It seems that people can't grasp the exponential rate of developments here. They're stuck in the GPT2 LLM narrative. Even with the amazing Veo 3 videos this week, people are still nitpicking and seemingly unable to remember the state of the art 2 weeks ago, 6 months ago, 1 year ago, etc.
I don't mean to say that scores on evaluation metrics will remain exponential but rather the developments, uses, integrations will (e.g., web search in ChatGPT), and people can't conceive or keep track of this, and therefore discussions on the area are always behind the times.
For example, I think it's inevitable now that TV/movie production will not exist as we know it in a short time, except as niche work, like fine art in the age of digital. It's also inevitable that fully personalised media will be predominant. I think this is obvious, but yet people are zooming in on the background of essentially perfect videos to spot minor and irrelevant coherence aberrations.
Nitpicking will also inevitably become a niche hobby, like people who complain about the colour grading on a movie remaster, while the rest of the world just watches the movie and doesn't notice or care about the issues.
You are saying all of this like it's a good thing.
I don't want to live in a world where reality doesn't exist (which tools like Veo3 will absolutely be used to distort the truth), but we are speed running ourselves to that destination.
Yes I agree, I don't think that it's good at all. It's just fascinating to me that people are criticising current LLMs using information they heard about LLMs 3 years ago, when right in front of their eyes are sci-fi-like results from the field.
> I don't want to live in a world where reality doesn't exist
Perhaps the end result is that the world turns away from digital completely and goes back to reality :) We see already that some universities are going back to written and oral assessments, for example.
> Should we focus on the 3 amazing papers this year, cited by hundreds, that resulted in clear progress or should we complain that 100 papers are useless?
Agree, and it seems that this is how fields naturally evolve anyway.
Overly pessimistic, and doesn't acknowledge that heads of steam only build behind promising findings, while the deficient (or 'fraudulent') work die on the vine, published or not. In other words the system tends to work.
Secondly, there are many ingredients required to successfully publish, communicate science, foster collaboration, etc., beyond technical brilliance. I'm sure we all know many technically brilliant people whose career never advanced because they lacked in some necessary area. People shouldn't be discouraged from improving in all areas because OP's delicate genius is offended by their technical ability.
Speaking of discouragement, it's a shame and a disgrace that you publicly called your colleague's work bullshit, including a first author that isn't yourself.
> Overly pessimistic, and doesn't acknowledge that heads of steam only build behind promising findings, while the deficient (or 'fraudulent') work die on the vine, published or not.
This might be true in hard sciences where a "head of steam" can only build based on real, replicatable results
But it's very common that public policy is proposed and adopted based on findings from soft sciences like psychology and sociology
If policy is adopted based on a research paper, I would count that as a "head of steam" being built.
And if that paper is fraudulent, then we are adopting well-intentioned policy on false pretenses
I did just mean AI and Computer Science per OP. By "head of steam" I mean to say that much research is built on it, think the likes of "Attention is All you Need". There isn't quite an equivalent of this in public policy in my experience.
Conversely, computer science/AI doesn't have an equivalent of the rigor that public policy research tends to go through. CS has e.g., benchmark datasets, typical evaluation metrics, but these are more like norms rather than requirements, whereas in public policy, instruments for validations are far more rigorously tested and enforced. Depending on the area.
I agree that outright fraud would be detrimental, but I think OP overblows this issue completely and should apologise to his co-authors.
It is tautologically obvious why expressing anger does not result in reduced anger in the long term, and why practicing subduing your anger is beneficial. Cathartic expression of anger is a long-debunked approach to coping with it.
> It is tautologically obvious why expressing anger does not result in reduced anger in the long term, and why practicing subduing your anger is beneficial.
It isn't obvious at all to me and doesn't match my experience.
On reflection, it depends on what was meant by 'express' in the previous post; if an expression is a constructive discussion then that may be good. If an expression is smashing things or shouting, that isn't good.
Aggressive expressions don't work long term because they reinforce the behaviour of outwardly and rapidly being angry.
Maybe you have some kind of healthy expression of anger, but multiple studies have shown that venting anger physically or verbally just exacerbates it, because it causes you to ruminate on the causes of the anger, which just feeds back into it [1]. Focusing on something else, like a logic puzzle, can diffuse it.
I thought we were talking about long term, not short term.
I agree, in the short term anger can have negative consequences. But longer term, as you ruminate on the causes of anger, you can be more objective and see that the reaction may have been an overreaction and unproductive. You might even have regret (I've certainly felt regret).
If I always subdued my anger, I'd probably end up like the dog in the this is fine meme (and I'm not an angry person!).
You're pointing out a single inconsistency and calling it a basic mistake, but are ignoring the thousands of advanced consistencies evident in the image, i.e., creativity that is beyond the ability of humans generally. Given this, the nose poking out is a trivial issue and isn't worth focusing on.
Would be interesting to see if the perceptive abilities of generative models are superior to human perception, when tested on optical illusions that humans are fooled by. E.g., do they correctly assess depth in a Ponzo illusion scenario
