These things with <think> and </think> tokens are actually trained using RL, so it's not like GSM8k or something like that where you just train on some reasoning.
It's actually like QuietSTaR but with a focus on a big thought in the beginning and with more sophisticated RL than just REINFORCE (QuietSTaR uses REINFORCE).
WW3 would not be sanctions rivaling those Russia are under. WW3 would be zero trade, submarines sinking the shipping etc. What Russia is under is just sanctions, not real war measures.
Maybe. It would certainly be good for Intel if Taiwan's and South Korea's microchip manufacturing were destroyed. They'd have a near-total monopoly, like int the 90s.
But it would be incredibly stupid strategically.
I think only sensible path is to give Taiwan nuclear weapons. It removes all the dangers for ever.
I don't see how this can be resistant to manipulation.
I don't see how it can be resistant to foreigners flooding the system with their text or to people creating multiple accounts to have an outsized influence.
I think if those problems could be solved it would be the best thing in the world, but the funnel procedure and the meritocratic promotion don't matter in the absence of those things. There are also many ways to do what this system proposes. A threaded discussion system is also good.
The problem is instead the lack of manipulation resistance from multiple accounts and people who shouldn't have accounts at all.
I agree that there's both cynicism and denial, but when I've explained my views I have usually been able to get through to the complainers.
Usually my go-to example for LLMs doing more than mass memorization is Charton's and Lample's LLM trained on function expressions and their derivatives and which is able to go from the derivatives to the original functions and thus perform integration, but at the same time I know that LLMs are essentially completely crazy with no understanding of reality-- just ask them to write some fiction and you'll have the model outputting discussions where characters who have never met before are addressing each other by name, or getting other similarly basic things wrong, and when something genuinely is not in the model you will end up in hallucination land. So the people saying that the models are bad are not completely crazy.
With the wrong codebase I wouldn't be surprised if you need a finetune.
Yes, that is why (most?) anarchists consider property that one is not occupying and using to be fiction, held up by the state. I believe this includes intellectual property as well.
You can't protect the environment without research. Without research you can't know what's dangerous.
Even in tiny countries, for example Sweden, when we notice a statistical uptick of health problems in a particular area, we have government organizations that go there to investigate and figure out the cause.
It's also worth noting that the mainstream German parties have actually supported ethnic cleansing abroad, in Nagorno-Karabach, etc. while the AfD opposed those things, so I don't think it's very clear cut that AfD is the dangerous party.
Personally I find the political inclinations of the German mainstream parties to be what appears to be dangerous, since what they're doing actually led to a large number of deaths and a large number of people being displaced, to the loss of sovereignty and to the expansion of a dictatorship.
I see very little difference between Aliyev and Hitler, and he is still tolerated (in fact, my perception of Azerbaijani hate attitudes is that they're actually more extreme that the Nazi hate attitudes, i.e. simply going further, the systematic teaching of this hatred to even younger children that the Nazis primarily targeted, etc.).
The thousands of peace-keeping troops in Armenia/Azerbaijan that looked the other way were not German, but Russian. By the way, both Russia and Armenia are members of the CSTO military alliance[1], while Azerbaijan is not.
Yes, I know that the Russians looked the other way.
However, Baerbock has absolutely monstrous statements and the German gas-guzzling contingent are the obvious culprits for the EU partership agreements with Azerbaijan and for the incorrect statements treating this whole thing as somehow restoring Azerbaijani territorial integrity and the numerous statements by the EC falsely claiming that Armenia had attacked Azerbaijani (i.e. these 'we call on both sides...' in the aftermath of Azerbaijani attacks). Furthermore, it is German influence on the EC that made the implementation of the ICJ decision subject to negotiation, and it is likely German influence on the EC that forced the agreement whereby mine maps were exchanged for the release of PoWs. These mine maps naturally enabled further Azeri attacks. It is also very apparent that there was government influence on media organizations to not report the starvation in the NKAO beginning after the Azeri blocking of the Lachin corridor-- for example in Sweden state television reported nothing, and reported of the ethnic cleansing itself only that 'Armenian separatists have agreed to leave Azerbaijan'. This shows co-ordination between Swedish government, Swedish state television (SVT) and Turkey or Azerbaijan, indicating a secret deal either for the sake the Swedish NATO entry or on the EU level. Certain phrases 'lightning offensive' which sound decidedly Turkish are also repeated in many newspaper articles, indicating a larger deal rather than something specific to Sweden.
The CSTO is absolutely irrelevant, as everybody who matters in any way knows completely. France would not be selling weapons to Armenia if they believed that their CSTO membership were relevant.
There were excellent opportunities to intervene even as the Azeri troops were rolling down the Agdam road to Stepanakert, and it's very unlikely that the Germans were unaware. SAR satellite imagery of the region is so readily available that unclassified images can be obtained on a commercial basis and I'm not even sure it was cloudy.
For the most part I can actually tell, but it also depends on the style of the art. A lot of anime-inspired digital images are immediately obvious - AI tends to add quite a lot of "shine" to its output, if that makes sense. And it's way too clean, sterile even. And it all looks the same.
But when the art style is more minimalist or abstract, I find it genuinely difficult to notice a difference and have to start looking at the finer details, hence the mental workload comment. Often times I'll notice an eye not facing the right direction or certain lines appearing too "repetitive", something I rarely see in the works of human artists. It's difficult to explain without actual inage examples in front of me.
That's a kind of survivorship bias though; sure, 100% of AI art that you identified as AI art turned out to be AI art, but what about the ones you didn't realise were AI art? It's the unknown unknowns. Was this comment written by AI?
The success rate of AI evading detection will only increase; the issue with "too many fingers" was solved years ago, and there's probably companies actively working on avoiding AI detection already. And on detecting it. It's the new spam / anti-spam, virus / anti-virus arms race.
There's no reason to say that it's raw footage and then put it through a video editing program, so I'd say what we know now is evidence of it being deceptively manipulated.
But I see you're quoting the article, so I guess you're not intending to say that it wasn't and that Wired are instead saying something ridiculous?
Easier than that, I do not often publish on hacker news, only when I find something that could be interesting to read the comments on. And I though the field there was for a subtitle, rather than that it was published itself as a comment.
If you ask me, I am on the "conspiracy" team. Too much coincidences, to much interests, and very few plausible explanations.
It's actually like QuietSTaR but with a focus on a big thought in the beginning and with more sophisticated RL than just REINFORCE (QuietSTaR uses REINFORCE).
reply