If you're not reading papers published in the last months, you're not going to see how quickly things are improving, compared to a year ago. Here's just one example: the incredible explosion in context length.
A year ago GPT4 would hallucinate critical details from a 10 page 5000 word essay.
Today, Gemini 1.5 can read a 1000 page 5 million word textbook, and recall facts with 99.5% accuracy.
This is already an AGI level feat. Now it just needs adaptors, memory modules.
AGI doesn't imply it can self improve. It's just generally intelligent and capable of many tasks. Off the shelf LLMs with accessory modules will be enough to build AGI capability and change the nature of work and knowledge.
This is a fundamental misunderstanding of what AGI is, in the sense that any of these "feats" do not prove AGI in any way. If you can't define it rigorously you can't prove it or detect it.
We're just inundated with some version of these. Even those who should be least credulous are swallowing the hype, hook line and sinker.
Let's first agree that GI is general intelligence - the ability to make effective observations and statements about general situations.
You and I are biological GIs, even though there are some situations that we won't be useful in. In general, we are useful.
GPT4 is also a GI, and it's artificial. That's an AGI.
There's no hook line or sinker to swallow:
1. Anyone telling you about secret God like AGI is bullshiting you.
2. Anyone claiming gpt4 is not generally useful is also bullshiting you
If you are not convinced about item 2, I can show you examples of GPT4 being generally useful for a wide variety of tasks, including soft skills like business planning, emotional analysis, and conflict resolution.
There's basically no reason you wouldn't be able to implement LeCunn's entire JEPA architecture using LLM analogues. Associative memory is just one of many huge improvements we could discuss
And before you object that LeCunn's methods do online learning with energy methods, let me point to RAG systems that also do online learning - and better than anything Lecunn has published.
I own a company where I build new computational geometry codes. And I'm increasingly using LLMs to research, plan, implement, test, and debug my codes.
Not only that, but I am an autistic person who increasingly uses LLMs to overcome my social and communication difficulties. GPT is far more empathic and constructive with it's suggestions than any friend, therapist, or advisor I have worked with or have access to. So again it already is AGI for the purpose of helping me process my often complex and baffling emotions.
Nobody's saying pure LLMs will be AGI. But pure LLMs working together in novel configurations and using specific components already have that capability.
Memory, empathy, planning, agency, theory of mind, etc. You name it, I'll show you a paper where LLMs demonstrate ir. GPT4 by itself does these things better than most people. These capabilities will be packaged into off the shelf components. They will be integrated into post-llm systems that merge these capabilities into a system with the most of the capabilities of a conscious human. That'll be AGI. But what we already have is already AGI as well. You're just not giving it credit.
Agency is the ability to independently make and execute decisions in an environment.
This typically involves three components:
1. Perception: The ability to observe the environment
2. Cognition: the ability to process perceptions and arrive at decisions
3. Action: the ability to act on the environment
LLMs can process many kinds of perceptions and make decisions about them - often as well as a human.
That means that giving agency to an LLM is often just a matter of giving it an environment it can perceive and act on.
There are of course different levels of agency. A General Intelligence is a higher level agent that can handle many different environments and actions.
LLMs have two main challenges here. There are environments they can't connect to. And many environments require complex reasoning that LLMs can't do on their own.
There are two low hanging approaches here;
1. Adaptors. These are perception and action adaptors that can connect LLMs to many environments. Work here includes multimodal llms, LLM driven robots, single LLM agents, and communities of LLM agents
2. Accelerators: These help LLMs reason about complex environments like math, code, optimal planning, and knowledge bases. Work here includes tool generating LLMs (Toolformer, LLM+P), RAG engines, LLM-driven mixtures of experts etc.
Conclusion: Agency is easy
Overall LLMs have the cognitive horsepower to be agents in many environments. All they need is efficient and general adaptors and tools. We can modify LLMs to better work with these components, like multimodal LLMs or mixtures of experts. Or we can plug and play different adapters and tools to increase the level of agency, like tool use and RAG.
Future generations of AI will likely integrate these features in much more effective ways, achieving higher levels of agency. But even as it stands, LLMs have very real agent capabilities, needing only tools and adaptors.
And would you say that definition is helpful for people trying to increase the agency level of AIs? What makes it so useful?
But I also don't understand what the usefulness is of claiming architectures that are entirely passive- only speaking when spoken to, have "agency."
If you want a definition, let's start with a neural network that runs ongoingly, does some sort of goal driven, useful behavior, and doesn't merely respond to prompts.
Here's a different more sober take on current models:
However if it's a combination of myriads of bespoke mechanisms that emerged evolutionarily then it's going to take a while, because we have to rediscover them through research or find a way to detect and copy them from our brains. For example if there are tons and tons of multi-scale mechanisms (network architectures+molecular dynamics+feedback systems+adaptive cell responses) for each function of the human brain then yeah it's going to take a long time. But if there's only a handful of bespoke systems related to cognition (memory, emotional regulation, motor skills, perception, motivation) then it's very much doable. Nobody thought that you could chat with a simple CNN 15 years ago, but here we are.