Perplexity Trains Its Computer Agent on Real Mistakes With Hint-Guided Self-Distillation
Perplexity has refined its computer-based agents by training them on historical user session data, specifically focusing on correcting past errors. By using a 'hint-guided self-distillation' technique, the company successfully reduced tool-call failure rates by over 20%. This method underscores a critical evolution in AI training: shifting from general pre-training to experiential learning, where systems iterate on specific task-based failures to achieve higher reliability in real-world environments.
What this means for you
Technical leads building agentic systems should implement feedback loops that explicitly train models on failed user interactions to improve reliability over time.
Put this to work
Try the tools in this story inside the AI Tool Lab, browse roles in the AI Job Portal, test your skills with Aptitude Tests, or read our in-depth guides.
Keep reading
More AI news
Scaling Indic Pretraining: 2.8 Trillion Tokens and an Open Data Generation Pipeline
Scaling Indic Pretraining: 2.8 Trillion Tokens and an Open Data Generation Pipeline
Read the briefWhen it came to Cal AI’s viral success, a small army of influencers paid off
Cal AI, a startup specializing in automated scheduling, achieved rapid commercial success by aggressively leveraging influencer marketing partnerships. Within two years, the platform scaled to $30 million in annual revenue, ultimately prompting an acquisition by MyFitness in March. The outcome underscores how targeted social media campaigns can propel B2C AI tools to lucrative exits, proving that strategic audience engagement often outpaces raw feature complexity when trying to establish a significant market foothold in the competitive health-tech landscape.
Read the briefChatGPT can peek at your Chrome tabs now
OpenAI has expanded the functionality of its ChatGPT Chrome extension, enabling the chatbot to access and interpret content directly from a user’s open browser tabs. By tagging Chrome during a conversation, users can provide the AI with immediate context from web pages, streamlining workflows like summarizing articles or drafting emails based on live data. This integration signifies a shift toward more seamless, context-aware AI assistants that interact fluidly with the user's primary workspace rather than existing in isolation.
Read the brief