Future Thinking: The Emerging Intelligence

Natural Selection: What do cells, plants, animals, humans all have in common? They all use future thinking to survive and anticipate changes in their environment, whether cells have to dye to preserve organs, plans need to grow roots deeper before droughts, animals need to foresee danger, or humans earn money to buy food. We all think about the future and try to predict the best actions based on acquired knowledge. This natural selection process made intelligence emerge in living organisms.

Biological Intelligence: The relation between prediction and intelligence has gained significant traction in cognitive neuroscience since Predictive Coding was introduced (Rao and Ballard, 1999) \cite{rao1999predictive, friston2005theory}. As Bubic et al. (2010) argue, prediction enables organisms to direct their behavior towards the future while remaining grounded in present. Recent theoretical work has proposed that intelligence itself can be approximated as the ability to produce accurate predictions (Tjøstheim and Stephens, 2022). This perspective aligns well with our focus on developing AI models specifically trained to predict future events in surgical contexts. Brown et al. (2022) suggest that over the lifespan, the brain becomes more effective at predicting (utilizing knowledge) compared to learning. In the surgical context, this may imply that experienced surgeons develop enhanced predictive capabilities, allowing them to anticipate and respond to surgical scenarios more effectively than novices. As surgeons gain experience, their ability to accurately predict upcoming steps, potential complications, and required instruments may improve (Brown et al., 2022). This predictive capability could be a key component of surgical expertise, contributing to better patient outcomes and more efficient procedures.

Emergent Intelligence

Complexity itself, rather than the quality of the dataset, can drive the development of sophisticated behaviours in models. The findings have broader implications for data curation, neural network training, and even understanding human cognition, where balancing predictability and randomness appears crucial for emergent intelligence.

Scaling Laws: Learning and Search

Scaling General Learning Methods Scaling laws show that pre-training large neural networks on more data becomes increasingly effective in constructing world models as the number of parameters grows. This approach has demonstrated emerging new capabilities in models with higher parameter counts. These deep networks learn hierarchical features—from low-level to high-level representations—by distributing information across many layers. This enables them to approximate complex functions efficiently, capturing much of the knowledge humanity has accumulated. As Richard Sutton wrote in The Bitter Lesson \cite{sutton2019bitter}, the actual contents of minds are tremendously complex, and attempts to simplify them can be limiting. Instead of hard-coding specific knowledge or structures, we should focus on developing general learning methods that can discover and capture this arbitrary complexity on their own. This philosophy underpins the success of large-scale deep learning models, which, through vast data and compute, can learn rich, generalizable representations of the world.

Scaling Solution Search Methods As Sutton highlighted \cite{sutton2019bitter}, there are only two techniques that scale indefinitely with compute: learning and search. The ability humans have to reason with an internal monologue (System 2, Kanheman) enables to search over ideas before responding. This cognitive process inspired Chain-of-Thought \cite{wei2022chain} which improves language models on reasoning tasks \cite{openai2024learning}. Interestingly, recent work has highlighted the potential of scaling inference-time compute i.e. search as an alternative to simply increasing model learnable parameters. Brown et al. (2024) and Snell et al. (2024) demonstrated that optimally allocating test-time compute can outperform much larger models on reasoning tasks [4],[5]. By leveraging the LLM as a simulator and exploring multiple potential futures, we may be able to achieve more robust and accurate predictions without necessarily requiring ever-larger models. What sets o1 (OpenAI, 2024) apart is its training approach—unlike previous GPT models, which were trained to mimic data patterns, o1 uses reinforcement learning to think through problems, step by step, using a “chain of thought.” This training approach, along with more “inference-time compute” that essentially allows the model to perform multiple rounds of reasoning prior to responding, allows o1 to outperform GPT-4o in reasoning-heavy tasks \cite{wei2022chain}. These findings suggest that developing efficient search during inference may be a fruitful direction for improving surgical anticipation systems.

General Intelligence vs. Human-Like Intelligence

• General intelligence is not synonymous with human-like intelligence; numerous conceivable types of general intelligence exist beyond human cognition.

Anthropocentric View of Intelligence

• We often assess intelligence based on human standards, leading to debates on when AI will reach "human-level" intelligence (Coley and Tanner, 2012).
• Ackermann (2018) states: “Before reaching superintelligence, general AI means that a machine will have the same cognitive capabilities as a human being.”
• Researchers extensively discuss the timeline for achieving general AI (Goertzel, 2007; Müller and Bostrom, 2016).

Differences Determined by Physics and Technology

• AI development is constrained by physics and technology, not by biological evolution.
• Like hypothetical extraterrestrial intelligence, AI may have different structures, strengths, and weaknesses compared to human intelligence.

Fundamental Differences Between Human and Artificial Intelligence (Bostrom, 2014)

Basic Structure

• Biological intelligence operates on neural "wetware," fundamentally different from silicon-based AI.
• In AI, hardware and software are independent, allowing learned skills (algorithms) to be directly copied to other systems (Kosslyn and Koenig, 1992).
• In humans, learned skills are bound to the individual's neural structure.

Speed

• AI systems transmit signals nearly at the speed of light.
• Human nerve conduction velocity is up to 120 m/s, much slower than computers (Siegel and Sapru, 2005).

Connectivity and Communication

• Humans communicate indirectly through language and gestures with limited bandwidth.
• AI systems can be directly connected, collaborating through integrated algorithms without miscommunication.

Updatability and Scalability

• AI systems can be rapidly updated, scaled, and reconfigured to suit task requirements.
• Humans have limited capacity for immediate structural expansion or improvement.

Energy Consumption

• Organic brains are millions of times more energy-efficient than computers.
• The human brain consumes less energy than a lightbulb.
• A supercomputer with comparable computational performance uses enough electricity to power a village (Fischetti, 2011).

Implications of These Differences

• These fundamental differences lead to distinct qualities and limitations between human and artificial intelligence.
• AI systems have response speeds thousands of times faster than humans.
• They can form integrated systems, reducing risks of miscommunication and errors.
• For example, autonomous vehicles can coordinate at intersections seamlessly as they operate under the same algorithms (Gerla et al., 2014).