Search the KHIT Blog

Sunday, July 7, 2024

OK, time to get back to work

Unreal smarts, this young scholar.
My follow-on observation:
To use a DNA analogy, genomic diversity is “adaptive” precisely because–mixing my metaphors–“you can’t ever step in the same river twice.” apropos, see @brianklaas’s killer book “Flukes.” #LLM inbreeding is as maladaptive as genetic inbreeding. House of Windsor, anyone?
OK, I was not hip to her until reading a new Science Magazine review of her current book The AI Mirror. Bought her prior release as well (I have no life).

...[M]ost commercial AI systems today are powered by a machine learning model trained on a large body of data relevant to a specific task, then fine-tuned to optimize its performance on that task. 
This approach to AI has made rapid progress in widening machine capabilities, particularly in tasks using language, where we have the most data to train with. Indeed, since so many kinds of cognitive tasks are language-enabled, most experts now regard the term “Narrow AI” as outmoded, much like its predecessor label “Weak AI.” Very large language models, like OpenAI’s various iterations of GPT or Google DeepMind’s Gemini, can now do an impressively wide variety of things: answer questions, generate poems, lyrics, essays, or spreadsheets, even write and debug software code. Large image models can generate drawings, animations, synthetic photos or videos. While such models have a considerable speed advantage over human performance of these tasks, the quality and reliability of their outputs is often well below the peak of human ability. Still, some see evidence of progress toward AGI in their widening scope of action and the flexibility of a single base model to be fine-tuned for many new tasks. While a large language model (LLM) can’t solve a problem unless the solution is somehow embedded in the language data it is trained on, multimodal models trained on many types of data (text, image, audio, video, etc.) are expanding the performance range of AI models still further. 
Even if it no longer makes sense to call these tools “narrow” AI, they remain below the threshold of general intelligence—AGI. But it’s a mistake to explain that in terms of the problems they can’t yet solve. The true barrier to AGI is that AI tools today lack any lived experience, or even a coherent mental model, of what their data represent: the world beyond the bits stored on the server. This is why we can’t get even the largest AI models to reliably reflect the truth of that world in their outputs. The world is something they cannot access and, therefore, do not know. You might think there’s an easy fix: pair an AI model with a robot and let the robot’s camera and other sensors experience the world! But to an AI model, a robot’s inputs are just another data dump of ones and zeros, no different from image and sound files scraped from the Internet. These ones and zeros don’t organize themselves into the intelligent awareness of an open and continuous world. If they did, the field of intelligent robotics—including driverless cars, social robots, and robots in the service industry—would be progressing much faster. In 2015, fully automated cars and trucks were predicted to be everywhere by the 2020s. Yet in 2023, robotaxis piloted in San Francisco were still driving over firehoses, getting stuck in wet concrete, blocking intersections during busy festival traffic, violating basic rules of the road, obstructing emergency vehicles—even dragging a helpless pedestrian.4 It’s not just driving: the real-world performance of most twenty-first-century commercial robots has lagged well behind AI tools for solving language-based tasks. So, what’s the problem? 
A world is an open-ended, dynamic, and infinitely complex thing. A data set, even the entire corpus of the Internet, is not a world. It’s a flattened, selective digital record of measurements that humans have taken of the world at some point in the past. You can’t reconstitute the open, infinite, lived, and experienced world from any data set; yet data sets are all that any AI model has. You might say, “But surely this is true of the human brain as well! What more do we have than data streams from our eyes, ears, noses, and so on?” But your analog, biological brain remains a far more complex and efficient system than even the most powerful digital computer. In the words of theoretical physicist Michio Kaku, “Sitting on your shoulders is the most complicated object in the known universe.”5 It was built over hundreds of millions of years to give you something no AI system today has: an embodied, living awareness of the world you inhabit. This is why we ought to regard AI today as intelligent only in a metaphorical or loosely derived sense. Intelligence is a name for our cognitive abilities to skillfully cope with the world we awaken in each day.6 Intelligence in a being that has no world to experience is like sound in a vacuum. It’s impossible, because there’s no place for it to be. 
We humans do inhabit and experience a world, one rich with shared meaning and purpose, and, therefore, we can easily place the outputs of our latest AI tools within that context of meaning. We call these outputs “intelligent” because their form, extracted entirely from aggregated human data, unsurprisingly mirrors our own past performances of skilled coping with the world. They reflect back to us images of the very intelligence we have invested in them. Yet accuracy and reliability remain grand challenges for today’s AI tools, because it’s really hard to get a tool to care about the truth of the world when it doesn’t have one. Generative AI systems in particular have a habit of fabricating answers that are statistically plausible, but in fact patently false. If you ask ChatGPT to tell you about me and my career, it usually gets a lot right, but it just makes up the rest. When my host at a festival I was speaking at used ChatGPT to write my bio for the live audience, the tool listed in a confident tone a series of fictitious articles I haven’t written, named as my coauthors people that I’ve never met, and stated that I graduated from the University of California at Berkeley (I have never studied there). 
Importantly, these are not errors. Error implies some kind of failure or miscalculation. But these fabrications are exactly what ChatGPT is designed to do—produce outputs that are statistically plausible given the patterns of the input. It’s very plausible that someone who holds a distinguished professorial chair at a prestigious world university received her degree from another prestigious world university, like UC Berkeley. This fabrication is far more plausible, in fact, than the truth—which is that, due to harsh economic and family circumstances, after high school I attended a local community college in-between full-time work shifts, and later received my bachelor’s degree from a low-ranked (but dirt-cheap and good-quality) commuter university that offered night classes. When I was offered a PhD scholarship at age 25, I became a full-time student again after eight years in the workforce. I first set foot in a college dorm in my 40s, as a university professor. My story isn’t common. And that’s precisely why ChatGPT selected a more “fitting” story for me; quite literally, one that better “fit” the statistical curves of its data model for academic biographies. Later, we’ll consider the cost of relying on AI tools that smooth out the rough, jagged edges of all our lives in order to tell us more “fitting” stories about ourselves. 
These systems can perform computations on the world’s data far faster than we can, but they can’t understand it, because that requires the ability to conceive of more than mathematical structures and relationships within data. AI tools lack a “world model,” a commonsense grasp and flowing awareness of how the world works and fits together. That’s what we humans use to generalize and transfer knowledge across different environments or situations and to solve truly novel problems. AI solves problems too. Yet despite the common use of the term “artificial neural network” to describe the design of many AI models, they solve problems in a very different way than our brains do. AI tools don’t think, because they don’t need to. As this book explains, AI models use mathematical data structures to mimic the outputs of human intelligence—our acts of reasoning, speech, movement, sensing, and so on. They can do this without having the conscious thoughts, feelings, and intentions that drive our actions. Often, this is a benefit to us! It helps when a machine learning model’s computations solve a problem much faster than we could by thinking about it. It’s great when an AI tool finds a new, more efficient solution hidden somewhere in the math that you’d never look for. But your brain does much, much better than AI at coping with the countless problems the world throws at us every day, whose solutions aren’t mathematically predefined or encoded in data...

Vallor, Shannon. The AI Mirror: How to Reclaim Our Humanity in an Age of Machine Thinking (pp. 22-26). Oxford University Press. Kindle Edition.
Dang. This old washed-up guitar player is majorly impressed.

I am briefly reminded of my June post "The Apple of my AI." Also, "The Coming Wave?"
Searching back through my blog turns up a lot of stuff under "Artificial Intelligence." Shannon would likely take issue with a lot of that stuff. 
One of my faves from a few years ago is "The Myth of Artificial Intelligence."


I like it.

Stay tuned. Tons to reflect upon here. Way more to come...

No comments:

Post a Comment