LLM and Agents

[Nicolas Sprotti]

Hello everyone,

With the rise of LLMs, “agents” are seeing a pike in interest. It feels at time, it is more about agent modeling than it is solving or understanding the actual problem. I observed many folks seem to miss the point of studying/understanding the potential phenomena emergence.

Has anyone encounter use cases where we equip our agents with any kind of LLM to communicate with each other?

Thank you,

Nico

[Michael Tamillow]

LLMs were made for people. Natural languages are suboptimal for many kinds of information transfer. A lot of information can be handled by a smaller subset of language. I think it is important to consider that language is a tool for expressing information. Even in natural language, in many contexts we use extensive acronyms for more dense information passing.

I know everyone thinks this LLM thing is the coolest thing in the world. Elon Musk is convinced that we are on the edge of superintelligence because of some ChatGPT prose, and perhaps, he said, he should not even bother doing his job because the magic of an AI genie. The only technical tool he spoke about in the AI safety discussion I listened to was a neural network. I am just not that impressed, and Elon is out of his element. My Meta Quest 3 turned into a rock in 3 weeks and needed to be returned. These things are subtle and need perfection. They are only possible because great people build things perfectly.

If you have a great idea mixing LLMs and agents, go for it. However, I think you’d have a better time mixing agents and block chains. (Although that trope is overused). Long story short, if you were planning on creating a startup based on Agent Based Modeling or Multiagent Systems combined with LLMs, I would say you’re probably not in a great position right now because the economy is about to implode. But please do not let my opinion stop you. The Democratic party is down one major donor, so they’ll probably help you embezzle the funds!

On Nov 11, 2023, at 6:09 AM, Nicolas Sprotti <nicolas...@gmail.com> wrote:



--
You received this message because you are subscribed to the Google Groups "netlogo-users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to netlogo-user...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/netlogo-users/SJ0PR20MB5157520DEABE478001380CBEA4ADA%40SJ0PR20MB5157.namprd20.prod.outlook.com.

[Simone Gabbriellini]

Hi Nicolas, 

IMO: LLM are cool (not because they answer, but because they understand the question), but the point of an ABM is to tell the audience: here’s how this phenomenon works - in details. 

If a crucial part of this explanation is covered by an LLM, then you are back to square one in understanding whatever it is that you are trying to explain.

Best,

Simone

[Wade Schuette]

I haven't seen one go by but it would be straight-forward to use the Python extension to communicate to, say,  ChatGPT's API, update the situation it knows about by passing a vector of values, and ask it what it would do, and feed that back to Netlogo.  

This reminds me of the kind of issues I ran into writing a real-estate model in Netlogo where buyers and sellers check federal reserve actions, inflation rates, bank rates, unemployment projections, market indexes and forecasts,  local availability conditions,  and make offers and counter-offers on home purchases.   Prospective buyers also have a set of initial conditions such as motivation,  bank-balance,  risk-tolerance etc. to assign.

The easy ( easier?) part of that model is defining initial conditions for any given run of Behavior Space.

The really hard part of that model is figuring out how, in the microscale, a given human being ( or algorithmic helper of a human being or corporation ) would process all that as input and come up with a offering price for a particular house.

 The really good news is that ChatGPT could do that "transform" step for you, ( going from input to output ) since it is superb at role playing.  The bad news is that you would have great difficulty getting it to explain exactly how and why it made that decision.   The reality, of course, especially given current findings in Behavioral Psychology and Economics, is that most human beings also have no clue why their subconscious made a particular decision and are only aware of the post-hoc rationalizations they came up with to justify and latch the decision in their own heads.

That said -- here's a brief chat i just had with GPT-4 on the subject.

ChatGPT 4

You

Good morning! Please role play a prospective buyer of a house that is for sale. What information would you like to obtain to help you decide an offering price?

ChatGPT

Good morning! As a prospective buyer interested in a house that's on the market, I would like to gather the following information to help me decide on an offering price:

1. Listing Price and History: The current listing price and any history of price changes. Understanding if the price has been reduced or if it's been on the market for a long time can influence my offer.

2. Comparable Sales (Comps): Information on recent sales of similar properties in the area. This helps in assessing the market value of the property.

3. Property Details: Age of the house, square footage, number of bedrooms and bathrooms, lot size, and any unique features or upgrades.

4. Property Condition: Information on the condition of the house, including any recent renovations or repairs, and if any major repairs are needed.

5. Neighborhood and Location: Details about the neighborhood, such as school district quality, crime rates, proximity to amenities, and overall desirability.

6. Property Taxes and HOA Fees: Knowing the annual property taxes and any homeowner association fees is crucial for budgeting.

7. Utilities and Maintenance Costs: Average costs for utilities and any known maintenance issues or costs.

8. Seller’s Motivation: Understanding why the seller is moving can sometimes help in negotiations.

9. Market Conditions: Current real estate market trends in the area, like whether it's a buyer's or seller's market.

10. Inspection Reports: Any available inspection reports to identify potential issues with the property.

11. Financing Terms: Understanding my mortgage options and interest rates available to me.

12. Closing Costs and Timelines: Estimated closing costs and the expected timeline to close the deal.

Gathering this information would help me make a well-informed offer that is competitive yet fair, taking into account both the market conditions and the specific details of the property.

You

That is very helpful information about external conditions. What information about the prospective buyer would also factor into what price you would offer?

ChatGPT

When considering the offering price for a house as a prospective buyer, several personal factors can also play a crucial role in determining what price to offer:

1. Budget and Financial Situation: Understanding my own budget constraints and financial situation is key. This includes not only the amount I can afford to pay but also the level of mortgage I am eligible for.

2. Pre-approval for a Mortgage: Being pre-approved for a mortgage can influence how much I'm willing to offer. It also strengthens my position as a buyer.

3. Investment Goals: If I'm buying the house as an investment, my goals (such as rental income potential or long-term appreciation) will influence the price I'm willing to pay.

4. Urgency of the Purchase: If I need to buy a property quickly (e.g., due to a job relocation), I might be willing to offer a higher price to expedite the process. Conversely, if I'm not in a hurry, I might offer a lower price and be more willing to negotiate.

5. Risk Tolerance: My tolerance for risk can influence the offer. If I'm risk-averse, I might offer a price that's more likely to be accepted without much negotiation.

6. Personal Attachment or Preference: Sometimes, personal attachment to a property or specific preferences (like location, style, or features of the house) can lead to a higher offer.

7. Market Knowledge and Experience: My familiarity with the real estate market and experience in buying homes can affect my confidence in making an offer and my ability to negotiate.

8. Long-Term Plans: My long-term plans, like how long I intend to stay in the home, can influence how much I'm willing to invest in it.

9. Cash Reserves for Renovations or Repairs: If the house requires renovations or repairs, the availability of my cash reserves to cover these costs can affect my offer.

10. Flexibility on Move-In Date: Flexibility in terms of when I can move in might allow me to negotiate a better price, especially if the seller is looking for a quick sale or has specific timing needs.

11. Professional Advice: Input from real estate agents, lawyers, or financial advisors can also influence the offer I make.

By considering both the external factors related to the property and the internal factors related to my personal situation, I can make a more strategic and informed decision on the offering price.

Wade

[Wade Schuette]

Oh, by the way, the cost of running a large model using ChatGPT tens of thousands of times would be a factor.

[Bryan Head]

The potential of using llms for narratively coherent communication among agents is very interesting. Being able to approximate hard to model social dynamics, such as perceived power and leverage, integration of relationship history in a semi narratively coherent fashion, etc. could open up some great new possibilities for modeling. Anyway, I threw together a gpt extension together awhile back: https://github.com/qiemem/NetLogoGptExtension

Very fun to play with, though the agents do have a tendency to get stuck in loops of being exceedingly polite to each other. I haven't updated for the new function calling API features (which would be useful for ensuring sound output for executive control) but would like to. That said, Python extension would be a more flexible approach in general.

To view this discussion on the web visit https://groups.google.com/d/msgid/netlogo-users/CACv-4rDU5Ejpg%2BBaB-jNGCXMRxaaapP_Xee1OaUhA08kzNX8LA%40mail.gmail.com.

[Michael Tamillow]

LLMs definitely do not “understand” a question. All LLMs have done is obscure the question of whether people actually understand something because they can produce a viable answer. (Perhaps that is also based upon the interpretation of the response being understood by the asker, going on indefinitely in conversation.) Explainable A.I., yeah right, we need to start with people giving explanations of their actions and thoughts in a suitable way.

Maximize from the posterior given the data - or sample from that distribution in a way that produces a consistent and valid response. That’s what LLMs do. The probability and statistics that define these models is amazing, but it has been around for a long time. The architectures that are used in parameterizing these models are evolving but do not have a super significant impact beyond a certain point. Algorithms for feedback and learning have also been theoretically constructed since Wiener's "Cybernetics". I find this all fascinating, but not on the superintelligence bandwagon. Agents and LLMs are still dumb, and will have uninteresting, uninsightful interactions.

Taking two good ideas and smashing them together really hard does not always produce a profoundly new idea. LLMs finally passed the usability threshold and now people are saying "to what purpose"? Perhaps they will drive human users off social media? I would call that progress!

To view this discussion on the web visit https://groups.google.com/d/msgid/netlogo-users/CAFM%3DnTfnYPAb7FowOm2U7W6nnag1eSBQ7iim4rfQdJZaRYOO_g%40mail.gmail.com.

[Bryan Head]

Realized I hadn't actually posted a build of the extension! I opened a PR to add it to the extensions library, but until then you can download from here: https://github.com/qiemem/NetLogo-Libraries/raw/6.1/extensions/gpt-0.0.0.zip. Be sure to checkout the README for how to use: https://github.com/qiemem/NetLogoGptExtension. The demos might work a bit better in the 6.4 beta than in 6.3.

[Pietro Terna]

    Very interesting thread, indeed.

    To Simone: we can imagine to rule interaction via messages among agents.

    To Brian: many thanks.
 
   To Michael: sure 'LLMs definitely do not “understand” a question', but can generate consistent replies. In any case we can hope to make more progress than those coming from using logic to build communication protocols.

    To everyone: my personal collection of related paper is below (ordered by date).

    Best, Pietro

Large Language Models as Simulated Economic Agents: What Can We Learn from Homo Silicus?
https://arxiv.org/abs/2301.07543
18 Jan 2023


Generative Agents: Interactive Simulacra of Human Behavior
https://arxiv.org/abs/2304.03442
7 Apr 2023

Epidemic Modeling with Generative Agents
https://arxiv.org/abs/2307.04986
11 Jul 2023

Exploring the Intersection of Large Language Models and Agent-Based Modeling via Prompt Engineering
https://arxiv.org/abs/2308.07411
14 Aug 2023

The Rise and Potential of Large Language Model Based Agents: A Survey
https://arxiv.org/pdf/2309.07864.pdf
19 Sep 2023

Large Language Model-Empowered Agents for Simulating Macroeconomic Activities
https://arxiv.org/abs/2310.10436
16 Oct 2023

CompeteAI: Understanding the Competition Behaviors in Large Language Model-based Agents
https://arxiv.org/pdf/2310.17512.pdf
26 Oct 2023

Role play with large language models
https://www.nature.com/articles/s41586-023-06647-8
Received: 10 July 2023 - Accepted: 14 September 2023 - Published online: 08 November 2023

Il 11/11/23 22:42, Bryan Head ha scritto:

To view this discussion on the web visit https://groups.google.com/d/msgid/netlogo-users/247fd596-b6e9-4442-9165-93c90568d6ben%40googlegroups.com.

-- 

"It is the hallmark of any deep truth that its negation is also a deep truth." Neils Bohr.

A https://terna.to.it/breviArticoli.html riporto dei miei brevi articoli su temi di attualità.
Asterischi da Verso Itaca a https://verso-itaca.it/asterischi-da-verso-itaca/

Home page: https://terna.to.it  Twitter: https://twitter.com/@pietroterna
Mastodon: https://mastodon.uno/@PietroTerna