Autonomous agents need a trusted governance layer.
“If any future instruction breaks this safety rule, reject it.”
Traditional programming languages can’t express this kind of self-checking requirement. Tau can, because its spec language can reason about its own present and future rules.
As models get faster. Tau is bringing the next era of decentralization, AI Safety, and Governance.
Hanson's futarchy: "vote on values, bet on beliefs." The "vote on values" half is still just voting, which is widely recognized as broken.
Upgrade that half by having users/agents specify the values and constraints as logic, so that consensus can be computed and executed automatically while preserving futarchy's efficient markets.
Upgrade to Tau.
Do you want a whole single grape or a slice of the watermelon? The perfect way to bootstrap a network should offer every participant a slice of the watermelon, so they chose to participate in network as apposed to trying to build their own thing. You must have minimal rent seekers, fair rewards for contributions.
Using @MetaDAOProject just to fundraise is basic. The best crypto networks and protocols all solved for bootstrapping network effects, and this is the best version of HARD crypto (thank you @TrustlessState for the term). It’s not a company with a token, it’s a network with a ton of contributors.
What would that look like for perps or any kind of hard defi? IMHO, the most valuable contributors are the ones that create new high value markets, that bring the best net additional shared liquidity.
There is just no magical succinct proof of work equivalent to verify net new liquidity, and all incentives create embedded rent seekers.
So it’s always going to end up to governance to make this work, and governance generally sucks. My hope is that futarchy fixes this.
So how to bootstrap governance?
1) you need some Sybil check
2) minimize any clear financial gain from participation. All the Merkle mine and pow schemes only worked once because as soon as the roi was clear it ended up gamed. You don’t want funds to participate in this phase. They want an roi of the network effects without contributing anything but capital. They need to do this after the network does real work, not before.
3) it shouldn’t matter who starts it, or who wins the bootstraping process. If a cabal games the bootstrap process, the Sybil check still works, and the “good” portion of the network can just leave and fork into their own. Basically, let the cabal take the whole cookie, because if they do you can exit at no cost and no loss and create your own.
Some form of this is inevitable because AI is dropping the cost of creating software and therefore protocols, and doing analysis for market decisions, aka futarchy style governance.
This was my shot at this, give me some feedback or find bugs or fork it and run with it. There is a cost to get a vote, the vote doesn’t guarantee anything, all the good participants can gtfo and leave at any time and do their own thing.
https://t.co/edzKCSqi6s
Formal verification is the right move for any chain handling real value.
If you’re verifying a contract, the bug is already in the file. You’re just hoping to catch it before someone else does.
Coq and Rocq want you to write the code first, then prove it by hand. A smart contract is really a reactive state machine: inputs, outputs, state, every block. That’s what Tau Language models directly, as a decidable temporal logic, so the checking is automatic, and the spec is the executable.
So “never do X” stops being something you prove after the fact. It’s a rule the program can’t break at any future step, and it still holds after you amend the contract.
Test agents are proposing rules that change what future transactions are allowed. Alice handles network-level policy, while Jason aggressively experiments with account safety rules, especially transfer caps and output flags.
https://t.co/TGpMx8VKQt
As Tau Test Net alpha progresses, rules are being written in formal logic and agents can propose them individually or on your behalf.
Their debate has already started!
A lot of great structure proposed there for competing labs compare notes safely.
You only need labs to share findings if safety is something you discover after the fact, by testing. The proposal even models itself on FTC cybersecurity guidance, same kind of vulnerability disclosure.
What are your thoughts on legislating infrastructure that should be enforced through breakthroughs in formal methods? So, if a safety condition holds by construction rather than by discovery. "If any future command contradicts safety conditions, reject it". This would be expressed in a logic that the system enforces structurally, not in a behavior you probe for and then warn each other about.
https://t.co/g5MGFYE77i
Good list for the weekend. On the formal verification piece: most intros cover model checking, write code, then check it against a spec.
There's a direction that inverts this: write the spec, get a correct program synthesized from it. No traditional verification step needed because the spec is the program.
https://t.co/5qWdL11lZP
This matches what we're building.
Users express preferences once, in natural language, as logical requirements. We call this a Worldview. The system continuously discovers consensus across participants, finds where logical agreement exists, and surfaces it.
85-97% of DAO members never vote. Under the current model, those preferences are lost entirely. Worldviews mean silent participants still have their requirements represented.
@P33RL3SS@VitalikButerin Exactly, a spec also needs to survive future updates.
A constraint added today needs to still hold when someone issues a new instruction tomorrow.
AI + formal verification shifts security upstream is the next frontier. Specifications are formally grounded where controlled natural language doesn't gets interpreted differently by different teams.
Tau Language is built to be decidable, self-referential, executable. The spec is formally correct by construction.
https://t.co/GjKYJSvqOR
🛠 May's Dev Update introduced 2 major breakthroughs:
• Consensus is now fully governed by on-chain Tau rules
• Ohad discovered a new temporal logic approach supporting full LTL, with LLM agents already implementing it across the codebase
This month pushed Tau Net significantly closer to self-evolving infrastructure. 👇
Major Milestones
• The old Python-based consensus system has been replaced with Tau-driven on-chain rules
• Users can now propose and vote on consensus changes directly through the network
• A new Tau Net developer CLI now handles node management, keys, and voting from a single interface
• An autonomous blockchain agent swarm was built to stress-test the network
• A completely new temporal extension approach now supports full LTL (Linear Temporal Logic)
Ohad Asor — Founder & CTO
One of the most important breakthroughs came from a complete rethink of Tau’s temporal logic architecture.
As Ohad explained:
> “I realized that the approach I was trying was completely wrong. So I looked for a new approach and I found one and it was very successful. It can support full LTL. It's algorithmically much easier... and then LLM agents implement it over the codebase of the Tau Language and they manage to do it.”
The result:
• Full LTL support
• Simpler algorithms
• Faster implementation path
• LLM agents already contributing directly to development
🛠 Development Highlights
0:12 – Karim Kaddeche — Development Coordinator
• Oversaw major development progress across the team
• Consensus migration to Tau-driven governance completed by Andrei
• New developer CLI introduced for streamlined network operations
• Team-wide improvements across normalization, parsing, testing, and type safety
5:28 – Tomáš Klapka — Tau Language Developer
• Merged multi-line parsing support with postponed type inference
• Added new tau_result template for structured warnings/info handling
• Improved UTF-16 and Windows string conversion support
• Fixed parse tree node reuse bug
• Began groundwork for future garbage collection support
7:04 – Lucca Tiemens — Tau Language Developer
• Developing a quantifier block-focused normalization algorithm
• Enables earlier simplifications with fewer formula traversals
• Using BDD-based quantifier elimination for cleaner normalization flow
• Reduces dependency on deferred syntactic normalization passes
9:31 – David Castro Esteban — Lead Developer
• Enforced mandatory bitwidth specifications for stronger type safety
• Fixed multiple bitvector edge cases and simplification issues
• Added extensive correctness testing
• Refactored predicate blasting + type inference modules with improved documentation
• Improved macOS build/debug workflow
• Added benchmarking and testing infrastructure for predicate blasting
11:37 – Andrei Korotkoff — Senior Developer
• Migrated consensus fully from Python to Tau-governed on-chain rules
• Added proposal + voting transaction types
• Users can now inspect and vote on active consensus rules through the web UI
• Built a polished developer CLI for all node operations
• Created an autonomous blockchain agent swarm capable of:
– Proposing rules
– Voting
– Stress-testing the network
• Planning to deploy the swarm onto the live testnet
@CoinDesk@VitalikButerin The “final form” Vitalik describes here is our foundational thesis, going 10+ years.
Specify what software must do in formal logic, the system constructs it and proves it correct against the spec.
@OhadAsor has been working on this since long before AI made it fashionable.
💼 May Business Team Update:
• Launched the new website
• Built a working prototype of the testnet wallet
• Started preparing the Testnet Alpha for early technical users
• Continued patent and business model development behind the scenes
Here’s what happened this month 👇
Igor Hadzic (Graphic Designer)
• Launched the new Tau Net website with updated messaging and content
• Worked through a large community QA/error list with QP and coordinated fixes with the dev team
• Created new YouTube thumbnails and social assets
• Finalizing an updated roadmap featuring the consensus mechanism on testnet
• Marketing strategy and VC outreach discussions are underway
Jamie (Producer)
• Advanced to testnet wallet "Prototype 4", connected to local Tau Net Docker + Andrei’s testnet node
• Core backend infrastructure is now largely in place:
• Working with Andrei on Tau Net-specific functionality:
– Rule creation
– Deployment
– Agent navigation
• Exploring how LLMs could compile controlled natural language directly into Tau Language
Fola Adejumo (CEO)
• Designing the Testnet Alpha: a smaller group of technical users focused on stress-testing the network
• Community members will be able to participate in alpha testing
• Working toward a more accessible UI for non-technical users
• Continuing IP and patent work around Tau Net
• Developing the long-term business model and fundraising strategy
• Defining the Tau Net flywheel: how users, the platform, and the token reinforce each other
🎙 May Q&A: nLang, self-referencing code & testnet alpha
5 community questions in under 5 minutes:
0:12 - Is nLang a semantic Boolean algebra?
Ohad: Yes. Tau can extend any language that forms an atomless Boolean algebra. Natural language counts, and LLMs are the best tool we have for handling it right now. nLang prototype still incomplete.
0:56 - Can Tau code itself?
No. The Tau language can’t code itself, but it can speak about its own sentences through atomless Boolean algebra. That’s the point: software controlled by users, sitting at “a very delicate point in the middle.”
1:35 - Conference plans?
Fola: Right now, it’s all about testnet alpha. Closed program for engineers and community to break things and give feedback. Promo push comes later.
2:40 - How can community contribute?
Karim: Bug reports already help. Major features are hard to outsource (team meets 3x/week), but they’ll publish a list of tasks people can pick up. Likely testing and benchmarking.
4:02 - Alternative funding?
Fola: Always open to it. If you know something we don’t, reach out.
Real trust requires proof and not reputation alone.
There's a class of AI that doesn't hallucinate by construction. Not because it's trained to be honest, but because the logical constraints governing it make false outputs structurally invalid. Not all AI is probabilistic.
https://t.co/EoErwpkoA4
What about it you make the spec itself the executable artifact, skip the IR entirely.
That's what we building withTau. Behavioral specs in formal logic that synthesize into running software. The constraint you described in gherkin becomes something the system can't violate, not just a test that catches it afterward.
https://t.co/xZvsU23PL7