Anthropic is pushed to the brink, and a related lobster accidentally gets exposed.

Claude Code is evolving at supersonic speed into Claude Claw.

When the whole internet is throwing a party over Claude Code being “open-sourced,” Karpathy is also watching this fiasco unfold.

And he even released a bold guess:

The lobster is exactly the direction Anthropic has been pushing hard internally during this period.

The reason it’s so decisive is also thanks to Anthropic’s epic-scale gaffe this time. Karpathy finally found the key evidence for an idea he’d had back in February.

Buried deep inside the source code is Anthropic’s nuclear weapon—KAIROS—and it’s now out in the open.

It’s almost certain that this is Claude’s native “lobster.”

A secret plan to raise lobsters hidden in 510,000 lines of code

This time, Anthropic may really have been pushed to the edge by OpenClaw.

When developers turned Claude Code’s 510,000 lines of source code upside down like archaeologists, a netizen named Ole Lehmann dug out a discovery that sent Karpathy reeling with something new.

This leak could be the ace card in Anthropic’s source code exposure that they least wanted people to see.

A house-bred little elf codenamed KAIROS.

Compared with excitement, Lehmann’s first reaction to this discovery was shock.

With this level of breaking news, how is everyone not rushing to report it, and take a few shots at Anthropic together?

So he wrote a long post, and between the lines it was all excitement at finding “new continents”: “I can’t believe nobody’s talking about this!”

The article is really huge—so packed it can’t even fit on a full screen.

After reading the analysis, even Karpathy couldn’t help but feel a flood of thoughts, calling it “true resonance.”

Because this is exactly the next evolution direction for AI he predicted.

A Claude Code “lobster version.”

Yes—KAIROS’s positioning is essentially a full-spectrum benchmark against OpenClaw.

OpenClaw’s three core strengths—proactiveness, personalization, and Skill—KAIROS has all of them.

And it even goes further, more extreme.

First, the part most like a “lobster”—that proactive “lobster claw” going in to strike.

KAIROS is a Claude that actively comes to you. You haven’t even spoken, and it might suddenly appear, pat you on the shoulder, and tell you what it just did.

It runs 24/7 in the background—whether you’re working or asleep, it’s always there.

The underlying principle is also identical to OpenClaw’s heartbeat mechanism.

Every few seconds, KAIROS receives a “heartbeat.”

Basically, it’s a Prompt roughly meaning: “Wake up, see if there’s any work worth doing right now.”

Then it decides based on the situation: take action, or keep quietly waiting.

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

Basically, if Claude Code can do it, it can do it too—but the biggest difference is that you don’t need to speak up yourself again.

Also, when KAIROS is doing work, it unlocks at least three additional proprietary skills:

• Push notifications: it can proactively send messages to your phone or computer, even if you don’t have a terminal open.

• File delivery: it can send the generated content directly to you, without you having to ask out loud.

• PR subscriptions: it can watch your GitHub, and once there are code changes, it automatically responds.

That’s right—these were features that previously required OpenClaw plus an instant-messaging app to pull off. Claude instead bundles and integrates them directly, turning them into a ready-to-use product right out of the box.

Next is personalization.

KAIROS writes a daily report every day—not the simple memory feature inside an LLM. It’s more detailed than that: it records what it saw, how it judged, what it did… everything.

These are all out in the open. Users can fully trace its actions, and the records only keep getting longer—append-only, so you can’t delete anything. So the longer you keep it, the better it becomes, because it persists across conversations.

But as everyone’s “raising lobsters” for longer and longer, people also发现 a problem: personalization comes with a cost.

The context will balloon exponentially. If you do nothing and wake up and just say “hi” in the morning, that alone might burn up a couple hundred thousand tokens.

Anthropic clearly discussed this pain point in detail internally, because this time KAIROS clearly has a dedicated solution. And it’s quite poetic—

Let it dream.

Yes—at night, KAIROS runs a process called autoDream, consolidating what it learned during the day and reorganizing its memory.

It’s hard not to marvel again at how brilliant human design is. Who would’ve thought that sleeping could be such a clever design for handling context bloat.

As for Skill, you probably don’t need me to elaborate. It’s a concept Anthropic originally made famous first, and it can be plugged directly into Claude Code’s existing ecosystem.

Now imagine combining all these capabilities—what could KAIROS do?

When you’re asleep and the website goes down, KAIROS detects it, automatically restarts the server, and then notifies you. By the time you see the message, everything is already back to normal.

At 2:00 a.m., you receive an email from a customer complaint. KAIROS reads it, helps you reply, and records the whole process. When you wake up, the matter is already resolved.

Honestly, I’m really looking forward to what Claude’s own “lobster” would look like. After all, OpenClaw’s so-called “chosen API” to date is still Claude Opus 4.6. If Anthropic gets hands-on and doesn’t need to do reverse engineering, it should be able to make the architecture even more extreme.

By then, it might not even be fair to call them an employee—it should be more like a sleep-optional co-founder.

The only question is, what level of token consumption are we talking about…

The key is that Anthropic’s usage design is just too anti-human: it treats Pro users like free users.

Last night, Claude Code ran a task and, halfway through, suddenly threw an error saying I’d already used up my weekly quota.

KAIROS wants to roll this out officially, and the top priority right now is to optimize token consumption.

After all, OpenClaw’s context is already terrifying. Without Coding Plan, your wallet simply can’t take it.

Karpathy’s prophecy comes true again

Claw is the next direction of AI’s evolution.

As early as February this year—back when OpenClaw just blew up into the mainstream—Karpathy made this prediction.

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

Put it this way: if Chat is like the user driving their own car, and Code is the user in the passenger seat using it for navigation, then Claw is finally the time to completely slack off, lie back in the backseat, and take a nap.

In simple terms, the autonomy keeps increasing and the proactiveness keeps getting stronger.

Who would’ve thought that just a month later, this prediction would be validated—and in such a dramatic way?

Like a nuclear-bomb-level strategy product such as KAIROS, which Anthropic was supposed to heavily promote with money—yet somehow it got released “casually” just because of a blunder.

Even at the very beginning, netizens didn’t really notice…

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

In fact, many people have already felt this in daily experience: after installing Claude Code and Codex on a new computer, the first thing you do is always to enable “full access” directly.

It’s just that now the trend is becoming even clearer.

We’re entering the “post-prompt” era.

Prompt is no longer the only trigger method, and the time AI quietly works in the background will become longer and longer.

AI no longer waits to get a Prompt before it does work. Instead, it does the work first, and only after it’s done does it come to you to ask for the next instruction.

OMT

Finally, I’m really curious: when Anthropic officially brings this lobster to market, what name will it come up with?

After all, when it comes to naming, Anthropic and lobsters still have some connection.

Some friends might not know: in just a few months, OpenClaw’s first open-source project on GitHub has already been its third name.

The reason is also kind of awkward. At first, the founder Peter’s original name was actually Clawdbot, which sounds a bit like Claude.

But they didn’t expect it to blow up and catch on. Anthropic got seriously annoyed, strongly suspecting Peter was riding the hype, so they called him up and ordered him to change the name.

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

Honestly, even today I still feel that the original Clawdbot was actually more catchy.

Now it looks like there’s a chance to fulfill that dream.

Since they made Peter change the name back then, now Anthropic has finally raised its own biological lobster—

So why not… just call it Clawdbot?