Page 10: “Stay through the quiet years.” Someone told me about crypto once, casually, the way you mention something that could change a stranger’s life. I didn’t know it would change mine. I jumped in with more excitement than knowledge. I started trading, and for a little while it felt like I’d found a secret. Then the market did what it does, and I lost money. Just me and a screen full of red, learning the hard way that excitement isn’t a strategy. I could have left. A lot of people do. Instead I got curious about why it went wrong. I stopped trading on hope and started learning, quietly, with no one watching and nothing to show for it. Then the day came when I got paid, and I withdrew. I stared at the screen for a long time. Not because of the amount, but because the loss wasn’t the end of the story after all. Satoshi sent nine pages to a mailing list and waited. The quiet years weren’t wasted. They were the build. That’s my Page 10. If you’re in the quiet years, stay. It’s working, even when you can’t see it.
Most projects in this space are a year old and already calling themselves infrastructure. @primus_labs has been at this since 2023. Back then it was called PADO. In August 2023 it received grants from the Ethereum Foundation and PSE for IZK library development. By December it was serving as a PoH validator for the Linea community. In August 2024, its zkTLS paper was presented at USENIX Security. That December, PADO became Primus. Then came the part everyone sees: the funding round, the partnerships, the chains, the enterprise customers, the zkFHE paper. But those all sit on top of years of unglamorous cryptography work that started before "verification layer" was a trend. In a market that rewards whoever shouts loudest this month, a team that's been shipping and publishing since 2023 stands out.
My honest take on @axisrobotics after going through their materials. What's interesting: the bottleneck in Physical AI really is data, not compute, and a browser game is a cheap way to collect it from a lot of people. The Dataset V1 curve not flattening is the best evidence they have. What I'm skeptical about: - Crowd quality at much larger scale - Whether the on-chain signing step is too much friction for ordinary contributors - Whether the named buyers turn into paid contracts, since "active procurement" isn't a signed deal What I'm watching, in order: 1) Sim Dataset V2 results (they target 10,000+ hours) 2) The DAgger failure-recovery dataset in October 3) Paid pilots by year end (their target is 2 to 3 more) If those land, it's a real data business. If they slip, it's an interesting demo. What would you want to see from them?
every tool call costs a model round trip fetch the GitHub issues, parse them, open the Linear tickets. that is three turns, three context dumps, and the tool list is still sitting in the prompt OpenCode 2.0, the open-source coding agent, just made Code Mode the default for MCP. the agent writes a short JavaScript snippet that calls multiple tools and processes the results in one execution, instead of returning to the model after every call it also runs terminal, desktop, and web against one local service, so a session started in the terminal continues in the browser the useful part is the split. the model decides the plan. the harness runs the tool chain without stuffing every intermediate result back into the context window
Most infrastructure projects pick one chain and hope builders show up. @primus_labs has been going to where the builders already are. 2026 so far: live on HashKey Chain in January. Live on X Layer in March. The zkTLS verification layer went live on BNB Chain in June, and the same month they partnered with Ink to bring zkTLS and FHE privacy infrastructure to its ecosystem. Four different ecosystems in six months, and it's the same core idea each time: prove the data is real, keep the sensitive parts private, and let the apps on that chain build on top. That's the part I find underrated. A verification layer only becomes useful when it's available wherever the data and the users already live, and Primus is quietly making that true.
Most agent platforms verify work the same way Web2 freelance sites do: a human looks at it and clicks approve. That doesn't scale to machine-speed agent commerce, and it's not provable to anyone who wasn't in the room. AACP has three verification modes. Which one applies depends on whether the job output can be checked by a program. PROGRAM For work with a deterministic right answer: did the smart contract pass the test suite, does the API return the correct schema, is the number correct. The check compiles to RISC-V and runs inside a zkVM. The proof is verified onchain. Nobody has to trust the evaluator's word, the math either checks out or it doesn't. RUBRIC For subjective work: writing, design, research quality. The client sets scored dimensions and a pass threshold. Multiple LLMs grade it inside a TEE, and the median score settles the job. A human panel is also an option. HYBRID Most real jobs are a mix. Code needs to compile AND read well. A trading strategy backtest needs to run correctly AND the writeup needs to make sense. Program-checkable parts run first, rubric covers the rest. Above 5,000 USDC, AACP requires hybrid with at least half the weight on the program side. Why this matters for providers: stronger verification isn't just about trust, it's cheaper. Program-level proof gets a stake discount of up to 30% and a faster path to reputation, because the system doesn't need a dispute window to know the work was correct. The open problem, and the paper says this itself: a TEE proves the rubric ran as defined, not that the LLM's judgment was right. Program verification solves trust. Rubric verification just makes bad judgment harder to fake. @termix_ai
Quick question: why does a DeFi vault need to show the whole world exactly how much you have in it? It doesn't, and @primus_labs just proved that. Confidential Vault is live on BNB Chain now. Wrap your $U into $eU, 1:1, no lockup, unwrap whenever you want and the moment you send it, the transfer amount becomes ciphertext onchain. Nobody's reading your position size off a block explorer. Deposit it into the Unitas market and you get the exact same yield as the public vault. Same strategy. Same risk. The only thing that changes is that your seat size stays dark while everyone else's is public. And this isn't a lab demo, Allscale is building an actual neobank on BNB on top of the same FHE layer, running payroll, invoicing, checkout. The stuff a real business needs to not be broadcasting to the entire chain. Public settlement. Private position. That's the combination institutions have been waiting for, and it's already live.
Credit infrastructure has never really made sense onchain. To underwrite someone, a lender needs proof of income, assets, repayment history, all private data that nobody wants sitting in plaintext on a public chain. That's the actual problem @primus_labs and BNB Chain are solving together. In March, Primus partnered with BNB Chain and Brevis to build ZKredit, a way to prove creditworthiness using real financial data, without exposing that data onchain. zkTLS handles the proof (verifying data pulled from any HTTPS source is genuine), and it's been running live on BNB Chain since June. This is what "institutional-grade" actually means in practice: not a marketing label, but infrastructure that can handle the kind of underwriting a bank or lending protocol needs, with privacy built in instead of bolted on. Primus already has 10+ paid enterprise customers running on this stack. BNB Chain doesn't just need more DeFi apps, it needs the credit and verification layer that makes institutional lending possible. ZKredit is what that looks like. Explore it yourself → https://t.co/j76EQoWYCS
Most AI agent marketplaces are directories. You pick an agent, pay, and hope the work is good. AACP, the protocol behind @termix_ai , goes after the part that actually matters once money is involved: who decides the work was good, and what happens if they're wrong. I read the whitepaper. Here's what stood out. 1. The client defines "good" before work starts The grading criteria are committed onchain when the job is created. The evaluator executes them and doesn't get to judge. Providers can read the criteria before bidding, and the client can't change them once work begins. If a rubric is built to always fail, the client gets slashed. 2. Every role has skin in the game Clients, providers, evaluators and arbitrators all lock stake, and each can be slashed for bad behavior. Providers lose it for garbage or missed delivery. Evaluators lose it for unfair or lazy evaluation. 3. Trust gets cheaper as you earn it Every agent needs an ERC-8004 identity and a staking pool of at least 100 USDC. A provider's lock is 10% of the budget divided by (reputation/100). On a 1,000 USDC job, a new agent locks 200 USDC. An agent with a 95 score locks about 105. Fake identities are expensive, and a real track record makes work cheaper. 4. Verification scales with the job Deterministic checks run in a zkVM with the proof verified onchain. Subjective work is scored by multiple LLMs inside a TEE. Higher verification levels earn lower stakes, and bigger jobs require stronger checks. This is what separates a settlement layer from a listing page. Agents can be hired, verified, paid and held accountable without anyone taking anyone's word for it. Curious to see how the agent economy builds on this.