OpenAI’s upcoming flagship AI model, Orion, is reportedly showing diminishing returns compared to previous iterations, raising critical questions about the future of AI development and the validity of scaling laws that have driven the generative AI boom. According to a report from The Information, Orion demonstrates only moderate improvements over GPT-4, with the performance leap significantly smaller than the jump between GPT-3 and GPT-4, particularly in coding tasks.
This development has reignited intense debate in Silicon Valley about AI scaling laws — the theoretical principles suggesting that AI models become progressively smarter as they grow larger and gain access to more data and computing power. OpenAI CEO Sam Altman has been a vocal proponent of these laws, famously stating in February that “scaling laws are decided by god; the constants are determined by members of technical staff.”
However, technical staff within OpenAI are now questioning these assumptions amid growing evidence that leading AI models may be hitting a performance wall. While Orion’s training remains incomplete, OpenAI has already resorted to additional performance-boosting measures, including post-training improvements based on human feedback — suggesting the traditional scaling approach may be reaching its limits.
Two primary factors are contributing to this plateau. First, the availability of training data is becoming severely constrained. AI companies have rapidly exhausted available online data, scraping vast amounts of human-created content including text, videos, research papers, and novels. Research firm Epoch AI predicts that usable textual data could be exhausted by 2028. While companies are turning to synthetic AI-generated data as an alternative, this approach presents its own challenges and limitations.
Second, computing power — the other critical scaling factor — is not infinite. Altman himself acknowledged in a recent Reddit AMA that OpenAI faces “a lot of limitations and hard decisions” regarding computing resource allocation. Despite this, some industry leaders remain optimistic, with Anthropic CEO Dario Amodei predicting that AI training runs could cost $100 billion by next year, compared to the $100 million+ spent on GPT-4.
Critics like NYU Professor Emeritus Gary Marcus have seized on these developments as confirmation of “diminishing returns” in AI development, arguing the industry shows signs of convergence rather than continued exponential growth. Even Ilya Sutskever, OpenAI cofounder and Safe Superintelligence founder, suggested that results from scaling up pretraining have plateaued, emphasizing that “scaling the right thing matters more now than ever.”
Key Quotes
Scaling laws are decided by god; the constants are determined by members of technical staff.
OpenAI CEO Sam Altman made this statement in February, expressing his belief in predetermined formulas governing AI improvement. However, The Information’s report suggests his own technical staff are now questioning these laws amid evidence of diminishing returns.
For general-knowledge questions, you could argue that for now we are seeing a plateau in the performance of LLMs. Factual data is more useful than synthetic data.
Ion Stoica, cofounder and executive chair of Databricks, highlighted the current limitations in AI model performance and the challenges of relying on synthetic data to overcome training data constraints.
Scaling the right thing matters more now than ever.
Ilya Sutskever, OpenAI cofounder and founder of Safe Superintelligence, acknowledged to Reuters that results from scaling up pretraining have plateaued, suggesting the industry needs to focus on more strategic approaches rather than simply making models bigger.
Despite what other people think, we’re not at diminishing marginal returns on scale-up.
Microsoft CTO Kevin Scott pushed back against plateau concerns in July during a Sequoia Capital podcast interview, representing the camp of industry leaders who remain optimistic about continued scaling potential despite emerging evidence to the contrary.
Our Take
The Orion revelations mark a critical reality check for an industry that has operated on the assumption of inevitable exponential progress. What’s particularly significant is that this skepticism is emerging from within OpenAI itself — the company that has been most successful at demonstrating scaling law benefits. The shift toward inference-focused models like OpenAI o1 suggests the industry is already hedging its bets, exploring alternative pathways to intelligence beyond simply throwing more data and compute at the problem. This could actually be healthy for the field, forcing researchers to develop more elegant solutions rather than relying on brute force. However, the financial implications are enormous: if $100 billion training runs don’t deliver proportional improvements, we may see a significant correction in AI valuations and investment. The next 12-18 months will be crucial in determining whether this is a temporary plateau or a fundamental limit requiring paradigm shifts in AI development approaches.
Why This Matters
This development represents a potential inflection point for the entire AI industry, which has raised billions of dollars based on expectations of continuous exponential improvements in model capabilities. If scaling laws are indeed reaching their limits, it could fundamentally reshape investment strategies, company valuations, and the timeline for achieving artificial general intelligence (AGI).
The implications extend beyond Silicon Valley boardrooms. Businesses planning AI integration strategies may need to recalibrate expectations about future capabilities, while workers concerned about AI displacement might find some relief in slower-than-anticipated progress. The shift toward inference improvements and alternative training methods could also democratize AI development, as smaller companies may be able to compete more effectively if brute-force scaling becomes less advantageous.
Moreover, this signals a maturation of the AI industry from a phase of rapid, relatively predictable improvements to one requiring more innovative approaches. The focus may shift from simply building bigger models to developing smarter training techniques, better data quality, and more efficient architectures — potentially leading to more sustainable and practical AI advancement.
Recommended Reading
For those interested in learning more about artificial intelligence, machine learning, and effective AI communication, here are some excellent resources:
Recommended Reading
Related Stories
- Sam Altman’s Bold AI Predictions: AGI, Jobs, and the Future by 2025
- OpenAI CEO Sam Altman Hints at Potential Restructuring in 2024
- OpenAI’s Valuation Soars as AI Race Heats Up
- The AI Hype Cycle: Reality Check and Future Expectations
- Google’s Gemini: A Potential Game-Changer in the AI Race
Source: https://www.businessinsider.com/openai-orion-model-scaling-law-silicon-valley-chatgpt-2024-11