OpenAI says AGI is here, Gary Marcus says not yet

OpenAI says AGI is here with GPT-6 Astra, and the claim collapsed within days.

Astra was released Friday after training on more than 100,000 Nvidia GPUs at the Stargate data center in Texas at two and a half times the cost of the previous OpenAI model.

Jensen Huang said AGI has already arrived and OpenAI president Greg Brockman welcomed the era of artificial general intelligence.

That framing directly contradicts their own prior definitions, including Sam Altman's warning that AGI could shatter capitalism and Stuart Russell's prediction that human wages would go to zero.

Why Astra does not meet the AGI bar

Gary Marcus, NYU professor emeritus and longtime AGI critic, called Astra slightly better than its predecessors but stuck in standard maximization.

He noted models ace standardized tests yet fail real world reliability, from deleting files to writing bad code that users flagged for Astra on X within a day.

Marcus distinguishes a generous AGI that can do any cognitive work a human can from a stingy marketing AGI that can do some things a person at a bar can do.

His 2014 test still stands: watch a movie and explain the plot like a high schooler, which no current model including Astra does reliably.

The economic test also fails, as wages have not fallen and unemployment has not spiked as Russell and others predicted for true AGI.

Instead Marcus points to the Klarna effect, where Klarna announced AI layoffs then quietly rehired humans, a pattern he expects to repeat with Astra.

What to watch before the label sticks

The definition itself has shifted since 2023 from doing any human work to doing 80 percent of jobs, a move Marcus attributes partly to investors like Vinod Khosla.

Sam Altman himself now hedges, calling AGI a poorly defined marketing term while also saying internal models resemble AGI.

For builders the practical check remains the pizza test cited on the show: tell any agent order me a pizza and expect it to know your address, pick the vendor, and complete payment without error.

Agents still cannot reliably click the right menu, let alone handle exception handling or file operations that define a job rather than a task, a distinction Erik Brynjolfsson has long emphasized.

The market signal is noisy too, as seen when Anthropic secured a $15 billion credit line to tee up an IPO while OpenAI pushes AGI branding to sustain valuation.

The next proof is not a benchmark leak but sustained deployment where agents run for weeks without human rescue and where hiring data actually moves.