Tomorrow’s Turing Test: An Ode to AGI (Artificial General Intelligence) — Navigating the Inevitable Encounter (1 of 3)
The question “Is AGI inevitable ?” has been answered “Yes” — with a touch of anticipation … even an air of reverence ! Now we are on our journey to recognize the unseen and identifying the inevitable arrival !
The next question for us is “Would we recognize AGI when we see one ?” — I will try to answer this question in a set of 3 blogs — very pragmatic, we have to use the tools we have now and project our progress to the inevitable unseen !
I am also curating the concepts and tools in a set of GitHub repositories. Checkout https://github.com/xsankar/Awesome-Awesome-LLM (Top Level), Awesome AGI Concepts and the Metric Minds AGI Evaluation GitHub ! Just starting and it is work in progress …
[Update 12.18.23] The NeurIPS’23 Conference has lots of very interesting papers and workshops in the AGI space ! I have a prelude blog [here] and working on an aftermath blog [done here] — just finished going through 3,584 papers !
1. OpenAI has made its business to be the Captain/Chaos Controller for AGI
In fact, our “new computer overlords” and OpenAI — their closest “Prime Facilitator/Supreme AGI Aide”, have staked their sights to the nearest future ! [Update : In an interview with Gates, Sam Altman has a lot more to say [Here]]
In a span of a few months OpanAI went from “Thoughtful, Unpretentious” to “Intense and scrappy” with a motto “Anything that doesn’t help with that (AGI) is out of scope” !
2. Google DeepMind - Levels of AGI: On the Path to AGI
A very interesting paper by Google Deepmind [Here].
- Note their levels. We have achieved narrow AI at all levels — surpassing human capability !
- But AGI by popular definition (Normative Philosophy of Computing by Seth Lazar [Here] and talk [Here] )is still a few years away :
— AGI = Human-level performance across a sufficiently wide range of tasks, integrated into a single entity that can make plans to achieve goals
— Super intelligence = AGI + significantly better-than-human performance.
- Also remember, going from 90% to 99% to > 100% is not linear at all. The last 1% can take decades, in many cases/tasks.
- But, even Level 4 will suffice as AGI, if it covers a broad spectrum of tasks — especially by a single entity
- On a fundamental level, note their 6th principle — it is a path to AGI not a single end point !
We are still far from the popular concept of super AI, the Skynet ! (Crossing fingers, one never knows, do one ? …)
3. What makes AGI AGI ?
Now we come to crux of our discussion, what are the intrinsic characteristics of the G in AGI (given the 5 levels & 6 principles) ?
Three references — a) Fireside Chat w/ Ilya Sutskever (Chief Scientist Open AI) by Jensen Huang (CEO, Nvidia) (Here) ; b) An excellent interview of Ilya Sutskever (Chief Scientist Open AI) by Dwarkesh Patel (Here); c) My own simple blog ChatGPT- The Smooth Talking Stochastic Parrot (Here)
While Jensen brings out the larger broader principles, Dwarkesh dives into the details ! I like Dwarkesh’s style. As usual Ilya is very insightful and detailed …
1. World Model
- Whether we like it or not, implicit or explicit, LLMs have a world model.
- They are stochastic parrots, compressing the training data to do the next token prediction. But as Ilya puts it “in the process they learn a representation of the underlying reality that produced the text (and images) i.e., a projection of our world !”
2. Reasoning
- Reasoning is am inportant step
- Still not good at multi-step reasoning; math problems were very difficult — getting better.
3. Emergent Behavior
- This is what gets us fascinated about the Generative AI and the promise of AGI. They can do counterintuitive things, create very interesting artifacts and in general have capabilities that blend the human creativity and the computational capabilities of the machines
- This results in the so called “Hallucinations” — either they don’t know enough or they know too much. BTW, hallucination is not a bug but a feature ! But we need to know how to tame these fantastic beasts !
4. Alignment
- Ultimately it is not whether a machine exhibits some super human capability, but the degree of reliability, that determines AGI (maybe 😇)
- Alignment is Reliability + Controllability
- All the training (in the world!), with 1000s of GPUs running months and months, do not specify the desired behavior we wish our neural network to exhibit (we need mechanisms like RLHF et al)
- Ilya says that at the current level of capabilities, we have a good set of ideas of how to align them (RLHF, fine tuning, RAG); but don’t underestimate the difficulty of alignment of future more powerful models
- Existential dread exists, but most probably won’t materialize
- An interesting read on Alignment [Here]
5. Other characteristics
- Interestingly other characteristics like conscience, sentience and ethics can be defined by the above 4 properties
Our next stop is to map the above ideas into pragmatics. That, of course, is the topic of our next blog, next week !
One more thing …
In the meantime I am working on the curation in the github repository — please take a stroll through the repositories and suggest improvements.
https://github.com/xsankar/Awesome-Awesome-LLM (Top Level), Awesome AGI Concepts and the Metric Minds AGI Evaluation GitHub !
Thanks to New Yorker, The Information/Kaval Desai for the very interesting pictures