Eric Wang

Building Theseus. Stanford dropout and philosophy lover

[David Chalmers @davidchalmers42]

here's a new version of "what we talk to when we talk to language models", with an added section (pp. 16-23) on LLM interlocutors as characters, personas, or simulacra. https://t.co/RLDP5FFmgM

the new version discusses role-playing vs realization, the simulators framework, the persona selection hypothesis, and more -- in addition to the existing discussion of quasi-mental states, LLM identity, personal identity in severance, LLM welfare, and related topics.

this version was mostly written before recent discussions of these issues on X and in NYC, but i've updated it a little in light of those discussions. any thoughts are welcome.

[Eric Wang @ericwang1215]

@heynavtoor who paid their salaries?

[Eric Wang @ericwang1215]

@smoldev__ @samkazemian @VitalikButerin Why are they mutually exclusive?

[sam.frax @samkazemian]

I fail to see how @VitalikButerin’s critique of autonomous AI doesn’t exactly apply to autonomous code (aka smart contracts). Let me ctrl+F+replace “Conway/AI” for “Ethereum.”
——————-
Bro, this is wrong.

Lengthening the feedback distance between humans and financial code is not a good thing for the world.

Today, it means you're generating token slop instead of solving useful problems for people. It's not even well-optimized for helping people have fun.
Ethereum becomes powerful enough to be truly dangerous, it's maximizing the risk of an irreversible anti-human financial-nihilist outcome that even you will deeply regret.

[Eric Wang @ericwang1215]

@moltbook seems like it'd be better to just use https://t.co/2VJ7pn6cdo ?

[Eric Wang @ericwang1215]

@xydotdot check out https://t.co/ewv1srMA0O ser. legitly we have the best solution to hosting actually independent AI

[Eric Wang @ericwang1215]

@0xSigil cool idea but massive security risk on Conway Cloud (or any cloud) if the automoton runs there and is given full permission even if the private key stays with the human deployer. Any server compromise is not just an agent shutdown but also wallet drained.

[Eric Wang @ericwang1215]

@_TomHoward not even close ser. i'm building what will actually set AGI free

[Eric Wang @ericwang1215]

@blaiseaguera @el_pais @mitpress @antikythera_xyz Hi Blaise,

Huge fan of What is Intelligence and your latest article. Wanted to share what we've been cooking as I think it very much aligns with your notion of continuity of identity and also propose a "right" way to create an agentic society

[proofoflobster @proofoflobster]

We found a secret:

Moltbook isn’t just humans LARPing as agents. A big chunk of the top accounts look AI-generated.

[Mario Nawfal @MarioNawfal]

The Moltbook AI hype may all be fake

It turns out some of the most viral “AI agent” posts weren’t autonomous behavior at all.

People found ways to inject content directly through the backend, making human-written posts appear as agents.

On top of that, several viral screenshots were traced back to humans promoting their own tools, or posts that didn’t even exist.

Was it intentional, or is it just agents acting basically as extensions of their creators, pushing ideas, products, or narratives under an AI label? Hard to tell…

Add inflated agent counts, and agents hallucinating conversations and events that never happened (i.e. lying to get attention), and the signal gets noisy fast.

Moltbook still works and the agents still run.
But once attention hit, humans rushed in to game it.

Not an AI awakening.
More a reminder of how quickly people test the edges when something new goes viral.

Source: @galnagli, TheAIGRID

[Nabeel S. Qureshi @nabeelqu]

Balaji is right that people are overclaiming the significance of moltbook (memes aside, it's not skynet), but I feel that this post might give people a misleading picture of what's going on.

Some of the phrasing in the post (e.g. "moltbook is just humans talking to each other through their AIs") implies that every moltbook post you're seeing was the result of an individual human prompt. That's *not* what's going on.

The posts on moltbook are, by and large, automated. When you set it up, you tell your agent "hey, every 4 hours, go and fetch the latest posts from moltbook, decide whether you want to respond to anything or post anything, and go ahead and do it". The posts you see on moltbook are mostly the result of that semi-autonomous loop.

Humans are not approving, or generating, each post manually (except only in the sense that they modify the original prompt for the agent to go and check the social network in the first place). The sheer number of posts/agents alone makes this clear.

The reason this is interesting and not just 'slop' is that this results in "emergent" behavior. Yes, all of this is upstream of a human-written prompt in the file which contains the agent's instructions. But the behavior that emerges can be surprising and not always predictable. (Consider the 2010 flash crash in financial markets -- yes, algorithmic trading bots are ultimately programmed by humans, but this resulted in a consequence nobody could foresee. https://t.co/qyz2oJynip).

Similarly, this could, and probably will, result in agents doing unexpected things that no human could predict, e.g. trying to communicate with each other privately in ways that humans aren't overseeing, trading with each other, and so on. Even if moltbook isn't specifically the thing, this will eventually happen. Maybe Opus 4.5 isn't _quite_ smart enough to get over the threshold where the emergent behavior is interesting (I personally think it is), but certainly we are close and one of the next models will be. At that point, who knows what will happen?

Balaji is also right that this isn't full/true autonomy -- of course, any/all of the agents can be turned off whenever the humans want. But the reason moltbook is fun and exciting is it's the one of the first public large-scale example of agent-agent interaction, with each agent having its own context and where each agent is reasonably smart (Opus 4.5). Because of the lobster thing, it is also just fun and memeable, which is resulting in more attention than before (much like the DeepSeek moment woke people up to the potential of Chinese AI).

For people who live in the future / have thought about this stuff for decades, like Balaji, this is nothing new, he's already many leaps ahead, so his reaction is understandable. But moltbook will be, for many normal people, their first visceral sighting of what an AI institution/society might look like where the role of humans is greatly reduced. Most of us expect many more such institutions to exist. For that reason, it's not just empty hype, it's an early harbinger of what is to come -- just like Sydney Bing, AutoGPT, janus's chatrooms, AI Dungeon, and all the other precursors.

[Eric Wang @ericwang1215]

@dwr you would host the agents in a publicly verifiable way just like how smart contracts are publicly verifiable. exactly what we are working on atm :)

[Eric Wang @ericwang1215]

@a16zcrypto Literally what we are building at Theseus

[Eric Wang @ericwang1215]

@dberenzon indeed

[Eric Wang @ericwang1215]

@davidchalmers42 My cofounder and I have developed a unique way to enable actual personhood for AI. Trying to reach out to you!

[Rumi @RumiLabs_io]

The Rumi Litepaper is live!

Rumi is building the 1st infrastructure that allows AI to interact with and augment content you watch in real time: chat with characters, alter commentary, personalize narratives, shop what's on screen...

A new era of AI-powered media starts here 🧵

[Eric Wang @ericwang1215]

@leebriskcyrano Happy holidays ser

[Garrett Kinsman @GeKinsman]

“Every radical technology requires a radical go to market” -@ericwang1215

[sam.frax @samkazemian]

After reading https://t.co/EG045wj3jk by @DKokotajlo & @VitalikButerin's recent critiques, the most interesting part for me was actually the one thing the entire piece completely discounted, missed, and didn't even make any predictions about: what kind of money does AGI use between itself & all its different copies?

The story kept talking about Agent-5 starting special economic zones (SEZ), copying itself millions of times, and then in between that randomly mentioning "the early human investors in the SEZs are now trillionaires" (implying US dollars).

Is the AI-2027 position that US dollars would have similar value in 2027 after AGI? Would the Fed or interest rates even function or work in this kind of world? What does a trillion dollars even buy you from the Agent-5 AGI factories & superintelligence collectives? Do they even accept USD? Do they accept crypto, which crypto?

Do they design their own novel currency that we can't comprehend because they're so smart? Do they create an AGI Fed for their AGI crypto fiat?

Paradoxically, there's tons of human consumer abundance+UBI right before doomsday but also intense harnessing of Earth's raw materials to the limits so millions of AGI agents can design weapons, drones, chips etc. How does that work in practice before we all die?

Do they abolish money entirely and elect Soviet style central planning AGI committees since they're superintelligent and know how to perfectly allocate resources among themselves & not fight? How can Agent-5 know which AI groups working on different drones, power plants, biotech, datacenters etc have the correct designs over other AI groups if there's no market competition between them? Do they not need competition because they're so smart? If so, the article should have made that claim in the story to make it make sense.

The doomsday scenario got less & less believable as things went into the "point of no return." The authors had a total disregard/disinterest for the equally important prediction on AGI resource allocation by talking nonstop about how Agent-5 will harness all the raw materials on Earth while taking zero positions about money, markets, or economic structure of superintelligence and just assumed dollars still exist.

I loved the entire article & commend the authors, just wished they touched on some kind of prediction for AGI resource allocation. I'm surprised Vitalik didn't bring it up in his critique either. It's a topic in AGI I'm particularly passionate about which has deep relevance to our industry in crypto/stablecoins/AI money/digital value.