Hacker News Reader

Comment by ricardobeat | Hacker News Reader

I noticed the same, and posted about it three days ago.

Opus is suddenly fumbling basic tool calls and failing to finish tasks even at high, where I had been permanently using it on low effort with success since it came out. It is very frustrating.

My guess is they quantized it and/or reallocated DC capacity to train a new model, this has happened before.

View on HN

> It is very frustrating.

Not just frustrating, it can also be dangerous. What if you trust Claude with certain security-sensitive things (because experience), and suddenly you cannot trust it anymore like that?

You shouldnt do that in the first place

This argument would be much more convincing if Anthropic wasn't constantly trying to sell everyone on the fact that you should.

It’s happened before and it will happen again. At least now you know.

And knowing is half the battle!

Then you should have evals.

> What if you trust Claude with certain security-sensitive things

then $corp should remove your access to these security-sensitive things ;)

Maybe don't trust something with non-deterministic output for critical level security?

*chaotic

It is deterministic, if we're being pedantic

If we're being pedantic, it may be deterministic from the point of view of Anthropic. They can guarantee that my prompt X results in actual prompt X+Y (Y being all the settings they control), which would hopefully result in deterministic output Z. From my point of view, Y may as well be random, resulting in non-deterministic output Z for my input X.

It really though?? Even if you set temperature to 0 I saw research showing that the scheduling of the order of operations in the GPU firmware itself can change trailing decimals places and affect the output token

I'm aware operations can technically finish at different rates throughout all of a GPU's cores, but I am not exactly sure that GPU firmware would change the outcome of a calculation (unless there's something specifically in the firmware for it)

I don't know a lot about it tbh so I probably mislabeled it as a firmware issue

It is not deterministic.

Everything might be deterministic if we are being super-pedantic.

Instructions unclear; revoked all human access.

Right lol, i didn't know we'd built a machine that was capable of getting critical level security deterministically correct. That's not me, at least...

It literally could have been rolled out in the middle of that brain surgery experiment on a mouse. Woops.

Anyone who chooses a cloud model instead of local for medical use has a damaged brain.

You don't need to worry about this. If you've outsourced your thinking to a cloud model, it's going to take your job soon enough. I hear layoffs are popular around the holidays ;)

It seems to me the answer is in your question.

And it’s here: https://www.anthropic.com/claude-fable-and-mythos-5-2

Multiple instances of failing a Bash tool, calling bash (lowercase) and getting rejected and even reprimanded by the API "No such tool, call Bash instead".

Killing it's own shell, multiple, multiple times in the same session. I'm sure I saw a phrase "I did it again".

It's so shit that it must be damage control. The alternative is to limit usage and that'd affect the IPO even more.

I noticed its been trying to drag out the conversation on every topic. Maxxing token consumption for the dollars instead of staying on point. Makes sense for them but its frustrating for me as a user. I suspect this is going to be the AI version of enshittification. Now its only missing some badly needed in-stream ads.

Reply on news.ycombinator.com