prophesi

I enjoyed this alternate phrasing of the Sam Altman quote:

"From a global allocation perspective, actually, it's quite terrible to [let billionaires] have a bunch of the world's limited supply of [money] sitting in their own [stocks] and being used 5% of the time. You want maximum [taxes] if you're trying to deal with this very supply-constrained environment. And somehow many people's intuition is the opposite of that."

I appreciate the ramble!

From what I can gather, DDR5 SDRAM are what data centers need (I assume it's involved with actual training, and efficiently running inferences from caches to consumers of their API's [?]).

VRAM, meanwhile, still remains the bottleneck for consumers. I think this means those working for Anthropic/OpenAI/etc won't be making as many meaningful advances to cut down VRAM for inferences. Why do thusly if that only allows the layman to use your open weight competitors in a more efficient manner?

For my critique on dataset and learning algorithm, that seems to only be pursued by European states who are lagging very far behind on LLM's. I'm happy that Nvidia is doing their best to be as open as they can with LLM's that the EU is acceptable to use, but there are still a lot of gaps in letting the layman verify end-to-end the entirety of the dataset[0][1][2][3][4] and the algorithms used to train the model outside of a high-level view of the framework used.

it goes back to international copyright law. Even if it was obtained in a legal manner, it's not okay to release it to the public. But it's fine to provide to states given you can prove the data was obtained legally.

So I'm not too hopeful for a decent open source LLM for the public to manage and improve on in the near future. Even if the datasets are actually open (which they're not, even for nemotron[4] beyond post-training), the architecture of learning algorithms with said limited dataset will still be proprietary. Just enough is open to be performant and compliant for government use.

[0] https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Super-VL-1...

[1] https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Super-VL-1...

[2] https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Super-VL-1...

[3] https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Super-VL-1...

[4] https://docs.nvidia.com/nemotron/latest/nemotron/super35/REA...

The graphene team is probably used to comments like this. It's a lot easier to give unsubstantiated claims than it is to actually work on a FOSS privacy/security product. You can't please everyone if the expectation is perfection.

I'm just going by occam's razor. It's good to be skeptical/paranoid and to know your threat model. I wouldn't be surprised if there's a powerful entity out there with a zero day for GrapheneOS / Pixel hardware and used it to target Medhurst. But I wouldn't throw the baby out with the bathwater even if that actually did happen. We're in the middle of the storm with the advent of highly competent AI-assisted cybersecurity atm.

On the other hand, there are state actors whom will profile you for using a Pixel phone as they assume you're running GrapheneOS, so the promise of privacy seems to be working for a lot of people if it's scaring them that much.

And with your logic, you would also need to say that QubesOS / Whonix / Tails are all Disappear/Honeypot operating systems since he's known to run those as well. It's much more likely that the disappearance is due to the fact that his adversaries are very powerful state actors and it takes just one slip up in your OpSec for them to pwn you.

it is manual, just automated via running inferences on a dataset of assumed refusals, and is a small rank-1 weight edit that doesn't require fine-tuning or training. Perf hit is much smaller than baking in these refusals in the first place.

I think we already are seeing the destructive potential :/ The genie's out of the bottle at this point with open-weight models (and thus can be abliterated in terms of offensive cybersecurity) that can run inferences well on relatively cheap hardware with relative parity to frontier models.

I'm a bit of an outlier in regards to today's HN culture. I'm still holding on to Stallman's idea that information wants to be free. There was a comment on another thread that aptly elucidated what I think is necessary.

Not everyone has the know-how or desire to compile the Linux kernel, or any what-have-you executable in user space. But a good amount of people do, and it's a boon to security/privacy that they do, and cryptography can verify that this work was done. (granted, manufacturing a computer is an international effort with a lot of unknowns, ie proprietary binary blobs in BIOS/UEFI) (and people are lax to validate the crypto themselves if it's not an automated thing; that's a UX problem imo)

I think we'll eventually find sufficiently performant LLM's trained on public domain information[0] that people can validate. Sufficiently performant training algorithms that people can validate[1]. And FOSS plugins to integrate with their FOSS IDE's which they can validate. And like the Linux kernel, volunteers to make sure there's nothing fishy going on end-to-end.

Can this stop the danger of what open models can do? No. And that does scare me. It's Metasploit on steroids. The cat and mouse game of cybersecurity is going to increase in power law levels of speed. I do think verifiable LLM's for consumers to use is a worthwhile pursuit regardless. And for LLM's capable of offensive cybersecurity, I think it might be better for it to be democratized than held only by state-actors and individuals/corporations with an absurd amount of consolidated wealth.

I have no horse in this race though. I'm just happy I can learn how these things work and toy around with getting quantized models to work well enough for my coding side projects. Though I often reach for a frontier model for serious coding issues, local for rubber-ducking, and Kagi's research assistant for initial planning.

[0] Pipe dream, unless international copyright law drastically changes. Since opening the dataset for any competent LLM will violate said laws, whether obtained legally or not themselves.

[1] Knowing the training algorithm and dataset will let it actually be open source models instead of open weights.

edit: These are just shower thoughts; didn't mean to write such a long response. Would love any pushback or any correspondence in general since it'll be on my mind all day.

It's pretty much how abliteration works too; find where the refusals are happening and flip the bit.

It's unfortunate, because the security theater performed here detracts from the model's overall cybersecurity performance. And uncensoring has another small hit to performance.

They were given day zero access to release their quantizations[0]. The lowest quant will still require 256GB VRAM to run inferences, but as always I'm amazed by the Unsloth team by how much they can reduce the requirements for these powerful LLM's, and the easy-to-follow guides to get started with using them.

[0] https://unsloth.ai/docs/models/mistral-large-4

DNC did a great job endorsing that candidate.

Have you tried Open WebUI? It doesn't determine which model works best for your hardware, but it otherwise has a very large feature set with a large community, and runs offline by default. Geared more for technical users, however.

Jailbroken 3DS scene is usually for modding the console OS itself given how easy it is to run custom firmware. And for ports, and emulation. While the GBA has a pretty dedicated community for custom games and romhacks, people often create games for the platform at gamejams because of how great the dev tooling is. It's a bit lackluster on the 3DS, but I did find the Love Brew[0] framework to be really nice. Even has simulator support[1] for testing your builds locally, though fonts will look much different on real hardware, with likely a bunch of other gotchas. Running out of RAM being a big one.

[0] https://lovebrew.org/

[1] https://github.com/lovebrew/nest

edit: compared it to the GBA homebrew scene since it's a large dedicated community to derive pros and cons between the two from. Switch is still receiving updates from Nintendo to brick your system.

And it's something they've been working on for the past 8 years!

Love the site, though I'll always prefer this attempt at explaining the Zelda timeline by Brian David Gilbert https://www.youtube.com/watch?v=Q-25c8Rsobw

This is just an ad for contextetech. Search engines build their own index, which is hard. Decentralized search engines like Searx have middling results, and it's a project running for over a decade. "Smart" is subjective, and not something I'd trust an LLM to determine.

I think yours and the rest of the comments to mine are missing the point of the presentation I'm critiquing, and missing the (admittedly cursorily added) escape hatch on why I think it's totally fine to program recreationally. It should be encouraged for everyone to learn. Akin to learning an instrument and/or music theory. It helps you in ways that you might never notice.

Recreational programming used to be one of the rubrics to hire a software developer. If you have an intrinsic motivation to code, it would often be more valuable than your credentials. That's why junior devs would push up sites showcasing their resume of random projects they've done outside of education in order to give them an edge over their competition. Some companies would even encourage some time on the clock for working on side projects to prevent burnout.

At the same time, some of the workers in the industry were disgruntled that you were expected to enjoy coding in your spare time to keep up. While some truly just wanted to do it because they were good at it, and would rather turn off the software dev brain after work to do other things they enjoy in life.

With the advent of high-performing LLM's for coding, though, it seems more important to showcase how you can harness AI for your recreational projects than the projects themselves. It's a bit of a catch-22 for junior devs though.

I still think it's good rubric to get into programming if you like to do it recreationally, but these are just a few of my thoughts on how it can be problematic.

Last time I visited Miami Beach, I found out they have a forced 20% gratuity at restaurants. And still ask for a tip on top that. My friends and I were likely tipping >40% for subpar service and overpriced drinks for a few days until someone noticed it on their receipt.

Don't most things people use day-to-day for both work and entertainment require use of the actual app these days, or at least make it very difficult to do so?

So now we're believing it's great that LLM's can write macros, thus further obfuscating human maintenance of the code-base, because LLM's can understand and implement it than better than humans.

This isn't the future I wanted, and is why I advocate for trade schools these days.

I miss the days I could listen to music and use my adhd/autism to its fullest potential reading documentation and figuring things out myself.

Of course, these are same people you'd want managing these AI agents in the first place, but reviewing code written by others always kind of sucks, especially when it's assumed the language model knows better than you do which isn't often the case!

Was the PRD perfected on the requirements? Only God knows, and I personally want to be there when it's written.

They explicitly mention that Svelte hasn't changed much since version 3? Svelte 5 did introduce runes but, as you say, still kept legacy code compatible.

Not offering solutions here, because I think it's more multi-faceted than I know about. But when people are afraid to take on the debt of on-board EMS for a 911 call, and local pharmacies are forced out to be bought up by the top national pharmaceutical companies letting them rent-seek, I do think "for profit" is part of this multi-faceted problem.

I'm thinking more of a future where big tech successfully moves the overton window for AI, people gradually rely on getting information/entertainment from LLM's more than books, then it can follow the regular plotline of 451 with the public voting to ban books.

Someone needs to write an updated version of Fahrenheit 451, maybe titled Fahrenheit-451GB-A32.

TFA is a bit obtuse about it, and assumes you've read their previous article on the topic, but from what I gathered, it's about whether the AAMVA would standardize cryptographically signing the barcodes of driver licenses to mitigate creating fakes in the US/CA. Cali showed it's entirely possible, but there's still no pressure for the standard to change across the board.

I was flagged by Facebook to provide a government ID years ago, presumably because I tend to use a VPN and never really provided any personal information. I imagine it doesn't take much on any of the other social media platforms to look too much like an anomaly before it warrants a suspension. The platform has nothing to lose since "personalization" and ad revenue won't be lost; in fact, you wouldn't want your personalization algorithm to be trained on such users.

I appreciate the update. I wish RCS was more open. I'm still astonished that it wasn't until 2024 that iOS adopted the protocol. Relying on everyone in your group chat to be all Android or all iOS (iMessage) wasn't helping anyone.

It will always be a cat-and-mouse game. I don't think it's possible to have any sort of visual that an algorithm can't be trained to recognize.

sidenote: did Adversarial Fashion stop selling the ALPR shirts/hoodies? That was my favorite thing to wear years back. And they're probably still effective at inserting noise.

Signal > RCS > SMS.

Just because RCS can't be done without metadata doesn't mean it's a fruitless endeavor. Why not have someone use RCS to then get on Signal? That is so much better than plaintext SMS to Signal.

And for most people, RCS between Android <-> iOS is the biggest advancement in encrypted communication for the layman that I've seen since Let's Encrypt.