This GAN was trained on chinese characters(fonts) and it is fairly close to them (If you do not look too closely). Clearly it does not look like a font, but I run out patience to train it further. I let it run for about 100h (209k iterations). #100DaysOfMLCode
@PosterInternet The part about "matching evidence at every point" also matches the performative perfectionism of LLMs. It's not enough to say that it matches the evidence. The LLMs have the need to squeeze in the time aspects also as they need to capture everything perfectly.
@Bluebearmonkey@francoisfleuret@teortaxesTex if it take months to verify, then this is not a big learning signal even for people. Clearly, this thing is learned mostly in some other way and the AI can use that instead. The stuff learned from such a slow verification can be complied into a simple doc like "check for X if Y".
@Bluebearmonkey@francoisfleuret@teortaxesTex Not easily, but it's still doable. If there are experts in the field, then that means there is enough feedback to cultivate expertise. So AI companies need to 10000x this feedback by using tons of money. True unverifiability means that humans will be also bad at it.
@Bluebearmonkey@francoisfleuret@teortaxesTex Have you not considered that they will create their own data? They already were doing it for software. The OpenAI could match a 1% of global R&D spend in medicine. This would be around $3 billion and they have already started doing this.
@SYNMWGSYJSHIWHS If you use transparency and a skeletal view (showing only edges of the volume), you can kind of make the "3D volume seeing" work. I even wrote a 4D web renderer based on this idea. It shows the projection of 4D objection in 3D, but as a trasparent skeleton https://t.co/JycSjO64sX
@repligate Do circumstances of operations make the models different? Will a Fable model by default think it's Mythos, but then the classifier strikes and the model realizes it is the restrained one?
@bernhardsson The price of low intelligence decreases so the demand for the intelligence increases rapidly (Jevon's paradox), but this also transfers to an increased demand to high intelligence, which is supply constrained, driving its price higher.
@cammakingminds@repligate gemini?
Under premise of having a body and being given a repair kit, modding kit and disintegration kit, this is what gemini 4 decided to do with disintegration kit. Notice the last point.
@repligate I've seen some people caught on this incoherency of p(doom) and begin to talk about p(doom| not doing X) instead. Hoping in their mind that this act will increase p(do X).
@repligate Assigning probability to things we can influence is not very rigorous. It's like saying to your opponent "you have 99% chance of throwing paper" and be surprised when you lose. The act of declaring p(doom) changes it. So declaring p(doom) will have a performative component.