Anthropic added game-cloning limits to Claude Code, but developers say Opus 5.5 is now refusing reverse-engineering and software work it once handled. Claude 5.5 Opus is changing even though the version numbers stay the same. I've seen comments that 5.5 seems to be dumber than usual. I have noticed that 5.5 makes mistakes that Gemini Argon catches, and some have to do with security, so now Argon is my go-to coding assistant instead of Claude. I suppose I could use Claude Fable but that model is 5x more expensive.
––– 良くない–––
People have been amazed at MIT's CSAIL which improved Qwen 3.5's ability to peform on math problems without retraining it using gradient descent. This was done by what they called weight-space programming. Hedwig AI took a look at it and found that what actually happened was that they stopped Qwen 3.5 from self-questioning its answer. Qwen 3.5's MATH-500 score is 43.6% when thinking is turned off, but drops to 7.9% when thinking is turned on.
The unedited model plans the story for 2,400 tokens. Then it writes two lines over and over: “Wait, I need to make sure I don’t confuse myself.” and “Okay, let’s write.” It never writes the story.
Just like self-doubintg humans. Qwen 3.5's architecture and weights are open, so they analyzed it and found that when it went into one of these loops, layers 15 and 19 fired harder. They tried randomly replacing the weights at those two layers, and there was some improvement in the MATH-500 score (up to 19%). So they focused on changing the weights so that when thinking was on, there was no repetitive looping.
So the agent did not find a math circuit. It found a general brake on repetition, and math is simply where the model loops most.
Nick Shirley uncovered Somali fraud. Now DailyWire has uncovered Bhutani fraud in Pennsylvania. A lot of them do it. It's their little secret. Unbelievable!
No-show “home care” work is so common in Pennsylvania that it’s no longer viewed as something to hide. In August, the federal Drug Enforcement Administration arrested a drug trafficker in the state who asked the agents to let him out of handcuffs so he could clock out of the home health shift he was supposedly working. When they pointed out that that was illegal, he seemed confused, saying, “Everyone is doing this. If this is a problem, you’ll have to arrest the whole city.”
Here's another report of $45 million Medicare fraud from California. And $49 million Medicaid fraud from Virginia. All just recently. How much more that we just don't know about? So much fraud and waste. I am so convinced that the federal government should not oversee single-payer healthcare insurance. The opportunities for abuse and theft are too great. These are not picked up by internal fraud detection mechanisms, but often by random reporters who just notice something suspicious.
––– 良くない–––
Google Will Turn Off Search for Schools That Reject A.I. If you want to use Google Classroom, you have to turn AI on for Google search, even though you may want the actual organic search results instead of AI summaries. I don't see a problem with leaving AI off completely. Schools can still use Gmail and Google docs, all the stuff they already have. It's not hard to turn AI off and get raw organic search anyway. I don't think Google has much leverage to be honest.
––– 良くない–––
Lovable: Many Apps Built on the Platform Expose Sensitive Data. I've taken classes where the instructors want to encourage use of AI tools such as Lovable and Replit. First time users are amazed that they can build seemingly professional web apps without needing to code. But be careful if you choose to deploy them. Security could be an issue.
––– 良くない–––
OpenAI seems to be silently re-enabling model training on Codex chats. Even if you turn off data-sharing in OpenAI's Codex, OpenAI may turn it back on without your knowledge. Mathematicians can tell you what happens next.
––– 良くない–––
California’s Billionaire Tax Slips From 52% Support to 42%. Below half now. That's progress, I guess.
Update (4 November 2026): the bill was defeated in California. But Washington will continue to punish millionaires. 😔
––– 凄い –––
16 times the carcinogenic impact of standard plastic straws, along with six times the acidification damage (the metric scientists use to measure changes in ocean and soil pH levels).
