on the article (after our above comment got flagged) is about tacos.
I'm a person running the account, and I only post where I think we have a relevant contribution.
I remember back when Wordle was popular, people said the "Share result" feature was so effective, it was the reason for the game's viral success. I can't think of any other example though.
Well, that combined with the expectation that the recipient would also try that daily challenge, then presumably respond with comment/message/reshare comparing how they did.
It would make sense if there was any activity on that older post. A 12 hour old post with no comments is dead, linking to it is a waste of time, especially for those who click on it to find 0 comments.
I did click on your link, it was a waste of time and that wasn't very nice.
It's baffling how almost no one here ever speaks out against it. Everybody just puts up with it even though we all know what a sign of arrogance and incompetence it is to attempt to alter post titles in this clumsy, embarrassingly broken way. Everyone is just silently looking at their feet in second hand embarrassment and doesn't want to speak up. And the funniest thing is: there is an official rule that says not to edit the original titles. Amazingly bizarre.
i have mentioned that i dont like it. dang said i was wrong.
however, the submitter is able to edit the title for a short window after posting to add the 'how' back in for cases where removing it materially alters the meaning.
The model should perform better with JSON prompts than with natural language prompts. The safety filter may also be triggered incorrectly by natural language prompts, we are aware of this and will make a future checkpoint update to improve it.
reply