andy99

Yes came here to say the same “look at us, our AI is also dangerous! Please ban our competition”

I would have said 0x to infinity-x. Some stuff I think it would be a good idea and waste a bunch of time and then give up and just write it myself. And like you say, there’s lots of stuff I would just never have done, basically any kind of front end for example, if I didn’t have an LLM to do it.

I also do mostly research or one-of script development, for which I find LLMs less useful, as I want simplicity and I want to understand exactly what is happening. The only time I find LLMs help really there is if there’s some part I can abstract away like writing and interface to an api or something.

So, there is no subliminal learning in this situation, under what conditions would we expect it. I find a transfer attack to be a bit far fetched but it’s definitely interesting.

If we trained from random initialisations on DeepSeek output (that didn’t explicitly contain the political questions) we would expect transfer? And if we fine tuned a model pretrained elsewhere on Deepseek output?

What is the line?

I didn’t see a gguf yet, going to be most interesting if there’s a quant that fits nicely into about 90FN so it can run in 128GB unified memory.

It’s 12B active so should hopefully be pretty fast at 2 bit quant if it fits, as in the ratio of total to active params is big compared to e.g. Qwen 3.5 122A10 which seems more commom

Edit: and it’s out - excited to try https://huggingface.co/unsloth/Inkling-Small-GGUF

There’s such an opportunity, it has never been easier to differentiate oneself now, all you have to do is ignore whatever online advice or LLMs say and do your own thing, it’s sad that almost nobody does it.

Depends what you mean by meal, the parent didn’t like pizza slices but I believe one of the thick granny slices at Outta Sight pizza is still under $10, I’d consider that a lunch.

It’s a fair point that generally it’s focused on single items rather than say a main, side, and drink for $10.

Cheaper to just buy a 60 OZ Kirkland vodka and invert it into the existing water cooler come party time.

Or these which sadly aren’t available anymore: https://www.cbc.ca/news/canada/edmonton/how-a-4-litre-jug-of...

Looks like they updated the article to say 20,000 litres of additional fuel, it wasn’t like that before as what I posted above was a copy paste.

Edit: it actually says at the bottom:

  (This story has been refiled to add the dropped word 'additional' in paragraph 8)

I’m glad he feels this way and is standing up for this, even if I didn’t agree with lots of metas practice in the past (they are the same company that wouldn’t let people have access to the original llama model - though to be fair Mark might have sanctioned the leak that ultimately kicked off the modern local AI ecosystem, you can’t say Meta didn’t play a major role).

I find it funny that this is basically a rebuttal to Dario’s thing, only because the Anthropic post was ostensibly saying “we too support open AI models” but absolutely nobody believes it to the point where someone like mark would come out and rebut the popular interpretation, ie that Anthropic wants to prevent access to open weight models.

> The first aircraft, which carries 20,000 litres of fuel and can seat 238 passengers

Pretty sure a zero is missing right? Wikipedia says normal A350 capacity is ~150k litres if I’m looking at the right thing.

200k litres means almost 1k per person, literally a ton of fuel.

Why would it obviously give attackers the upper hand? If we accept the premise that AI helps find vulns more efficiently, shouldn’t this be at least equally helpful in securing against attacks and make that increasingly efficient?

Right, it’s really a foundation of post enlightenment society. These people, Dario et al, would have wanted to ban sharing information about calculus or Newtonian physics because of “safety” - it’s trying to go back to the dark ages where only priests could read

> If "Mythos-class" models are such a problem, then... why not just let it fix everyone's code?

Because it doesn’t really confer the advantage they claim, especially compared to e.g. paying an equivalent amount of money to do traditional security scanning.

It’s much better to play of FOMO and hype than to let everyone use it and be underwhelmed.

Anyone who calls it “safety” probably has a certain world view and is more aligned with the big 2 (and stuck in 2023).

There is a growing industry of commercially focused risk evals that has a broader customer base.

You forgot “what is the definition of ‘sufficiently capable’”. Presumably it’s anything that competes with Anthropic. If they’re around in a year, presumably they won’t care about Fable level and will only think that whatever competes with Claude 7 or whatever needs to be restricted.

> I want AI as an assistant that helps me make decisions, and ensures that I'm in the driver's seat.

Right now we’re still in the benchmaxxing phase, I feel like most AI models of combination of models and harness are not really a “product” yet in the sense of having a particular and opinionated UX engineered to be helpful at a specific task, there much more still in a race to be the smartest.

There’s a lot of room for diversification of the actual character of this interaction, for example in the direction you mention. I think once that starts we’ll get into the real “useful” phase, right now we’re really just showing off and exploring how much it can do, even if it’s incidentally really useful in some cases.

Personally I agree with the parent post, I hate it when AI tries to lead me “would you like me to ...” I’d much prefer something awaiting my instructions than wasting tokens tryitn or steer me to doing something inevitably dumb.

Definitely agree, at a big company anyway there is nothing to be gained by sending these, or by any communication not directly related to the process. It doesn’t come across as genuine, it’s neutral to slightly annoying and has no impact on anything, it’s basically a chance at rubbing someone the wrong way with no upside.

I think there’s a kind of laundering that goes on. Like if leadership just told devs to go build something (gave them a prompt) and the devs picked some defaults, leadership wouldn’t like it, they’d want some different interpretation of the prompt, there’d be lots of back and forth.

Somehow when it’s the LLM that makes the choices, everyone is impressed with what AI did. It’s really just whatever defaults have been trained in, but somehow we’re ok with this.

Part of it is better marketing and communication. Basically the defaults of OpenAI and Anthropic are better than what a random dev will pick. But it’s not really that natural language is a better interface, it’s more that having “AI” for now somehow intermediates responsibility so everyone is ok with what it picked, when they probably wouldn’t accept the same if the internal team came up with it. It’s not too different from hiring consultants.

Nothing personal, if someone told me they’d generated a 15 page paper based on a two week LLM discussion I wouldn’t want to read it at all. My prior is that it would be the result of some sycophantic echo chamber that the LLM had reinforced.

Edit: missed your insult in the first read, I wouldn’t have replied if I’d realized the first comment was in bad faith but will leave the above for posterity.

Just send me the prompt if you’re going to write with AI. No point in trying to convince me it’s great, the marginal cost for me to re-run the prompt is zero, just send the prompt - it’s like unzipping a file before you send it to me, it doesn’t help, it just decreases information density.

#1 in a very close race is way less useful when you have to walk on eggshells to avoid triggering censorship (“safeguards”) that either refuse or knock it down to another model. I’ve almost completely stopped using Claude (except some legacy workflows) for this reason, reliability matters more than scoring 61 instead of 57. To me Claude is the most compromised and unreliable model (between the censorship and the id checking - which I have not experienced personally), it’s not worth whatever slight benchmaxxing they did for the latest release.

You should assume the weights have been poisoned, the model has been prompt injected, etc and design software around it accordingly. You can’t really trust the model and it makes sense to always have controls around it and not depend on it behaving a certain way.

I’m aware this doesn’t happen, just saying. There are more examples of a model randomly hallucinating and deleting something than of a deliberate compromise.

The open weights models would be even more effective by incentivizing people and companies to switch jurisdictions. It might even break SFs monopoly on AI, or at least weaken it, everybody isn’t there exclusively to funnel money to openAI and Anthropic.

It would also boost research in non US jurisdiction. Who is going to be wooed by “come to our lab where you’re only allowed to work with closed models!”

I’m in a bubble so I may not see the big picture, but I feel like these guys are completely destroying any kind of credibility they had.

Anthropic especially has this arrogant, “we’re smarter than you”, communication style, and acts like they’re serving some higher purpose (to the point of absurdity in some cases like “model welfare”) - I think part of all this might be the EA roots but I’m not sure. But then they get a little competition and suddenly they’re a little tattle-tale running to the government looking to artificially preserve their advantage. I’ve seen nobody who doesn’t see the naked self interest. There’s no mission, not higher purpose, not even competitive spirit, just some people that were on top and can’t handle the idea that they weren’t as special as they thought. It’s going to bite them in the long run, we’ll probably get some politically motivated legislation, but the cats out of the bag, they’ve lost all credibility and don’t even have a uniquely competitive model anymore, they’ll end up as just another also-ran.

Yeah this makes it unreadable for me, immediately distracted by how stupid it is and lose interest in the article. As someone else noted, it’s people who think they are writing for an “executive” audience that have never been within 100 feet of an executive.

Rumored M5 ultra bandwidth is apparently 1100 GB/S (M3 Ultra is something like 880). This is 276 GB/S so they’re not really in the same league unfortunately. This will be considerably cheaper though.

> RAM prices weren't so f*cked I think we'd be seeing 256GB and even 512GB unified memory systems becoming quite common

I agree but unfortunately that speaks to the depth of demand right now. Even as more capacity comes online (if it does) I suspect it will all get used up without prices needing to drop. Hopefully on the other side of this 10 years from now it means RAM will be insanely abundant and we’ll all have machines with a couple TB as the same inflation adjusted price as say 16GB a few years ago. Considering single digit MB of RAM was the norm in the early-mid ‘90s its not so far fetched.

I use it exclusively as a server for LLM inference - first more for research but have been using it for coding now that there are sufficiently capable models that run on. For that I’m very happy.

Try poolside that came out yesterday https://news.ycombinator.com/item?id=49004937

or Qwen 3.5 122B A10B, both use more memory and still have experts sized for decent speed at the 395’s memory bandwidth at 4bit quantization

I have a framework desktop w/ 128GB that I bought last Christmas and if I’m looking at it right it costs $2000 (CAD) more now because of the RAM shortage (and in any event is apparently out of stock). Would love to have 192 GB but I’m not sure I can justify buying another.

Does anyone have a sense of how this might progress, e.g if I can get a 256 or 512 GB in a year if I wait. In any case I’m jealous this exists and I don’t have one.

One last thing, I assume this isn’t exclusive and there will be other builds with this same config same as current Strix Halo?