loader image

Ajla Karajko

Karpathy gives reality check on AI agents

One of the most renowned researchers in artificial intelligence, Andrej Karpathy—former member of the OpenAI and Tesla teams—gave a candid assessment of the current state of AI agents in a conversation with Dwarkesh Patel. Instead of adding to the hype, Karpathy made it clear: autonomous AI systems are still far from what the industry promises.

According to him, today’s agents often produce mere “AI slop” — shallow results without true contextual understanding. He believes the models “just aren’t there yet,” lacking sufficient intelligence, multimodal capabilities, and continuous learning. In other words, agents that claim to perform tasks autonomously still rely heavily on human supervision and limited scenarios.

Karpathy was particularly critical of reinforcement learning, calling it “terrible” and “noisy,” though he admitted it still appears impressive because “everything we had before was much worse.”

On X, Elon Musk jokingly challenged him to test his AI ideas against the Grok 5 model, but Karpathy replied that he would rather collaborate with the model than compete.

His remarks serve as a reality check for an industry that declared 2025 the “year of AI agents.” Still, even if agents fall short of top researchers’ standards, they already bring significant time and productivity gains for most users. The difference lies in expectations — between what AI can do today and what it will likely achieve within the next decade.


In Brief: Tech World Highlights

  • Opera has launched Neon, a new AI browser capable of performing actions on behalf of users, available via a premium waitlist subscription.
  • Meta has acquired chip startup Rivos to accelerate the development of its own AI chips and reduce reliance on Nvidia.
  • According to The Information, OpenAI generated $4.3 billion in revenue during the first half of 2025, while spending $2.5 billion on research and computing resources.
  • OpenAI’s new social app, Sora, surged to third place on the Apple App Store—right behind Google Gemini and ChatGPT—after a viral, invite-only launch.
  • Hume AI has released Octave 2, a new multilingual text-to-speech model supporting 11 languages, featuring voice conversion and phoneme editing capabilities.


AI Trending Tools:

  • Apps SDK – enables direct conversations and app creation within ChatGPT.
  • Hunyuan-Vision-1.5-Thinking – Tencent’s advanced vision-language model.
  • PromptSignal – a tool that visualizes how large language models (LLMs) rank your brand.

Podijeli objavu:

Preporučeni blogovi