top of page
Writer's pictureHugo Pinto

Discussion: AI cannot understand reality

I've been researching for progress being done in how AI systems can learn to interpret reality around them using the equivalent of all of our senses, and with the ability to decide how to combine the use of the different sense-equivalent models in both inputs and outputs.


I came across this interview / vide clip with Yann LeCun and Lex Friedman debating it, and it just resonates with key points that dispel a misconception that AI understands reality just like humans do, and that it has the potential to physically interfere with it.



Here are some highlights form that conversation, and what I see as takeaways for anyone trying to incorporate this technology in everyday businesses.



1. The Reality Check: Grounding AI in Real-World Data


The debate on whether AI needs real-world experience or if language alone is sufficient for developing true intelligence and understanding is crucial for businesses. Grounding AI in reality involves ensuring that AI systems are connected to real-world data and experiences to prevent them from generating inaccurate or misleading information. This means they need data that might have been capture or even digitised because we haven't made it available to be processed.


Takeaway: Companies must prioritise the integration of real-world data into their AI systems to ensure accuracy and reliability in business applications. This means investing in both your data - operations, finance, performance, etc. - as well as finding complementary datasets that can help expand the context surrounding activities, consumer needs and impact, and business areas.


Further Reading:



2. Embodied AI: Bringing Intelligence to the Physical World


Embodied AI emphasises the importance of AI systems interacting with the physical world to develop more comprehensive intelligence, similar to human learning processes. This approach helps AI systems understand real-world contexts better, which is particularly relevant for businesses in industries like manufacturing, robotics, and logistics, but ultimately, the vast majority of services businesses provide insights or inputs to many interactions with either the physical world, or other humans. Learning is always better in context, and if there are any doubts just look at the impact Covid had on incoming business graduates, which missed out on learning from being in the office or at least co-located with other peers and senior team members. Maybe these learning systems can help us "see reality" better.


Takeaway: Businesses should consider incorporating embodied AI approaches in scenarios where physical interaction is crucial, such as in automated manufacturing, smart warehousing but also in Point of Sale scenarios.


Further Reading:



3. Navigating Moravec's Paradox in Business AI


Moravec's Paradox highlights the discrepancy between AI's ability to excel at complex cognitive tasks and its struggle with simple physical tasks that humans find easy. This paradox has significant implications for businesses implementing AI in various operational contexts. This could lead Human BIAS to deciding to attribute the wrong tasks to machines and systems, as well as preventing quick and easy impact of what these technologies far surpass human abilities.


Takeaway: Companies should be aware of the limitations of AI in physical tasks and plan for human-AI collaboration in scenarios where both cognitive and physical capabilities are required and allocated in an efficient manner



4. Beyond Words and Images: The Promise and Pitfalls of Multimodal AI


Multimodal AI aims to integrate vision and language in AI systems, but current approaches often fall short of true understanding. This limitation is particularly relevant for businesses looking to implement AI in customer service, content moderation, or product recommendation systems. The issues are often linked with the lack of comprehensive data that covers all senses. Example: an AI system can read a complaint on an email, but if the contextual data and metadata isn't included in the inputs, and there aren't examples of how each of those data points impacts the decision, logic or outcome, then the system will simply be processing text and potentially using voice or images as an output transforming them from the text, rather than using images and conversational signals relevant to the input.


Takeaway: While multimodal AI offers exciting possibilities, businesses should be cautious about overestimating its current capabilities and should invest in robust testing and validation processes. This also reinforces the need for quality data and gold-standard examples of how information impacts a sequence or outcome.


Further Reading:



5. Coding Gravity: Intuitive Physics in AI Systems


Developing AI systems that can understand and reason about physical reality in the way humans do intuitively is a significant challenge. This capability is crucial for businesses implementing AI in fields such as autonomous vehicles, robotics, and smart city infrastructure. Though in these areas it is the core challenge at hand, in any real world application these issues come to life. What's the best way to store an object, how to balance it, how to assess size and fit, and behaviour across a user experience. Human behaviour might adjust to AI systems, but physics won't.


Takeaway: Businesses should invest in research and development efforts to improve AI's understanding of intuitive physics, particularly in applications where safety and reliability are paramount.


Summary and Business Implications

  1. Prioritise real-world data integration for accurate and reliable AI systems.

  2. Consider embodied AI approaches for physical interaction scenarios.

  3. Plan for human-AI collaboration to overcome limitations in physical tasks.

  4. Exercise caution with multimodal AI and invest in data as well as robust testing.

  5. Invest in improving AI's understanding of intuitive physics for safety-critical applications.


For companies looking to develop and deploy AI-powered solutions, these takeaways underscore the importance of a balanced and realistic approach to AI implementation. It also brings to life the need to having diverse teams addressing all dimensions of an AI use case. While AI offers tremendous potential, it's crucial to understand its current limitations and the areas that require further development as well as human intervention.


Businesses should:

  • Invest in a mature data capability, as well as high-quality, diverse datasets that reflect real-world scenarios.

  • Develop strategies and design for effective human-AI collaboration.

  • Implement rigorous testing and validation processes, especially for multimodal and physically interactive AI systems.

  • Stay informed about the latest advancements in AI research, particularly in areas like intuitive physics and common sense reasoning.

  • Foster a culture of continuous learning and adaptation as AI technologies evolve, and Generative AI spikes every other week.


By acknowledging the challenges and actively working to address them, companies can position AI effectively in the minds of both employees and clients, creating reliable and trustworthy solutions that deliver value, not fear and uncertainty.


What are your thoughts? Do you agree or disagree?


14 views0 comments

Comments


Post: Blog2_Post
bottom of page