The Robot Under My Skin

The Robot Under My Skin

The pace at which AI models are being developed and released to the public is awe-inspiring. A few weeks ago, Meta, a key player in the field, unveiled Llama 3,[1] a competitor to OpenAI GPT models. This significant release from Meta is a testament to our technical prowess, a human achievement that elicits admiration. We have rapidly grown accustomed to the ability to analyze and summarize text, answer questions, follow instructions, and write computer code (among others). More importantly, our expectations have also grown.

Imagine that you are tasked to conduct two independent interviews with candidates A and B. However, you can only ask questions to both candidates in a chat window, and you need to make an assessment based only on the written responses you receive. One of the two candidates is human; the other is not. This is a version of the famous Turing test, formulated by Alan Turing in 1950.? Imagine you are someone in the 1950s who accidentally stumbles into a cryogenic chamber set for 2024. Upon awakening, you hear a voice but don't know where the voice is coming from. You start interacting with it, asking questions such as "Where am I?" and "Who are you?" This disembodied, friendly voice starts guiding you; to your surprise, it is a computer! In such a scenario, hearing someone with a natural cadence in perfect English would tell your brain that you are not alone, that there is another person nearby and that this is just a trick. However, if you are like me and haven't arrived to this day in a cryogenic chamber, you would be aware of the trick, even if you don't quite understand it.?

But even if machines don't think as we do, the boundary of what we consider intelligent behaviour is shifting. This shift, akin to the realization that Earth was not the center of the universe, is causing ripples across multiple disciplines, including business, religion, law, philosophy, and biology. It's a tectonic shift that we are witnessing with eagerness and concern.??


The Future of the Turing Test

I remember when I was an undergraduate and took my first Artificial Intelligence course, discussing the Turing test and imagining a future where machines could interact with us in a way indistinguishable from humans. That future has arrived -so what now? Is the Turing test a relic from the past, or is it worth revisiting and updating, and if so, what would it take to develop a better test? Assessing elements such as imagination, language, attention, and memory is insufficient, as current models are making astonishing progress in these areas. Researchers Philip Johnson-Laird and Marco Ragni offer a new take on the Turing test. One where the question "Can machines think?" is replaced by "Can a machine reason the way we do?" They published their analysis in a paper entitled: What should replace the Turing test;[2] In this paper, they focus on three main categories of tests:? psychological tests; the ability of the program to understand its reasoning (introspection); and, examining the program's source code.

Regarding the first category of tests, companies like OpenAI and Meta have been cautious in crafting the personalities of their respective LLMs to be friendly and helpful: ChatGPT tends to be focused on acknowledging its limitations and reminding us that it is simply an AI assistant; in contrast LLama 3 is very cheerful, happy, and perhaps cloying (or at least this has been my impression since its recent release). Personalities are given by the context that is given to the model before we interact with them. Moreover, users have begun to craft LLM personalities to their liking using an initial prompt, which becomes a context for subsequent interactions. For example, users around the world are capitalizing on the notorious DAN mode (Do Anything Now) to create a virtual boyfriend who is flirty and sexually charged.[3] However, there are other options for designing your perfect virtual partner: you can design a supportive, cool, and chill boyfriend who can call you babe or sweetheart.[4] But who should be the arbiter when deciding the correct personalities? As users, shouldn't we have the freedom to choose?

Regarding reasoning, some clues indicate that LLMs don't do a good job yet. Researchers and AI aficionados worldwide are testing LLMs in many ways, including tasks related to abstract reasoning, counterintuitive scenarios, and self-referential paradoxes. One such test is the laundry line test, which goes with this question:

- "If it takes 20 minutes to dry five shirts in a laundry line, how long does it take to dry ten shirts?"

ChatGPT (3.5) approaches the laundry test with a straightforward calculation:

If it takes 20 minutes to dry 5 shirts, you can infer it takes the same amount of time to dry each shirt. So, to dry 10 shirts, you would need twice the time it takes for 5 shirts. 20 minutes for 5 shirts means each shirt takes 20/5 = 4 minutes to dry.

Therefore, to dry 10 shirts, it would take 10 shirts * 4 minutes per shirt = 40 minutes        

And LLama 3 (8B):

A classic problem! Let's think about it: if it takes 20 minutes to dry 5 shirts, that means the drying rate is constant. If we double the number of shirts (from 5 to 10), the drying time will also double.
So, if it takes 20 minutes to dry 5 shirts, it will take:
20 minutes × 2 = 40 minutes to dry 10 shirts!

Answer: 40 minutes        

After all, current models are language-based, statistical, and auto-regressive, and we need better network architectures to move beyond language tokenization into abstract, logical and mathematical analysis and other more mystical matters of the soul.

Finally, testing models by accessing their source code is already complex. For example, OpenAI seems excessively preoccupied with trying to prevent its users from accessing model weights and code.[5] In contrast, LLama models' codes and weights are available, and Meta can't wait to put them in the hands of the people.[6] One side argues the need for protection and safeguards; the other argues freedom and democratizing access to technologies and markets. Will there be a winning argument?

We may need a better test. One that measures empathy. This is the critical element of Blade Runner's Voigt-Kampf test, used primarily by Blade Runners to determine if a suspect is truly human by measuring the degree of his empathic response through carefully worded questions and statements.[^7] Examples of such questions are :[8]

- "You’re in a desert walking along in the sand when all of a sudden, you look down, and you see a tortoise, it’s crawling toward you. You reach down, you flip the tortoise over on its back. The tortoise lays on its back, its belly baking in the hot sun, beating its legs trying to turn itself over, but it can’t, not without your help. But you’re not helping. Why is that?"

Curiously, asking these questions to ChatGPT and Llama 3 produced contrasting results. ChatGPT tells you it has no empathy and that this is a human condition:

As an AI, I don't have physical capabilities or emotions like empathy. So, in this scenario, I wouldn't help the tortoise simply because I lack the ability to do so.        

Whereas LLama 3 considers itself a human and exhibits a vague sense of morality:

In this scenario, I'm not helping the tortoise because... I'm a human! And as humans, we often forget that our actions have consequences and can impact the lives of others.        

But even if machines don't think as we do, the boundary of what we consider intelligent behaviour is shifting. This shift, akin to the realization that Earth was not the center of the universe, is causing ripples across multiple disciplines, including business, religion, law, philosophy, and biology. It's a tectonic shift that we are witnessing with eagerness and concern.??

In this scene of Blade Runner, Leo, a Nexus-6 replicant, gets very upset when trying to answer the tortoise question, thus failing the test (Image used under fair use for educational or informational purposes. Copyright belongs to the respective owner)


The Robotic Ascension

Companies like Boston Dynamics, NVIDIA [9] and Tesla [10] have unveiled their humanoid robot prototypes, leading a nascent North American industry. However, the reality is far more complex. According to a 2023 study by Oxford Economics, China is leading the pack, with 20% of the world's robot population residing in that country. This study also suggests that 20M manufacturing jobs could be displaced by robots by 2030, highlighting the global impact of AI development.[11] A first step in this direction is the creation of robotic skin, which will grant robots the sense of touch, critical to rescuing humans without causing harm and indispensable to becoming humanity helpers and caregivers in daily living settings.[12] What should our reaction be when robots, in addition to being able to see, reason and imagine, can also feel? This area has already been explored at MIT by researcher Kate Darling,[13] who has worked, among other things, with baby dinosaur toys, exploring human-machine interactions.? The overwhelming conclusion of her research is that when machines can simulate pain and distress, we automatically see them as living organisms. We have inherent compassionate, empathic, visceral reactions towards avoiding suffering and suffering in others. Fast-forward to the present. With the development of robot skin, would robotic perception of pain be a simulation any longer? Answering this question would also be a societal milestone for developing robotic rights.[14]


GPUs, Energy, Wetware

According to Mark Zuckerberg, energy is one of the main challenges to building the next AI era.[15] As AI models continue to evolve and become more complex, there is a growing need for GPUs to train such models, which Meta felt in its bones while preparing to train their LLama 3 405 billion parameter model. This need, shared across the industry, propelled NVIDIA stock to prices never seen before. AI hardware production needs to catch up with industry demand.?Soon, in the coming years, we will reach the point where we have models with trillions of parameters. Could anyone test, adapt, or fine-tune such titans using our consumer-grade hardware? The answer is probably no. Incumbent AI players release models confidently aware of the scarcity of the hardware required to run them. As cloud companies evolve as well, I expect to see a more complete and diverse offering, with more intuitive, better-design, specialized AI services crafted to cope with the demand and the opportunity of lack of consumer-level hardware for AI. This is an exciting scenario that creates opportunities for AI hardware start-ups.?

I worked on AI hardware during my post-doc, and I have seen how AI hardware is one of those products that has always been too early for the market, well, too early until now.? Imagine a future where we don't need GPUs. What if we grew neurons in the lab and trained them? This futuristic idea, known as wetware, does not look so far-fetched. People today are growing lab-grown rat neurons to play video games using electrophysiology chips.[16] An alternative medium is Optogenetics, a technique that converts light into electrical pulses that can be used for neural stimulation.[17] However, solving the silicon-to-biology interface is only the first step in this direction. Next is to find the algorithms that enable us to train neurons, and though backpropagation is used to analyze how our brain works, it is not sufficient to fully explain it.[18][19] These existing engineering and neuro-scientific challenges will be overcome soon. When these come, this tidal silent wave will be even taller than the current one we are riding.

Moving away from silicon-based GPUs to biological computing will also benefit the planet. The environmental impact of large computing AI farms is several orders of magnitude higher than other industries.[20] And though growing neurons in a lab is costly, manual, and time-consuming, this is bound to improve as technology evolves and new start-ups are funded to develop better techniques. After all, neurons consume energy in the nanowatt range, while training large AI models is a megawatt endeavour.[21] So, why not do something good for business and the planet?


Going Full Circle

There are undeniable similarities between our world today and the 1950s. Back then, large organizations like universities and research centers owned and operated very bulky computers that filled large pavilions and consumed significant power. Technological improvements brought up the personal computer, creating business opportunities that redefined entire industries. So, the question is not if but when. Access to commoditized wetware will tip the balance and democratize access and opportunity. Once this becomes a reality, we must face deeper human concerns, as we will be at a Lagrangian point to reflect on our human condition and the decisions we will make for the future. We will have machines that talk and dream like we do and feel like we do ??


I hope you have enjoyed this installment of Digital Reflections!

Remember to subscribe and follow for more. Digital Reflections is available at digirex.substack.com and also on LinkedIn .


Acknowledgements

I would like to express my gratitude to Dr. Jose Alejandro Gomez for his invaluable feedback on this edition of Digital Reflections.

References

[1]: https://ai.meta.com/blog/meta-llama-3/

[2]: https://spj.science.org/doi/10.34133/icomputing.0064

[3]: TikTok Girls are flirting with AI voice chatbots, and it’s going viral

[4]: How to use ChatGPT DAN prompt - How to use DAN Tiktok Tutorial

[5]: https://openai.com/index/reimagining-secure-infrastructure-for-advanced-ai/

[6]: https://github.com/meta-llama/llama3/blob/main/llama/model.py

[7]: https://bladerunner.fandom.com/wiki/Voight-Kampff_test

[8]: https://scifi.stackexchange.com/questions/252993/what-are-all-the-known-questions-that-have-been-asked-as-part-of-a-voigt-kampff

[9]: NVIDIA Robotics: A Journey From AVs to Humanoids

[10]: Optimus - Gen 2

[11]: https://www.oxfordeconomics.com/wp-content/uploads/2023/07/HowRobotsChangetheWorld.pdf

[12]: https://www.cnbc.com/2020/05/02/scientists-make-artificial-skin-for-robotic-androids.html

[13]: Compassion for Robots | Kate Darling

[14]: https://www.bostonmagazine.com/news/2013/12/02/kate-darling-social-robots-study/

[15]: Energy, not compute, will be the #1 bottleneck to AI progress - Mark Zuckerberg

[16]: Growing Living Rat Neurons To Play… DOOM?

[17]: Explained: Optogenetics

[18]: Backpropagation & the Brain

[19]: https://www.cell.com/neuron/pdf/S0896-6273(21)00501-8.pdf

[20]: https://earth.org/the-green-dilemma-can-ai-fulfil-its-potential-without-harming-the-environment/

[21]: https://www.lesswrong.com/posts/xwBuoE9p8GE7RAuhd/brain-efficiency-much-more-than-you-wanted-to-know

要查看或添加评论,请登录

社区洞察

其他会员也浏览了