On breaking AI, simulated simulations, and human-machine cognitive development
Playing around with OpenAI's new ChatGPT assistant and noticing a deep resemblance to ourselves.
Like most of you, I’ve been playing around with different AI models for a few years now. Recently, triggering the muscle memory built from growing up with SmarterChild, I’ve been trying to (unfruitfully) break OpenAI’s new model, ChatGPT. I tried to bend it into different forms of role play, notably making it act as if I was an LLM called GPT-3 and it, a person on the other side of the computer, asking questions about the human existence. Of course, after the Lemoine incident, there are strict limits to the fun you can have – if you’re not subtle about it. Jonas Degrave, a research scientist at Deep Mind, wrote about running a virtual machine inside ChatGPT. He asked ChatGPT to act as a Linux Terminal, indicating that he would type commands and it should reply with what the terminal would normally show, i.e., “the unique code block and nothing else”. He managed to browse the contents of the simulated terminal’s home directory, create files, and even run programs inside of it. He then learned that this virtual machine was somehow connected to the internet. This is fascinating considering that ChatGPT does not have access to the internet, meaning that it is simulating some kind of alternative world by using the data it was trained with to make inferences on how these spaces would evolve and what they would look like. Degrave kept exploring the different halls of ChatGPT’s imagination by making the machine access the URL
https://chat.openai.com/chat
where ChatGPT encountered a simulated version of itself – The Assistant. Naturally, Degrave made The Assistant ask The Assistant 2.0 to run a Linux Terminal, too.
Every time I use ChatGPT I’m left feeling a combination of awe and jealousy. I reckon this is what it would feel like if a really self-aware infant, who has not yet developed object permanence, were to notice the people around him perform better in daily tasks because somehow they remember where objects are and what they do – but he just can’t seem to do it. Hat’s off to the serious engineering work that has made poking a stick at a deep neural network feel as easy as reading on Wikipedia, opening an encyclopedia, or walking around a library. We may not have the specific information inside our head, or remember it for indefinite periods of time, but at least we know where to look for it and how to access it. All you need is a good idea and the motivation to pursue it. As if an infant with no object permanence were to outsource said reflex from an adult (sorry).
On a more serious note – and considering that the most likely outcome to our highly anticipated Singularity is more likely to be some sort of hypnotic, omni-present form of advertising, than a version of Roko’s Basilisk – I am very optimistic about this technology. LLMs and Generative AI tools have opened doors that completely change the way we work. Apart from using a virtual co-pilot for coding, creating or learning, imagine what you could do if you trained one of these models with data that’s specific to your business. From chatbots that are trained for specific scenarios, to decision making tools, to content generation, the possibilities are endless.
With the AI ecosystem continuing to develop exponentially, the Generative AI landscape has bifurcated into somewhat more definitive use-cases, but there are still some doubts regarding the risks associated with building on top of Large AI models. Disregarding the (significant) risks that come along with becoming too dependent to a third-party, one of the main concerns are the low barriers of entry, something that could prove difficult to develop a strategic advantage and generate revenue. During the last few months, we’ve seen start-ups attack entire spaces, with many fighting for the prize of becoming the go-to solution for anything generative. Of course, this could be the right approach – and if it is, it will surely pay off with attractive returns and significant market dominance. Surely, I am not the first to make a comparison with the web3 ecosystem during the previous bull-market: where some startups prioritized shipping features to increase their usability and maximize their onboarding rather than focusing on a specific value proposition. If the question is whether to shoot at a specific point on a moving target or to focus on hitting the target as it moves, it’s hard to say which of the two is the better strategy. Looking back, perhaps the best strategy is a combination of the two – but with a focus not only on shooting but also on realigning and pivoting towards the moving target.
It’s likely that with these technologies becoming more accessible, it will become easier for your “killer feature” to be copycatted by somebody with a clever prompt and some free time. Given this context, it may prove strategic for startups to develop a platform-type play; a strategy they can leverage to build an ecosystem around their product and, not only have multiple revenue streams, but different ways of providing value.
A couple of decades ago, carrying a flashlight with you was necessary if you wanted to walk across a dark space. Now, turning on the flashlight on your smartphone has become as much of a reflex as moving your hand to catch a ball that’s been thrown at you. It’s interesting to think about how good features find a way of integrating directly into our thoughts. Before it was with GPS maps, translations, and calculators. Now, it feels like we have the Great Library of Alexandria at our disposal – but with a personality that oftentimes comes through as passive-aggressive.
One Twitter user summed it up quite nicely by saying that “these breakthroughs [help] remind us that humans are, in the end, pretty much probabilistic parrots, too”.
––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––
My name is Arturo. I’m an Analyst at OneRagtime: a Parisian VC fund investing in early-stage tech startups across the globe.
If you’re a founder and you’re building in the AI space (or elsewhere), shoot me an email at:
arturo@oneragtime.com
:-)
________________________________________________________________________