What if AI Doesn't Get Much Better? Kairos Future's Take
In a recent New Yorker article, Cal Newport asks a provocative question: What if AI doesn’t get much better than this? The piece examines the rise and recent slowing of large language models (LLMs) such as OpenAI’s GPT series. After years of hype around “scaling laws” – the idea that ever-bigger models would inevitably deliver ever-bigger breakthroughs – the release of GPT-5 has led to a wave of disappointment. Instead of a new leap, we have seen incremental improvements, fuelling scepticism that current approaches can truly deliver artificial general intelligence (AGI).
The Debate
The article highlights how optimism peaked when GPT-3 and GPT-4 vaulted ahead, inspiring talk of superintelligence and radical economic disruption. But since then, progress has plateaued. Incremental post-training tweaks have replaced the dramatic scaling gains of earlier years. Critics such as Gary Marcus argue that LLMs replicate patterns of text rather than “reason” in any meaningful way, while proponents like Sam Altman continue to predict near-term superintelligence.
Our Perspective
At Kairos Future, we recognise the importance of distinguishing between hype and reality – and this article is a useful reminder to recalibrate expectations. Several insights resonate strongly with us:
- LLMs are not true intelligence. Today’s models are brilliant mimics but lack real reasoning or understanding. They are extremely useful – but we should not mistake them for minds.
- Expectations need adjusting. Future Strategist and social anthropologist Axel Gruvaeus points out, the intellectual legacy of books like Nick Bostrom’s Superintelligence still shapes public discourse. Many people either overestimate or underestimate current capabilities. Yes, being able to “talk to your computer” is transformative, but no, this doesn’t mean AI will seamlessly replace how work gets done today.
- Domain-specific applications matter most. Senior Consultant Daniel Lindén argues that some of the most exciting opportunities lie not in general-purpose chatbots but in domain-tailored solutions: AI that can generate architectural designs, model traffic flows, or optimise garden planning. These applications show real-world value beyond generic text generation.
- Scientific breakthroughs inspire. Axel reminds us of AlphaFold and other DeepMind projects, where AI is not just summarising text but helping to explore and create entirely new biological and material structures. This frontier work may prove far more consequential than chatbots.
- Practical business cases already exist. Kairos Futures founder Mats Lindgren adds that while LLMs alone are not enough, they play a critical role in systems that remove repetitive “mass-trawl” tasks. For example, replacing teams of people who manually scan thousands of Instagram accounts or sift through scattered risk data with automated research agents that integrate LLMs as one component.
Looking Ahead
From our vantage point, the truth lies between extremes. The evangelists promising imminent superintelligence overstate what today’s models can do. The harshest sceptics risk overlooking the very real and growing impact of AI in specific domains and in reducing tedious work.
At Kairos Future, our role is to help organisations navigate between hype and reality – testing pilots, identifying high-value use cases, and building the competence to make AI work in practice. The future of AI is unlikely to be a straight line towards AGI. Instead, it will be a patchwork of domain breakthroughs, productivity gains, and gradual but real change in how we work.