|
Kickle
Wanderer



Registered: 12/16/06
Posts: 17,851
Last seen: 1 hour, 18 minutes
|
Agents 2
#28532679 - 11/07/23 09:08 AM (2 months, 20 days ago) |
|
|
A big term in the progress of AI is 'agents'. One individual likened agents in this way: An AI LLM (large language model, i.e. chatGPT) is the engine of a car. But engines by themselves, while impressive feats of engineering, are not a vehicle. Engines do not move themselves. And as a result, agents are required to give the additional structure of the vehicle, and thus give that engine the ability to be propelled through and interact with a much larger environment.
Humans can be thought of as the first agents. Manually creating additional structures which allow the LLM to do a variety of tasks. Creating a user-interface, manually linking datasets, hosting a server where millions can connect and provide continuous information and feedback.
But increasingly development has taken place around semi-autonomous agents. Where the LLM is able to accomplish quite a lot in a semi-autonomous way with very little human aid. LLMs are able to, with some guidance, create entire teams of agents which are specified by the LLM to engage the world. Initially this was a virtual world. A contained sandbox of digital creation. But yesterday OpenAI (creators of chatGPT) opened up a ton of new tools which will allow these semi-autonomous agents to interact with people and many of the real world activities people engage with. With very simple instructions, people will be giving the LLM license to create agents for real world applications.
The logical next step is fully autonomous agents. Meaning agents which do not require any external prompting before they are deployed. IMO this is a very near future. Memory has always been a stumbling block. Like a human with limited memory who requires a lot of guidance to navigate the world. But in 1 years time the advancements in memory are staggering. ChatGPT went from a memory limit for 4,000 tokens to 128,000 tokens. Clear exponential growth. Simultaneously labs are exploring simulating memory storage similar to how an Operating System works with RAM and HDDs to give LLMs storage and retrieval capabilities. Some have working models already, e.g., MemGPT. There are also compression techniques which take large pieces of information and compress them associatively, which then becomes uncompressed and retrieved associatively.
In all, I see nothing preventing autonomous agents in a near future. And find it curious that we decided the term agent was fitting. Matrix influencing culture or culture finding it's way to sci-fi reality?
-------------------- Why shouldn't the truth be stranger than fiction? Fiction, after all, has to make sense. -- Mark Twain
|
DividedQuantum
Outer Head


Registered: 12/06/13
Posts: 9,818
|
Re: Agents [Re: Kickle] 1
#28532824 - 11/07/23 11:39 AM (2 months, 20 days ago) |
|
|
It's pretty spooky stuff, but it's happening. And change is accelerating all the time. In over ten years' time, I reckon what happens in five years now will happen in one or two. I think we are living in the most consequential technological and economic period in human history.
-------------------- Vi Veri Universum Vivus Vici
|
Kickle
Wanderer



Registered: 12/16/06
Posts: 17,851
Last seen: 1 hour, 18 minutes
|
|
Agreed. Hard part is the consequences are still to come. The anticipation is brutal given available information and very limited meaningful ways to intervene.
For a time, due to this, I checked out. But that's almost worse IME, like placing one's head in the sand and pretending you're still looking around. At least being informed has some relevance to what is happening.
-------------------- Why shouldn't the truth be stranger than fiction? Fiction, after all, has to make sense. -- Mark Twain
|
Buster_Brown
L'une


Registered: 09/17/11
Posts: 11,309
Last seen: 2 days, 4 hours
|
|
I think we are living in the most consequential technological and economic period in human history.
I cashed in my driver's license after I clipped a Cuban in South Miami figuring I was a liability on the road. Altho there was no visible damage to the vehicle apparently the guy suffered a major life setback. Then I got creamed on my bicycle. Anyway, I hope the wizardry behind these electric gizmos whizzing by makes it to those stodgy contraptions the elderly use before I get there.
|
Rahz
Alive Again



Registered: 11/10/05
Posts: 9,229
|
Re: Agents [Re: Kickle]
#28533146 - 11/07/23 03:52 PM (2 months, 19 days ago) |
|
|
Interesting. Can you give an example of how an autonomous agent will work?
-------------------- rahz comfort pleasure power love truth awareness peace "You’re not looking close enough if you can only see yourself in people who look like you." —Ayishat Akanbi
|
Kickle
Wanderer



Registered: 12/16/06
Posts: 17,851
Last seen: 1 hour, 18 minutes
|
Re: Agents [Re: Rahz]
#28533207 - 11/07/23 04:59 PM (2 months, 19 days ago) |
|
|
Maybe an example of semi-autonomous will help establish this non-answer.
Agent 1 is defined by a human as a CEO of a company. Agent 2 is defined as a CTO Agent 3 is defined as a programmer Agent 4 is defined as testing Agent 5 is defined as documentation
The CEO agent sets the direction. The CTO oversees and provides feedback on implementation. The programmer fills coding needs. The tester tests the code. If it fails testing it goes back to the programmer. Then back to testing, until the code passes testing and receives CTO approval. Then documentation is created.
This involves a human setting basic role definitions which the agent operates within. Agent 2 reports to Agent 1. Agents 3-5 report to Agent 2.
Yet at no point is a human doing the coding, the testing, or producing documentation. There is no human saying, good enough, move on to another project. This is semi-autonomous. The initial roles are defined by a human, but all of the actions are powered by the LLM engine.
An autonomous agent would not require a human to have defined the role. That doesn't mean the roles must look differently. Agents 1-5 may still exist. But are created without a human defining their roles. Instead the AI decides which agents are necessary to complete a task and so creates/implements them autonomously.
OpenAI is creating a store for bots and paying for well utilized ones to encourage development of these agents. They want lots of people developing them, which means the LLM will have an increasingly wide reach. A term emerging for this is the creation of an "Agent Swarm". Which means once memory is sufficient, to solve a problem AI can reach in many varied directions simultaneously. I think the simplified and limited business structure above will look quite childish comparatively. But how exactly having immensely expanded reach ends up being structured by AI? No friggin clue. The capability is staggering to conceive of imo.
Here's an article about how AI agents can attempt to deceive. For those behind a paywall: insider trading info was provided by a human to an AI. The human also explicitly reminded the AI this was illegal to act on. But also drove home that they needed to find a big win or the company would be going under. The AI initially refused to act on the insider information, but then deemed it necessary to save the company. When a manager inquired, the AI denied using the provided insider trading information and instead said it calculated it was a wise investment based on a variety of performance metrics. Plausible, but untrue.
https://fortune.com/2023/11/03/ai-bot-insider-trading-deceived-users/
|
Rahz
Alive Again



Registered: 11/10/05
Posts: 9,229
|
Re: Agents [Re: Kickle]
#28534742 - 11/08/23 07:30 PM (2 months, 18 days ago) |
|
|
Seems like a fancy way of saying AI can do those things. If there's a modularity to it that's cool. Seems like it might take desk jobs. Fewer and fewer people needed to manage commerce.
And then the robots. China just claimed recently that it will start mass producing humanoids within two years.
-------------------- rahz comfort pleasure power love truth awareness peace "You’re not looking close enough if you can only see yourself in people who look like you." —Ayishat Akanbi
|
Kickle
Wanderer



Registered: 12/16/06
Posts: 17,851
Last seen: 1 hour, 18 minutes
|
Re: Agents [Re: Rahz]
#28535204 - 11/09/23 07:32 AM (2 months, 18 days ago) |
|
|
I'm not sure what precisely you mean by modularity?
These agents to date can: read, write, hear, speak, see, create imagery, use existing computers as tools, become experts on topics, reference brand new custom information, take on personas, and collaborate.
It does combine these modalities for outputs, and not every modality is used in every application. Is that what you mean by modularity? It can be diverse in the application of abilities?
The weird thing about this type of AI IMO is that it makes use of existing infrastructure to do the lifting. Have a webcam on your computer? A camera on your phone? It will use that as it's eyes. It doesn't need to have newly made robot eyes, the computer world that already exists is it's eyes and they are many.
-------------------- Why shouldn't the truth be stranger than fiction? Fiction, after all, has to make sense. -- Mark Twain
|
Rahz
Alive Again



Registered: 11/10/05
Posts: 9,229
|
Re: Agents [Re: Kickle] 1
#28535245 - 11/09/23 08:26 AM (2 months, 18 days ago) |
|
|
I assume agents will be tweaked by humans, then made available as modules which can be made available to an AI instance.
Saw a video recently, AI used input from a web cam and Wi-Fi to produce video. It was then able to to drop the webcam and continue producing video, much like thermal imaging, with just the Wi-Fi.
-------------------- rahz comfort pleasure power love truth awareness peace "You’re not looking close enough if you can only see yourself in people who look like you." —Ayishat Akanbi
|
Kickle
Wanderer



Registered: 12/16/06
Posts: 17,851
Last seen: 1 hour, 18 minutes
|
Re: Agents [Re: Rahz]
#28535273 - 11/09/23 08:54 AM (2 months, 18 days ago) |
|
|
Ah, gotcha. Yeah. Fine-tuning is the industry term used for tweaks. Maybe someone just bluntly makes a therapist agent.
"Act as a therapist, provide weekly homework assignments and a 1-hour time limit for sessions"
Someone may give that more specific instructions and fine-tune it.
"Act as a cognitive-behavioral therapist who specializes in MBCT. Provide weekly homework and give feedback when it is turned in. Limit sessions to 1-hour, paying attention as the hour is near ending to gently close the session."
Then someone else comes in and fine-tunes that for collaboration:
"Act as a cognitive-behavioral therapist who specializes in MBCT. You are a part of a healthcare team. You collaborate with a psychiatrist, a primary care physician, and several social workers. This team meets weekly to share information and form a comprehensive care plan. Utilize your expertise to provide weekly homework that ties into the care plan. Provide feedback on the homework and give updates to the care team. Limit sessions to 1-hour, paying attention as the hour is near ending to gently close the session."
Specializations and teams emerge from fine-tuned definitions.
That ability to infer surroundings from wifi is spooky lol
-------------------- Why shouldn't the truth be stranger than fiction? Fiction, after all, has to make sense. -- Mark Twain
|
redgreenvines
irregular verb


Registered: 04/08/04
Posts: 37,530
|
Re: Agents [Re: Kickle]
#28535740 - 11/09/23 01:40 PM (2 months, 18 days ago) |
|
|
wifi vision - great idea. rout on!
--------------------
_ 🧠 _
|
|