Is GPT-5 Just an Incremental Improvement Over GPT-4?
GPT-5, the latest advancement from OpenAI, has reignited discussions about the future of artificial intelligence . Many had hoped for a radical leap in capabilities, but the common notion surrounding its improvements has been described as something akin to “better than GPT-4.” This lack of specificity raises questions not only about the model itself but also about the direction of AI development . Will we reach a plateau where enhancements become marginal rather than transformative?
The Climb Works, but Less
The journey of AI has been a remarkable one, characterized by notable milestones . In 2020, OpenAI researchers published a paper titled “Laws of Climbing for Neuronal Language Models,” suggesting that by increasing data and computational power, AI models would consistently improve. The release of GPT-3 supported this hypothesis, as it was ten times larger than its predecessor and markedly better in performance.
<img alt="AI Competition Landscape" width="375" height="142" src="https://i.blogs.es/f059dc/gini/375_142.jpeg"/>Deceleration in Performance Gains
However, as Gary Marcus , a prominent AI researcher and professor, pointed out in a 2022 article, this perspective may not hold indefinitely. He stated that the so-called “laws of climbing” are merely observations, not universal truths. Even tech leader Satya Nadella echoed this sentiment during the Ignite 2024 event, confirming that while improvements continue, the pace of progress is slowing down .
GPT-5 Shows Notable Improvements
Despite concerns about deceleration, GPT-5 does showcase improvements in specific metrics. Researchers from Epoch AI assessed its behavior in a challenge called Frontiermath, noting that GPT-5 performed better than its predecessor, O4-mini. However, the conclusion was that while the changes are positive, they are not revolutionary . The model tackled a unique problem as though it truly understood it, which was a notable milestone in its development.
<span>In mathematics, GPT-5's behavior shows improvement, yet complex problems still pose challenges. Source: Epoch AI.</span>Enhanced Reasoning Abilities
Another analysis conducted by ZVI Mowshowitz pointed out that while the base model of GPT-5 did not particularly stand out, its advanced variants (like GPT-5 Pro and GPT-5 Thinking) exhibited significant improvements in comparison to their predecessors. Particularly, these variants are lauded for their ability to mitigate hallucinations, leading to a conclusion that the base version of GPT-5 might appear underwhelming, especially compared to the free ChatGPT plan.
Shifting Towards Symbolic AI
In light of current challenges, the conversation about symbolic AI —which uses rules and logic for problem-solving—has resurfaced. This approach dominated the AI landscape until the 1990s but stagnated due to a lack of advancements. The field of neural networks emerged, giving rise to generative AI but now appears to be facing limitations itself.
<img alt="AI Benchmarking Challenges" width="375" height="142" src="https://i.blogs.es/4e26b0/robot-pizarra/375_142.jpeg"/>Persistent Skepticism Regarding AI’s Future
Skeptics like Ed Zitron and Gary Marcus have consistently warned against overly optimistic expectations surrounding generative AI. Even those involved in ChatGPT’s creation, such as engineer Ilya Sutskever , have highlighted inherent scaling limitations. The rise of reasoning models provides a potential remedy to the stagnation of conventional models, but many still feel AI is unlikely to evolve significantly from its current capabilities.
The Futility of Achieving AGI?
Thomas Wolf , co-founder of Hugging Face, raised concerns about this situation, stating that today’s AI often resembles “a country of men who say yes to everything on servers.” He expressed that creating more advanced AI doesn’t necessitate systems with all the answers, but rather ones that can ponder novel ideas or questions. This insightful critique highlights how the current models fail to generate new knowledge, instead only filling in the gaps of existing information.
“To create an Einstein in a data center, we do not need a system that has all the answers, but rather one that is able to wonder about things that nobody had thought before.”
This sentiment aligns with thoughts from Yann Lecun , a pioneer in AI, who noted that current generative AI often lacks the intellect we generally associate with smart systems.
Realigning Expectations in AI Development
Given these insights, it appears that the massive investments in data centers and foundational models may not yield the groundbreaking advancements many anticipated. As Zitron suggested, the market for AI might be much more modest than the billions currently being invested. The idea that we could face stagnation poses serious implications not just for developers, but for society at large.
Implications of Stagnating AI Innovations
If the stagnation hypothesis bears out, we can expect AI to remain a useful tool that enhances certain tasks but stops short of delivering the transformative societal changes that some tech leaders claim. This could mean a powerful tool for efficiency, akin to the early days of personal computing or the Internet, yet falling short of revolutionary potential.
For many, these realizations may lead to frustration as expectations shift. As time progresses, the focus will likely shift toward pragmatism in AI developments, reassessing what is truly feasible and beneficial for society.
Image | Levart Photography
In Xataka | There are too many AI models. That raises a true death sentence for Anthropic and Claude.