September 9, 2025 JBS
Beta Version
The AI realists' perspective serves as a pragmatic antidote to Ai ANGST by shifting the discussion away from fears of a vague, omnipotent future intelligence and toward the development and regulation of practical, reliable, and trustworthy AI tools.
AI realism acknowledges that AI (Artificial Intelligence) is a powerful tool but also recognizes its current limitations and the potential for a slowdown in the kind of rapid, exponential growth seen with models such as Generative pre-trained Transformers (GPTs) .
The promise of artificial intelligence once seemed boundless.
Each new release of OpenAI’s large language models promised leaps in fluency, reasoning, and coding.
GPT-3 impressed with human-like writing, GPT-4 expanded into logic and mathematics, and many believed Artificial General Intelligence (AGI) was within reach.
But GPT-5 has slowed that momentum, raising doubts about whether scaling alone can push AI much further.
In 2020, researchers at OpenAI identified the scaling law: the bigger the model and the more data it was trained on, the better its results.
GPT-3 (2020): Validated the scaling law by producing fluent and stylistically convincing text.
GPT-4 (2022): Expanded capabilities into coding, math, and reasoning, fueling expectations that AGI was just a few iterations away.
The results astonished even the engineers who built them.
GPT-3 stunned observers with its ability to write essays and mimic human style.
GPT-4 expanded into mathematics, coding, and logic with striking accuracy.
But attempts to extend that curve further have faltered.
Despite massive increases in computing power and the construction of new data centers, the performance gains of GPT-5 proved incremental at best.
The disappointment has tempered the once-fervent belief in Silicon Valley that Artificial General Intelligence was only a few iterations away.
Start-ups and major firms alike have shifted their focus from scaling to refining, adjusting outputs, adding safety layers, and tuning behavior for specific tasks.
These post-training methods make the tools more practical but do not alter their underlying architecture.
GPT-5 broke the pattern. Despite massive increases in compute power and training data, it offered only incremental improvements.
The scaling law that defined early progress stalled
Engineers were surprised by the limited gains
The confident belief that AGI was near has softened
Large language models face structural constraints:
They are word-prediction machines, not true reasoning systems.
They lack memory, planning, and simulation abilities.
They fail on tasks slightly more complex than their training examples.
They make simple mistakes, like illegal chess moves, showing they lack clear rule-based models.
These weaknesses reveal why scaling further no longer delivers the dramatic leaps of earlier models.
With the scaling curve flattening, researchers are exploring multimodal and neurosymbolic AI.
These systems combine language models with other components such as planning engines, rule-based logic, and simulations.
Neurosymbolic AI combines language models with planning engines, rule-based systems, and memory structures.
Early examples include AIs that excel in strategy games like Diplomacy and poker, where negotiation and prediction are vital.
These approaches focus on hybrid systems rather than brute-force model expansion.
This shift represents a recalibration of AI research, toward integration, not just scale.
The slowdown also reframes how AI will affect employment. Predictions of mass job automation have quietened down.
Experts compare AI’s likely effect to the internet’s impact, disruptive in some industries but incremental for most.
Tools with clear value (just like spreadsheets or email from the early Internet era) will endure.
General chat interfaces become social media platforms that also contain advertising if they do not deliver measurable productivity gains.
AI is more likely to complement work than replace it outright.
Education systems are adapting. Universities are moving away from take-home essays toward:
In-class assessments
Oral examinations
Socratic dialogue methods
These changes aim to evaluate genuine human understanding rather than outputs generated by AI.
AI is a powerful technology, but not a near-term path to AGI.
The trillion-dollar promises of mass job loss and total automation are overstated.
The future of AI lies in incremental advances and specialized tools, not in sudden leaps.
The diminishing returns of ChatGPT highlight a turning point in artificial intelligence.
GPT-3 and GPT-4 soared on the power of scaling, but GPT-5 revealed its limits.
The next breakthroughs will likely emerge from hybrid approaches, multimodal systems, and smarter integrations, not from simply making models larger.
For now, AI’s promise lies in complementing human intelligence, offering targeted improvements rather than replacing it altogether.
AI realists advocates a more grounded, pragmatic, and critical perspective that urges a move away from the over-exaggerated promises of a rapid, human-like AI future.
They promote focus on the practical, reliable, and grounded applications of current AI capabilities.
Inside China’s AI Revolution
Open Standard for Internet-Native Payments
Access Your Essential Glossary!
It refers to the slowdown in progress from GPT-5 compared to GPT-3 and GPT-4. The newer model delivered only incremental improvements, suggesting scaling large language models has reached its limits.
GPT-3 validated the scaling law by producing fluent, human-like text at a level unseen before. It marked a major leap in natural language processing and fueled new interest in AI capabilities.
GPT-4 expanded beyond text fluency into logic, coding, and mathematics. It performed more reliably across complex reasoning tasks, solidifying confidence that scaling models could bring Artificial General Intelligence (AGI) closer.
Despite huge increases in compute power and training data, GPT-5 showed only small improvements. This signaled the breakdown of the scaling law, proving bigger models alone no longer deliver dramatic advances.
They predict words rather than reason. They lack memory, planning, and world models. They often fail on unfamiliar problems and even make simple errors, such as producing illegal chess moves.
The scaling law is the principle that larger AI models trained on more data perform better in predictable ways. It held true for GPT-3 and GPT-4 but broke down with GPT-5.
Neurosymbolic AI combines language models with symbolic reasoning systems, planning engines, and simulations. This hybrid approach aims to overcome the limits of pure word-prediction models by integrating memory, logic, and structured decision-making.
Experts suggest AI’s impact will mirror the internet: disruptive in some fields but incremental overall. Most workers will see AI tools integrated into tasks rather than wholesale job replacement.
Universities are shifting from take-home essays toward in-class exams, oral assessments, and interactive learning methods to ensure authentic student performance and reduce reliance on AI-generated work.
The AI realist view sees AI as powerful but limited. It rejects exaggerated claims of imminent AGI or mass unemployment, focusing instead on steady, practical advances through hybrid and specialized systems.