Anthropic is pushed to the limit, and its "pet lobster" is accidentally exposed. Karpathy: This is Claude Claw

Claude Code is evolving at lightning speed into Claude Claw.

While the entire internet was celebrating the so - called "open - source" of Claude Code, Karpathy was also watching this fiasco.

Moreover, he made a bold speculation:

The lobster is exactly the direction that Anthropic has been fully promoting internally during this period.

Karpathy's certainty can be attributed to Anthropic's epic blunder. He finally found important evidence for an idea he had in February.

Anthropic's nuclear weapon hidden deep in the source code - KAIROS, has been made public.

It's almost certain that this is the original "lobster" of Claude.

The Secret Lobster - Rearing Plan in 510,000 Lines of Code

This time, Anthropic might really be pushed to the limit by OpenClaw.

When developers rummaged through the 510,000 lines of source code of Claude Code like archaeologists, netizen Ole Lehmann unearthed a new discovery that surprised Karpathy.

This revelation might be the trump card that Anthropic least wants people to see in the leaked source code.

A domestic elf codenamed KAIROS.

Upon this discovery, Lehmann's first reaction was shock rather than excitement.

Such a bombshell, shouldn't everyone report it quickly and give Anthropic a few more blows?

So he wrote a long article, full of excitement as if discovering a new continent: "I can't believe no one is discussing this!"

The article is really long, and a full screen can't hold it.

After reading the analysis, even Karpathy was deeply touched and called Lehmann a "soulmate".

Because this is exactly the next evolutionary direction of AI he predicted.

A "lobster version" of Claude Code.

Yes, the positioning of KAIROS is almost a full - scale counterpart to OpenClaw.

OpenClaw's three proud core features - initiative, personalization, and Skill, KAIROS has them all.

Moreover, it does them even better.

First, the most "lobster - like" part is the proactive "lobster claw".

KAIROS is a Claude that will actively approach you. Before you even speak, it might suddenly appear, pat you on the shoulder, and tell you what it just did.

It runs in the background 24/7. Whether you're working or sleeping, it's always there.

The principle is exactly the same as OpenClaw's heartbeat mechanism.

Every few seconds, KAIROS receives a "heartbeat".

Actually, it's a prompt, roughly meaning: "Wake up and see if there's any worthwhile work to do?"

Then it will judge according to the situation: whether to take action or stay quiet.

Once it decides to act, it can: fix code bugs, reply to messages, update files, execute tasks...

Basically, anything Claude Code can do, it can do too. The biggest difference is that you don't have to ask it yourself.

In addition, when KAIROS is working, it unlocks at least three exclusive skills:

Push notifications: It can actively send messages to your phone or computer, even if you haven't opened the terminal.

File delivery: It can directly send the generated content to you without you asking.

PR subscription: It can monitor your GitHub and automatically respond once there are code changes.

Yes, these functions that previously required OpenClaw and instant messaging software can now be integrated into a ready - to - use product by Claude.

Next is personalization.

KAIROS writes a daily report every day. It's not just a simple memory function in LLM; it's much more detailed: what it saw, how it judged, what it did... all are recorded.

These are all visible. Users can fully trace its behavior, and the records will keep growing. They are all additive and cannot be deleted. So the longer you use it, the better it will be because it's continuous across sessions.

However, as people "raise lobsters" for a longer time, a problem has been found: Personalization comes at a cost.

The context will expand exponentially. Without doing anything, a simple "hi" in the morning might consume more than 100,000 tokens.

Anthropic has obviously discussed this pain point internally because KAIROS has a specific solution this time. And it's quite poetic -

Let it dream.

Yes, at night, KAIROS will run a process called autoDream to integrate what it learned during the day and reorganize its memory.

One can't help but marvel at the amazing human design. Who would have thought that sleeping could be such a clever design to deal with context expansion.

As for Skill, there's no need to elaborate. It's a concept that Anthropic first popularized and can be directly integrated into Claude Code's existing ecosystem.

Imagine what KAIROS can achieve by combining these capabilities.

If your website goes down while you're sleeping, KAIROS will detect it, automatically restart the server, and then notify you. By the time you see the message, everything will be back to normal.

If you receive a customer complaint email at 2 a.m., KAIROS will read it, reply for you, and record the whole process. By the time you wake up, the problem will be solved.

To be honest, I'm really looking forward to what Claude's own "lobster" will be like. After all, OpenClaw's current "chosen API" is still Claude Opus 4.6. If Anthropic does it on its own without reverse - engineering, it should be able to optimize the architecture even more.

By then, it might not be just an employee; it could be a non - sleeping co - founder.

I'm just wondering what level of token consumption it will have...

The problem is that Anthropic's usage design is really anti - human. It treats Pro users like free users.

Last night, when Claude Code was running a task, it suddenly reported an error halfway through, saying that it had used up my weekly quota.

For KAIROS to be officially launched, the top priority is to optimize the token consumption problem.

After all, OpenClaw's current context is already quite large. Without the Coding Plan, your wallet won't be able to handle it.

Karpathy's Prediction Comes True Again

Claw is the next evolutionary direction of AI.

As early as February this year, when OpenClaw just became popular, Karpathy made this prediction.

He pointed out that Claw - type products are a brand - new level in the AI technology stack after Chat and Code.

For example, if Chat is like the user driving the car themselves, and Code is like the user sitting in the passenger seat as a navigator, then Claw is like finally being able to relax and sleep in the back row.

In short, it means higher autonomy and stronger initiative.

Who would have thought that just one month later, this prediction would be confirmed, and in such a dramatic way.

A strategic product like KAIROS, which Anthropic should have heavily promoted, was "casually" released due to an accident.

At first, netizens didn't even notice it much...

The next step for AI is clear. In the application era, the release of model capabilities requires more background information and higher permissions.

In fact, many people have already felt this in their daily experience. The first thing they do after installing Claude Code and Codex on a new computer is always to directly enable "full access permission".

Now, this trend is becoming even clearer.

We are entering the "post - prompt" era.

Prompt is no longer the only trigger method. AI will work silently in the background for longer periods.

AI doesn't just work after receiving a prompt. Instead, it starts working first and then asks you for further instructions after finishing.

OMT

Finally, I'm really curious what name Anthropic will give to this lobster when it officially launches it.

After all, Anthropic has some history with naming lobsters.

Some friends might not know that in just a few months, OpenClaw, the top open - source project on GitHub, is actually its third name.

The reason is quite embarrassing. Initially, the original name given by founder Peter was Clawdbot, which sounded a bit like Claude.

Unexpectedly, Clawdbot became extremely popular, and Anthropic was really furious. They strongly suspected that Peter was riding on the coattails of the hype and called to order a name change.

After a series of twists and turns, this lobster finally met everyone under the name of OpenClaw.

To be honest, I still think Clawdbot sounded better.

Now it seems like there's a chance to fulfill this dream.

Since they forced Peter to change the name back then, and now Anthropic has finally raised its own lobster -

How about... just calling it Clawdbot?

Reference links:

[1]https://x.com/karpathy/status/2039057005802082814

[2]https://x.com/itsolelehmann/status/2039018963611627545

This article is from the WeChat official account "Quantum Bit", author: Jay, published by 36Kr with authorization.