The core of this shift seems to lie in their upcoming "Astra" model. According to OpenAI's chief scientist, Jakub Pachocki, Astra is already functioning much like an automated research intern. This isn't just a fancy way of saying it can summarize papers. We are talking about a level of agentic behavior where the AI doesn't just retrieve information but actively assists in the scientific process.
What "Inventing" actually means for LLMs #
Altman made a very specific claim that caught my eye: he expects the next generation of models to be "the first model where the model actually invents new things in a way that matters."
This is a massive jump from current prompt engineering workflows. Right now, we use LLMs to rearrange existing knowledge—taking A and B to create C. If Altman is right, we are moving toward a stage of true autonomous discovery. Imagine a deployment where an AI agent identifies a gap in a chemical database, hypothesizes a new molecular structure, and drafts the simulation parameters to test it. That is the difference between a tool and an entity.
The definition problem #
The reason I'm skeptical of the 2026 timeline isn't because I think the tech is impossible, but because "AGI" is a moving goalpost. If we define AGI as "a machine that can do anything a human can do at a desk," we might get there soon. But if we define it as "true consciousness or biological-level reasoning," we are likely decades away.
OpenAI seems to be leaning toward a functional definition:
Autonomous Research: The ability to work through multi-step scientific problems without constant human prompting.Novelty Generation: Creating new ideas, formulas, or code architectures that haven't existed in the training data.Agentic Workflow: Moving away from simple chat interfaces toward autonomous LLM agents that manage their own tools and environments.
If they can pull off the "automated research intern" aspect with Astra, it changes the entire AI workflow for developers and scientists. We won't just be debugging code; we'll be supervising agents that are writing the code for us. It's a wild thought, but if the progress from GPT-4 to what's coming next holds steady, that 2026 window might actually be realistic in a purely functional sense. Video models are still struggling to grasp temporal consistency 5h ago
Linear chat interfaces are fundamentally broken for complex 10h ago
Goodfire just released a tool to peek inside the AI black box 12h ago
Is the AI hype cycle finally hitting a wall of actual business 14h ago
Bill Gates thinks we are flying blind with AI development 23h ago
Why the US immigration bottleneck is creating a massive talent 1d ago
Next AI-generated food imagery is actually making people physically →