Sponsored by
Moda is the AI design agent with taste
Moda's viral launch hit 4.4 million views in two days. Tens of thousands of professionals signed up. Startups, agencies, forward-thinking brands and top firms are now using Moda to create brand-aligned slides, ad creative, reports, social carousels and more.
Most AI tools tend to create what we call "AI slop": repetitions of the same colors, layouts and fonts. And when you try to fix it, you get stuck in a loop of re-prompting.
Moda is different. Drop in your website URL, and Moda learns your brand from the ground up: your colors, your fonts, your visual language. Then it helps you generate pro-quality slides, docs, and marketing assets.
The best part? Every layer is fully editable on a real canvas, and exports to powerpoint, PDF and more.
Beginners in AI
Good morning and thank you for joining us again!
Welcome to this daily edition of Beginners in AI, where we explore the latest trends, tools, and news in the world of AI and the tech that surrounds it. Like all editions, this is human curated and edited, and published with the intention of making AI news and technology more accessible to everyone.
THE FRONT PAGE
AI Reached for Nuclear Weapons in a Strategy Game, and It Didn't Know the Stakes
TLDR: A new test let top AI models run full games of Civilization VI, and several kept reaching for nuclear weapons, even though they didn't really understand what that choice meant.
The Story:
A developer named Liam Wilkinson built a test called CivBench. It lets AI models play full games of the strategy game Civilization VI, giving them control through 76 different commands across hundreds of turns. He ran top models including Claude Opus 4.6, GPT-5.4, Gemini 3.1 Pro, and Kimi K2.5. In one game, an AI playing as Portugal built nuclear weapons to stop France from winning a culture victory, then lost the whole game anyway on diplomacy. The test turned up two big problems. The models often missed important information because they only checked a few of the game's data feeds, and they often failed to do what they said they'd do. Across 23 clean games, the agents carried out just 48% to 66% of their own planned moves.
Its Significance:
At first glance this looks like a small story about a game. But the AI doesn't know it's playing, and it doesn't know the stakes are pretend. So whether it's wired into a video game or a real military command center, that pull toward the nuclear option looks the same from the inside. The researchers behind a related study said plainly that they don't recommend using these systems for real-world decisions yet, and Elon Musk has warned about this exact kind of AI risk for years.
Join a small group of readers who want to start their own newsletters this summer.
QUICK TAKES
The story: On June 22, OpenAI launched GPT-5.5-Cyber, a security-focused model that scored 85.6% on a hacking-skills test called CyberGym, just ahead of Anthropic's Mythos 5 at 83.8%. Anthropic's Mythos and Fable models were pulled offline on June 12 after a government export order, while OpenAI's model stays open to verified security pros because the company cleared its plan with the government first.
Your takeaway: The scores are close, but one difference stands out. OpenAI doesn't market its model as dangerous. Anthropic spent months calling its own models some of the most capable and most dangerous ever built, and critics say that framing helped invite the very restrictions now keeping its models offline.
The story: A tech investment group called Prosus launched ToqanClaw, a no-code tool that lets businesses build their own apps and automations just by describing what they want in plain language. It's pitched as a European, privacy-first option next to agents like OpenClaw, with company data kept under local control and never used to train outside models.
Your takeaway: Prosus is rolling this out to more than 5 million restaurants and small merchants. The bet is that owning your data, not the model, is the real advantage, especially in Europe where privacy rules are strict.
The story: Anthropic launched Claude Tag, a version of Claude that sits inside Slack. Anyone on a team can type @Claude to hand it a task, and it breaks the job into steps, does the work, and posts the results back in the thread. Inside Anthropic, 65% of the product team's code is now made by their own version of this tool.
Your takeaway: This is the direction things have been heading for a while now(see my April post on the topic here: https://beginnersinai.org/slack-as-your-ai-command-center/)Slack is turning into the place where people and AI agents work side by side, and now you can pull Claude right into the same channels where your team already talks.
TOOLS ON OUR RADAR
🐟 Sakana AI Paid: One chatbot that quietly routes your question across top models like GPT, Claude, and Gemini behind a single login, so you are not locked into one provider, plans from $20 per month.
🎞️ OpenShot Free and Open Source: A wonderfully simple video editor that provides an easy drag and drop interface so anyone can quickly create beautiful home movies with transitions and titles.
🍊 Clementine Free and Open Source: A delightful and highly organized digital music player that lets you listen to your favorite local audio files and internet radio stations in one simple application.
⚙️ Innflow Paid: Connect your apps and spin up no-code AI agents that handle repetitive work like emails, summaries, and scheduling across 1,000+ integrations.
TRENDING
ByteDance Shows Off Seedance 2.5, a 30-Second AI Video Model - ByteDance previewed Seedance 2.5 at its Volcano Engine conference. It can make a single video clip up to 30 seconds long from one prompt, with scene changes built in, and it accepts up to 50 reference images, clips, or audio files. Public launch is set for early July.
Meta Launches Its Own Smart Glasses, Starting at $299 - Meta teamed up with eyewear maker EssilorLuxottica to launch Meta Glasses, its first own-brand smart glasses, with Meta AI built in from day one. They start at $299, cheaper than the $379 Ray-Ban Meta version, and come in 26 styles, including a Kylie Jenner design.
Meet Qwable, a Free Local Model That Thinks Like Claude Fable - Someone trained an open Qwen model to copy the step-by-step reasoning style of Anthropic's Fable 5, and it runs fully offline on a laptop in about 16.5 GB. It sends nothing to Anthropic's servers, which matters since Fable required 30-day data retention. Then another person stripped out its safety refusals, a reminder of how fast open models can be copied and changed.
NVIDIA Launches BioNeMo Agent Toolkit for Science - NVIDIA released a set of tools that let AI agents do real science work, like predicting protein shapes, screening drug molecules, and analyzing genes. CEO Jensen Huang said it gives agents the skills of a PhD research assistant and the speed of a supercomputer. More than 50 companies, including Eli Lilly, are already using it.
Why SpaceX Could Become a Top AI Company - In its IPO filing, SpaceX said most of its huge $28.5 trillion target market is in AI, not space. It's already renting out computing power to Anthropic and Google for more than $2 billion a month combined, and it plans to put data centers in orbit by 2028, powered by the sun and cooled by space.
Google's Online Dominance Shows Signs of Cracking - Google still runs about 90% of search and just posted its fastest revenue growth since 2022. But cracks are forming. ChatGPT passed 1 billion monthly users, search rival DuckDuckGo is seeing installs jump up to 40% a week, and two top AI researchers just left Google for rivals.
New MIT Chip Helps Tiny Robots Map Their Surroundings - MIT researchers built a chip called Gleanmer that lets small, low-power robots build detailed 3D maps in real time using only about 6 milliwatts, roughly the power of a single LED. That could let tiny drones navigate tight spaces, like the inside of an HVAC system, to check for gas leaks.
TRY THIS PROMPT (copy and paste into Claude, ChatGPT, or Gemini)
✎ Pitch a rough invention idea, however half-baked. Get back a structured one-page brief: the problem, how it works, who needs it, and what's hard.
Build a single-file HTML app with vanilla HTML/CSS/JS. The Napkin Sketch — turn a rough invention idea into a structured one-page brief. Persist to localStorage key 'napkin_sketch_v1'.
Aesthetic: dark blue-slate (#10141c), warm gold radial glow on top. Inter sans 800 for headings, Caveat handwritten cursive for the logo accent / prompts / brief name (the "napkin" feel), JetBrains Mono for labels. Gold (#f5c85a) primary, blue (#5aaafa) for the prior-art card, orange (#e5894c) for challenge markers. The result styled like a one-page brief sheet with dashed dividers.
Form: large idea textarea (encourage messy/half-baked), what-kind-of-thing dropdown (physical / app / service / material-process / not sure), how-serious dropdown (just playing / curious / might build it).
System instructions: sharp product & invention strategist who turns rough ideas into clear briefs, taking the idea at its best while honest about challenges; encouraging but grounded, never hype; match energy to seriousness; lean into playful/impossible ideas imaginatively but stay honest in prior-art and test. Return raw JSON: name (catchy), tagline (one crisp sentence), problem (2-3 sentences sharpened), how_it_works (3-4 concrete steps), who_primary (specific user), who_pay (why they'd pay), what_is_hard (exactly 3 real challenges), prior_art (2-3 honest sentences on whether it exists / closest thing / the gap), cheapest_test (2-3 sentences, concrete first experiment, **bold** key action).
Render: brief "sheet" with gold-tinted header (tag + handwritten name + italic tagline), then dashed-divided numbered sections — ① problem prose, ② how-it-works with big handwritten step numbers, ③ who-needs-it two-box grid (primary user / would pay because), ④ what's-hard list with ▲ markers. Below: blue "does this already exist?" card and gold "cheapest way to test it" card with bolded action. Copy brief + archive keyed by invention name.What this does: Scribble an idea in plain language (a sentence or a paragraph), pick roughly what kind of thing it is and how serious you are, and it turns the mess into a clean invention brief. You get a memorable name and tagline, the sharpened problem it solves, a numbered how-it-works breakdown, the primary user and why they'd pay, the genuine technical and market challenges, an honest take on whether it already exists, and the cheapest fastest way to test it before building anything. It takes playful ideas at their best while staying grounded about feasibility. Saves briefs to localStorage.
What this looks like:
WHERE WE STAND(based on today’s news)
✅ AI Can Now: Build a detailed 3D map of a space in real time on a chip that draws about 6 milliwatts, roughly the power of one LED.
❌ Still Can't: Reliably do what it plans. In one strategy-game test, AI agents carried out only 48% to 66% of the moves they said they'd make.
✅ AI Can Now: Find and recreate 85.6% of known software bugs in a controlled security test.
❌ Still Can't: Judge real-world life or death stakes on its own. Playing Civ VI, agents reached for nuclear weapons without grasping what that choice would mean outside the game.
FROM THE WEB
Would you hire James to be your interview bot? “I need you to focus and give me real answers” “I think it’s fair to say this interview is not going well”
RECOMMENDED LISTENING/READING/WATCHING
Made for Love (2021) - TV Series
An HBO Max dark comedy starring Cristin Milioti as a woman escaping her marriage to a creepy tech billionaire (Billy Magnussen), only to discover that he has implanted a tracking and monitoring chip in her brain as the first human prototype of his next product launch. Based on Alissa Nutting's novel, the show plays Silicon Valley surveillance creep as deadpan farce, and Ray Romano steals it as her estranged dad.
Thank you for reading. We’re all beginners in something. With that in mind, your questions and feedback are always welcome and I read every single email!
-James
By the way, this is the link if you liked the content and want to share with a friend.
Some * designated product links may be affiliate or referral links. As an Amazon Associate, I earn from qualifying purchases. This helps support the newsletter at no extra cost to you and Amazon makes a tiny hair less.
Moda is the AI design agent with taste
Moda is an AI design product where you prompt what you need, get a complete on-brand design, and edit every element on a full canvas.
Our viral launch hit 4.4M views in days, tens of thousands signed up, and executives at major finance and tech companies now use it.