Who decides when AGI happens, and why does it keep turning out to be the guy selling the chips?
On March 24, 2026, Nvidia CEO Jensen Huang posted on X: “From ChatGPT to o1 to Astra in 4 years. AGI has arrived. Congratulations @OpenAI team.” He repeated the sentiment in a conversation with Lex Fridman, saying he thinks we’ve achieved AGI. Forbes cut a clip of it. The clip has pulled roughly 47,800 views and 461 likes at the time I’m writing this, which tells you something about how the general public receives the announcement of a new form of intelligence. Fewer likes than a decent cat video.
I test AI tools for a living. I run agents against real repositories, real ticket queues, real customer data. So when the most quoted executive in the industry declares that machine intelligence now matches or exceeds our own, my first instinct is not awe. It’s to open my terminal and check whether the thing I was working with yesterday remembers what we did.
The declaration versus the desk test
Huang’s framing is a timeline: ChatGPT, then o1, then Astra, four years, done. OpenAI unveiled Astra on the Thursday before his post and described it as the world’s most capable system of its kind. That’s a genuinely fast arc. Nobody serious disputes that the capability curve since 2022 has been steep in a way that embarrassed most predictions.
But “the curve is steep” and “we have arrived” are different claims, and only one of them is measurable. AGI has never had an agreed definition, which is exactly why it makes such a convenient announcement. There’s no benchmark to fail, no board to certify it, no auditor who can look at Astra and say no. It’s a vibe with a press cycle.
Meanwhile, in the same week Huang was congratulating OpenAI, one of the more grounded complaints circulating among developers was about memory. Agents lose completed work through context compaction. You spend an afternoon building something with a model, come back, and the collaborator you were working with has no idea what happened. As one person put it, being able to remember things is a fairly important quality in a teammate.
That’s the gap I keep hitting. Not intelligence. Continuity. Reliability. The unglamorous stuff.
What actually breaks when you use these things all day
My working list of what still fails, consistently, across the tools I review:
- Memory that survives a session boundary without a bolted-on vector database and a prayer.
- Knowing when to stop. Agents will confidently finish a task the wrong way rather than ask one clarifying question.
- Reporting their own failures. A system that quietly skips a step and calls the job done is worse than one that errors out.
- Cost predictability. The same task can vary wildly in token spend depending on how the model decides to think about it.
None of those are intelligence problems in the way the AGI conversation frames intelligence. They’re engineering problems, product problems, and in some cases incentive problems. A model that reasons through a hard math proof and a model that can be trusted with your production database are not the same achievement, and the industry keeps using the first to sell the second.
Follow the incentive
I’m not going to pretend Huang is lying. He has spent more time close to these systems than almost anyone, and his read on the trajectory deserves weight. But he also runs the company that supplies the compute for every lab racing toward this milestone. When the person whose revenue depends on continued buildout announces that the buildout has reached its destination, the announcement functions as marketing whether or not he means it as marketing.
Notice the structure of the post, too. He congratulated OpenAI. It reads as generosity, and it might be. It also positions Nvidia as the arbiter who gets to hand out the trophy. That’s a strong place to stand.
What I’d want before I agree
I’d change my assessment the day an agent can hold a multi-week project in its head, tell me when it’s uncertain, refuse a task it shouldn’t attempt, and produce the same result twice on the same input. That’s a lower bar than human-level intelligence. It’s just a bar you can actually check.
Until then, treat the AGI headline as what it is: one influential executive’s opinion, stated on a social platform, about a category with no definition. Astra may well be extraordinary. I’ll review it like I review everything else, by pointing it at real work and writing down what happens.
The systems we have are useful. Some are remarkable. They are also forgetful, overconfident, and expensive in ways nobody has solved. You can hold both of those thoughts. The people building your tools would prefer you only hold the first.
đź•’ Published: