OpenAI cofounder and researcher Andrej Karpathy poured cold water on the tech world's “agentic AI” hype, describing today’s generation of autonomous AI systems as “slop”.
Speaking on The Dwarkesh Podcast, hosted by Dwarkesh Patel, a tech-focused podcaster, Karpathy said most of these systems don’t yet work as promised. The researcher, who made major contributions to deep learning applications at OpenAI and Tesla, argued that many developers and investors have overshot the technology’s actual capabilities.
"I feel like the industry is making too big of a jump and is trying to pretend like this is amazing, and it’s not. It’s slop. They’re not coming to terms with it, and maybe they’re trying to fundraise or something like that,” he said. “We’re at this intermediate stage."
This comes at a time when much of Silicon Valley has declared 2025 the “year of agents”. In contrast, he called it a “decade of agents”. Current AI agents, which are marketed as autonomous coders or virtual coworkers, still struggle with reasoning, multimodal perception, and memory retention. Karpathy said it would be a decade before they actually work as intended.
The OpenAI cofounder also critiqued reinforcement learning (RL), saying it was “terrible, but everything else is much worse”. He argued that while RL has produced short-term successes in games and simulations, it remains “noisy and inefficient” when applied to real-world intelligence systems.
Karpathy said there must be collaboration between humans and AI rather than replacement. He also urged developers to focus on building systems that help users reason and learn, not models that generate automatic results with diminishing value.
Highlighting what he sees as deep deficits in current large language models (LLMs), Karpathy said they “do not learn the way humans do”. He added that modern systems perform at prediction but fail at cognition and reflection, producing fluent output without any genuine understanding.
“Currently, the state-of-the-art model that I go to is the GPT-5 Pro, and that’s a very powerful model,” he said.
Karpathy is currently focused on his AI education company, Eureka Labs, and producing educational videos on AI for his popular YouTube channel. He left OpenAI for Tesla in 2017, and then rejoined in February 2023, only to leave a year later.
AI 'bubble'
The AI specialist’s comments come at a time when speculation is rife about an AI bubble, with AI giants sealing mega deals with each other in the past month alone.
On September 10, OpenAI and Oracle signed a $300 billion deal for cloud and computing capabilities. Later that month, OpenAI signed a $100 billion deal with leading chipmaker Nvidia. Earlier this month, OpenAI and AMD signed a multi-billion-dollar deal that sent the latter’s stock surging.
There is a growing debate within the tech industry and investment circles over whether the current surge in enthusiasm and funding around AI could be a speculative bubble. Many have questioned whether AI technologies, especially generative AI and LLMs, are being overhyped relative to their actual capabilities.
After the podcast dropped, conversation on social media focused around the same subject. "i liked the @karpathy interview on @dwarkesh_sp because in a moment when all incentives point one way, “agi is around the corner,” karpathy sounds like someone with their feet on the ground and a hard hat on," one user wrote. Another said if his interview doesn't "pop the AI bubble, nothing will".
One user lauded the podcast for being realistic. "Karpathy by far has the best and most sober takes on AI progress in the community. super careful with what he says and how he articulates it and he isn’t too bearish or bullish," the post read.
Speaking on The Dwarkesh Podcast, hosted by Dwarkesh Patel, a tech-focused podcaster, Karpathy said most of these systems don’t yet work as promised. The researcher, who made major contributions to deep learning applications at OpenAI and Tesla, argued that many developers and investors have overshot the technology’s actual capabilities.
"I feel like the industry is making too big of a jump and is trying to pretend like this is amazing, and it’s not. It’s slop. They’re not coming to terms with it, and maybe they’re trying to fundraise or something like that,” he said. “We’re at this intermediate stage."
This comes at a time when much of Silicon Valley has declared 2025 the “year of agents”. In contrast, he called it a “decade of agents”. Current AI agents, which are marketed as autonomous coders or virtual coworkers, still struggle with reasoning, multimodal perception, and memory retention. Karpathy said it would be a decade before they actually work as intended.
The OpenAI cofounder also critiqued reinforcement learning (RL), saying it was “terrible, but everything else is much worse”. He argued that while RL has produced short-term successes in games and simulations, it remains “noisy and inefficient” when applied to real-world intelligence systems.
Karpathy said there must be collaboration between humans and AI rather than replacement. He also urged developers to focus on building systems that help users reason and learn, not models that generate automatic results with diminishing value.
Highlighting what he sees as deep deficits in current large language models (LLMs), Karpathy said they “do not learn the way humans do”. He added that modern systems perform at prediction but fail at cognition and reflection, producing fluent output without any genuine understanding.
“Currently, the state-of-the-art model that I go to is the GPT-5 Pro, and that’s a very powerful model,” he said.
Karpathy is currently focused on his AI education company, Eureka Labs, and producing educational videos on AI for his popular YouTube channel. He left OpenAI for Tesla in 2017, and then rejoined in February 2023, only to leave a year later.
AI 'bubble'
The AI specialist’s comments come at a time when speculation is rife about an AI bubble, with AI giants sealing mega deals with each other in the past month alone.
On September 10, OpenAI and Oracle signed a $300 billion deal for cloud and computing capabilities. Later that month, OpenAI signed a $100 billion deal with leading chipmaker Nvidia. Earlier this month, OpenAI and AMD signed a multi-billion-dollar deal that sent the latter’s stock surging.
There is a growing debate within the tech industry and investment circles over whether the current surge in enthusiasm and funding around AI could be a speculative bubble. Many have questioned whether AI technologies, especially generative AI and LLMs, are being overhyped relative to their actual capabilities.
After the podcast dropped, conversation on social media focused around the same subject. "i liked the @karpathy interview on @dwarkesh_sp because in a moment when all incentives point one way, “agi is around the corner,” karpathy sounds like someone with their feet on the ground and a hard hat on," one user wrote. Another said if his interview doesn't "pop the AI bubble, nothing will".
One user lauded the podcast for being realistic. "Karpathy by far has the best and most sober takes on AI progress in the community. super careful with what he says and how he articulates it and he isn’t too bearish or bullish," the post read.