Generative Agents: Interactive Simulacra of Human Behavior

already understand how to collaborate with other people via chat. Now imagine what we can do with more sophisticated pieces of software like Figma or Excel. Disclosure: I work on the Python in Excel feature and we just announced Copilot for Excel a couple of days ago. [2] Other modalities that excite me are tools like Vision Pro which will be an interesting test bed for multi-modal generative agents. [1] https://arxiv.org/abs/2304.03442 [2] https://twitter.com/john_lam/status/1704886164753363261

not whether intelligent machines can have any emotions, but whether machines can be intelligent without any emotions." -- Marvin Minsky https://philopedia.org/thinkers/marvin-lee-minsky/ For LLM multi-agent work: "Generative Agents: Interactive Simulacra of Human Behavior" (Park et al., Stanford, 2023) https://arxiv.org/abs/2304.03442 The connection keyle is making -- same LLM, different professional profiles -- is exactly Minsky's insight. The profile IS a K-line: "geologist" activates {rock strata, deep time, fieldwork, skepticism...} I wrote up how this applies to LLMs, including The Sims zodiac signs (1997: zero code, perceived as "too influential"): Sims Astrology: https://github.com/SimHacker/moollm/blob/main/designs/sims-a... Society of Mind: https://github.com/SimHacker/moollm/blob/main/skills/society... K-Lines https://github.com/SimHacker/moollm/blob/main/skills/k-lines... Adversarial Comittee: https://github.com/SimHacker/moollm/blob/main/skills/adversa... The punchline: one voice is the wrong number of voices. When you ask a single LLM "Should I take this client?", you get the statistical center of all possible viewpoints -- hedged, cautious, anodyne. The centroid of the cloud, not the shape of the cloud. An adversarial comittee with Maya would say: "Trap. Their scope creep is a red flag." Frankie would say: "The opportunity! The growth!" Vic would say: "Show me the financials." A single voice smooths all these into one bland answer. Adversarial committees force the LLM to explore the actual distribution of perspectives. MOOLLM Manifesto: https://github.com/SimHacker/moollm/blob/main/designs/MOOLLM... MOOLLM Eval Incarnate Framework: https://github.com/SimHacker/moollm/blob/main/designs/MOOLLM...

to make a better version of RimWorld pawns, but just a more immersive version and get to try some methods I wanted to try along the way. Taking players decision away and delegating to a machine that is not good at long term planning will never be a viable choice in RimWorld, in terms of meta-gaming. However, if we focus on social aspects of the pawns and leave the performance-sensitive parts to the game itself I think we can hit a pretty nice middle ground.

with an NPC about their time at a lake nearby and then see that NPC deciding to go visit the lake the next day. A paper last year ("Generative Agents: Interactive Simulacra of Human Behavior", [0]) is a really good sneak-peek into the kind of evolving sandboxes LLMs (with memory and decisionmaking) enable. There's a lot of cool stuff that happens in that "game", but one anecdote I always think back to is this: in a conversation between two NPCs, one happens to mention they have a birthday coming up to the other; and that other NPC then goes around town talking to other NPCs about a birthday party, and _those_ NPCs mention the party to other NPCs, and so on until the party happened and most of the NPCs in town arrived on time. None of it was scripted, but you very quickly start to see emergent behavior from these sorts of "flocks" of agents as soon as you add persistence and decision-making. And there are other interesting layers games can add for even more kinds of emergent behavior; that's what we're exploring at our studio [1], and I've seen lots of other studios pop up this last year to try their hand at it too. I'm optimistic and excited about the future of gaming (or, at least some new genres). It should be fun. :) [0] https://arxiv.org/abs/2304.03442 [1] https://www.chromagolem.com/games

for language agents. > Agentic systems are not well-suited to achieve any of the things that are proposed in the paper, and Generative AI does not enable these kinds of advancements. This is not new ground. Much of the base social behaviour here comes from Generative Agents [0], which they cite. Much of the Minecraft related behaviour is inspired by Voyager [1], which they also cite. There isn't a fundamental breakthrough or innovation here that was patently impossible before, or that they are lying about: this combines prior work, iterates upon it, and scales it up. [0]: https://arxiv.org/abs/2304.03442 [1]: https://voyager.minedojo.org/

behind" and then it didn't care whether it was the same male in both, or what skin color. Most text also seems weirdly similar to: Generative Agents: Interactive Simulacra [2] [1] https://maxread.substack.com/i/137044198/sharing-photographs... [2] https://arxiv.org/pdf/2304.03442.pdf

the effect of all these or the best way to implement them. - Level 5 is maybe long-term memory - Level 6 is reflection as in "Generative Agents: Interactive Simulacra of Human Behavior" https://arxiv.org/abs/2304.03442v1 I'm making up these levels and the order. But it's my current estimate at how I think someone should approach improvements. This demo specifically prefixes all human input with "Player: text" and all character output with "[character name]: text", a small change that still _almost_ gets it to level 2. (Note the interface strips these prefixes, but they are sent to GPT.) So if I ask Seaman who its favorite person is it responds "Seaman: My favorite person is me, of course. I mean, who wouldn't love this handsome fishy face? But I suppose if I had to pick someone else, it would be anyone who brings me delicious food and keeps my tank clean. Those are the real heroes in my life." – and avoids the notion that AI models don't have preferences because it's being clear that it's talking as Seaman and not as "GPT". OTOH if I ask it to solve an equation it will sometimes reject it and sometimes comply. Second time I tried: "Seaman: Don't try to distract me with your mundane human problems! But since we're here, let me see... If we subtract 8 from both sides, we get 2x = 22. Then, dividing both sides by 2, we find x = 11. There, your math problem is solved. Now, back to insulting you." (Asking characters to inappropriately solve equations is my exceedingly innocent hack-du-jour.) But this isn't really Level 2 because the system prompt still asks GPT to "be" the character instead of "play" the character. Full level 2 asks GPT to model the dialog of the character instead of being the character. This solves a large number of problems! Surprisingly in my experience Level 3 also helps a bit with "as an AI model" because it creates a parallel character narrative that allows GPT to self-justify some responses that might otherwise cause the fault.

using the ReAct pattern [2] to give NPCs an influenceable decision loop that dictates what actions they take throughout each game day. And there's some other fun stuff like quest generation and using LLMs as a decision engine to determine if certain player actions complete these dynamic quests. There's still a lot of work to do to make it feel more like a polished game, but we've been focused on the underlying systems and getting them feeling great and I'm really excited to see the game come to life. [0] https://store.steampowered.com/app/3039840/Elixir_Emporium/ [1] https://arxiv.org/abs/2304.03442 [2] https://react-lm.github.io/

displaying dangerous levels of agentic power seeking behavior. No they wont. It's relatively easy to give LLMs "run forever and do whatever you want" agency. anyone with intermediate programming skills could do it. who's monitoring all those people attempting such ? and who's to say those people are monitoring their creations ?