Water from the Temple of Penang
QbitAI is public.
In the meantime, I'm going to have to go to the bathroom.
I bought the Claude Code quota for real gold and silver, and every day I eat chicken with a lot of fear.Backstage bigfoot crazy take my quota as wood burning。
Just now, Officer Claude Code tweeted in a proud tone, announcing that they had made a major performance improvement:
Backstage agents (subagent) update the status line in the terminal, cost directly down by 99%.。
It was supposed to be a technological breakthrough, and as a result, everyone looked at the details, and the moment was blinking.
For half a day, backstage agents in the previous terminal updated the status bar every 30 seconds.
Every time(Note is every time)It is necessary to re-read the whole context of the entire dialogue from beginning to end.
I've spent my precious amount of money trying to figure out a few words at the bottom of the terminal to give me a state of mind!
You're upset with me?
It's all over the internet and it's all over the place.
30 seconds to read the full text.
It's not normal for a complex project to run for decades when you're writing code, right?
It's the first time I've ever heard of a token in tens of thousands.
So what did Claude Code do before that?
As long as the backend agent is running, it has to update the user's status below the terminal every 30 seconds.
In order to generate this short sentence, it directly reads the entire record of all conversations from the beginning to the present.
30 seconds at a time, 2 readings in a minute, 120 readings in an hour.。
Who's heart is bleeding!
This is where the code is written, which makes it clear that there is a crusher in the terminal.
The official solution to this problem is now replaced by Haiku 5.5 to generate the status bar based on a short note.
If you don't need to read the context, look at a few short notes and write them down.
According to official accounts, the cost of updating a single state line directly decreased by about 99%.
When I see the 99% figure, I think a lot of people have the same first reaction as me:
Wouldn't it be better to save half of it later?
Don't be happy too early.
The development team added in the comment section:
The sub-agent itself retains the original model, with sessions with the backstage sub-agent, reducing the consumption limit by about 5%, and the same amount is likely to allow you to run more than 6% of the Prompt.
The translation is that the backstage agent maintains the model it was originally called for while it is actually running the task, and this operation is only a state line.（status line）It's a question of self-discovery.
Compared to the overall actual meetings with backstage agents, the quota consumption was reduced by about 5%.
The same amount, discounted, will probably only allow you to run more than 6% of the Prompt.。
Lydia Hallie, who is in charge of the Claude Code developer experience, also ran out to explain:
This game only affects the CLI command-line interface, and the desktop-side app doesn't write this sub-proxy status line at all, so it doesn't need to be changed at all!
The case is closed.
Desktop users have been quiet for years, and command-line heavy-duty developers are all the rage.
I'm going to be like:
Game development misintercepts, followed by
Claude Code's riotous operation is much more than that.
It's the same thing with the glasses and the neighbors.Incidents of misinterception in game development。
Developers who used Claude Code to write the game earlier found that AI motionlessly refused to respond.
One reason for crying is not laughing:
A previously added a rule of thumb to the system prompts, intending to prevent users from having Claude duplicate an existing well-known game at a malicious pixel level.
As a result, the rules of safety were overly forceful and directly evolved into one that could kill a thousand people and not let one go.
The developers simply used a game name to describe the game's mechanism, such as helping me make a jumping mechanism similar to a flappy bird mouse, or referring to a falling logic mouse that darkened the puzzle, which triggered the sound of a system mouse, which was deemed to be a mouse infringement copy mouse, and refused to write the code.
There is no longer any direct access to the people who are doing the original development.
It was later estimated that it was also scorned, so the official had to urgently patch the following version of v2.1.317:
Now that the developer's name description mechanism would no longer be directly rejected, Claude would build the original game of the genre based on it; if the user's will required a complete copy of the original, Claude would turn to the original version of a similar mechanism instead of a stick to kill.。
In a blog post, Andrew Curran commented:
At the heart of these rules is how big models understand them. The user's vocabulary, phrases, and information already in the above context window can lead to a huge bias in the model's judgement. Anthropic's findings are alarming and may lead developers who want to develop games directly to OpenAI.
OpenAI can't help but say that the netizens have gone numb.
The most outrageous thing about this is that the authorities have been releasing these issues as a major version of the optimization flag-drum.
You repaired an amazing bug, and you wanted to do it, so you just leaked the bottom.
The spit of Mason Geloso is directly in the throat of the netizen:
This game was supposed to be a silent patch, and I'm ashamed of you all when you blow up your flags.
Payment users in the comments section were even more completely laughed at.
This is the first time I've ever heard of this kind of thing.
Is this what you're going to do to give the guys 10 free reinstatements? This is just an amazing abuse of Token use.
Hashim is also on fire:
In our real gold and silver paid subscription plan, this is just a ridiculous over-the-top token waste.
We all use tools, who don't know how to calculate.
Even many people who encounter a cartoon, a limit, have their first reaction to find out why:
Am I writing too long or am I doing too much at the same time?
Who would have thought that underneath the terminal, that stunning state of affairs had been stealing my tokens.
To tell you the truth, sometimes I think the ecosystem of AI tools today is too magical.
The development of the front-end field has been rapid, there are endless layers of cool concepts, one measurement benchmark is far ahead, and as a result, even the most basic contextual duplication and light monitoring are not clear.
This rough re-reading of the full text makes the ROI feel uncommon....
On the contrary, I think that even if the A-wave doesn't give you money back, high and low should give the affected users a free refund.
The link is as follows:
[1]
[2]
[3]
[4]