The AI world got rocked (or perhaps tickled) yesterday with the reveal of a new leak from OpenAI. According to inside sources, OpenAI’s next big language model—likely called GPT-5, though insiders suggest the codename "Orion"—has hit some unexpected hurdles in its development. The news? Diminishing returns. The response from Sam Altman, OpenAI's CEO? A casual, meme-worthy “lol.”
But let’s go beyond the “lol” and take a closer look at what’s really going on with OpenAI's progress and why people in the AI industry are both skeptical and excited. Beneath the social media memes and carefully curated quotes lies a nuanced reality about how AI models improve—and sometimes, don’t.
AI’s Progress: Are We Hitting the Wall?
According to this latest leak, OpenAI's internal tests reveal that Orion (let’s call it GPT-5 for simplicity) is indeed more capable than GPT-4—but the difference isn’t as staggering as we’ve come to expect from prior model updates. Apparently, some OpenAI researchers worry that GPT-5, despite showing early promise, is struggling to consistently outperform its predecessor in complex tasks like coding. Others argue the new model performs better with language tasks but point to a slow-down in more technical domains.
Is this surprising? Perhaps not. After all, the foundations of large language models (LLMs) rely heavily on a massive trove of web data. As companies like OpenAI exhaust the available high-quality data online, squeezing more “intelligence” out of a new model becomes less about sheer scale and more about new methods to train models. This is where we start to see talk of “diminishing returns.”
Scaling has worked well for AI so far—making models bigger has consistently made them better. But that strategy is, as OpenAI’s co-founder Greg Brockman notes, becoming costly—potentially reaching “hundreds of billions of dollars.” And with costs ballooning and available data dwindling, how much longer can the scaling strategy hold out?
The ‘lol’ Heard Around the Internet
Back to Sam Altman’s “lol” reaction. This isn't the first time Altman has downplayed rumors and skepticism around OpenAI’s advances. Over the last few weeks, he’s made some very optimistic statements, hinting that OpenAI is getting closer to achieving Artificial General Intelligence (AGI)—an AI that could theoretically perform any intellectual task a human can.
Altman’s soundbites could fill a hype reel of their own: he’s claimed that OpenAI knows “what to do” to reach AGI, and that recent research results are “breathtakingly good.” He’s even mused about AI’s potential to solve the mysteries of physics, a comment that led many to roll their eyes but did spark some serious intrigue. So, when Altman says “lol” to the rumors of stagnation, he’s doubling down on his confidence in OpenAI’s mission to keep pushing boundaries. But, as the saying goes, the proof of the pudding is in the eating—or in this case, in the model's performance.
What Does the Data Say?
The reality of AI scaling is, unsurprisingly, not as clear-cut as either hype or gloom might suggest. The latest internal tests of Orion allegedly show that it’s struggling with some tasks but not others. Coding, for instance, remains an Achilles heel, while language-based tasks are showing greater improvement.
And there are the benchmark tests. OpenAI's new model reportedly underperformed in Frontier Math tests, a cutting-edge math benchmark that challenges even the smartest humans, including the likes of Fields Medalists (the math equivalent of Nobel Prize winners). If GPT-5 can barely score 1-2% on this high-level benchmark, it might indicate a real slowdown in the advancement of AI’s reasoning abilities. In other words, we’re still some ways off from achieving AI that can handle complex, creative problem-solving tasks as well as human experts can.
But on simpler benchmarks, Orion’s performance is still impressive, as Altman and co. like to remind us. Does this mean Orion is stagnating, or is it simply the nature of these models to struggle with some kinds of logic and reasoning tasks while excelling in others?
The Data Bottleneck and AI’s Next Big Problem
The limitations we’re seeing here point to a new challenge for AI: data efficiency. Early models like GPT-2 and GPT-3 benefited from OpenAI scraping vast portions of the internet. But now, with most accessible data already incorporated into models, there’s only so much more web data to go around. Scaling to new heights, therefore, will likely mean improving how efficiently AI can learn from less data. More intelligent models will have to be more selective in training, zeroing in on the most relevant information rather than simply consuming all available data.
One approach OpenAI is investigating, called “test-time compute,” could let models retrieve specific answers from a broader data pool rather than generating everything from a pre-trained knowledge base. Theoretically, this could mean more accurate answers in complex domains, and it’s a promising direction for future advancements. But it’s also a reminder that to move forward, AI research must break out of its current paradigms.
Looking Beyond Text and Language Models
While Orion/GPT-5’s language capabilities might seem stalled, other areas of AI are still seeing rapid progress, especially in domains like image and video generation. For instance, OpenAI’s upcoming video generation model, Sora, is due to be released soon and promises to bring some jaw-dropping capabilities. Why the difference? Video and image data is vastly more plentiful than high-quality textual data, and as long as data remains plentiful, progress tends to follow.
As OpenAI and other companies build multimodal models, we might see improvements in tasks that blend vision, language, and audio. After all, humans don’t process information in isolation; we rely on our senses together. If AI can mirror that approach, progress may continue—albeit in fits and starts.
What’s Next?
In this landscape of leaks, hype, and speculation, we’re left in a uniquely ambiguous position. While GPT-5 may not bring a leap as monumental as GPT-4 did over GPT-3, OpenAI and others seem committed to finding new ways to push the frontier of AI.
If this leak has taught us anything, it’s that AI’s future will not unfold in a straight line, and the road to AGI remains full of twists, turns, and Sam Altman’s “lol”s. Whether we’re talking about text, video, or math problem-solving, progress will depend on how effectively these models use data and the breakthroughs yet to come.
For now, the best answer might be patience. And perhaps a sense of humor.
Comments