Gary Marcus just earned himself the 2025 Avocado Award f951 for what might be the fastest aging tweet in AI history. His bold proclamation about AI models failing on International Mathematical Olympiad problems lasted exactly six hours before OpenAI dropped a model that solved 5 out of 6 IMO problems under full competition rules.
The original tweet read: “All the models underperform humans on the new International Mathematical Olympiad questions, and Grok-4 is especially bad on it, even with best-of-n selection? Unbelievable!”
Six hours later, OpenAI released their breakthrough model, essentially turning Marcus into a real-time fortune teller in reverse. As the joke goes, this tweet didn’t age like milk. Milk lasts longer than six hours.
The Six-Hour Expert
Let’s give Gary Marcus some credit here. He managed to be spectacularly wrong faster than anyone thought possible. This is efficiency at its finest why wait months or years to be proven wrong when you can do it in the time it takes to watch a Lord of the Rings movie?
For context, Marcus has built his reputation as AI’s most prominent critic, particularly when it comes to large language model capabilities. His central thesis: LLMs are sophisticated text compression algorithms that lack true reasoning. He’s made a career out of explaining why neural networks can’t do symbolic reasoning, mathematical thinking, or achieve real intelligence.
The IMO problems were supposed to be his smoking gun the type of mathematical reasoning that requires genuine insight and creative problem-solving. The kind of stuff that would expose LLMs as the pattern-matching pretenders he claims they are.
Oops.
Perfect Timing, Perfect Failure
What makes this particularly beautiful is the timing. Marcus didn’t just make a vague prediction about AI limitations sometime in the future. He made a specific claim about current model performance and got immediately dunked on by reality.
The International Mathematical Olympiad represents one of the most challenging benchmarks for AI reasoning. These problems require deep mathematical insight, creative approaches, and rigorous proof construction. OpenAI’s model achieving 5 out of 6 correct solutions under full competition rules isn’t just impressive it’s exactly what Marcus said couldn’t happen.
This wasn’t a close call or a marginal improvement. This was OpenAI essentially saying “Hold my beer” to one of AI’s most vocal skeptics.
The Symbolic AI Prophet
Marcus has spent years arguing that current AI approaches are fundamentally limited. His pitch for hybrid symbolic-neural architectures makes sense in theory. The problem is he keeps making specific predictions about current limitations that get immediately contradicted by new releases.
It’s becoming a pattern. Marcus declares something impossible for current neural networks, then watches those same networks do exactly that thing within days or weeks. At this point, OpenAI should just set up Google Alerts for his tweets and use them as product launch timing guidance.
Timeline of the fastest aging AI prediction in 2025
The Real Lesson
Marcus’s Avocado Award isn’t just about one wrong prediction. It’s about the dangers of making confident negative claims in a field moving at light speed. When you’ve built your brand on explaining AI’s limitations, every major breakthrough makes you look a little more out of touch.
The irony is that Marcus probably has valid points about AI’s current limitations. But his delivery method confident proclamations that immediately get contradicted undermines his credibility. It’s like being a weather forecaster who confidently predicts sunny skies right before every hurricane.
For those working in AI, Marcus serves as a perfect example of what not to do: don’t make absolute statements about current AI limitations in 2025. The field moves too fast, and you’ll end up winning awards you don’t want.
My Take: The Six-Hour Truth
I’ve seen enough AI cycles to know that extreme predictions in either direction usually age poorly. But Marcus has turned this into an art form. His six-hour turnaround from confident skeptic to immediately-proven-wrong might be a new record.
The broader lesson is simple: in AI right now, humility beats confidence every time. The moment you declare something impossible, someone will make it happen just to prove you wrong. Marcus earned his Avocado Award fair and square, but at this rate, he might need a bigger trophy case.
At least milk has the decency to last a few days.