Anthropic is leaning into agentic coding and heavily so. It makes sense to use swe verified as their main benchmark. It is also the one benchmark Google did not get the top spot last week. Claude remains king that's all that matters here.
Maybe somewhere in the original comment it would have been fair to mention you can barely see the house in the original photo. This is actually a hilarious complaint
That cannot be a valid excuse. Other than adding extra windows to the clearly visible wall, it's obvious that model perfectly capable to "see" the house. It just cannot "believe" that there can be a big empty wall on a garden house.
It leads on arc-agi-1 with Gemini 3.0 Deep Think, which uses "tool calls" according to google's post, whereas regular Gemini 3.0 Pro doesn't use "tool calls" for the same benchmark. I am unsure how significant this difference is.
Very interesting to hear two technologists at a tech business conference say things along the lines of: "our tools do not merely extend us, they transform us", followed up with "we've become numb to the devastating consequences of technology".
(I know I'm somewhat selectively reading but still)
Interesting because games are exactly the kinds of RL environments that models can effectively learn - but the catch is that they must do this learning on the fly in test-time. Very exciting to see this.
Right, like in math, not all infinite sequences contain every finite subsequence. For example, a non-repeating sequence of 2's and 7's contains no sequence "4". The further condition is that the number be normal[1].
Also, TIL we don't know whether π is normal thus the popular claim that "every string of numbers eventually occurs in π" is not known to be true
Nice post. With respect to maximizing future options, I find the ideas expressed in the following quotes are interesting counter-points.
From '4,000 weeks': "Not only should you settle; ideally you should settle in a way that makes it harder to back out, such as moving in together, or having a child. The irony of all our efforts to avoid facing finitude -- to carry on believing that it might be possible not to choose between mutually exclusive options -- is that when people finally do choose, in a relatively irreversible way, they're usually much happier as a result."
From 'Zero to One': "When people lack concrete plans to carry out, they use formal rules to assemble a portfolio of various options. ... A definite view, by contrast, favors firm convictions. Instead of pursuing many-sided mediocrity and calling it "well-roundedness," a definite person determines the one best thing to do and then does it."
It hasn't worked perfectly though, as the vaste majority of American bisons (all but 4 herds in fact) aren't true bisons but have cattle genes due to hybridation.
There are second hand accounts from indigenous people, excavated bone sites with tools and spear tips, but reportedly no first hand recorded observations from Europeans of buffalo running off cliffs.
Buffaloes aren't exactly Main Battle Tanks you know: it's not that difficult to build fences that are able to stop them (it just won't be the light fences you have in mind).
Also, not voluntarily pushing the beasts to your fences is going to help a bit, since buffaloes don't run to cliff on their own…
