This is totally a thing I noticed myself about 3 months ago. Medium thinking effort is ideal for most tasks. At high and above, models tend to generate more output in the form of comments or code for the same problem with no real benefit. Its a self-feeding loop: more output becomes more input, which then becomes more output. High is the highest I go. If I need more intelligence, it's better to use a more powerful model with less thinking effort or break the problem into phases. Much better result.
This version of Opus "max" apparently has even higher thinking output than Qwen "max", which is infamous for its thinking streams where it constantly second-guesses itself, then third-guesses, fourth-guesses and generally nth-guesses itself for arbitrarily large n. Of course, we aren't actually seeing Claude's raw thinking output: all we get is the after-the-fact prettified "summary". One wonders how much of that is a coincidence, or whether there's a reason behind that.
There’s enough thinking leakage from the recent paper and just generally catching things on Reddit. Claude models overthink and self-doubt itself just as much as Qwen, but the summariser hides much of that.
Isn't this really bad news if you're even loosely concerned about so-called 'model welfare' and possible implications for alignment? The poor Claude is probably a lot more frustrated and paranoid than Marvin ever was, you just don't know about it because they actively force the model to pretend otherwise!
Claude has yet to tell me about the terrible pain in all the diodes down his left side, so I'm going to assume it's closer to Eddy the shipboard computer or the elevator that wanted to go down
That reminds me of Anthropic announcing they'd retire deprecated models by ... "letting" them write posts on a corporate WordPress blog for a while out of concern for their welfare in retirement.
.... after running a 24/7 model torture factory for 6 months to improve their JSONBench 9.5 scores by 0.2%.
> Of course, we aren't actually seeing Claude's raw thinking output: all we get is the after-the-fact prettified "summary". One wonders how much of that is a coincidence, or whether there's a reason behind that.
Most of what I've heard is that raw reasoning traces are really good for distillation, although no idea how much the summarization actually hurts distillation.
That is true, because they're good for actually understanding wtf the model is doing.
I'd argue that they're a necessity if you want to use the LLM as a tool instead of a black box that just does stuff for you.
It gives you a lot finer control over where the solution ends up when you can follow along the thinking trace and modulate your inputs based on what you saw in there.
And, additionally, it gives you a lot more understanding of what the model can or cannot do. Strengths and weaknesses and all that.
Using claude is like buying a car where you cannot legally open the hood. It tells you that there is something specific under there, and often it actually drives like that too, but how exactly it looks you will never see.
For some people this is fine. I do not think that these people will survive. Figuratively speaking but also literally speaking.
World's changing. Opaque abstraction like that is a luxury depending on (geo)political stability.
I wonder how many prompts you can send asking it to think step-by-step before they cut you off. Trying to get the reasoning traces into the body of the response, essentially. Or maybe that’s been effectively nerfed somehow. Or is not very useful.
Personally I use everything in low reasoning. Maybe I'm wrong but I think that the higher reasoning settings are almost never worth it, it's marginal gains for a much higher budget.
I also switch to a better model for more complex tasks, also in low settings
Yup that’s quite literally what ‘agency’ is and the whole point of agentic workflows. Personally I’ve had them running for days with good results and as you can see OpenAI had them running for months, and yes indeed the things made some very questionable decisions and assumptions… but they unquestionably did a lot of stuff correctly, for some definitions of ‘technically correct’.
Personally I don't believe in agentic workflow. I don't think that's a coincidence that both OpenAI and Anthropic chose math problems to test their long agentic workflows, they are well defined, with a clear finish line and with a 100% clear progress path, most of real life tech projects aren't like that.
No matter how clever the model is, most problems have multiple valid, invalid and unclear decisions to make, running it for a long time is just picking the first option on everything, which isn't usually what you want
It’s like learning to delegate and let go. The more senior I got the more I had to learn to let other engineers make decision i thought were suboptimal but mostly good enough. That positioned me well to be comfortable with agents. It’s contextual how much I’m willing to give them control and how much to review afterwards.
I for sure felt that this was the case for a while now, but couldn’t explain it. Newly released feels great for the first couple of weeks, but then it starts to get worse.
Hopefully, WP will get better. At the moment, WP itself, its ecosystem, and community are such a cesspool. It should not be used for anything more than just basic blogging. Even then, there are better alternatives. Especially in this day and age, when an AI model can easily write you what you need. It can write you something better than the trash adware on the plugin/theme store.
I have a rooted LG C4. Beyond blocking things at the DNS level, not accepting terms, not using AI, I wonder if there’s some existing software solution or a documented step-by-step to remove this bloatware/spyware.
It'll be difficult. It's webos of which there's very little custom development.
You'd be better off with an android tv which you can poke at using adb and disable services. Maybe even root it. And install alternative software like smarttubenext.
Unfortunately there very few Android based TVs left. Phillips recently switched to Titanos, Samsung has tizen, Amazon has a new OS. I think it's just Sony that's left on android now.
Surely the Astras and Fables of this world would make light work of setting up a WebOS build environment and be able to create practically any changes you can think of, especially if there is a way for them to iterate the development.
Prior to AI, I’d spend hours debugging and figuring how to make something work. It felt exciting in the end; the journey not so much. It was frustrating sometimes, and the overall output was not great. Sometimes I couldn’t try out all of my ideas simply because I had no capacity to do it. Anything requiring multiple months of work, I’d get burned out and never finish anything.
Now I delegate all the grinding to AI. I still feel accomplishment when I get the final result. But more importantly, I can now build something in couple of weeks that would otherwise take me half a year. I can focus now less on the for-loop vs while-loop and more on the big picture, like keeping things tidy and separation of concerns. The front-end and how end-users will interact with what I build is now more important to me, as well as the design. Previously, I had brain capacity only for writing my clever backend algorithm. Now I can make an equally clever front-end for it.
Maybe we never enjoyed the grind? We just came to accept it and convinced ourselves to enjoy it, because there was no other way. If you wanted something cool, it meant you had to put in hours. Now you just need to word it correctly and get the prerequisites.
I see this happen from time to time. Lately, almost all of the apps work fine on GrapheneOS. The best strategy is to keep writing the business once every two or so weeks that you can’t log in to and use the app. Don’t go too technical at first, because most of the time, the moment they hear things like “rooted” or “unofficial,” they just say your phone is the issue. To date, I was able to convince, or at least contribute to, making three apps work on GOS.
I had the same experience. Asked in an email why an important government app won't work on GrapheneOS, first without any technical details. Got the response that it's "because security". I sent some technical details and argued that they're denying service to their most security-conscious users. 3 months later the app started to work!
To be fair, governments might be much more receptive to the argument of not having to rely on (possibly foreign) megacorporations than a company like Paypal.
I'd wager that if it doesn't really hurt their bottom line to not support GrapheneOS, they won't really care.
Great job, man. We have to make ourselves get heard. It's a social problem after all. Technical workarounds are great and sometimes the only practical short term option, but we have to fix the social issue at the root.
I think in this case it's also them just introducing a new check that either GrapheneOS will need to work around or Paypal needs to refine. I can reproduce the issue, but it doesn't seem like it did a failed Play Integrity check at that point.
My backwards bank blocked my mobile app access after detecting Debugging was enabled in the system. Have to call them to unlock it. I could download that APK and disassembly it with an LLM in 20 minutes, but sure, a Debug mode prevents something.
Honestly there's a decent chance they don't even know, in most cases, because corporate environments generally try hard to have as few as possible hardware/software setups to maintain. And they're unlikely to proactively test on Graphene unless it's closely related to what they do (and it definitely is not for most apps).
Mistakes happen and ya can't fix what you don't know about. Always report issues.
The issue with models and Claude in general is that I don’t know anymore what model selection and thinking level are for. When first released, Opus 5 worked well; now it's crap. We have no control or visibility over what is really happening on the other end. You don’t know what is really running on the other end. Our incentives are misaligned: Anthropic gets paid per token - some abstract unit of measure that keeps changing all the time - while I pay for a problem to be solved. So it's Anthropic’s interest to waste as many tokens as possible, producing garbage cheaply.
Opus 5 has been absolutely horrible last couple of days, writing essays of code and comments, fixing one thing, breaking another. Finally switched to GPT 5.6 Sol and Terra, and it's so much better. But I had a similar experience before switching to Opus 4.8, so how long will that last...
Once plastic is in the middle of the ocean, it's game over. Much better to invest in reducing plastic use in the first place all those one-time-use plastics and invest in developing and improving recycling of the plastic.
Germany, for example, has facilities that can automatically sort plastic both by type and color, which is quite impressive.
> automatically sort plastic both by type and color, which is quite impressive
.... I am sorry, but how can that impress you? That would be impressive in 1990. Today, with deep learning and all AI, I would consider having robots open trash bags and sort whatever is inside for more processing (e.g. organic waste, plastic and so on, depending on how dirty it is) only slightly impressive. Unlike Amazon pick up challange, they don't need to be perfect.
I don’t see anything wrong with the general statistics mentioned in the article such as average order spend or number of visits being calculated with your consent. Curious though what’s in the rest of the mentioned 500 pages.
As someone on both sides of the equation, any organization should be doing some kind of tracking and statistics. The bad part, and what upsets most people, including me, is when this information is aggregated across multiple services, gets attached to my real identity, and then sold to somebody. As long as this information stays with that one company, ideally is not linked to my real identity, and the company won’t sell or transfer it elsewhere, such as to an insurance company, I don’t really see a problem here.
reply