jwitthuhn

If you have 20 hours to spare I highly recommend this youtube playlist from Andrej Karpathy https://www.youtube.com/watch?v=VMj-3S1tku0&list=PLAqhIrjkxb...

It starts with the fundamentals of how backpropagation works then advances to building a few simple models and ends with building a GPT-2 clone. It won't taech you everything about AI models but it gives you a solid foundation for branching out.

Rosetta 2 is going to be EOL'd within the next few years. A more permanent solution would certainly be welcome.

You are describing Windows 11 LTSC which is a product that exists because Microsoft knows people want to turn this crap off.

It is of course only available in volume licensing to keep it away from normal users. Only businesses get to control their computers.

Yes the security of every Windows computer was much better then, any software that automatically updates itself without user consent is obviously a massive security risk because the user is no longer in control of what software they run.

I would think it is pretty silly if I needed some sort of verification to drive people I personally know around because other people were getting their car hijacked after choosing to pick up strangers they found on the highway.

I see them going for $150-300 on ebay, just don't ask where they came from.

It is wrong to release something unreliable even while acknowledging it is unreliable? The product performs as advertised. If people want accurate information an LLM is the wrong tool for the job.

From the Gemma 3 readme on huggingface: "Models generate responses based on information they learned from their training datasets, but they are not knowledge bases. They may generate incorrect or outdated factual statements."

So you can only use it if Canva has your email and they can send you exciting new offers.

Worst part about this is that these days the error message simply lies to the user. "Whatever.App is damaged and can't be opened. You should move it to the trash."

If you remove the 'quarantine' attribute that gets added to downloaded files it runs great.

Somewhat related: I wrote up a MTG card generator based on nanoGPT a while ago that I think produces pretty good results for being 1m parameters.

The real neat thing about this is that WotC makes a few thousand new cards each year, so my training data set just grows over time and the model gets better with no effort spent on my part.

https://github.com/jlwitthuhn/TCGGPT

If they are worried about the cost of initial ingestion then a gate on enabling would make a whole lot more sense than a gate on disabling.

Interesting, will give that a try next time. I guess I never looked because I had assumed that option made you connect to an existing domain.

Enterprise might but Professional sure doesn't.

The readme has this to say about hosting your own blog on it:

"Bear Blog has been built as a platform and not as an individual blog generator. It is more like Substack than Hugo. Due to this it isn't possible to individually self-host a Bear Blog."

"It isn't possible" is obviously not true but a plain reading of both that combined with the license would suggest you can't use bear at all for anything.

Not "as a paid service" just "as a service". The license does not allow me to stand up a Bear instance and let people blog on it for free.

Anyone pointing to that as proof that AI slows developers down has not actually read it. See appendix B

"We do not provide evidence that: AI systems do not currently speed up many or most software developers"

Haven't tried myself but it looks like it probably does. The weight files total 13.8 GB which gives you a little left over to hold your context.

An extension I trust is by definition trusted code. What is trusted is for the user to decide, not the broswer developer.

Pytorch also runs great on apple silicon, though it is hard to directly compare because Apple's high end GPUs can't compute anywhere near as much as nvidia's high end stuff.

e: I'll also add that pytorch does still have one oddity on apple silicon which is that it considers each tensor to be 'owned' by a particular device, either a cpu or gpu. Macs have unified memory but pytorch will still do a full copy when you 'move' data between the cpu and gpu because it just wasn't built for unified memory.

For anyone who never saw his show, it is (all?) now available on this youtube channel.

https://www.youtube.com/@MagicSecretsRevealed/videos

Is it easily solvable by humans without code? I suspect if you asked a human to write down all the steps in order to solve a Tower of Hanoi with 12 disks they would also give up before completing it. Writing code that produces the correct output is the only realistic way to solve that purely due to the amount of output required.

It would have made sense for Mozilla to mention that in the article. Instead they just lie about how it works.

You might be interested in this video, which is about a physical guitar pedal someone made using a chatgpt-based design. Same idea but all analog.

https://www.youtube.com/watch?v=J-PTzq1bv9M

If you're interested in a FOSS MTG clone check out XMAGE and Forge. They don't have every card but they do have a whole lot.

https://github.com/magefree/mage https://github.com/Card-Forge/forge

The DAW Reaper also uses a similar approach and for that alone I would highly recommend anyone interested in audio production give it a try.

Everyone else in the industry uses online activation. Whenever they take those servers down you lose your ability to install and use the software you bought.

With programs like WinRar and Reaper, even if the company producing it disappears tomorrow and takes all their servers with them, you can continue to make full use of the software you bought in 'free trial' mode and that is huge.

This seems like purely a good thing to me.

As established when Phil Zimmerman published PGP, software is speech and any restrictions on that speech must be enacted in a way consistent with the first ammendment.

State governments ignoring the constitution to prevent ai-related speech seems like a problem worth solving.

There are some but not a lot, current AI is a lot better focusing on smaller more well-defined problems than stuff like "add this feature" or "fix this bug".

A good example is this PR for llama.cpp which the author says was written 99% by AI. https://github.com/ggerganov/llama.cpp/pull/11453

When the problem statement can be narrowed down to "Here is some code that already produces a correct result, optimize it as much as possible" then the AI doesn't need to understand any context specific to your application at all and the chance it will produce something usable is much higher.

Maybe it will eventually be better at big picture stuff but for now I don't see it.

Love the concept of BSOL, might give that a try.

The difference is that SSPL is a license that was written in bad faith with the explicit intent that people can't comply with the parts about running a hosted service.

https://www.mongodb.com/legal/licensing/server-side-public-l...

See section 13, "Offering the Program as a Service". To comply with that you need to release all software used "to make the Program or modified version available as a service" under the SSPL.

For something like redis this includes: redis itself, the os you are using to host redis, the drivers and firmware for the hardware you are using to host redis, and more. Also your whole deployment stack up to and including the mouse drivers you use to click on the "deploy" button.

It is an absurd condition that effectively makes section 13 say "you can't offer a hosted version" and the OSI and FSF are right to reject that fig leaf of "it is like the agpl, but more".

It sort of can but all non-adobe software I know of, even commercial stuff like Affinity Photo, has spotty support for some PSD features.

Basically any given PSD will certainly load correctly in photoshop, but you're rolling the dice if you want to load it into anything else. More so if you are using more modern features.