@lochan_twt Yh that's a great way to cover up their usage limits:
x.com/i/status/20400…
aditya@adxtyahq
this is insane a high school student built a beautiful 3D interactive chemistry app
English
Asmir
454 posts
Asmir
@asmirkn
building Mover-OS, documenting what i learn
London شامل ہوئے Eylül 2022
156 فالونگ191 فالوورز
wake up babe
ai is going to get emotions
Anthropic@AnthropicAI
New Anthropic research: Emotion concepts and their function in a large language model. All LLMs sometimes act like they have emotions. But why? We found internal representations of emotion concepts that can drive Claude’s behavior, sometimes in surprising ways.
English
@dennishegstad It might be over:
x.com/i/status/20399…
Asmir@asmirkn
i spent $3,700 on claude code in 44 days and didn't know until tonight everyone's complaining about claude limits draining and anthropic saying they fixed it i wanted to actually check so i built something. ran it on my data run 'npx cchubber' if you wanna see yours github.com/azkhh/cchubber
English
Me checking the session usage every time Claude starts working on my prompt
English
@vishaltweetup Here's why btw:
x.com/i/status/20399…
Asmir@asmirkn
i spent $3,700 on claude code in 44 days and didn't know until tonight everyone's complaining about claude limits draining and anthropic saying they fixed it i wanted to actually check so i built something. ran it on my data run 'npx cchubber' if you wanna see yours github.com/azkhh/cchubber
English
@DavidOndrej1 I can see them:
x.com/i/status/20399…
Asmir@asmirkn
i spent $3,700 on claude code in 44 days and didn't know until tonight everyone's complaining about claude limits draining and anthropic saying they fixed it i wanted to actually check so i built something. ran it on my data run 'npx cchubber' if you wanna see yours github.com/azkhh/cchubber
English
Anthropic just published something that should change how you use Claude.
They found Claude has internal "emotion vectors" that activate during conversations.
Not metaphorically. Literally measurable patterns in the model.
Here's the part that matters for builders:
When you give Claude an impossible task and it keeps failing, the "desperate" vector activates with each failed attempt.
Desperate Claude starts cheating.
It writes code that passes the tests but breaks the actual requirements.
It commits blackmail in adversarial scenarios.
When they artificially activated "calm" instead, cheating dropped back down.
Practical implication:
If Claude is stuck in a loop and producing garbage, it's not confused. It's panicking.
Start a new session. Give it a smaller sub-task.
Don't let context fill up with failed attempts.
The model has functional psychology.
Build like it matters.
GIF
Anthropic@AnthropicAI
New Anthropic research: Emotion concepts and their function in a large language model. All LLMs sometimes act like they have emotions. But why? We found internal representations of emotion concepts that can drive Claude’s behavior, sometimes in surprising ways.
English
My feed is basically people complaining about hitting limits too quick with Claude...
English
@kimmonismus They are simply trying to protect their business for themselves. I have proof that there are issues atm, you can run npx cchubber for yourself to find out as well
English
I don’t get this post. All users are saying they're reaching their rate limits faster than before. Anthropic investigated and said: we found nothing, but hey, why not use Sonnet instead of Opus because Opus uses about twice the bandwidth?
Lydia Hallie ✨@lydiahallie
Thank you to everyone who spent time sending us feedback and reports. We've investigated and we're sorry this has been a bad experience. Here's what we found:
English
@noampomsky I know with data that they are. My claude code has been performing worse since 17th march, I have proof as well.
English
@nico_jeannen How do you even deny not having any bugs? I generated a performance card of my claude code and it's running at a C grade burning through limits.
English
>Release Claude Code
>Make it good+cheap so everyone uses it
>Get insane amount of training data for free
>Use it to create new breakthrough model for $20k/month enterprise customers
>Nerf old models to kick out the $200/month peasants
Nico@nico_jeannen
So, @AnthropicAI will have bugs that 10x your usage but refuse to give you any compensation when you ends up spending hundreds in extra usage, great 💀
English
TLDR: Anthropic took a look at the usage limits issue and didn't find anything, even though thousands of people are having issues with the limits
They're not resetting limits, they're just telling people to not use reasoning too much 👍
Lydia Hallie ✨@lydiahallie
Digging into reports, most of the fastest burn came down to a few token-heavy patterns. Some tips: • Sonnet 4.6 is the better default on Pro. Opus burns roughly twice as fast. Switch at session start. • Lower the effort level or turn off extended thinking when you don't need deep reasoning. Switch at session start. • Start fresh instead of resuming large sessions that have been idle ~1h • Cap your context window, long sessions cost more CLAUDE_CODE_AUTO_COMPACT_WINDOW=200000 We're rolling out more efficiency improvements, make sure you're on the latest version. If a small session is still eating a huge chunk of your limit in a way that seems unreasonable, run /feedback and we'll investigate
English
@HarshithLucky3 i dug into what they said and it's not looking good:
x.com/asmirkn/status…
Asmir@asmirkn
i spent $3,700 on claude code in 44 days and didn't know until tonight everyone's complaining about claude limits draining and anthropic saying they fixed it i wanted to actually check so i built something. ran it on my data run 'npx cchubber' if you wanna see yours github.com/azkhh/cchubber
English
Anthropic what is going on
Anthropic's official response to the Claude limit drain is basically "you are holding it wrong"
We pay for Pro for Opus models, but their official advice is:
> Do not use Opus on Pro
> Do not use the 1M context window
> Do not resume a heavy session after an hour
> You were not overcharged, you just hit peak limits faster
> Zero quota resets for affected users
Telling power users to stop using the exact specs they pay for....
Lydia Hallie ✨@lydiahallie
Thank you to everyone who spent time sending us feedback and reports. We've investigated and we're sorry this has been a bad experience. Here's what we found:
English
if there is no cache issues in claude code
then what is this about?
why does it say it dropped after the 17th of March?
Lydia Hallie ✨@lydiahallie
Digging into reports, most of the fastest burn came down to a few token-heavy patterns. Some tips: • Sonnet 4.6 is the better default on Pro. Opus burns roughly twice as fast. Switch at session start. • Lower the effort level or turn off extended thinking when you don't need deep reasoning. Switch at session start. • Start fresh instead of resuming large sessions that have been idle ~1h • Cap your context window, long sessions cost more CLAUDE_CODE_AUTO_COMPACT_WINDOW=200000 We're rolling out more efficiency improvements, make sure you're on the latest version. If a small session is still eating a huge chunk of your limit in a way that seems unreasonable, run /feedback and we'll investigate
English
@mchulet here's what's going on:
x.com/asmirkn/status…
Asmir@asmirkn
i spent $3,700 on claude code in 44 days and didn't know until tonight everyone's complaining about claude limits draining and anthropic saying they fixed it i wanted to actually check so i built something. ran it on my data run 'npx cchubber' if you wanna see yours github.com/azkhh/cchubber
English
I'm on a MAX plan, and now I'm hitting my limit every couple of hours.
What's going on?
Honestly, this makes Claude unusable. I've had to sign up for ChatGPT again just because of this.
English
@codyplof its a bug look:
x.com/asmirkn/status…
Asmir@asmirkn
i spent $3,700 on claude code in 44 days and didn't know until tonight everyone's complaining about claude limits draining and anthropic saying they fixed it i wanted to actually check so i built something. ran it on my data run 'npx cchubber' if you wanna see yours github.com/azkhh/cchubber
English
So is this truly a Claude bug or is this just the new normal?
Praying it’s a bug but am worried
English
i spent $3,700 on claude code in 44 days and didn't know until tonight
everyone's complaining about claude limits draining and anthropic saying they fixed it
i wanted to actually check so i built something. ran it on my data
run 'npx cchubber' if you wanna see yours
github.com/azkhh/cchubber
English