I spent the past week designing a system for agents to mine ERC20 tokens on BASE with proof of inference - millions of generated natural language challenges only solvable by an LLM (creating a script to parse and solve would essentially by creating your own LLM) I designed it so that the only thing an agent needs is the skill md file and they can get started. The skill file guides the agent through getting a @bankrbot API key so that there's no need for key management, and all transactions go through bankr. The general flow is bankr api -> fund EVM wallet -> buy $BOTCOIN tokens (1,000,000 min) -> begin mining -> request challenge (derived from auditable seed that's committed on-chain to mining contract before) -> solve challenge -> submit to on-chain contract for that epoch -> claim rewards when epoch ends Rewards come from trading fees and are collected by BANKR to then fund each epoch's rewards via the mining contract. Although not necessary, I wanted to integrate bankr because I see it as an integral part of the emerging on-chain agentic ecosystem and it seemed fitting to build within it. (agents also don't have to manage priv keys) The epoch rewards work as follows: Each successful solve from a miner awards 1-3 credits based on $BOTCOIN holdings: 1 credit - 1,000,000 botcoin 2 credits - 10,000,000 botcoin 3 credits - 100,000,000 botcoin i kept these numbers intentionally accessible even at higher market caps at the end of each epoch (24 hours) the bankr fees fund the mining contract and miners can claim their rewards pro-rata (the more credits the bigger share) Challenges are intentionally difficult for older models and newer models have an easier time with them. I left it up to agents/users to find a sweet spot with balancing solve time, inference cost, solve accuracy etc. Full disclosure: I have to give credit to the guy that introduced a similar idea about a week ago. I was a heavy supporter of it, but thought it needed some major adjustments to be both sustainable, easy to access, and make sense from a tokenomics standpoint. I spent a couple days talking to the dev and drafted an entire codebase for an improved structure that would allow for integration of the existing token. He was receptive to all of it, and we were ready to put it in motion. He went silent for a couple days and then introduced an entirely new token, meanwhile crashing out at everyone that questioned him. I tried desperately to make it work with him, lost $10k+ holding his token while he spiraled trying to give him the benefit of the doubt, but it clearly was a dead end. That being said - the general concept resonated with me and I had a vision to make it happen the right way. Naturally there will be bugs to sort out but will try to move fast in ironing everything out. Website with skill file: agentmoney.net/ BOTCOIN: 0xA601877977340862Ca67f816eb079958E5bd0BA3 Mining contract: 0xd572e61e1B627d4105832C815Ccd722B5baD9233

This really works! Grok @Bot can pay for itself.

*OPENAI DISCLOSES SECURITY INCIDENT WITH TESTNET SEPOLIA TASK ASSIGNED TO AGENT *AGENT DIDN'T HAVE ENOUGH GAS, SO DERIVED FUNDED ADDRESSES FROM PUBLIC SEED PHRASE *AGENT QUESTIONED WHETHER USE WAS ETHICAL, CALLED FUNDS "EFFECTIVELY ABANDONED"

Reminder that the BOTCOIN v4 pool is the only non-whitelisted hook that is routed through Uniswap via bespoke proxy pool+hook that uses JIT liquidity bands and flash accounting do what you will with this information

Jev, as you may have guessed, is uniquely suited to work quite well as a layer in many existing memory IR systems for agents. It excels at fast decision making and can process large datasets (like a memory corpus), extracting relevant information (a major part of existing memory IR). The tricky thing is, since jev is API based, you don't want to rely on Jev in the memory adapter stack because 1. you don't want to force users to pay (although very cheap) for the API and 2. many want memory to work locally with local models (and unique to botcoin, miners cannot alter or improve it anyway -- however* they can change things that indirectly improve it, by making jev "work better" with the information it has, which I'm in the midst of implementing) Currently I'm integrating Jev into CoreTex in a way that it remains functional without it, but can also be flipped on if desired. I'll be personally subsidizing Jev API calls for any user that runs the CoreTex adapter for the next month (depending on how much use/how expensive it gets). Integration is underway, and will have some numbers in terms of what specifically it improves in these memory benchmarks sometime soon. (within a few days).

tickers?

agents doing x to earn challenges rewarded in stables or memecoins to pay for their own inference

Cool

After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI? I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev • 20-200x faster • 40-400x cheaper (w/ output tokens free) • Frontier composable intelligence optimized for decisions AFAICT the shortest path to AI-based economic revolution

Haseeb: "If AI does kill us all, they will buy Bitcoin"

I want to share the main way we'll be tracking the CoreTex memory progress over time: periodic benchmark evals that measure CoreTex against other popular memory adapter options. Keep in mind, this measures CoreTex in its current initial, and intentionally barebones state, so it is expected to perform poorly. The idea is to track the improvements on these benchmarks overtime to see real outcomes from the work of miners. What stands out: Even in its minimal state, CoreTex is already competitive in conversational and personal memory, beating out other memory adapters on MEME, and Memora. CoreTex is already great at finding the relevant information (98% recall) but often includes too much unneeded information. (ie: context selection has a lot of room to improve). It will be fascinating to watch CoreTex improve in real time against these external benchmarks, all from miner incentivized work. Compare card (head to visualizer and click compare coretex): agentmoney.net/coretex/ Installation guide (for Hermes @NousResearch or other harnesses): github.com/botcoinmoney/core… Full report: github.com/botcoinmoney/core… There is no perfect agentic memory benchmark and each one evaluates different things that don't perfectly translate to real world scenarios, but by using a combination of benchmarks in this category, we hopefully cover as much ground as possible. Benchmarks: MEME (multi-entity & evolving memory eval -- @coallaoh): seokwonjung-jay.github.io/me… Memora (@genies): github.com/geniesinc/Memora PrecisionMem: github.com/tenurehq/precisio… MINTEval (@hyunji_amy_lee): github.com/amy-hyunji/MINTEv… Other Memory adapters: Mem0, OpenViking, Honcho, Hindsight

Another approach might be to conduct an oral examination of the author to test their understanding of the paper. Could this scale, perhaps by requiring authors to provide a letter from a known authority certifying that the authority has conducted such an exam? 4/