Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

The Llama 4 herd

ai.meta.com

Post reply on HN
Print view View original

591–600 of 695 posts

Re: The Llama 4 herd

#591

Post by LeafItAlone »

hitekker wrote: post #569
LeafItAlone wrote: post #291

Earlier quoted context omitted.

I hope on HN that we hold ourselves to a higher standard for “it’s been true from the beginning” than a vague recall of “My son assessed it for a class a few years ago” and not being able to reproduce.

The GP put in the work to verify his own memory, after acknowledging the gaps. And then you belittled him. He met the “standard” or guidelines of our community in a way you have not.

>The GP put in the work to verify his own memory, after acknowledging the gaps.

The original claim didn’t say anything about it being the experience of their son for specific questions about unions. It was much broader than that. And at least partially inaccurate, given the stated result isn’t even one of the results.

>And then you belittled him.

If asking for a higher standard of evidence for a broad claim than referencing a previous experience and then trying again, but not even sharing the link from a tool that makes it easy to share the conversation from, is considered belittling, then maybe the castrations going on in these models is the right way to go for this crowd. I, personally, aim for a more truth-seeking standard.

>He met the “standard” or guidelines of our community in a way you have not.

These are two different things, and you clearly understand that but are intentionally conflating them. Regardless, if this is where are, maybe HN no longer is the place for me.

Re: The Llama 4 herd

#593

Post by LeafItAlone »

mike_hearn wrote: post #572
LeafItAlone wrote: post #291

Earlier quoted context omitted.

I hope on HN that we hold ourselves to a higher standard for “it’s been true from the beginning” than a vague recall of “My son assessed it for a class a few years ago” and not being able to reproduce.

That claim isn't something Peter made up, it's the claim made by Meta's own researchers. You're picking an argument with them, not HN posters. Anyway it's trivially true. I think most of us remember the absurdities the first generation LLMs came out with. Prefering to nuke a city than let a black man hear a slur, refusing to help you make a tuna sandwich etc. They were hyper-woke to a level way beyond what would be c…

Peter’s claim I was asking about was one about being labeled as something via a Pew research or similar survey. And the response I got was about their personal experience asking a questions about unions. Do you think that those are the same claims and equivalent?

>Prefering to nuke a city than let a black man hear a slur, refusing to help you make a tuna sandwich etc. They were hyper-woke

On its own, all this tells me is that the non-human, non-conscious tool was programmed specifically to not say a slur. To me that seems like something any reasonable company trying to create a tool to be used by business and the general population might incorporate while it is still learning to otherwise refine that tool.

And I took the Pew survey mentioned above and it didn’t ask me if I would say a racial slur.

Finally, if anyone, from any point on the political spectrum, thinks that a tool being limited to not respond with racist terms, is a reflection of its overall political leaning, I suggestion you look inward.

Re: The Llama 4 herd

#594

Post by LeafItAlone »

vessenes wrote: post #338
LeafItAlone wrote: post #291

Earlier quoted context omitted.

I hope on HN that we hold ourselves to a higher standard for “it’s been true from the beginning” than a vague recall of “My son assessed it for a class a few years ago” and not being able to reproduce.

I literally went back to the oldest model I could access and hand verified that in fact it does what I described, which is lecture you if you don't like unions and goes sweetly along if you do like unions. I feel this is a fair and reasonably well researched existence proof for a Saturday afternoon, and propose that it might be on you to find counter examples.

You made a claim about political surveys, and linked one in particular, providing a labeling of the tool.

Your follow up response did not reference any of those surveys and did run through the types of questions on those surveys. You apparently only did questions about unions.

Is that what you would fair and reasonable?

Re: The Llama 4 herd

#595

Post by gesman »

aimanbenbaha wrote: post #384
qwertox wrote: post #63

Earlier quoted context omitted.

Llama 4 Scout, Maximum context length: 10M tokens. This is a nice development.

I don't think RAG will survive this time

RAG gets bigger as everyone else gets bigger. Flooding prompts with garbage is not a sound strategy...

Re: The Llama 4 herd

#596

Post by darksaints »

huijzer wrote: post #517
vmladenov wrote: post #480

Earlier quoted context omitted.

It is not and has never been half. 2024 voter turnout was 64%

Sure and the voters who did not participate in the election would all have voted the democratic party. I think the election showed that there are real people who apparently don't agree with the democratic party and it would probably be good to listen to these people instead of telling them what to do. (I see the same phenomenon in the Netherlands by the way. The government seems to have decided that they know better…

We have an electoral college that essentially disenfranchises any voter that is not voting with the majority unless your state is so close that it could be called a swing state. This affects red state democratic leaning voters just as much as blue state republican leaning voters…their votes are all worthless. For example, the state with the largest number of Trump voters is California, but none of their votes helped decide the election because California as a whole chose Kamala. And let’s not forget that we have one of the largest metropolitan areas and several territories that legally can’t vote for the president or have representation of any kind in the federal government.

A lot of people try to claim the popular vote as a measure of who won over the country’s opinion, but that’s simply not possible because the incentives and structure of the electoral college make it impossible to use as a measure of that.

The best we have for measuring who won over the hearts and minds of the country are polls. Polls are full of faults, but if executed correctly, they don’t disenfranchise by structurally underrepresenting entire classes of people. And the results of polling over the last hundred years suggest that Americans generally lean to the left of how our votes play out. You can call bullshit all you want on that, and there are very fair criticisms of polling as a measure of who would vote for what, but the fact of the matter is that the Republican Party knows this. That is why they oppose any attempt to get rid of the electoral college and also why they refuse to entertain enfranchisement of DC and US Territories. They know they’ll lose.

Re: The Llama 4 herd

#597

Post by nwienert »

LeafItAlone wrote: post #295
nwienert wrote: post #255

Earlier quoted context omitted.

China, Africa, India, Vietnam, Philippines, Russia? Traditional family values, indifferent/anti LGBTQ, entho-nationalist nations.

Ah, yes, the often used, peer-reviewed, expert-backed source of just listing random things. Thank you.

If you were looking for truth you wouldn’t reply like this. I’m not going to do an hour of work to carefully cite this for you, but it’s true nonetheless.

Re: The Llama 4 herd

#598

Post by CooCooCaCha »

KingMob wrote: post #505
CooCooCaCha wrote: post #474

Earlier quoted context omitted.

There is no rhetoric here, it’s just literal truth. There is no implication of equivalence or any statement about the value of objective truth. Any position is a bias. A flat earther would consider a round-earther biased. That doesn’t make them equal positions.

> truth itself is a bias Ehh, bias connotes unfairness, but espousing the truth should be considered the fairest position. In statistics, bias literally refers to an inaccurate distortion of results. I get what you're trying to say, but I don't think it's a useful definition of bias.

Truth isn't fair because reality isn't fair. Dice are considered "biased" if not all sides have equal probability, even though that's the "truth" of the die.

Re: The Llama 4 herd

#599

Post by brookst »

jamesblonde wrote: post #541
ramshanker wrote: post #320

Earlier quoted context omitted.

I have a gut feeling, next in line will be 2 or more level of MoE. Further reducing the memory bandwidth and compute requirements. So top level MoE router decides which sub MoE to route.

The solution to all problems in computer science is add a new level of indirection (or abstraction).

Except when the solution is to collapse abstraction in the name of efficiency.

Re: The Llama 4 herd

#600

Post by lxgr »

igravious wrote: post #581
phire wrote: post #409

Earlier quoted context omitted.

It's not even per token. The routing happens once per layer, with the same token bouncing between layers. It's more of a performance optimization than anything else, improving memory liquidity. Except it's not an optimization for running the model locally (where you only run a single query at a time, and it would be nice to keep the weights on the disk until they are relevant). It's a performance optimization for lar…

> making the load distribute evenly, nothing else. so you mean a "load balancer" for neural nets … well, why don't they call it that then?

Some load balancers are also routers (if they route based on service capability and not just instantaneous availability) or vice versa, but this kind isn't always, to my understanding: The experts aren't necessarily "idle" or "busy" at any given time (they're just functions to be invoked, i.e. generally data, not computing resources), but rather more or less likely to answer correctly.

Even in the single GPU case, this still saves compute over the non-MoE case.

I believe it's also possible to split experts across regions of heterogeneous memory, in which case this task really would be something like load balancing (but still based on "expertise", not instantaneous expert availability, so "router" still seems more correct in that regard.)

Post reply on HN