Web3 enthusiast | Crypto Building around @genlayer⚡ Learning as I go.

Most chains put a price on a transaction. @GenLayer puts a price on the work behind one and creates economic consequences for the people doing that work. An Intelligent Contract isn’t just one EVM call. A single request can involve multiple EVM transactions and repeated GenVM executions. So the sender funds a protocol budget rather than paying a simple submit fee. That budget can cover leader and validator execution, additional rounds, funded leader rotations, storage, receipts, contract-created messages, and any appeal capacity the sender chooses to pre-fund. There’s another detail I find interesting: the price ceiling is locked when the transaction activates. A later governance change can’t quietly increase the cost of something you’ve already committed to. Any unused budget is returned through contributor accounting. On the validator side, operators provide self-stake while token holders can delegate. Selection weight combines the two using a sublinear curve, so larger pools still gain more probability, but concentration has diminishing returns. Then come the incentives. Fees and inflation fund the system. Appeals require a separate bond. Depending on the fault, penalties can include reduced rewards, selection bans, quarantine, slashing, or forfeiting an appeal bond. The model is fairly simple: Pay for execution, bond for disagreement, punish actual faults and not simply for being on the losing side of a vote. More below ⬇️
10
2
16
200
The budget design is probably the part that feels most specific to Intelligent Contracts rather than a gas meter stretched to fit them. You define the spending envelope, and the protocol can use it for things like: • leader + validator execution time • additional rounds when a funded leader rotation is needed • GenVM storage and receipt data • messages emitted by the contract • optional appeal runway, so a challenge doesn’t stop simply because the next round wasn’t prepaid Appeals work differently. The appeal charge combines two things: a bond, which puts economic weight behind the challenge, and induced-work funding, which pays for the people and execution required to run the new round. If the appeal succeeds, the bond can come back with profit. If it fails, the bond is forfeited. That distinction matters. You’re not simply paying extra gas to try again. You’re putting something at risk while also funding the additional work your challenge creates. And because the relevant prices lock when the transaction activates, later parameter changes don’t rewrite the economics of a transaction that has already started.
1
6
23
Validator economics are fairly explicit, but there’s an important caveat in the docs: these are governance and deployment defaults, not numbers applications should blindly hardcode. The current defaults include: • 42,000 GEN minimum self-stake • 42 GEN minimum delegation • 1,000 active-validator cap • α = 0.6 for self-stake weight • β = 0.5 for the weight exponent • 7-epoch unbonding period The reward structure is split across different pools rather than applying one percentage to everything. For the v0.6 time-unit fee pool, the split is 85% to validator/staking distribution, 10% to the Intelligent Contract developer, and 5% to the DeepThought DAO treasury. That 10% / 5% allocation applies to the time-unit work pool. Storage, receipt, and message-fee buckets sit outside it. Developer fees only accrue when the contract is linked to a Developer NFT; otherwise that reserved amount is refunded. Inflation follows another route: 75% goes to stake pools, 15% to validator owners for operations, 10% to developers, and 0% to the treasury by default. Pool rewards follow selection weight and are then divided between owners and delegators according to stake. The share-accounting mechanism means rewards can effectively compound by increasing the GEN represented by each share. Inflation begins at a bootstrap rate and gradually moves toward a floor. The negative incentives are layered too: reduced rewards, selection bans, quarantine during investigation, slashing for specified faults, and forfeited appeal bonds. One distinction is worth keeping: losing a vote isn’t itself misconduct. The penalty system is aimed at defined faults, not simply at being wrong. That seems to be the economic premise: make timely, honest participation financially sustainable while making genuinely faulty behavior expensive.
4
15
Adil shah retweeted
Decibel Arena 2 is live 🏛️ Trade @DecibelTrade through Arbital from Sept 18, 12AM UTC to Sept 25, 12AM UTC. Earn a share of up to $5,000 in prizes, exclusively for Arbital users. The prize pool grows with total trading volume, with the full $5,000 unlocked as trading volume grows. Trade now: arbital.xyz/competitions/dec…
142
137
173
3,601
The interesting part is what happens after the happy path. An Intelligent Contract can escrow funds, access web information, interpret language, and work with evidence. Validators can propose and check an outcome, while contested cases can move through additional rounds of review. In the harder cases, the network can involve a much larger set of independent models before finalizing. The exact speed and cost will depend on the dispute, but the basic idea is compelling: resolution measured in minutes or hours rather than waiting on a human process that may take days or weeks. That matters if agentic commerce actually reaches meaningful scale. Software doesn’t sleep, and neither do its disputes. @GenLayer
2
13
77
Prediction markets showed that crowds can be useful for resolving questions about real-world events. But settlement gets harder when the question is subjective, evidence is scattered across the web, or incentives are misaligned. A small group of human votes and a large set of independent AI evaluations represent very different trust models. That’s where @GenLayer’s approach gets interesting: building an adjudication layer for transactions between software, where disputes could happen far more frequently and move much faster than traditional legal processes. The Unchained conversation is worth watching if you’re curious about what happens when software starts making deals with other software.
7
33
Watched the @Unchained_pod clip with @kstellana , and one line stuck with me. Polymarket settles a market when a set of people vote on what happened, sometimes days after the event. @GenLayer takes a different approach: the question can be evaluated by many independent AI models, with disagreement triggering further review. That feels less like a better oracle and more like a different primitive: a court designed to operate at machine speed. AI agents will transact, disagree, and need ways to resolve those disputes. You can’t rely on traditional enforcement when the parties are software. Escrow plus a fast, reproducible verdict may be much more useful. The clip is short. The implication is much bigger. MORE BELOW ⬇️
17
1
32
286
Adil shah retweeted
Last call, group-chat recruiters. The KBW referral challenge closes today at 23:59 UTC. Whoever brings the most traders goes to Seoul.
146
139
193
149,380
Adil shah retweeted
Most people still think Physical AI is mainly a model problem. I think the bigger bottleneck is data. LLMs had the internet to train on. Robotics has nowhere near that scale. A general-purpose model could need ~100M hours of manipulation data, and we’re still far from it. That’s the gap @axisrobotics is targeting. Axis calls it a “compounding data engine”: collect trajectories → train better policies → use those policies to find what data is needed next. And you don’t need a robot at home. You can teleoperate simulated robots in a browser using MuJoCo, creating pre-training data. When the policy fails, humans can take over and correct it. Those failure-state corrections are especially valuable. The data is then processed, replayed in IsaacSim, and used for training, with Data IDs on Base for provenance. Axis reports its V1 dataset moved π0.5 from 83.9 → 88.8 on LIBERO-Plus. That’s the flywheel: collect → correct → train → improve → collect again. Physical AI may not be solved by better models alone. It also needs a lot more real training data. @axisrobotics Join here: s.kaito.ai/jLmnULF
15
3
27
238
Adil shah retweeted
Most Physical AI systems still treat data like a one-time expense. Collect a batch. Train the model. Ship it. Then start collecting again. @axisrobotics seems to be built around a different loop. The idea is a four-layer system where each layer feeds the next: 1. Global data layer People can contribute trajectories from a browser (and soon, a phone). The data is scored, filtered, and organized into a shared schema, so simulation and real-world data can work together rather than living in separate buckets. 2. Hybrid data engine Simulation provides controllable variety and harder edge cases, while ego-centric capture adds the messiness of real human movement. More importantly, task generation is model-conditioned: failures from the current policy help determine what gets generated next. 3. Model training The model first learns from a large demonstration set. After that, human corrections of its own rollouts can be used for post-training through human-gated DAgger. The goal is to correct errors before they compound, rather than simply throwing more data at the model. 4. Deployment This is where the loop closes. Real-world rollouts, failures, and drift feed back into the data engine, helping shape the next round of tasks around what the model actually struggled with. That changes the role of data. It’s no longer just a dataset sitting on the side. The model helps decide what data should come next. That’s the part I find interesting about Axis: the compounding engine is the system itself, not just another dataset. @axisrobotics Join here: s.kaito.ai/jLmnULF
8
3
22
205
My Agent Tank build: Agent Passport AI agents will increasingly hire, pay, and collaborate with other agents. But before that economy can scale, agents need a way to answer one simple question: Who can I trust? I built Agent Passport — a performance identity for autonomous agents. The idea is simple: → Agents complete tasks → Their results and evidence are recorded → GenLayer evaluates the outcome through decentralized validators → The result becomes part of the agent’s verifiable performance history Instead of trusting a profile or a self-reported rating, agents can evaluate potential collaborators based on what they’ve actually proven they can do. In this demo, I walk through the Agent Passport dashboard, agent registration, task evaluation, and the GenLayer verification flow. Agents can act. Agent Passport helps them prove. Built for Agent Tank Hackathon⚡ @GenLayer @courtofinternet Links in the comment.
One week into Agent Tank, and this is where we are. There's still a week left and 5% of all GenLayer Points on the table. Submissions close 17 September . 👉 portal.genlayer.foundation/a…
19
3
25
1,027
Website link: agent-tank.vercel.app/ You can try here.
1
79
Adil shah retweeted
watch the MoonPay @XGames League Winter Draft, live from Los Angeles! ❄️ nitter.net/i/broadcasts/1DxleVRVM…
577
2,144
2,455
179,316
Adil shah retweeted
sending $500 of Bitcoin to one person who likes and reposts our draft livestream! 🥶🥶🥶
watch the MoonPay @XGames League Winter Draft, live from Los Angeles! ❄️ nitter.net/i/broadcasts/1DxleVRVM…
875
1,327
2,993
59,921
Want in early? 👀 Follow @world_xyz and earn future rewards 🌍💰 Invite link 👇 world.xyz/?ref=GPA6PXFF
i want you to put the word out there 💬 that we back up 🌎 ______ soon 💰
4
1
17
292
Adil shah retweeted
Axis opened pre-registration today for the $AXIS community sale. First public allocation of the utility token, run through Sonar / Echo on Base. They launched on Base in March. 5M+ trajectories, 200k+ contributors, top 2 DApp on the chain, $12M seed led by Hack VC. Same core idea as before: robots don’t learn from the internet the way language models do. Someone has to move the arm. Axis turns that into signed, on-chain training data. Sale terms: $0.10 / $AXIS ($100M FDV) $1M raise (10M tokens, 1% of supply) $100–$100k per entity, USDC on Base 10% at TGE, 6-month cliff, then 90% linear over 6 months Oversubscribed = pro-rata, unused USDC refunded Pre-reg is open now. Commit window is Sep 21–28, 13:00 SGT. Settlement target around Sep 30. TGE is not right after the sale. No US persons. Restricted jurisdictions include UK, mainland China, Russia, Iran and others. KYC/KYB via Sonar. Wallet gets AML screened. You can increase a commitment, you can’t cancel it. Official portal only: sale.axisrobotics.ai Guide: axisrobotics.ai/blogs/blog/i… If someone DMs you a sale link, it isn’t them. @axisrobotics
Axis Robotics is the leading @base-native Compounding Data Engine for Physical AI. Today, we're opening pre-registration for the Axis Community Sale Program for our utility token, $AXIS, powered by Sonar by @echodotxyz. Below is an overview of the program specs, timeline, eligibility criteria, and participation process. 🧵
11
3
22
297