yorwba

6 ms·

1.
by yorwba 3d ago
A hacking model is aligned if it hacks when you ask it to hack, but when you ask it to play chess, it just plays chess instead of looking for weaknesses in the evaluation setup, as in the article. I presume you would also be less enthusiast
2.
by yorwba 5d ago
It would hardly have become a popular slang term on the Chinese internet if it were only about the worries of a small minority of capital owners. Any kind of excessive competition dynamics may be described as neijuan . I would like to give
3.
by yorwba 5d ago
It's a bunch of rules flagging specific patterns that could be expressed more succinctly: https://github.com/gabeorlanski/scb-check/tree/a8618228939de...
4.
by yorwba 5d ago
Last I checked, they control most of the population centers, but not most of the country by area: https://en.wikipedia.org/wiki/Template:Yemeni_Civil_War_deta...
5.
by yorwba 5d ago
The next sentence is "On a personal level, neijuan reflects a life of being overworked, stressed, anxious and feeling trapped, a lifestyle where many face the negative effects of living a very competitive life for nothing."
6.
by yorwba 5d ago
The people building the wheel of suffering and the ones complaining about it may both be Chinese, but they're not the same person.
7.
by yorwba 5d ago
For agentic tasks where the model outputs tool calls that run on the customer's computer, you can't just store and eval later, because then the execution environment is no longer available.
8.
by yorwba 5d ago
Claude's reasoning traces are encrypted, but there was a design flaw that made it possible to extract them: https://stolen-thoughts.com/ And it makes perfect sense for them to route some requests to Claude, as it lets
9.
by yorwba 6d ago
The 1996 paper mentioned in the NASA article: https://dstretch.com/DecorrelationStretch.pdf
10.
by yorwba 6d ago
Weibo's VibeThinker manages with half of that: https://arxiv.org/abs/2511.06221 (They finetuned Qwen2.5-Math-1.5B for reasoning.)
11.
by yorwba 6d ago
OpenAI released their first reasoning model (o1-preview) https://openai.com/index/introducing-openai-o1-preview/ several months before DeepSeek's R1 https://arxiv.org/abs/2501.12948
12.
by yorwba 7d ago
Sure, there are idiosyncratic preferences and some people really want that one famous work, but there are also idiosyncratic preferences when it comes to houses, and particularly famous houses command a premium, but that doesn't render
13.
by yorwba 7d ago
Owning the copyright to a work grants the exclusive right to profit off that work, but so does owning a house grant the exclusive right to profit off that house. House ownership is not a monopoly because there are other houses, and you can
14.
by yorwba 7d ago
Well, if you could somehow guarantee that your book would sell enough copies in that two-year window to make a million dollars profit, it would make sense to pay you a reasonable salary for the five years it takes you to write it. It's
15.
by yorwba 7d ago
Have you tried to read the Lean proofs produced for any of the recent high-profile results? They're extremely long, terribly structured, and don't indicate which parts are restating known results from literature and which are uniq
16.
by yorwba 8d ago
How sure are you that the phase of the moon is not an input to the system somewhere? http://www.catb.org/jargon/html/P/phase-of-the-moon.html
17.
by yorwba 8d ago
Buddhism didn't start from a blank slate either and incorporates various supernatural beings like deva (linguistically linked to Latin deus ) or asura (similarly linked to Norse æsir ), but they're relegated to minor roles,
18.
by yorwba 8d ago
Yes, here's the graph of weekly downloads: https://stats.documentfoundation.org/downloads#week,version The 26.8 peak is within the normal range of week-to-week variation.
19.
by yorwba 8d ago
It's intended as a "test to see how effective agents are when guided by someone with no expertise in testing who's maybe heard that you should apply certain techniques or use certain libraries." What do you think would b
20.
by yorwba 8d ago
If you're referring to the few-point difference between 2012 and 2018 in the first chart here: https://www.oecd.org/en/publications/pisa-2025-results-volum... I think that's more likely to be statistical
21.
by yorwba 8d ago
From the paper: "To generate sufficient amount of training data for our model in simulation, we use KanjiVG [5], a database of Japanese Kanji characters. This database consists of both images of the character as well as an ordered list
22.
by yorwba 8d ago
He did remark on it: "Although mutation testing is a standard programming term, agents generally didn't actually do mutation testing and instead did normal testing with some small amount of mutating things in a way that isn't
23.
by yorwba 8d ago
The setup is described in an earlier post linked in the second paragraph. Agents are given the zstd RFC and told to implement it, greenfield. Testing frameworks were used however the agents decided, which is to say, as mentioned repeatedly
24.
by yorwba 9d ago
You can get exponential growth from completely ordinary feedback loops. You start with some amount of stuff, you do a series of steps and you end up with more of the same stuff you started with. As you keep going through the loop, the stuff
25.
by yorwba 9d ago
ARC-AGI-3 consists of various games specifically made for the competition, and the goal is to clear the game with as few moves as possible. So there's no answer key per se, but Astra was probably trained on game-like reinforcement-lear
26.
by yorwba 9d ago
Yes, contrary to reporting, the map doesn't represent Africa more accurately, but instead it represents regions near the poles more accurately to eliminate size distortion, at the expense of more shape distortion near the equator. But
27.
by yorwba 10d ago
(2019) Paper page: https://www.atsunobukotani.com/robotdrawing
28.
by yorwba 11d ago
This is damage control, not a PR stunt. They spammed someone's wiki without telling anyone, but then it was discovered and traced back to them. Now that it's in the news, they need to put out some kind of statement.
29.
by yorwba 11d ago
If you discard reinforcement learning sessions where a sandbox escape was discovered, sure. Because that creates a reward gradient in favor of avoiding the honeypot and remaining undetected. But if you reward triggering the honeypot after a
30.
by yorwba 11d ago
1. Include an official communication channel in the prompt. 2. Configure the sandbox to prevent access. 3. As soon as models start communicating on the official channel, stop everything and figure out how they escaped the sandbox. A honeypo

More ›