Key Takeaways

  • Traditional scientific papers are an inefficient relic, especially for complex knowledge like theoretical physics, says Alex Lupsasca from OpenAI.
  • The future of research communication isn't static PDFs, but "interactive papers" living inside LLMs, complete with attached chatbots for dynamic explanations.
  • AI's superhuman capabilities, demonstrated in solving gluon and graviton scattering amplitudes, create a massive bottleneck for human verification.
  • Future AI models must not just provide answers, but explicitly indicate their confidence levels, moving beyond simple guessing.
  • This shift forces builders to rethink how internal knowledge is stored, accessed, and trusted, anticipating a wave of AI-generated content requiring automated scrutiny.

Your Next Scientific Paper Is An LLM Chatbot

For anyone building at the edge of AI, you already know static documents are dying. Alex Lupsasca, a research scientist at OpenAI, isn't just seeing this trend; he's actively frustrated by it in his own work. He studies theoretical physics, a field bursting with complex, interconnected ideas. But when it comes to sharing discoveries, he feels stuck in the past.

“I spend so much of my time writing papers,” Lupsasca said on the Latent Space podcast. “And the way I think now is so far from papers it just feels like not the right way somehow to store and communicate knowledge.” Think about it: you pore over a complex idea, distill it into a formal paper, and then someone else has to deconstruct it again, perhaps even feeding it back into an AI. “Why are we doing this?” he asks.

Lupsasca's vision isn't just about efficiency; it's about making knowledge alive. He sees a world where “Maybe some kind of interactive paper which lives in some LLM. Maybe your whole paper is some chat GPT page. And you know, you there's a chat bot attached to the paper and you can say explain the big picture and like zoom into this fact.” Imagine your product documentation, research findings, or even legal briefs not as files, but as dynamic conversations, always ready to explain themselves on demand.

The AI Slop Problem: Why Verification is the New Bottleneck

AI isn't just doing grunt work anymore. Lupsasca highlighted how GPT models have achieved superhuman capabilities in theoretical physics, solving problems like gluon and graviton scattering amplitudes that stumped human experts. This is incredible, but it creates a new, urgent problem: how do you trust it?

“The second thing is verification,” Lupsasca explained, “because we're now in this new regime where the models are so capable that for very hard computations at the frontier of knowledge they they can just do the whole thing, but you know, is it correct?” As AI generates more sophisticated content and solutions, the human bottleneck for checking its work grows tighter. This is the "AI slop" problem, where the sheer volume and complexity of AI-generated output overwhelm our ability to verify its truth.

The solution, Lupsasca argues, lies not just in better human oversight, but in getting the AI to police itself. “I think improving verification or even just having the model indicate more directly how confident it is in its answer,” he said. He believes models are “smart enough to know” when they're truly confident versus just guessing. Getting AI to be explicit about its self-assessment is key to building reliable, cutting-edge research and products.

What to Do With This

Stop thinking of internal documentation and product knowledge as static artifacts. This week, challenge your team to prototype an internal 'interactive paper' within an LLM interface, where the knowledge is a chatbot conversation. Design specific prompts that force the AI to not only provide answers but also express a confidence level or list potential caveats, building a feedback loop for better output before 'AI slop' becomes your team's biggest operational drag.