- AI researchers at NeurIPS 2025 say current scaling approach has reached its limits
- Despite Gemini 3’s good performance, experts say LLMs still cannot reason or understand cause and effect.
- AGI remains far from a fundamental overhaul of how AI is built and trained
The recent successes of AI models like Gemini 3 don’t hide the more sobering message this week at the NeurIPS 2025 AI conference: We could build AI skyscrapers on intellectual sand.
As Google celebrated the improved performance of its latest model, researchers at the world’s largest AI conference sounded a warning: As impressive as the current crop of large language models may seem, the dream of artificial general intelligence recedes even further unless the field rethinks its entire foundations.
All agreed that simply scaling current transformer models, providing them with more data, more GPUs, and more training time, no longer provides meaningful results. The big jump from GPT-3 to GPT-4 is increasingly seen as one-off; Since then, everything has become less like breaking glass ceilings and more like simply polishing glass.
This is a problem not only for researchers, but also for anyone who believes AGI is imminent. The truth, according to this year’s scientific participants, is much less cinematic. What we built were highly articulated models. They are good at producing answers that seem right. But looking smart and being smart are two very different things, and NeurIPS made it clear that the gap isn’t closing.
The technical term going around is “climbing wall”. This is the idea that the current approach – training ever-larger models on ever-larger data sets – faces both physical and cognitive limitations. We lack high-quality human data. We burn huge amounts of electricity for tiny marginal gains. And perhaps most troubling, models still make the kinds of mistakes no one wants their doctor, pilot, or science lab to make.
It’s not that Gemini 3 didn’t appeal to people. And Google has devoted resources to optimizing the model architecture and training techniques, rather than simply adding more hardware to the problem, making it incredibly performant. But Gemini 3’s dominance only highlighted the problem. It’s still based on the same architecture that everyone now quietly admits is not designed to accommodate general intelligence – it’s simply the best version of a fundamentally limited system.
Manage expectations
Among the most discussed alternatives were neurosymbolic architectures. These are hybrid systems that combine the statistical pattern recognition of deep learning with the structured logic of older symbolic AI.
Others have advocated “world models” that mimic the way humans internally simulate cause and effect. If you ask one of today’s chatbots what happens if you drop a plate, they might write something poetic. But he has no internal sense of physics and no real understanding of what happens next.
The proposals aren’t about making chatbots more charming; it’s about making AI systems reliable in environments where it matters. The idea of AGI has become a marketing term and fundraising pitch. But if the smartest people in the room say we’re still missing the fundamental ingredients, maybe it’s time to recalibrate expectations.
NeurIPS 2025 may be remembered not for what it presented, but for admitting that the industry’s current trajectory is incredibly profitable but intellectually stuck. To go further, we will have to abandon the idea that more is always better.
Follow TechRadar on Google News And add us as your favorite source to get our news, reviews and expert opinions in your feeds. Make sure to click the Follow button!
And of course you can too follow TechRadar on TikTok for news, reviews, unboxings in video form and receive regular updates from us on WhatsApp Also.
The best business laptops for every budget
