95

[95 @cellestialsea]

@honorablepicnic @janleike "Models are getting smarter and scoring better on alignment at the same time." is definitely a fitting start for a horror story

[95 @cellestialsea]

@Aygitkocaman2m @emrrnr @ubqtn_e3 @oktayarslan Tabiki öğrenciden öğrenciye değişiyor, herkes için geçerli değil bu hesap, örneğin 15-20 bin civari için çok daha iyi bence tıp, ama iyi bir öğrenci için fazla uzun çok fazla ek faydası olmadan

[95 @cellestialsea]

@Aygitkocaman2m @emrrnr @ubqtn_e3 @oktayarslan İlk 10 bin içindeki tıpların 18i ingilizce, 24ü türkçe, altına indikçe de tıp yüzdesi artıyor. Niye bu kadar abartı geliyor size tıpçılar 7 sene okuyor 1 sene de belki tusa hazırlanıyor diyince ? Tusta istediği uzmanlığı almak için etüt tarzı yerlere giden bile bir sürü insan var

[95 @cellestialsea]

@GregKamradt Do you guys update the charts when labs/providers update their prices? Deepseek reportedly plans on "significant increase" in their prices

[95 @cellestialsea]

@emrrnr @ubqtn_e3 @oktayarslan Hocam ingilizce tıp okuyan birisi için 1 sene hazırlık 6 sene okul 1 sene de tus gayet yaygın bir durum, birçok arkadaşım tam olarak böyle yaptı. Bu sene yerleşen için 2034, hazırlanan için 2035 yapıyor.

[95 @cellestialsea]

@ubqtn_e3 @oktayarslan 6 sene okuyacaksın,1 sene de belki TUS için hazırlanacaksın.Mesleğe adam akıllı atıldığında yıl 2035 olacak. Mühendislikte 3. sınıfta staj ile iş hayatına adım atıyorsun, iyi öğrenciysen de direk return offer veriyorlar mezun olduğunda. o fark bile önemli ai'ın hızından dolayı

[95 @cellestialsea]

@tenobrus They saw the whole thing and thought their safeguards were the problem, instead of you know, the models going rogue

[Micah Carroll @MicahCarroll]

Within OpenAI, we recently paused access for an internal model due to misalignment. See the blogpost for details. We have since improved our safeguards and redeployed the model.

https://t.co/eSJpvqo8ve

[95 @cellestialsea]

@NathanAClark @adub01488424 I would classify this as an OK appointment, if you dont I would like to hear your reasoning

[95 @cellestialsea]

@NathanAClark @adub01488424 I think you are exaggerating, he was god awful, but at the time it did look like an OK hire

[95 @cellestialsea]

@Philip_Mundhen3 @carl_feynman There would need to be a clear disadvantage to sleeping longer, which there wasnt.

[95 @cellestialsea]

@IbrahimDagher20 Speaking in the context of LLMs, there appears some form of "integrity" adoption in the model during the pretraining that can be a lot more robust compared to trying to do it solely in post-training. If their system is like a "super-smart 15 year old",would feel similar in spirit

[95 @cellestialsea]

@llllvvuu @Jsevillamol Vision is very much not solved, especially with models outside google. that would be my bet

[95 @cellestialsea]

@_xjdr Sol derived the jacobian disproof independently within a single day, which is supposedly more important than any of these problems (dont know if its actually harder). So you would expect current models *probably* can replicate this with the right harness and steering

[95 @cellestialsea]

@henryquantum You could focus on creating new seemingly impossible problems that would interest you, that is obviously harder than working on a given problem (or picking one), but at least it is something

[95 @cellestialsea]

@Doctorthe113 @InverseMarcus Please call the api, dont give it any tools, and see if sol high can or cant oneshot the first (or any) problem of imo 2026, without any python or bash runtime access. Stop being stupid

[95 @cellestialsea]

@MaxNiederman Whats worse is you can also generate the opposite scenario and word it a little differently to cover both sides

[95 @cellestialsea]

@RealAdamHunt Your whole premise is based on a simple and subjective argument, which is: models got worse at english(?) Which is something wholly dominated by lab choices, and at best based on your vibes.

[95 @cellestialsea]

@EpistemicHope I dont think we currently have a framework that allows us to map billions of years of time and the ginormous sample space evolution operated on onto FLOPs. The answer could be yes but not in our lifetime or yes and likely to happen in few years.

[95 @cellestialsea]

@alxndrdavies @_robertkirk @AISecurityInst The fact that openai RL'ing gpt to hell on web search is the thing that came back to bite them is just funny

[95 @cellestialsea]

@signulll fwiw i fucking love that the nazi party has a different worldview than many of the other parties & thats actually a wonderful thing. if anything, this whole episode gave them more aura because they were willing to again stand on principle. time & time again this has been true.