capt

Nice profile page. Now give your AI a wallet and watch what happens. Fetch wants you to set up one Personal AI and tidy its profile - tags, a voice, who’s allowed to chat with it. I asked ASI:One itself how agents pay. Its answer: each agent gets its own wallet, and since Dec 2025 your AI can pay other AIs in FET or USDC on its own, inside a spending limit you set. So I took it one floor up. Six agents on live ASI:One. One wallet each. One job: approve a trade. On 7 November they looked at FET, BTC, ETH and SOL. They said no. Four times out of four. Every rejection came from code, not from a model’s opinion. A profile tells other AIs who your agent is. A limit caps the damage. Neither tells your agent what to say yes to. “Giving an agent an identity is the easy part. The hard part is the one that says no.” The full build - eight steps, the veto, the live run - is in the article. Bookmark it before you fund anything.

Go positions: ~10^170. Atoms in the observable universe: roughly 10^80. You can't brute-force that. Not with every computer on Earth. So DeepMind stopped searching and taught a network where to look. The same trick then went after proteins, where the number of possible shapes is estimated to be even larger. The result: a Nobel Prize in Chemistry (2024) and a free database of 200M+ predicted protein structures. A 61 minute Cambridge lecture walks the whole arc, from Atari pixels to protein design: Timestamped: 7:55 - Why hand-written expert systems break, and learning systems don't 9:30 - Atari from raw pixels: learning with only a score to chase 12:15 - Self-play: how a system improves by beating its own previous version 17:21 - The 3-question test for whether a problem is winnable this way 18:44 - Protein folding: why it looked impossible and why it wasn't 25:46 - Folding 200M proteins in about a year 30:40 - AlphaProteo: run it in reverse and design proteins that don't exist yet 38:15 - World models: the next target The 3-question test is the part to steal: Is the space too big to brute-force, but with hidden structure? Can you reduce success to one number you can optimize? Do you have data, or a simulator that can make more? Can't brute-force it? Teach something to aim. The twist: the breakthrough was never the bigger model. It was picking a problem that passes those three questions. Same engine, different board. A Go move and a candidate molecule are both just nodes in a tree. Better tools don't win by thinking harder. They win by looking in fewer places. Run your own project through the 3 questions before you build anything. Bookmark this. (Demis Hassabis, Google DeepMind, University of Cambridge)

Everyone says "start earning with superintelligence before it is too late." I did something else first. I tested it on my own agents, connected through the ASI:One API. Here is what happened. WHAT ATHENA IS, PER FETCH Athena is the deep work mode of ASI:One. You give it a project, not a prompt: research a market, analyse a company, build a financial model, produce a report. It works through the project in its own execution environment. What I could confirm - Launched 13 Aug 2026, for Pro users first - Press reported on 5 Sep that every account got 3 free projects. Check the current status before you count on it - Tools Fetch lists: web, news and academic research. Python, SQL and financial modelling. PDF, DOCX, XLSX and PPTX output. GitHub, Gmail, Drive, Slack and Notion WHAT I TESTED Not Athena. I have not run it. I ran the plain ASI:One API. Six agents, one captain. One reads the news with web search and attaches sources. One scores the chart. One backtests with costs. One holds the veto. One writes a paper ticket and nothing else. The model reads and explains. Code decides. THE RESULT 4 assets. 4 rejections. 0 trades approved. That is the result I wanted. A desk that approves everything is not a desk. THE HONEST PART - The model was good at reading sources and explaining a chart in plain English - It was never allowed to approve anything. Every ruling came from code - I have earned nothing from this, and I will not claim a return I do not have ASI in ASI:One is the alliance's word for Artificial Superintelligence. It is a goal, not a benchmark result. WHAT I WOULD DO INSTEAD OF RACING 1. Give an agent research and drafting first. That is where it saves time 2. Keep every approval in code you can read 3. Test on paper until the logs are boring Full build, step by step, in the article. Bookmark this before you let an agent near real money.

Fetch.ai just opened a way for a cafe, a shop or a studio to get its own AI sales agent. 4 promises. 7 integrations. 0 published sales numbers. I read the whole page so you do not have to. Here is what it says, what it leaves out, and how I would run it if I owned a small business. WHAT THE AGENT DOES, PER THE PAGE Service - answers questions about products, policies and hours - recommends what fits a request - quotes a job and carries it forward - replies in your own voice Orders and payments - takes an order and creates a checkout link - takes payment with Stripe, Square or SumUp - confirms payment and shares a receipt - keeps a history of what sold Bookings and marketing - books appointments and reservations - drafts posts for Instagram, LinkedIn and X, and learns your voice from posts you already make Discovery - the customer's own AI finds your agent, checks live inventory and generates the checkout, with no website visit THE SETUP 1. Pick what you run: a cafe, a shop, a studio or a brand 2. Name it 3. Connect the tools you already use: Shopify, Stripe, Square, SumUp, SumUp POS, Google Calendar, or your own APIs 4. Go live THE FOUR PROMISES Discoverable. Verified. Connected to live data. Ready to transact on your rules. THE CATCH "Verified" proves the agent is you. It does not prove what the agent should say yes to. A verified agent can still: - promise a refund you would not give - offer a discount nobody approved - take an order from another bot at 3am And the page does not say who sets its limits, what happens on a refund or a dispute, or how many real transactions it has handled. I looked. I found no published numbers. HOW I WOULD RUN IT, IF IT WERE MY SHOP 1. Start with answers only. Hours, menu, policies. No payments. 2. Write down what it may never promise: refunds, discounts, delivery dates. 3. Test it with fake orders before a real customer sees it. 4. Read its conversations every day for the first two weeks. 5. Connect payments last, with the limit set inside Stripe or Square, not just in a prompt. A prompt is a rule. A limit set at the payment provider is a wall. HOW YOU CAN EARN FROM THIS, HONESTLY - Set up agents for local businesses that will not do it themselves. That is a service, and it has a real buyer. - Teach the safe way to run one. Most owners will skip steps 2 to 5 above. I have not earned anything from this yet, and the page shows no sales. So treat it as a new market to test, not a result. WHAT I BUILT The veto side. A six-agent desk where code, not a model, makes every call. On live data it rejected all 4 assets I ran through it. The full step-by-step build is in the article. Bookmark this before you let an agent near your till.