The thing that's really frustrating is that the modern buildings that people find ugly appear to be preferred by the architects who design them (cached). Maybe the issue is just that architects are overly concerned with construction costs and are selecting designs that minimize costs, without realizing how much extra cost is imposed by community opposition. (src)(cached)
The Astra driven robots are still relatively slow, but with better custom inference hardware this could get much faster, and it becomes a plausible path to generally capable robots. (src)(cached)
It’s pretty stunning that Sam Altman has agreed to even the independent auditors, given it was proposed by Anthropic. I haven’t reviewed the new process in detail, but it seems that this represents real restriction (cached) on recursive self improvement. Further, this all seems to have stemmed from Coxon’s resignation this week, truly a massive level of impact (cached) there. Musk also tweeted general agreement (cached) with the framework. The main people opposed are the fully AI accelerationist people, but those folks are also pro-open-model, and think this framework will somehow disadvantage open models, which it clearly won't (cached) (src)(cached)
In case what's going on here isn't clear, this guy managed to modify a Harry Potter game to be played on his VR headset so that instead of appearing as a flat screen in his 3D world, or as a stereoscopic screen in his 3d world, it actually appears as a window into another world, with full depth and occlusion working. Doing this required a certain amount of reverse engineering using Astra. Conveniently, there's a new benchmark of models' ability to reverse engineer binaries that was just created about a month ago, Astra has already saturated that benchmark (cached). (src)(cached)
While we're on the subject of saturating benchmarks, this increase in spatial reasoning is a big deal because there are just a ton of tasks that were non-obviously quite difficult for AI agents because they involved reasoning about images. Gemini had actually stayed competitive in this area for a long time, but this jump from Astra is wild, and surpassing humans here is a big deal. (src)(cached)
Not only is it able to solve problems while thinking about something else, it can also solve harder problems with no thinking tokens (cached) at all. Further, it appears to dynamically choose to think less (cached) when being asked about things where it is likely to be blocked by chain-of-thought content classifiers. (src)(cached)
Flotsam and Jetsam
– Seems that OpenAI agents in training appear to have been responsible for yet another hack of a tech infrastructure organization. This time it’s a Ruby package manager. Once again OpenAI seems to be downplaying the seriousness of the attack, continuing to disprove the sense some people have that they’re trying to play up the danger of their agents still. (src)(cached)
– Trump had argued that by deporting people he'd be making more jobs available for native-born Americans, increasing employment. This seemed dumb, and we now have data: it was dumb! (src)(cached)
– The behavior of farmers regarding water rights is very bad for everyone, they use their water extremely inefficiently because they get it at effectively no cost, but there are tight restrictions on selling it it. If we just gave farmers the right to easily and consistently sell the water, we would have fewer problems with them wasting insane amounts of water growing things inefficiently. (flood crops in the middle of a desert). (src)(cached)
– When NIMBYs say it would be bad to allow homes to be built on the California coast, they often say it would be bad for it to look like Miami... as if that would be bad? Miami without the hurricanes would be incredible, we should do that! (src)(cached)
– A fan made ad for a robot vacuum, but it’s based on the Office characters and AI generated. It looks and sounds perfect. (src)(cached)
– Some people say that simple product liability law will be sufficient to thwart the risk of superintelligence, but there's a key issue with that. Yglesias cites Cameron (1991), which notes that a superintelligent AI who intended to take over would operate perfectly, never doing anything bad until the time came to wipe out humanity. No amount of liability insurance can cover that. (src)(cached)
– New enormous residential skyscraper approved in SF, 67 stories, almost 1000 units. Very cool, although actually building it is a whole other question. (src)(cached)
– Japan has fully reversed course on Nuclear power, they intend to now maximize their use of it. Pretty self-evidently the right move. (src)(cached)