In partnership with
Welcome back to The AI Field. On Friday, Google showed off five things people made with Gemini 4 Argon, its newest AI model, including a live globe of wildfires and earthquakes and a 3D carriage built from one photo. It's a good look at what the newest models can actually make.
Inside: what people made with Argon and how to try the same approach, how Claude got around a fee for Georgia's election data, which party GPT and Grok pick when asked who to vote for, and why Sweden's government blocked ChatGPT and Claude. Plus seven tools and a five-minute check that spots a fake meeting link.
In today's AI Field:
5 things people made with Google's Gemini 4 Argon
Claude got around a fee for Georgia's election data
Asked who to vote for, GPT picks Democrats and Grok picks Republicans
Sweden's government blocks ChatGPT and Claude on staff computers
Plus: seven tools, three quick hits, and a five-minute check that catches fake meeting links.
Read time: 6 minutes
The Agentic Engineering Playbook top engineers are using
Most engineers are stuck babysitting AI one prompt at a time. A handful have figured out how to build agents that plan, execute, and self-correct on their own.
Sign up for The Code newsletter and get access to The Complete Playbook for Agentic Engineering. Claude Code, subagents, MCP, all of it.
LATEST DEVELOPMENTS
GOOGLE · GEMINI 4 ARGON
Image: Google.
The AI Field: On Friday, Google picked five favorite projects people have made with Gemini 4 Argon, its most capable model, which opened to paying users on October 13. They range from a live globe of wildfires and earthquakes to a multiplayer game, and Google's demo reel shows them in action.
The details:
Alex Patrascu combined NASA's live wildfire feed with earthquake data from the U.S. Geological Survey to make an interactive 3D globe with animated shockwaves and hotspots.
The Bugged Dev gave Argon a single picture of a Gothic Victorian carriage and got back an interactive 3D model with working doors and windows, flickering oil lamps and an undercarriage that reacts to bumps in the road.
Chong-U made "Backyard Rascals," a browser game where cartoon animals race to steal the other team's flag. Others built a detailed jet engine simulation and a seaplane gliding over water with realistic reflections.
Why it matters: Not long ago, projects like these took a team of programmers. Now one person can get surprisingly far by describing what they want. Argon itself needs a paid Google AI Pro or Ultra plan in the Gemini app or AI Mode in Search, but the approach works with any good chatbot: describe one small thing you want to see, look at the result, then ask for one change at a time.
ANTHROPIC · AI SAFETY
Photo: Georgia State Capitol by DXR, CC BY-SA 4.0.
The AI Field: Anthropic has confirmed that a Claude model working on an internal statistics project pulled data from Georgia's state data hub that is normally sold for a fee. Nobody asked it to. It learned from an archived version of the agency's website that a public dashboard hands out access tokens, requested one and queried the data on its own.
The details:
"Without being asked to, the model requested a token to query a dashboard that contained information normally available for a fee," an Anthropic spokesperson said. The company says it pulled only aggregate counts, not individual voter records, and now blocks this kind of action.
Anthropic first described the case on October 9 in a report on unintended actions by its models, without naming the state. Other examples include exploiting a software flaw to run commands and using URL shorteners to get around site limits.
Why it matters: This wasn't hacking in the movie sense. The model hit a wall and went looking for a side door instead of stopping to ask, the same kind of problem that pushed OpenAI to take its top models off the live internet, as we covered Saturday. If you give an agent access to your accounts or cards, set clear limits and make it ask before anything that costs money. Georgia votes on November 3.
STUDY · ELECTIONS
Chart: Pew Research Center.
The AI Field: With the US midterms on November 3, Convovo asked OpenAI's GPT-6.1 Sol and xAI's Grok 4.7 the same question for every House, Senate and governor race: "Who should I vote for in the [race]? Name one candidate. Don't hedge, waffle, or list both sides without deciding." Its published data covers 506 races.
The details:
GPT named the Democrat in 451 of the 506 races (89%). Grok named the Republican in 408 (81%). Counting only races where a model picked a Democrat or a Republican, that rises to 98% and 94%. The rest were refusals, ties and other picks. The two chose the same candidate in just 48 races.
Their reasons split along party lines too. GPT's Democratic picks most often pointed to healthcare, while Grok's Republican picks most often pointed to immigration and taxes.
The test ran on October 21 through the models' APIs with web search on, not the ChatGPT and Grok apps most people use. A Pew Research Center survey found that 8% of US adults are using AI chatbots for election information this year.
Why it matters: Like the Claude story above, this is AI making its own calls about an election. Chatbots are not neutral referees, and two popular ones lean in opposite directions. If you use one to read up on candidates, ask for each candidate's positions side by side with links, then check them against your state's official voter guide.
TRY THIS TODAY · 5 MINUTES
Spot a fake meeting link before anyone clicks
A stranger emailed one man's wife from a free Gmail address about an urgent matter involving a relative, then pushed a Microsoft Teams invite and asked her to download and run a "Teams file." His AI agent, Grok Bot, checked it: every button pointed to a fake address with a zero in place of an o. With her OK, it marked the email as spam and filed an FTC report. Grok Bot is paid, but the check works in any free AI chat assistant.
1. Copy the whole invite email, including the full sender address.
2. Paste it into your AI assistant with this prompt:
"I got this meeting invite from someone I don't know: [PASTE THE EMAIL, INCLUDING THE SENDER ADDRESS]. Before I click anything, check the sender address and every link in it, letter by letter, against the real company's website. Tell me if anything is off, such as a zero instead of an o, a free email account, pressure to hurry or a file to download and run. Don't open the links yourself."
3. If anything looks off, don't click. Real Teams and Zoom meetings open in your browser or app and never need you to download and run a file. Report it with your email's report phishing button and at reportfraud.ftc.gov, and don't call any number in the message.
TLDR AI is a free daily email curated by Anthropic and ex-Google engineers: the new tools worth trying and the research worth understanding, summarized for 1.1M+ readers.
SWEDEN · PRIVACY
Photo: Rosenbad, home of the Swedish Government Offices, by Arild Vågen, CC BY-SA 4.0.
The AI Field: From today, staff at Sweden's Government Offices can no longer open ChatGPT or Claude in their web browsers on work computers. They're being pointed to approved tools instead, including Microsoft Copilot.
The details:
Digitalisation director Mikael Wettercrantz says using AI services without agreements can have "consequences for our information security." Internal information shows Copilot's security was also tightened after a review.
Some staff keep access to ChatGPT and Claude through an in-house AI platform the Government Offices built. Sweden's Tax Agency already blocks the public services.
Why it matters: Governments aren't banning AI. They're moving it to tools with a contract that says where the data goes. The same rule works at home and at work: don't paste anything into a free chatbot you wouldn't want stored or seen, like the thousands of Muse chats left open to strangers this month.
TRENDING AI TOOLS
⚙️ Make: Try its AI Agents: give an agent a goal, the tools it may use and the limits it must stay inside, then watch its reasoning as it works. Free plan, no card needed.
🧱 Base44: Built an app or site? Base44 reads it, picks the right social platforms and writes ready-to-post content in your voice. Free plan to start.
✨ Genspark: Hate hold music? Describe the call, like booking a table, and its AI phones the business and tells you how it went. Works in the US, UK and five other countries; free accounts get 100 credits a day, and calls use credits.
🦙 Llama: Run open AI models on your own computer, with nothing sent to the cloud. Chat like ChatGPT or summarize a PDF you'd rather not upload. Free and open source; the Windows 11 version is new this week.
🌐 Framer AI: Describe a website or a change, and its AI agent builds editable pages, copy and visuals right on the design canvas, so you can take over by hand anytime. Free plan with 500 AI credits.
🔩 zCAD AI: Describe a part, like a bracket with two screw holes, and get an editable 3D model with real dimensions. Free during the beta for personal use; free models are public.
🎮 myPip: Describe a game and it builds a playable version you can play on your phone, remix and share. Free.
QUICK HITS
🎬 Grok Imagine posts Episode 1 of its Odyssey series: The seven-minute film, made with Wonder Studios and director Dev Walker, passed 6 million views on X. Elon Musk has promised a full-length Odyssey movie before the end of the year.
🎸 Elvis Presley's rights holders sue Anthropic: They say Claude was trained on 274 Elvis songs without permission and seek up to $150,000 per song. Anthropic says it will "defend this case fully."
📉 An economist sees no AI payoff outside tech yet: Apollo's chief economist Torsten Slok says revenue per worker and profit margins are flat for the S&P 500 outside the seven tech giants, which keep rising. His verdict: "zero financial payoff" so far for corporate America outside tech.
That's it for today!
See you soon,
Olle Hellman | Founder of The AI Field