Is anyone here using phi-4 multimodal for image-to-text tasks?
The phi models often punch above their weight, and I got curious about the vision models after reading https://unsloth.ai/blog/phi4 stories of finetuning
Since lmarena.ai only has the phi-4 text model, I've tried "phi-4 multimodal instruct" from openrouter.ai.
However, the results I get are far below what I would have expected.
Is there any "Microsoft validated" source (like https://chat.qwen.ai/c/guest for qwen) to easily try phi4 vision?
Sometimes, you can't find the average because it's undefined: it can happen with a Cauchy and a few other statistical distributions: the wikipedia page has a nice plot of how the first 2 moments don't converge https://en.wikipedia.org/wiki/Cauchy_distribution#History
When in doubt, don't use the mean: prefer more robust estimates, as even with degenerate statistical distributions, there are still some "good numbers to report" like the mode or the median.
And if you don't know statistics, just use a plot!
> it's like systemd trading off non-determinism for boot speed, when it takes 5 minutes to get through the POST
That's a bad analogy: if a given deterministic service ordering is needed for a service to correctly start (say because it doesn't start with the systemd unit), it means the non-deterministic systemd service units are not properly encoding the dependencies tree in the Before= and After=
When done properly, both solutions should work the same. However, the solution properly encoding the dependency graph (instead of just projecting it on a 1-dimensional sequence of numbers) will be more flexible: it's the better solution, because it will give you more speed but also more flexibility: you can see the branches any leaf depends on, remove leaves as needed, then cull the useless branches. You could add determinism if you want, but why bother?
It's like using the dependencies of linux packages, and leaving the job of resolving them to package managers (apt, pacman...): you can then remove the useless packages which are no longer required.
Compare that to doing a `make install` of everything to /usr/local in a specific order, as specified by a script: when done properly, both solutions will work, but one solution is clearly better than the other as it encodes more finely the existing dependencies instead of projecting them to a sequence.
You can add determinism if you want to follow a sequence (ex: `apt-get install make` before adding gcc, then add cuda...), or you can use meta package like build-essentials, but being restricted to a sequence gains you nothing.
> Military authorities in California requested census data to identify the Japanese-American population. Then in 1942, president Franklin Roosevelt issued an executive order to authorise their removal.
zfs in mirror mode offers redundancy at the block level but scrub requires plugging the device
> All error correction has a limit. If too many errors build up, it becomes unrecoverable errors
There are software solutions. You can specify the redundancy you want.
For long term storage, if using a single media that you can't plug and scrub, I recommend par2 (https://en.wikipedia.org/wiki/Parchive?useskin=vector) over NTFS: there are many NTFS file recovery tools, and it shouldn't be too hard to roll your own solution to use the redundancy when a given sector can't be read
> Cosmopolitan is neat hack, but it's not a viable multiplatform executable format, runtime sysytem, or distribution mechanism.
It's the best we have to ensure programs will be kept running. The multiple payloads are like a rosetta stone.
If I had 1 wish, I would make wine support sixel output + VNC to a localhost port + webgl or equivalent to extend this redundancy to GUIs
> Something like WASM + WASI seems much more likely to fulfill this function in the future.
Time will tell, but given 2 binaries (Sun Java vs Windows i386) from the same time period, the one that was "much more likely to fulfill this function in the future" is much harder to use
p-code is a neat technology, but failing at what it was supposed to achieve
I wanted to limit the space needed on the EFI, and keep as much as possible on the Windows partition because there's also the question of where the bitlocker key would go.
With something like a chainloader using a special part of the NTFS partition posing as bad blocks, Windows could "remove" this access easily, without having to touch the EFI partition or the boot variables.
> Could you stick the Linux kernel and initramfs on the EFI boot partition as a UKI
I considered that, even if it would go against the idea of having everything inside the Windows partition. I'd rather have had a shim in the EFI, with the UKI in C:\
The difficulty was bitlocker: my approach was a UKI with a small kernel and a few binaries to open the bitlocker volume and kexec the bigger kernel.
I was also exploring how to mark part of the NTFS volume as unusable to stick a different payload there.
The "ISO on a thumbdrive" was done to get baremetal boot working and out of the way, to see if I needed deeper changes to what had started as a 2 stages boot process, or if it was good enough as-as.
> Linux supports NTFS.
The kernel module is great!
I wish there was a linux distribution that could be run from either WSL or baremetal, to get more people familiar with baremetal linux.
You would have to add support for the peripherals in the kernel, and have some kind of init system. You would also need a filesystem supported to boot the computer.
I was doing something similar for the entire OS a few years ago: cosmopolinux, a distribution of cosmopolitan binaries: https://github.com/csdvrx/cosmopolinux
My idea was to replace the WSL binaries to have a Linux distribution living on C:\, but that could also be booted baremetal if you didn't want to use Windows
I had to put together a multi stage init system for that: if you get the ISO, you can put in on a thumbdrive and boot it: https://gitlab.com/csdvrx/cosmopolinux
The only difference between them is the kernel and the filesystem: the github NTFS has a firecracker linux kernel, the gitlab ISO has a regular kernel with many modules.
I wanted to do a full NTFS solution but I couldn't find a bootloader I liked that would support booting from a NTFS partition.
However, a 2003 study estimated only 45% of the Sahara could be covered by vegetation, and a 2022 study found that it may not be sufficient to start another AHP: it just "lowers the threshold for orbital changes to induce Sahara greening"
> That attack requires continuously monitoring a given device or area though, right?
The "randomization" seems to be a pseudo-randomization: with the seed and the timestamp, you should be able to deduce the future "randomized" addresses.
> Also — secondhand books are generally outdated, undesirable, and/or damaged — do collectors still find diamonds in the rough?
Secondhand stores offer in general "better" products because of the double curation:
- someone found the good interesting enough to buy it in the first place (1st curation)
- the store found it good enough to buy it again from the first owner (2nd curation)
> Aren't brick-and-mortar bookshops, generally speaking, as viable as Apatosaurus today?
No, it's even better because of the limited space they have to display the goods they want to sell: while online stores can show their full inventory, brick-and-mortar need to select what's most likely to sell.
This adds yet another level of curation: the store found the good valuable enough to be exposed to buyers, instead of keeping it in the back (3rd curation)
I find great music by randomly buying second-hand CDs from brick-and-mortar secondhand stores, thanks to this triple-curation,