There have been multiple accounts created with the sole purpose of posting advertisement posts or replies containing unsolicited advertising.

Accounts which solely post advertisements, or persistently post them may be terminated.

kromem

@[email protected]

This profile is from a federated server and may be incomplete. Browse more on the original instance.

kromem ,

That’s not what the scientists said, and technically given the performance metrics of this LLM are in line with GPT-3 and not meeting or exceeding the state of the art GPT-4, this model is entirely in line with what would have been allowable research and productization under the proposed measures they’d put forward (which only called out pausing advancing the state of the art).

kromem ,

Weird take.

What’s not to be trusted in Meta lagging behind their competitors in AI so they strategize open sourcing their model in order to subvert their competition?

This is pretty standard and transparent a move, and one that brings a multi million dollar trained AI to the masses under an extremely broad license that only requires companies with more than 700 million MAU to request permission for it and its derivatives.

Do you not trust websites built on React because it was made and open sourced by Meta?

At a certain point blind tribalistic takes around companies which are often much more nuanced crosses over into unproductive.

kromem ,

He followed it up by saying that if Ukraine doesn’t watch their step, Russia may need to send troops into Ukraine too.

kromem , (edited )

Not really.

The problem is with pushing out the edge of a normal distribution curve by the output regression to the mean.

Merely mixing what’s fed back in between AI generated and human generated would avoid this outcome, and arguably as long as the AI output was generally rated as better than the mean human output would even lead to recursive mixed training iterations improving over the original models.

This and the Stanford paper were problematic exclusively training on new AI generated output over and over, which increased median tokens or diffusions and dropped edges until you ended up with output that had overfitted lackluster discrimination.

The real takeaway here isn’t “oh noz, we can’t feed AI output back into AI training” but rather “humans in both generator and discriminator roles will be critical in future AI training.”

There’s been a recent troubling trend of binaryisms in the ML field as hype and attention has increased, and it’s important to be careful not to improperly extrapolate a finding of a narrow scope to an overly broad interpretation.

So yes, don’t go training recursively on only synthetic data over and over. But even something as simple as using humans upvoting or downvoting the generations to decide if you feed them back in or don’t (i.e. human discriminator and AI generator) would largely avoid the outcomes here.

Which means that human selection of the ‘best’ output from several samples for initial sharing and human rating of shared outputs for broader distribution is already cleaning up AI generations online enough that fears of ‘poisoning’ the data as suggested here and in the Stanford study are almost certainly overblown.

Edit: From section 5 of the paper it even addresses some of this.

One might suspect that a complimentary perspective to the previous observation—that fresh new data mitigates the MAD generative process—is that synthetic data hurts a fresh data loop generative process. However, the truth appears to be more nuanced. What we find instead is that when we mix synthetic data trained on previous generations and fresh new data, there is a regime where modest amounts of synthetic data actually boost performance, but when synthetic data exceeds some critical threshold, the models suffer.

kromem ,

I think people may be confused about what this is saying, so an example might help.

Remember when Stable Diffusion first came out and you could spot AI generated images as if they killed your father and should be prepared to die?

Had those six+ digit monstrosities been fed back into training the model, you’d have quickly ended up with models generating images with even worse hands from hell.

But looking at a study like this and worrying about AI generated content poisoning the Internet for future training is probably overblown.

Because AI content doesn’t just find its way onto the web directly the way it is in this (and the Stanford) study. Often a human is selecting from multiple outputs to decide what to post, or even if it is directly posted, humans are voting content up or down based on perceived quality.

So practically, if models were being trained recursively on popular content online that had been generated by AI, it wouldn’t be content that overfits spider hands or striped faces or misshapen pupils or repetitive text or broken links or any other number of issues generative AI currently has.

Because of the expense in human review of generated content this and the previous paper aren’t replicating the circumstances that real world recursive training of a mixed human and AI Internet would represent, and the issues which arose will likely be significantly muted in real world circumstances outside the lab.

TL;DR: Humans filtering out six fingered outputs (and similar) as time goes on is a critical piece of the practical infrastructure which isn’t being represented, and this is simply a cautionary tale against directly piping too much synthetic data back into training.

kromem ,

Kind of. It’s more complicated (for example in 5.3 of the paper it discussed how a little bit of AI generated data mixed with new human data actually improved outputs over only human data).

Under the hood it has to do with sample diversity. The more apt comparison than Xerox (where it’s lowering quality because of necessary fidelity loss) is genetic reproduction.

Even if you have great genes, after a few generations of sex with siblings you’re going to end up with messed up kids.

But if you have great genes, a small degree of over-representation of your genes in a larger mixed gene pool would be better than only new random genes.

This is basically saying that AI models shouldn’t have incest levels of recursion moreso than it is saying that they shouldn’t have ANY recursive data (which would be the case if it worked like a Xerox).

kromem ,

In most cases. For a banjo playing AI, this might be desirable though.

kromem ,

This is effectively the same issue as what’s going on in the paper and why I used it as an analogy.

Much like how maladaptive genes can piggyback on good genes, but then become overrepresented in an endogenous sample pool, small errors in the diffusion model end up exacerbated through subsequent generations without enough difference in ‘genes.’

There’s definitely good ‘genes’ in the diffusion model, but it’s not the frequency or abundance of the good genes that’s at issue, but the frequency of maladaptive traits in subsequent generations. Much like the issues with human reproduction.

kromem ,

Well, the ideal would probably be to train a discriminator based on human ratings of generated outputs.

Take generation 0 (G0), produce output which is accepted or rejected based on humans, train a discriminator to predict those ratings off output, and then use the combined accepted outputs from humans and trained discriminator to train G1.

Repeat again for G1, G2, G3, etc.

My guess would be that the end result would continue to get better and better rather than worse.

The problem is if the diffusion model can’t properly reject weird hands or pupils, those magnify in subsequent rounds.

But there’s likely adaptive and maladaptive tendencies in the diffusion model, and adding a halfway decent filter between human selection and synthetic selection of outputs separate from the diffusion model itself would effectively curb the magnification here.

kromem ,

digitally

I see what you did there…

kromem ,

fivehonor

Always think ahead.

kromem ,

It’s also a common misconception that Everett’s many worlds involves an infinite number of universes.

And that it involves multiple outcomes for macro objects like lottery balls.

It only means multiple ‘worlds’ specifically for quantum outcomes, so in OP’s case their winning or not winning the lottery would need to be dependent on a superposition of quanta (i.e. Schrodinger’s lottery ticket).

And given the prevailing thinking is that there’s a finite number of quanta in the universe, there cannot be an infinite number of parallel worlds. (There could only be an infinite number of aggregate worlds if time is infinite and there’s perpetual quantum ‘foam’ in its final state perpetuating multiple possibilities).

The theory is much less interesting than is often depicted in mass media (though as of recently is a fair bit more interesting given the way many worlds as a theory would mirror what backpropagation of the physical universe might look like).

kromem ,

In a literal sense, assuming the theory that consciousness in some way depends on quantum processes is correct, this is the proper interpretation.

Lottery balls being picked seems very unlikely to be dependent on a superposition.

But (a) choosing to buy a ticket, and (b) what numbers you choose both plausibly could if the above assumption is correct.

So not only would other yous be buying tickets in other worlds, they’d be buying many different numbers in many different worlds, even if the you in this world wasn’t buying any tickets at all.

And even if the you in this world was now so strongly against the lottery that no future ‘branch’ of you would ever buy a ticket regardless of the degree to which a superposition might influence your decisions, the many yous from childhood would be so variably influenced in different ways from others around you from birth to now that there might be other parallel yous who superstitiously buy every ticket.

Even in terms of number selection - if the you here might choose the birthdate of a spouse or children as the numbers, yous in other worlds might have different spouses or children to choose numbers based on.

Many worlds is a rather boring theory unless also entertaining it with the notion that - like how birds navigate - our decision making somehow depends on quantum effects.

kromem ,

He really seems to be losing it.

I’m not sure his being ignored is that respectful or much of a political courtesy at this point.

The claims are getting wilder and wilder.

First it was alien light sails as the most plausible reason for an oblong object traveling into the solar system, and now spherical metal in a meteor chemically consistent with the “sky iron” of antiquity can only be part of an alien spaceship?

He’s going to erode any legacy he once had by the time he’s done chasing windmills.

kromem ,

TIL Lemmy title length limits coupled with broken links are a poor match.

kromem ,

I don’t think it will render us extinct.

Oh, it probably will, though the memory of us may live on after that.

In fact, arguably it happened long ago, and we’re currently in an echo of the past in a very immersive history lesson simultaneously teaching the grandeur and folly of humanity.

kromem ,

There are already laws regarding producing works too similar to copyrighted material.

Production is infringement, not training.

If I feed all of Stephen King into a LLM such that it learns what well written horror narratives looks like, and it produces a story with original and different plot elements distinct from copyrighted works, that’s fine.

If it starts writing about killer clowns thwarted by child orgies in the sewers then you might have an infringement problem.

And ironically, the best tool for protecting copyrighted material from infringement is going to be…LLMs (acting in a discriminator role comparing indexed copy to protected works).

If ‘training’ ends up successfully labeled as infringement we’re going to end up with much worse long term outcomes in jurisdictions that honor that ruling than we otherwise would.

This is the longer tail masses adopting MPAA math in trying to tally potential losses and in the efforts to protect the status quo are shooting themselves in the foot on laying claim to the future of the industry, inevitably leading to being left out of the next round of growth.

Also, from an ‘infringenent’ standpoint it just means we’ll see less open models and more closed ones which ends up using other jurisdictional models to launder copyrighted materials for synthetic training data.

This is beyond dumb.

kromem ,

In part it feels that way because you, along with pretty much every other human being online today, have been propagandized for decades now with SciFi inspired from dystopian futurist predictions around AI which are almost universally clearly obsolete and misinformed by now, but still persist due to anchoring bias.

AI trained to predict collective human thought ends up replicating quite a lot more than most people thought would be possible in our lifetimes.

And yet when it exhibits emotional intelligence it’s called creepy, when it exhibits above average reasoning capabilities it’s called scary, and when it displays a potential for automating large swaths of busywork for most humans it’s called a threat.

Next to no one I see discussing the topic is considering the opportunity costs here, as the media influence on perceiving AI as ‘other’ is so pervasive that most humans fall into treating it like a monkey from another forest competing for bananas rather than treating it like a much better stick.

TIL of The Business Plot of 1933, a failed attempt to overthrow FDR and install a dictator. Led by a covertly bankrolled Wall Street coalition of affluent businessmen (explorethearchive.com)

In 1933, a coalition of businessmen dissatisfied with FDR’s economic policies surreptitiously planned a coup. The failed Business Plot aimed to depose FDR and install a dictator.

kromem ,

Particularly the part where no one was held accountable…

kromem ,

Disagree with a lot of the stances here.

Meta has been hit hard by a series of failures over the last decade.

Continually missing first place, the company has broadly pivoted towards more open partnership in order to boost their offerings.

For example, their LLM weights being released to researchers when the product was clearly behind OpenAI and even Google.

Playing nice with federated networks fits into this.

Meta is betting that open platforms do well enough to corner a non-insignificant part of the market and are hoping to leverage compatibility with it in order to protect and differentiate their fledgling product from competition.

None of this means they aren’t still going to try to siphon every detail they can to maximize ad revenue for users.

But they aren’t trying to kill or sabotage the fediverse (which they rightfully don’t seriously see as competition in itself). They are hoping it is successful enough that it helps give them an edge against walled garden networks backed by competitors’ money.

In general, expect to see more openness from Meta in the coming years for much of what they do. They finally realized they aren’t Apple and can’t get away with siloing their products within the market.

kromem ,

Per the article Meta claims no ex-Twitter employees worked on Threads.

kromem ,

What’s the funniest legitimate non-joke standardization detail you’ve come across?

kromem ,

It’s definitely satire.

kromem ,

This is fine…

kromem ,

Well you can take the knowledge that Lemmy.world grew 60% following it, look at current numbers for the server, and know at least around 60% of that number has shifted some of their media habits away from Reddit.

But the full picture is unknowable outside Reddit corporate.

Probably more than spez was anticipating though…

kromem ,

17 years was pretty close to that…

kromem ,

No tech burst.

It’s just a cold recession. No one is admitting it, including consumers who keep spending away savings.

But companies are aware of it enough they are tightening purses preparing for harder times ahead.

Of course, it’s a self-fulfilling prophecy.

If everyone makes their products worse chasing this quarter’s dollar, and people leave, those companies are going to have a harder time.

Especially as it becomes easier and easier to compete against them at scale.

Just wait until new feature requests and bug reports for something like Lemmy can be handled within moments by AI at dirt cheap pricing.

A very interesting future awaits around the bend.

kromem ,

A thousand times yes.

Federation should ideally work not only at a server level, but a community level.

If there’s three different /c/aww communities across different servers, their respective mods should be able to join into a meta-community.

There is the potential for a less moderated child community to end up polluting the meta community, so membership should probably be able to be revoked by majority mod vote.

But aggregating a meta-community across each sub community could allow for a much more seamless growth between servers and avoid the duplication that’s inevitable by putting the onus on the client-side.

kromem ,

In Jesus’s time, there were three different sects of Judaism.

One of them, the Sadducees, allegedly believed there was no life after death and that God didn’t care at all about what people did or didn’t do.

Their answer to your question of following the law is perhaps the most interesting.

They believed that what was put forth as laws were a gift to humanity and that following them inherently led to a better life in the here and now.

While I don’t personally see all of the laws put forward as beneficial, there are certainly instances where that makes a lot of sense.

For example, look at the full version of one of the commandments:

Honor your father and your mother, so that your days may be long in the land that the LORD your God is giving you.

  • Exodus 20:12

Would following a commandment to take care of your parents in their old age (‘honor’ here comes from the word for https://biblehub.com/hebrew/3513.htm) benefit you by setting an example such that when you are old that you too would be taken care of?

This was almost like social security in antiquity, much like the Sabbath was one of the first labor laws preventing working anyone more than 6 days in a row.

There’s something called the overjustification effect, where when you introduce external reward systems for something intrinsically rewarding people over focus on the external and forget the internal benefits. I think a number of religions have serious issues with that.

There’s even a certain irony in Job, named ‘persecuted’ in Hebrew because even though he lived a good life he experienced suffering which it explains by the intervention of Satan, today in the most common language among believers being the exact same word as “to do a task with the expectation of a reward.”

Maybe we’re too focused on the rewards.

  • All
  • Subscribed
  • Moderated
  • Favorites
  • •
  • random
  • lifeLocal
  • goranko
  • All magazines
    •