adtac

Two people? The department in charge of government efficiency has redundancy at the leadership level? The irony.

so that the next company won't sue

Is there any message from the Flipper Zero people that this is actually their CEO? :)

if you're on your phone and can't read the pdf, I ran it through an OCR: https://gist.github.com/adtac/d4baaa5b23344689cff42b9bf16554... (no photos)

I haven't verified that it's correct (there are a few benign artifacts at the very least), but it looks right content-wise

>WRONG! ENJOY MORE PENALTY, SCRUB!

Is that true tho? During training, the model predicts {"wall": 0.65, "fence": 0.25, "river": 0.03}. Then backprop modifies the weights such that it produces {"wall": 0.67, "fence": 0.24, "river": 0.02} next time.

But it does that with a much richer feedback than WRONG! because we're also telling the model how much more likely "fence" is than "wall" in an indirect way. It's likely most of the neurons that supported "wall" also supported "fence", so the average neuron that supported "river" gets penalised much more than a neuron that supported "fence".

I agree that distillation is more efficient for exactly the same reason, but I think even models as old as GPT-3 use this trick to work as well as they do.

You forgot to set the washing machine to "Whites"

Too early to tell. It's not useful as a primary monitor, but it's great as a secondary. In its current form, the frame rate is good but the latency is awful. This makes for a bad typing experience, but there are plenty of usecases where I don't need to type. And even in situations like a chat window, where low latency typing feedback is critical, I can just have the input text box on my LCD monitor and have the chat history on the Kindle since it's mostly static. It needs some work to get there tho, but for now, this is a good start.

That would indeed be much simpler! I'm skeptical that the browser is powerful enough to handle a usable frame rate since every frame would have to go from the network -> DOM -> browser app memory -> frame buffer. The browser can barely keep up with the Kindle store / Goodreads, but it'd be nice to be proven otherwise since it'd make it much easier to get this working ootb on brand new Kindles.

Jailbreaks are highly firmware and hardware specific. I've only seen one specific combination (5.13.5, PW3), but the LanguageBreak authors have seen hundreds -- including mine -- so I have no additional insight or instructions that the extremely thorough LanguageBreak thread doesn't have.

> random forum bulletin board post

As opposed to a random gist? :)

dang, can we merge this thread into https://news.ycombinator.com/item?id=41155177 or vice versa?

Thanks!

I should stop writing code in files like /tmp/x.go no matter how throwaway I think the code might be. I still have the linux/armv7 binary if anyone wants it lol.

updated the gist: I'm getting close to 3-4 frames per second! this is only possible because of the partial screen refreshes since most pixels don't change between consecutive frames

Writing a quick Github gist, sorry haha, I'll post a link here in a few minutes.

Edit: https://gist.github.com/adtac/eb639d3c707b55a28f0ee9a420aa7e...

Are there enough trees from 300 BCE that still survive today?

In the ones that do, wouldn't there be enough thermodynamic diffusion within the tree trunk to obfuscate the true isotope levels in each ring?

The target image would need to reset that field so that a second reboot puts you back into the bootloader because otherwise you'll be stuck booting that image forever.

virtually every battery technology in existence today got created in a non-production environment

a less blunt critique would've specifically identified why this breakthrough is similar to the last one that didn't make it past the lab

surely there's a large enough sample set of past failures to choose from since there are claims like this every week

And only once every four years.

Unfortunately, they award it on the 4n+2 years. As someone born on a 4n+0 year, I’ll have just 38 years, which is too severe a disadvantage for me to stomach, so I didn’t bother with it.

4n+2 people, you have no excuses.

It's a common euphemism in ransomware and protection rackets in general. One of my favourites is the message the akira group leaves in infected machines that goes something like:

    Congratulations, you have passed a surprise information
    security audit and become a victim of ransomware.

    [...]

    We offer:

    1) full decryption assistance
    2) evidence of data removal
    3) security report on vulnerabilities we found
    4) guarantees not to publish or sell your data
    5) guarantees not to attack you in the future

They're just a security consulting company that you didn't know you had on payroll!

Btw I looked at what they provide as evidence of data removal (2) and it's literally just the stdout of `rm -vrf data` lol. I mean, I get that it's impossible to provide evidence of absence, plus the victims have no leverage anyway, but I dig the theatrics.

Like often, both will probably happen

yes, but wireguard is undetectable because the server won't reply to packets that fail auth whereas sshd will respond (with a banner!) to anyone who comes knocking

stealth is basically impossible with TCP

everything works fine with just tmux + mosh, but I like to bind sshd on the wireguard interface so that there aren't any listening ports accessible from the public internet

tmux over mosh over wireguard is all you need

The aphorism typically says cache invalidation is hard. Not because you don't know what index to invalidate but because it's hard to invalidate the thing at the right time.

Caching itself is quite easy, just ask the designers of speculative execution at Intel :)

> triple the training

From what I understand, not quite. It looks like the cost of training might be similar, but less parallelisable within a specific token sequence. This is because they have to compute the KV of token T before they can use it in T+1 whereas in a regular training process you can compute the KV at each layer for every subsequence. You're right that it took 2.7x longer to train the smallest model but I wouldn't be surprised if the GPU utilisation was proportionally lower too.

The KV cache is just another tensor to be used with matmuls. Unlike the model weights which are fixed, the KV cache is uniquely constructed for every input. Think of it as the model growing new weights to represent the new knowledge it learns about the user's input at inference time because not everything can be baked into the pretrained model.

You want to store your KV cache in the same processor that does the rest of your matmuls.

1. LLMs want to mimic conversations on the internet

2. Disagreements on the internet usually end up toxic

3. Anything perceived to be toxic has been hammered out with RLHF

Play sycophantic games, win sycophantic prizes

>Token volume needs to double just for revenue to stand still

Profits are the real metric. Token volume doesn't need to double for profits to stand still if operational costs go down.

Maybe I'm missing something basic but that still doesn't explain why it's 30% and not 50%.

I don't think purely qualitative arguments work here.

DCF and other tools to value companies make sense when the valuation is somewhat stable, but the real world often isn't. MSFT was worth $1T a few years ago because the world presumably expected $1T in profits over the lifetime of the company. But OpenAI came around and suddenly they're worth $2T. It wasn't because their lifetime doubled, it was because most people perceive AI as a major advancement. I believe this is the primary thrust behind the S&P500's 7-10% annualised returns in the last 20 years because most of the gains have come from the top.