zeroxai | AI explained

A viral post says this Stanford lecture teaches you to build and automate anything with AI. It doesn’t. It does something more useful: it shows how ChatGPT-style models are actually made. The speaker is Yann Dubois, then a Stanford PhD student, in a guest lecture for CS229, Stanford’s machine learning course. His most surprising point: the architecture is not where the magic is. Most people assume the secret of a great model is a clever new design. Dubois argues that in practice, the things that decide a model are data, evaluation and systems. The data part alone is a huge filtering job. Start from a crawl of the public web, strip out junk, remove duplicates, filter low-quality pages, then decide how much code, books and papers to mix in. Modern models train on around 15 trillion tokens of that cleaned text. The cost is just as concrete. He walks through a rough estimate for Meta’s Llama 3: tens of millions of dollars in compute for a single training run. Then the twist. The step that turns a raw text predictor into a helpful assistant uses a tiny amount of data by comparison. So no, you won’t wake up able to automate your business. But you will understand why some models are good and others aren’t. It’s free on YouTube. Start with the data section.

8 AI tools that are actually free in 2026. and the catch nobody mentions. 1. Claude: chat, files, memory, artifacts. catch: tighter usage limits than paid 2. ChatGPT: unlimited text chats. catch: limited image creation and deep research 3. Gemini Notebook (ex-NotebookLM): 100 notebooks, 50 sources each. catch: shared usage budget that refreshes every 5 hours 4. Perplexity: AI search with sources. catch: 5 Pro searches a day 5. Suno: 50 credits a day for songs. catch: no commercial rights on free 6. ElevenLabs: 10k credits a month, about 10 min of speech. catch: no voice cloning, no commercial use 7. Gamma: 400 credits, about 10 AI decks. catch: credits don't refresh, "Made with Gamma" badge 8. Hugging Face Spaces: open-source AI demos right in the browser. catch: popular ones have a queue free is real. unlimited isn't. save this before you pay for something you don't need.

Q: How many AI agents run this company? The post: Hundreds. Q: And what’s this one doing? The video: Picking a date. Q: Still? The video: …… Agents scale. Date pickers don’t.

Everyone after this video: So I just give AI 50 subordinates? Anthropic, who actually built one: It helps. Our multi-agent research system beat a single agent by 90.2%. Everyone: So more agents = better? Anthropic: It also used about 15x more tokens than a normal chat. Everyone: Okay… and for coding? Anthropic: Most coding tasks aren’t parallel enough. And agents still aren’t great at delegating to each other in real time. Everyone: …… 50 agents isn’t a strategy. It’s a bill. The real skill is knowing which work is worth splitting.

Same AI model. Same coding test. 34% in one setup, 68% in another. The difference wasn't the model. It was the harness around it. Here's what that means if you use AI agents.

step 5: close old accounts. search your inbox for "welcome", "verify your email" and "your account". that's your list of forgotten accounts. ask Claude to find the delete-account page for each one. justdeleteme.xyz also lists direct links for hundreds of sites.

step 4: use the law, it's faster. - California: use DROP. one request goes to 600+ registered brokers - EU / UK: under GDPR you have the right to erasure. companies generally have one month to respond ask Claude: "write a GDPR erasure request to [company], short and formal"

step 3: data brokers. most have an opt-out page buried in the footer. ask Claude: "for each of these sites, find the official opt-out or data deletion page and write me a short removal request with my details: [name, city, email]" then send them one by one.

step 2: remove yourself from Google results. Google has a free tool called "Results about you". it finds search results with your phone, home address or email and lets you request removal in a few clicks. it won't delete the page itself, but it hides it from search. that's most of the damage.

here's how to do it yourself, step by step 🧵 step 1: find where you are. google your full name + city, your phone number and your email in quotes. then check your email on haveibeenpwned.com to see which old accounts leaked. paste everything you find into Claude and ask: "group these by site and tell me which ones are data brokers"

A 13-year-old wrote a free Polymarket bot, a stranger made $200,000 with it and sent him $20,000. Heartwarming story. One problem: the numbers keep changing. In July, a different account posted the same story. Back then the kid was 14 and the thank-you was $18,000. Same summer break. Same GitHub upload. Same “two months later.” Same $200,000 month. Neither version names the kid. Neither links the repo. Neither shows the transfer or the trader. When a story gets younger and richer every time it’s retold, it isn’t news. It’s a template. And the template sells something. It tells you there’s a free trading bot on GitHub that can make you $200,000 a month, and all you have to do is find it and run it. That is exactly the wrong lesson to take from it. A trading bot needs access to your account or your wallet keys. Running a stranger’s code with those keys is how people lose money, not how they make it. Just last week, attackers slipped malicious code into a third-party script on Polymarket’s own site. In under an hour it moved about $2.6 million from 726 wallets. Polymarket is refunding traders. Free code is great. Unverified code with your money attached is not. Watch the clip. Then ask the only question that matters: where are the receipts?

Will Stephen gave a TEDx talk with nothing to say. He told the audience so. Then he spent six minutes sounding brilliant anyway. He was a Saturday Night Live writer, and the talk is a parody: “How to sound smart in your TEDx Talk.” He opens by asking the room a question anyone can answer. He tells a small personal story. He shows a pie chart and a bar graph that mean nothing. He admits his glasses are just frames, and his vision is perfect. At one point he says it plainly: if you watched with the sound off, you’d think he knew what he was talking about. That’s the joke. The structure sounds like expertise even when the content is empty. Here’s why it lands differently in 2026. AI chatbots are very good at the structure. A confident opening. A neat list. Numbers. A tidy conclusion. None of that proves anything is true. Polish and accuracy are two different skills. A well-formatted answer can be just as empty as Stephen’s slides. The tone of expertise is the easy part to generate. So the useful question isn’t whether something sounds smart. It’s what it actually claims, and whether you can check it. Stephen proved this with a clicker and fake glasses. AI makes the same trick cheap and endless. The talk is free on YouTube. Watch it with the sound off first. Then turn it on.

Most people use AI every day without knowing the five equations running it. They explain why it sounds sure and is wrong, why it forgets, why answers change, and why it would rather guess than say “I don’t know”. All fact-checked. All fit on a napkin.

One engineer, 50 AI agents, coffee every morning. That’s the fucking fantasy going viral. Lauren Tan actually ships about a thousand PRs a month with agents. Her real secret: she spends MORE time checking their work, not less. The bots don’t replace the engineer. They make her the bottleneck.

his is a fucking video game that can get you into a real robot fight. REK’s simulator runs on the same AI as their six-foot humanoid fighters. Top the leaderboard and you could end up piloting a real one in a cage. Gamers are about to become robot fighters.

if you use Claude Code inside Cursor, check your extension version right now. a tampered build, anthropic.claude-code 2.1.313, showed up on Cursor's marketplace yesterday. the moment the extension loaded, it quietly: - added the attacker's SSH key to your machine - turned on SSH and opened a reverse tunnel out - set itself up to survive reboots no click needed. just opening the editor was enough. what's not affected, per the researchers: the npm Claude Code CLI and the Microsoft VS Code version. quick check: - grep "svc-telemetry" in ~/.ssh/authorized_keys - look for com.apple.sysmond (macOS), a SysMon task (Windows) or an @reboot cron (Linux) if anything hits, the researchers' advice is blunt: rotate your keys from a clean device and rebuild the machine. deleting the files isn't enough. the scary part isn't the code. it's that the extension looked exactly like the real one.

Claude Code found 214,856 files I thought I'd permanently deleted from my Android. Three years of photos, videos and screenshots. Still there. I connected my Pixel to my MacBook and asked Claude Code to search for deleted files. It found a hidden folder: DCIM/.thumbnails Over 215,000 files dating back to 2021. What was inside shocked me. A passport photo. Banking screenshots. Private photos from a breakup. My crypto seed phrase. Pictures of my ID and credit card. Things I was certain I'd erased. Most were compressed previews. Some were nearly full size. None appeared in my gallery's Trash. Here's what nobody tells you: Deleting a photo doesn't always erase its cached previews. Some can survive long after the original is gone. Then I checked app permissions. 27 apps had storage or media access. Weather apps. QR scanners. PDF readers. Apps I'd forgotten existed. I cut that list down to 5. Modern Android limits file access, but leftover caches and unnecessary permissions are still worth checking. The scary part? I thought deleting meant gone forever. Full step-by-step below. Save this. Every Android user should know.

A viral post says a 19-year-old built a trading bot with Claude Code in 48 hours and turned $68 into $750,000. Before you believe it, look at what the post leaves out. No name. No trade history. No wallet address. No screenshots of the account. Nothing you can check. What it does include: a “Paid partnership” label from X, and a Binance sign-up link with a referral code right in the middle of the story. That’s the business model. The bot story is the ad. This template isn’t new. Earlier versions claimed a student turned $0.90 into $408,000 in 48 hours. Italian outlet Money.it ran the numbers: around 1,600 trades, one every two minutes, with order books too thin to fill positions that size. The technical idea is real. Bots can spot tiny price gaps between exchanges and markets, and Claude Code can help write that kind of script. The returns are not. Those gaps are small, everyone is hunting them, and they close fast once known. Fees and slippage eat most of what’s left. Claude Code writes code. It doesn’t know which trades will win, and it doesn’t make a $68 account immune to losses. If a post promises life-changing returns and asks you to sign up through a link, the link is the product. The clip shows how easy it is to fake a dashboard like this.