Decoded Media

I've always wondered what it's like to actually let an ai agent handle money. the version where you say "send 10 usdc to this address", it comes back with "done!", and you just... believe it. the more i thought about it, the more that "done" bothered me. an agent's "done" is usually two things stacked on top of each other: a wallet tool that returned "success", and a model that repeated it back to you. nobody in that loop actually looked at the chain. and the chain is where things go wrong: - the tx can revert, so nothing moved, but the tool still said success - the tx can go through but pay a different wallet, or a different amount - the model can just make up a hash that looks right even "the tx succeeded" isn't enough. a receipt with status 1 means the transaction executed. it doesn't mean the payment you asked for happened. so i built Proofly for the @NousResearch hermes agent hackathon. Proofly is a hermes plugin that sits between the agent's payment tools and you. when a payment tool runs, it takes the exact arguments the tool was called with (recipient, token, amount), pulls the transaction and receipt from base mainnet, and checks: - did it execute? - did the right token move? - to the right wallet? - for the exact amount? - with enough confirmations? if everything matches, the reply goes through with the proof attached: block, confirmations, basescan link. if it doesn't, the "sent!" sentence gets rewritten before you ever see it, with the reason. the video shows three cases: 1. send 5 usdc: the wallet says success, but the tx reverted onchain. nothing moved. Proofly corrects the reply. 2. send 10 usdc: a real tx that executed, but it paid a different wallet ~2.1 usdc. Proofly flags the mismatch and shows expected vs observed. 3. send 30 usdc: right wallet, right token, exact amount. verified, reply kept, proof attached. then /proofly shows every verdict from the chat in a local ledger. to be clear about the demo: the wallet is simulated and the model's replies are scripted, so i could show the worst case on purpose, a model that always says it worked. hermes is real, the tx hashes are real base mainnet transactions, and every check runs live against base rpc. a few choices i cared about: - read-only. no keys, never signs, never sends. it can't move your funds. - it only trusts structured intent (the tool's real arguments, or numbers you give it), never the model's own words or a random hash in the chat. - when it doesn't know, it says so. rpc down means "unverified", not "failed". weird cases like batch transfers get "unsupported", not a guess. limits: base mainnet only, plain eth and erc-20 transfers. and it checks what the tool was asked to do, so if the model put the wrong address into the tool call, Proofly verifies that wrong address. for that, /proofly check lets you check a tx against your own numbers. imo this is where agent payments have to go. not "trust the agent", but "trust the receipt". a model saying done costs nothing. a receipt that matches what you asked for is the part worth paying attention to. try it: hermes plugins install 0xileri/proofly --enable repo: github.com/0xileri/Proofly catalog pr: github.com/NousResearch/herm…

i reverse engineered wispr flow, then used deepseek to fine-tune LFM3-1.7B. the result is on par with the wispr flow models you pay for, and it's free, open source and runs entirely on your mac. key-up latency is 65 ms on my m4 max vs ~520 ms for wispr flow on the same clips. a 6 minute yap session pacing around my office still lands in about a quarter second. > rust backend with a native metal engine: each decode step runs as one fused megakernel > a dflash2 speculative drafter is built into the engine, trained on my local 4 x rtx pro 6000 server > parakeet transcribes, LFM3-1.7B refines. both stream while you talk, so key-up only waits on the last chunk > never leaves your macbook, so slow internet or bad ping is no longer an issue > import your own dictionary of terms for your workflow > i picked every default myself. it feels amazing out of the box > adapts to the app you're in: discord is lowercase casual, slack and email are formal, claude code gets clear prompts > training data: 77k transcripts of public audio and synthetic dictations, each rewritten by deepseek v4.1 flash 1017 following a written spec of what "what you meant" means, in 4 writing modes. none of my own dictations went into it > blind judge (claude opus) on an hour of lectures: phonon's text preferred 90 times, wispr's 23 > built with claude opus 5.5 [xhigh] brew install --cask infatoshi/tap/phonon

i'm approaching altcoins differently this cycle than i did in previous ones we've seen assets like HYPE and ZEC perform independently of BTC and ETH, even during periods when the broader market was struggling i think crypto is gradually moving toward a market where individual fundamentals, capital inflows and actual business performance matter more that makes identifying mispriced assets and understanding which teams are building increasingly important, rather than relying entirely on the direction of the majors