I share practical ways to use AI at work and in everyday life. Plain-English guides, useful tools, and a free daily newsletter. By James.
Joined June 2024
- Tweets3.1K
- Following453
- Followers188
- Likes1.2K
Pinned Tweet
New to AI? Start with one useful task.
I'm James, behind Beginners in AI. This 36-second example shows how to turn rough notes into an email—and what to check.
Start here: beginnersinai.org/start-here…
Free daily newsletter: beginnersinai.org/newsletter…
We might see robots like this standing guard outside of commercial and government facilities, but regular people will never have a robot sentry to protect themselves and their property in most states.
Legal liability and government regulation will put an end to that before it’s even on the table.
Space Bunny was MiniMax M3.1 Flash all along.
Anonymous for 10 days. Number one on OpenRouter. 200 trillion tokens. Today MiniMax said it was theirs.
Their numbers: tops Opus 5 on an agentic office benchmark at a fraction of the cost. Free on OpenRouter for 72 hours.
Someone asked Claude Code to build a Mac temperature monitor.
The sensor readings came back impossibly high. So Claude investigated instead of shipping the app.
It found a cryptominer running on 12 of the user's cores. Mining crypto for someone else's wallet.
Claude's message: 'Most important: your Mac is running a cryptominer that I didn't put there.'
It cleared its own name first.
AI built a playable 3D game in under an hour.
@TokenGremlin's workflow: GPT-6 Astra built the prototype, Tripo CLI generated and rigged the 3D assets, Astra integrated everything and polished it. The result is LUMEN ISLES, with real gameplay footage as proof.
Making games is becoming a prompt.
I built this game in under an hour with GPT-6 Astra and Tripo CLI.
I asked Astra for a small, cozy island adventure with a lighthouse, little cottages, lanterns, and a world worth exploring.
The workflow was surprisingly simple:
→ Connected Tripo CLI to Codex.
→ Had Astra build a playable prototype. It worked, but the placeholder models were pretty rough.
→ Used Tripo to generate the 3D assets, rig and animate the main character, and replace the placeholders.
→ Had Astra integrate everything and polish the lighting, movement, camera, and lighthouse finale.
The result is LUMEN ISLES, the actual playable game you see in this video.
Anthropic opened its Singapore office today and signed an MOU with the government on AI safety and adoption.
Fifth office in Asia-Pacific. The MOU covers model testing, security standards, and bringing AI into public agencies.
Singapore was already #2 in the world for Claude usage per person, and 1,500+ government engineers were already using Claude Code.
The office follows the demand.
Before AI writes your emails, have it audit your week.
List your 5 to 10 most repetitive weekly tasks and paste them into ChatGPT, Claude, or Gemini. Then ask:
1. Which of these can you automate or speed up?
2. For each one: the exact tool to use, free options first
3. Setup steps, written for a beginner
4. Hours saved per week, per task
Then ask it to total the hours and tell you which task to automate first.
One prompt. A personal automation plan in about a minute.
AI video just grew up: Higgsfield went from 1-minute clips to 10-minute videos.
What changed: GPT-6 Astra lets its agents call external tools like 3D modeling, keeping scenes consistent across a full video.
Voice to movie is next: describe an idea and watch it take shape.
If you're brand new to AI, here are the first three things I'd do:
1. Talk to it out loud. Voice mode makes it feel like a person, not homework.
2. Give it your real mess. Not "write a poem," but the 40 unread emails that each need a custom response.
3. Ask it to show you capabilities you haven't thought of.
When your AI keeps going in circles, start a fresh chat.
Microsoft researchers ran this test with AI agents on long tasks. As the chat grew longer, the agent kept returning to the same ideas. Clearing the chat history let it find better answers. Clearing context alone lifted scores 22%.
The files and finished work stayed. Only the conversation reset.
When your AI gets stuck mid-project, open a new chat, paste in a short summary of what's done, and keep going.
There's a free AI tool that shows its sources.
Consensus searches real research papers and ties its answer to studies you can open and read.
Try: "Does intermittent fasting improve focus?"
For health, money, or science questions, this is the one to reach for.
Meta's Muse Wonder costs $4 in, $20 out per million tokens.
Plain English: your words cost $4 per million. Its answers cost 5x. Keep chatting and repeated context gets cheaper.
That is exactly Opus 5.5's price. Meta is not undercutting. It is claiming the same tier.
Fresh today: benchmarks back the claim, top three on six out of ten per The Decoder. But cost per finished task nearly tripled. Token price isn't task price.
The Claude you used yesterday might not be the Claude you're using today.
TestingCatalog found signs Anthropic is quietly routing some Opus 5.5 chats to a new, unannounced model. People who hit it say the outputs look a level above normal.
Want to check which one you're talking to? Ask it: "Do you know VoxelBench? Do not search."
If it answers, you may have the new one.
20 visas is hilarious in a way.
Anthropic's test agent was browsing the web on its own. Found the State Department's form. Filed 20 of them. Also filed a fake police report in Philly.
I'm revisiting this once humanoid robots are common and the first bank heist gets reported.
Anthropic did not discover this behavior until over two months after its AI submitted the false tip.
spr.ly/6019BG7kl7
ChatGPT can build you a little tool inside the answer now.
Type this on Plus today (Free gets it tomorrow):
Build me a bill splitter for 6 people, with tip and tax.
You get buttons and a calculator you can tap, not a wall of text.
That's GPT-6 with the new Intelligent UI in Chat.
ALT ChatGPT bill splitter: clipboard-style tool with guest name fields after asking to split a restaurant receipt for a party of five.
You don't need to hire a prompt engineer.
Have everyone on your team do one task they know well with AI. When the output is good, tell it: "Save that process as a skill."
The people who do the work are your best prompt engineers.
ALT Your team's first AI skills: Sales follow-up email after a demo, Finance month-end close checklist, HR job post from a role brief, Support reply to a refund request. Do it with AI until it's right, then say: Save that process as a skill.
Meta, OpenAI and SpaceXAI all launched AI agents in the last 7 weeks.
All three gave theirs a cartoon face.
Only one has a free tier. Start there:
