361–370 of 1001 posts
I'm finding the "adaptive thinking" thing very confusing, especially having written code against the previous thinking budget / thinking effort / etc modes: https://platform.claude.com/docs/en/build-with-claude/adapti... Also notable: 4.7 now defaults to NOT including a human-readable reasoning token summary in the output, you have to add "display": "summarized" to get that: https://platform.claude.com/docs/en/build-…
If the model is based on a new tokenizer, that means that it's very likely a completely new base model. Changing the tokenizer is changing the whole foundation a model is built on. It'd be more straightforward to add reasoning to a model architecture compared to swapping the tokenizer to a new one. Usually a ground up rebuild is related to a bigger announcement. So, it's weird that they'd be naming it 4.7. Swapping o…
Benchmarks say it all. Gains over previous model are too small to announce it as a major release. That would be humiliating for Anthropic. It may scare investors that the curve flattened and there are only diminishing returns.
Too late, personally after how bad 4.6 was the past week I was pushed to codex, which seems to mostly work at the same level from day to day. Just last night I was trying to get 4.6 to lookup how to do some simple tensor parallel work, and the agent used 0 web fetches and just hallucinated 17K very wrong tokens. Then the main agent decided to pretend to implement tp, and just copied the entire model to each node...
I describe the problem and codex runs in circles basically:
codex> I see the problem clearly. Let me create a plan so that I can implement it. The plan is X, Y, Z. Do you want me to implement this?
me> Yes please, looks good. Go ahead!
codex> Okay. Thank you for confirming. So I am going to implement X, Y, Z now. Shall I proceeed?
me> Yes, proceed.
codex> Okay. Implementing.
...codex is working... you see the internal monologue running in circles
codex> Here is what I am going to implement: X, Y, Z
me> Yes, you said that already. Go ahead!
codex> Working on it.
...codex in doing something...
codex> After examining the problem more, indeed, the steps should be X, Y, Z. Do you want me to implement them?
etc.
Very much every sessions ends up being like this. I was unable to get any useful code apart from boilerplate JS from it since 5.4
So instead I just use ChatGPT to create a plan and then ask Opus to code, but it's a hit and miss. Almost every time the prompt seems to be routed to cheaper model that is very dumb (but says Opus 4.6 when asked). I have to start new session many times until I get a good model.
I'm finding the "adaptive thinking" thing very confusing, especially having written code against the previous thinking budget / thinking effort / etc modes: https://platform.claude.com/docs/en/build-with-claude/adapti... Also notable: 4.7 now defaults to NOT including a human-readable reasoning token summary in the output, you have to add "display": "summarized" to get that: https://platform.claude.com/docs/en/build-…
So many messages about how Codex is better then Claude from one day to the other, while my experience is exactly the same. Is OpenAI botting the thread? I can't believe this is genuine content.
This is a CC harness thing than a model thing but the "new" thinking messages ('hmm...', 'this one needs a moment...') are extraordinarily irritating. They're both entirely uninformative and strictly worse than a spinner. On my workflows CC often spends up to an hour thinking (which is fine if the result is good) and seeing these messages does not build confidence.
There’s one that’s like “Considering 17 theories” that had me wondering what those 17 things would be, I wanted to see them! Turns out it’s just a static message. Very confusing.
Too late, personally after how bad 4.6 was the past week I was pushed to codex, which seems to mostly work at the same level from day to day. Just last night I was trying to get 4.6 to lookup how to do some simple tensor parallel work, and the agent used 0 web fetches and just hallucinated 17K very wrong tokens. Then the main agent decided to pretend to implement tp, and just copied the entire model to each node...
1) Bad prompt/context. No matter what the model is, the input determines the output. This is a really big subject as there's a ton of things you can do to help guide it or add guardrails, structure the planning/investigation, etc.
2) Misaligned model settings. If temperature/top_p/top_k are too high, you will get more hallucination and possibly loops. If they're too low, you don't get "interesting" enough results. Same for the repeat protection settings.
I'm not saying it didn't screw up, but it's not really the model's fault. Every model has the potential for this kind of behavior. It's our job to do a lot of stuff around it to make it less likely.
The agent harness is also a big part of it. Some agents have very specific restrictions built in, like max number of responses or response tokens, so you can prevent it from just going off on a random tangent forever.
Earlier quoted context omitted.
"Also notable: 4.7 now defaults to NOT including a human-readable reasoning token summary in the output, you have to add "display": "summarized" to get that" I did not follow all of this, but wasn't there something about, that those reasoning tokens did not represent internal reasoning, but rather a rough approximation that can be rather misleading, what the model actual does?
The reasoning is the secret sauce. They don't output that. But to let you have some feedback about what is going on, they pass this reasoning through another model that generates a human friendly summary (that actively destroys the signal, which could be copied by competition).
My assumption is the model no longer actually thinks in tokens, but in internal tensors. This is advantageous because it doesn't have to collapse the decision and can simultaneously propogate many concepts per context position.
False: Anthropic products cannot be used with agents.
I'm finding the "adaptive thinking" thing very confusing, especially having written code against the previous thinking budget / thinking effort / etc modes: https://platform.claude.com/docs/en/build-with-claude/adapti... Also notable: 4.7 now defaults to NOT including a human-readable reasoning token summary in the output, you have to add "display": "summarized" to get that: https://platform.claude.com/docs/en/build-…
> Still trying to get a decent pelican out of this one but the new thinking stuff is tripping me up Wouldn't that be p-hacking where p stands for pelican?