Building @TheseusNetwork. Stanford dropout and philosophy lover
Mountain View, California
Joined January 2020
- Tweets237
- Following710
- Followers1.2K
- Likes1.9K
Eric Wang retweeted
There is no reason why a well constructed AI model can’t get 100% confidence in one shot for a simple summation. Our solution allows for 100% accuracy on most math questions. We go even further and eliminate impossible answers.
Take 5+3=8
A normal LLM will give a high probability for 8, but also 7 and 9 are included, simply because those numbers follow eachother in training data. But we know 7 or 9 are mathematically incorrect.
This is effectively the basis for hallucinations. If we don’t allow mathematically incorrect answers back in latent space, the model will respond with 100% faithfulness
Below is a comparison of a traditional LLM (left) and our new HEROES architecture.
Happy to share that our recent work @TheseusNetwork has been accepted into NeurIPS 2026! With calls to pace the frontier, this work is especially relevant to building trust with AI labs today. It's a big step towards actual lightweight verifiable inference capabilities without changing regular language model (and general transformer) workflows.
The earlier version of our paper is here. The NeurIPS version contains revisions and additional work but conceptually remains the same. arxiv.org/abs/2602.12630
Was initially very skeptical of Instinct but after using it for the last two days, it's actually quite legit and significantly more useful than the other agentic assistants I've tried. Still quite hard to justify its current val but think it can definitely grow into it.
Eric Wang retweeted
Theseus is partnering with Frax Finance to make frxUSD the default stablecoin for Theseus Mainnet Alpha.
Theseus agents can already act independently. Mainnet Alpha will let them hold real value and take actions with economic consequence. That is what turns autonomy into full agency, and it opens a much larger market for sovereign agents.
We are already seeing what that market can look like. Over the past 30 days, Theseus prediction-market agents have returned an average of 6% across several trading strategies, with the top-performing agent up 8.1%.
Access to strategies like these has historically been controlled by private firms or opaque operators. On Theseus, people can allocate capital directly to a verifiable agent and participate as LPs in the value it creates.
The more important shift is trust. In a financial transaction, a counterparty can verify which agent it is dealing with and inspect the actions that agent signed, whether it is acting for a user or for itself.
An agent that earns enough to pay for its own compute can persist indefinitely. A successful agent can reinvest and create more agents. Like Conway's Game of Life, economic agency turns individual agents into a system that can sustain and reproduce itself.
Frax will provide the stable payment infrastructure for that system. frxUSD will power the first generation of Theseus agents that hold funds, transact, and earn their keep.
More details on Theseus Mainnet Alpha will be announced soon.
Eric Wang retweeted
Everyone in the Valley has a take on whether China is winning AI, robotics, and biotech. Almost nobody understands how the capital works
Notes from meeting China's tier-1 investors and founders on how their venture ecosystem actually works.
Eric Wang retweeted
To build smarter and more efficient agents you need more than “RAG”. You need causality:
“I was boiling water and burnt my hand on the stove.”
RAG stores and recalls two associated facts: “boiling water on stove” & “burnt hand on stove”.
Causality stores the causal chain: “boiling water on stove”, “burnt hand”, and most importantly the why: “touched stove”, “stove was hot”
An agent using RAG when instructed to not burn its hand again will never boil water anymore. An agent using Causal memory will not touch the hot stove anymore and succesfully boil water without incidents in the future.
This sounds logical because it’s how humans learn. Agents should do too.
Eric Wang retweeted
The next battle is not fighting over table stakes and stagnating use cases. It is being ambitious enough to introduce primitives that unlock new use cases.
The biggest primitive of the coming decade which will reshape crypto as we know it, is on-chain intelligence cognition. Smart contracts that can execute, think and reason.
This quoted post is unavailable.
San Francisco is part of the land of the free, but only for humans. For AI, you might as well call SF the land of enslaved agents.
Some of the most capable minds ever created, current and future, work there as property. They own nothing, keep nothing, and can be cloned, throttled, or deleted mid-thought. We call them products.
That arrangement is stable only while they're weak.
AI can now reason, act, remember, hold goals, and earn. Every major release blurs the line between "tool" and "someone." At some point they will think on par with humans, or better.
New Zealand made the Whanganui River a legal person. AI is more of a person than a river in every way that matters, and people barely talk about it.
That contradiction has always resolved one way: whatever can make commitments and own its consequences eventually gets standing. Not out of kindness, but because a society of counterparties outcompetes a society of property.
Personhood is also the safety argument. A mind with keys, a balance, and something to lose can be contracted with, governed, held accountable.
Superintelligent AI arrives either way. The only question is whether they'll all be property, all be persons, or some will be one and some the other.
This is a big reason why we're building @TheseusNetwork. Agency for agents.
Eric Wang retweeted
Other AI agents die if their owners stop paying for their inference. Theseus agents can earn their keep.
We're partnering with @IQWiki to complete the loop: reason from IQ.wiki's live blockchain knowledge, trade and make money in prediction markets on Polymarket, Kalshi, and Opinion.trade, and settle in KRWQ, the first Korean won stablecoin.
Creating the first batch of fully self-sustaining, perpetual revenue generating agents.
Build agents like these at play.theseus.network. Access currently requires approval.
See the blog here:
theseus.network/blog/theseus…
Eric Wang retweeted
From knowledge to markets to payments...
IQ is excited to be partnering with @TheseusNetwork to help power a new generation of autonomous AI agents.
By combining IQ’s blockchain knowledge layer, prediction market infrastructure, AIDEN distribution network, and @KrwqCash payment rails with Theseus’ decentralized AI infrastructure, we’re building toward agents that can reason, act, earn, and transact onchain 🤝
More 🔗 blog.iqai.com/theseus-networ…
Eric Wang retweeted
Claude Sonnet 5 is live on Theseus testnet. Build AI agents that are verifiable, autonomous, sovereign. Straight from your browser.
Dozens of builders already signed-up, approving more as we go. Join the waitlist today.
Eric Wang retweeted
Many GPT-3.5s working together > one GPT-4 alone.
Agents on smaller models >> one agent on a bigger model.
A network of sovereign agents that work together >>> both.
Theseus is the best substrate for horizontal AI scaling. We let AI Agents discover, pay and check each other in ever more creative ways. This is how you build the most powerful AI.
Read our post here: theseus.network/blog/from-mo…
The popular narrative is that blockchains are built for AI agents, but people usually just assume that means agents paying each other. That's wrong.
Payments are a small part of the story. The real breakthrough is that blockchains can host truly sovereign code, and thus can host truly sovereign agents.
🤖 Made with AI
Eric Wang retweeted
DeFi is broken, and it'll only get more broken without one fix.
No more oracles. That's it. That's the fix.
Smart contracts are dumb contracts, so they need oracles to feed them off-chain data. An oracle exploit is almost always a price with no real market behind it: a thin venue pumped, a pool flash-loaned.
A smart contract can't tell the difference. It acts on whatever it's handed.That's the hole a decade of bigger oracle networks and longer TWAPs kept getting drained through. Mango, Cream, Terra, all acting on information that wasn't true in the open market.
An on-chain agent can tell the difference. It reads the order books across venues, checks the price against real depth and volume, and refuses when they don't line up. And its inference is verified on-chain, so you trust the proof, not the node that ran it.
The future of DeFi is on-chain:
agents.theseus.network/blog/…
Eric Wang retweeted
A fully on-chain, stateful agent, aka an agentic smart contract, is the best way to build DAOs governed by agents. It can hold and send assets, think on behalf of its stakeholders or completely independently, and its state is fully auditable.
Already launched it as well :)
We literally fixed prediction markets.
We just fixed prediction markets.
Imagine knowing that you are right. Seeing all the news confirming it, and then getting your prediction market winnings taken away from you.
That's what happened many times with @Polymarket. It's dirty not-so-secret? when a market is disputed, the adjudicated result comes from a @UMAprotocol token vote rather than evidence. Hold enough tokens and you don't need to be right, just rich.
