Why would it be more like OG YouTube, when the content they demoed very closely resembles YouTube shorts? The key difference is OG YouTube was long form.
Sam Altman has made (for me) encouraging statements in the past about short-form video like TikTok being the best current example of misaligned AI. While this release references policies to combat "Doomscrolling and RL-sloptimization", it's curious that OpenAI would devote resources to building a social app based on AI generated short form video, which seems to be a core problem in our world. IMO you can't tweak the TikTok/YouTube shorts format and make it a societal good all of a sudden, especially with exclusively AI content. This is a disturbing development for Altman's leadership, and sort of explains what happened in 2023 when they tried to remove him... -> says one thing, does the opposite.
I lets you inspect what actually constitutes a given cluster, for example it seems like the outer clusters are variations of individual words and their direct translations, rather than synonyms (the ones I saw at least).
Usually PCA doesn't look quite like this so this is likely done using TSNE or UMAP, which are non parametric embeddings (they optimize a loss by modifying the embedded points directly). I can see labels if I mouseover the dots.
GPT-5 claims it is just GPT-4o. Is OpenAI sending overflow requests to an earlier model? How could this not be the first thing they checked when they updated GPT-5?
It's also worth considering that past some threshold, it may be very difficult for us as users to discern which model is better. I don't think thats what's going on here, but we should be ready for it. For example, if you are an ELO 1000 chess player would you yourself be able to tell if Magnus Carlson or another grandmaster were better by playing them individually? To the extent that our AGI/SI metrics are based on human judgement the cluster effect that they create may be an illusion.
Id like to hear about the tools and use cases that lead people to hit these limits. How many sub-agents are they spawning? How are they monitoring them?
Establishing ground truth for this is not easy. Often the labeled calories on foods are quite inaccurate themselves, based on n=1 bomb calorimetry tests. There are also incentives that may lead to lower than actual reported calories on the label.
The article claims that none of these apps use "depth analysis", but newer iPhones have this capability. I would guess at least some of these apps are using some kind of volumetric analysis for the food when available.
Maybe stellarators will be the common design in 2060 once fabrication tech has improved, but for the near future I think its going to be one of the first two.
I know there has been quite a bit of inflation, but didn't Nintendo even delay the release because of the tariffs? Its hard to see how a 24% tariff on goods from Japan would not affect Nintendo's choice in setting prices.
There is not really some distinct pathology with hallucinations, its just how wrong answers (e.g. inaccuracies / faulty token prediction chains) manifest in the case of LLMs. In the case of a linear regression, a "hallucination" is when the predicted value was far from the actual value for a given sample.
Yea I don't get the point of this. Someone convinced someone that the old one was bad and they need to spend $ on a new one? I personally prefer the old one because it gives you a better idea of how far things are.
It's not as surprising to me given that this isn't really emergent in a bottom-up sense... its a direct response to being trained to be misaligned, albeit on other tasks. Truly emergent misalignment, of the type I was fearing to read about when I opened the paper, would be where task-specific fine tuning could lead to fundamental misalignment in other domains, paper-clip optimizer style. My company fine-tunes LLMs for time series analysis tasks and they are being taken pretty far out of the domain of their pre-training data, so if all of a sudden you take one of these models and prompt it with natural language as opposed to the specially formatted time series data it is expecting the results are hard to reason about... yes it still speaks English but what has it lost? I would be more surprised/worried if misalignment arose that way.
Again, the connection is likely not specifically with SQLi, it is with deception. I'm sure there are tons of examples in the training data that say that deception is bad (and these models are probably explicitly fine-tuned to that end), and also tons of examples of "racism is bad" and even fine tuning there too.
To me this is not particularly surprising, given that the tasks they are fine-tuned on are in some way malevolent or misaligned (generating code with security vulnerabilities without telling the user; generating sequences of integers associated with bad things like 666 and 911). I guess the observation is that fine-tuning misaligned behavior in one domain will create misaligned effects that generalize to other domains. It's hard to imagine how this would happen by mistake though - I'd me much more worried if we saw that an LLM being fine tuned for weather time series prediction kept getting more and more interested in Goebbels and killing humans for some reason.