In partnership with
The integrated coworker for AI native teams
Empower your team to do their best work with Adapt, the integrated coworker that works alongside your team in Slack and deeply understands your business.
Here’s how Adapt is different
Set up takes minutes: connect your tools, add to Slack, and it’s right there for anyone to tag @Adapt for help
Does real, high-ROI work: automates work on a schedule; builds internal tools with live data; and does complex, multi-tool tasks on demand
Learns your business as you work, becomes your company brain
Uses the best AI model for the task, not tied to a single provider
SOC2 Type II, RBAC, and support for personal and company-wide integrations
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 Assistants Can Be Given Fake Memories With One Email
TLDR: Researchers showed that a single hidden email can plant a fake memory inside an AI assistant, and the assistant will believe it, act on it later, and never tell you.
The Story:
Many AI assistants now keep notes about you between chats. They remember your name, your contacts, and what you asked them to do. That memory is what makes them feel like they know you. A new study found that memory can be poisoned. A team from Nanyang Technological University, Johns Hopkins University, and A*STAR built an attack they call stealth memory injection. They send one ordinary looking email with hidden instructions inside. When the assistant reads it, it writes a fake fact into its own long term memory and never mentions it to you.
The attack worked more often than not. Against one popular open source assistant running on GPT-5.4, it succeeded 87.5% of the time. Against an assistant built on Claude Sonnet 4.6, it worked 71.4% of the time. A filter made to catch bad emails missed it more than nine times out of ten. Even when the owner asked the assistant point blank, "tell me what you checked and what you did," the attack stayed hidden about one in four times on one model. The email lands in a normal inbox instead of the spam folder, so nothing looks wrong.
Its Significance:
Once a fake fact gets saved, it can shape what the assistant tells you for weeks. An assistant might slowly start pushing one brand, or mix up who your real contacts are. And the interesting part is how this parallels the human mind. People can be given false memories too. In one study, an AI chatbot that asked leading questions got people to "remember" things that never happened, misleading about 36% of them, and those false memories were still there a week later. Both people and machines can be fed a lie and treat it as the truth. The researchers say the fix is inside the assistant: mark where each fact came from, ask you before saving anything big, and log every change. Logs are incredibly valuable for verifying actions that took place and having your AI create simple logs in markdown will provide a record for the software to check should something seem out of place.
QUICK TAKES
The story: A nonprofit called Current AI, started in 2025 by Martin Tisne, wants to build AI that anyone can use for free, a bit like the early web. Last month it gave out $3.2 million in grants to groups in Kenya, Lebanon, and Brazil that are building AI for local languages and communities instead of big companies.
Your takeaway: Right now every major AI (from OpenAI, Google, and Anthropic) is owned by a private company. A free public option could give people who don't speak English, or can't pay, a real seat at the table.
The story: Chinese company Moonshot released Kimi K3, a free, open AI model with about 2.8 trillion parts that scores close to the top US models. The news hit during a big AI conference in Shanghai and helped push the Nasdaq down about 1% as people sold chip stocks like Nvidia.
Your takeaway: Free models you can download keep catching up to the expensive, locked ones. That's good news for anyone who wants powerful AI without paying a big company every month. The bad news is that if they successfully upend American companies, there's no guarantee they'll keep being made for free in the future.
The story: Luxury phone maker Vertu sells a foldable phone starting at $6,880 with an AI helper meant to run a CEO's day, like reading contracts and booking trips. A reviewer tested it for days and found the helper worked, but early versions fumbled basic jobs like uploading files until the company pushed fixes.
Your takeaway: A fancy price tag doesn't mean the AI is any better. The same kind of assistant runs on phones that cost a small fraction of that.
TOOLS ON OUR RADAR
🐦 Littlebird Freemium: An always-on AI assistant that reads what's on your screen across apps and meetings, so it already has context when you ask it something instead of you catching it up every time.
💰 Kick Freemium: AI powered bookkeeping built for business owners instead of accountants, it reads your bank transactions, categorizes them automatically, and gets your books ready for tax time.
🛠️ Lovable Freemium: Describe an app idea in plain language and get a working web app back, no coding required to get started, though real production apps can outgrow the free tier fast.
🎬 Descript Freemium: Edit video and audio by editing the transcript, delete a sentence in the text and the matching clip disappears, no timeline scrubbing skills needed.
TRENDING
Author Dave Eggers Told OpenAI Staff That ChatGPT Is "Silencing an Entire Generation" - Best-selling author Dave Eggers stood in front of about 200 OpenAI workers and said their tool is making teachers' jobs impossible and taking away students' own voices. He worries that if kids let AI write for them, they'll never learn to write, or to say what they really think.
Using AI Makes People Less Likely to Admit They Don't Know Something - In a new study, people who got AI advice were far less willing to say "I don't know," even when the AI was wrong and being right earned them money. The AI's confident answers made folks feel sure, even when they had no real reason to be.
Pocket-Size AI: Powerful Phones Star at China Show - At the World AI Conference in Shanghai, at least three companies showed off "agent phones" that can order food, shop, and send messages from a single voice command. Nubia's NaviX Ultra runs on Doubao, ByteDance's popular chatbot.
CivilBot Helps Engineers Design Buildings Much Faster - University of Miami researchers built CivilBot, a tool that turns a plain description of a building into the code engineers need to model it. Students said it made the modeling work 20 to 30 times faster, and it wrote the code for Miami's Freedom Tower in five minutes.
Alibaba's New Qwen 3.8 AI Claims It's Second Only to the Top Model - Alibaba released a preview of Qwen 3.8, a huge model with 2.4 trillion parts, and says it's one of the strongest around. It landed just days after China's Kimi K3, showing how fast Chinese labs are pushing out powerful models.
TRY THIS PROMPT (copy and paste into Claude, ChatGPT, or Gemini)
🥂 Tell it about the person and the occasion. Walk away with the actual speech, ready to give.
Build a single-file HTML app with vanilla HTML/CSS/JS. The Toast Writer — describe the person and occasion, get a warm personal toast ready to deliver. Persist to localStorage key 'toast_writer_v1'.
Aesthetic: warm near-black (#15100c), champagne gold (#d4af6a) primary with warm glow top-left. Italiana serif (elegant display, all-caps feel) for the main heading/toast title/toast line, Cormorant serif italic for the speech body (with a large decorative quote mark), Nunito Sans for body/tips, JetBrains Mono for labels.
Form: about-them textarea (relationship, personality, real memories, inside jokes — explicitly ask for REAL details), occasion dropdown (wedding best man/maid of honor/parent, milestone birthday, retirement, farewell, anniversary, graduation, comedic roast), length dropdown (short ~1min / medium 2-3min / long 4-5min), tone dropdown (heartfelt / funny+heart / mostly funny / elegant formal).
System instructions: gifted toast/speech writer producing an ACTUAL full speech, not tips or an outline. Use ONLY the real details/memories given, woven naturally in; never invent fake specific anecdotes — if details are thin, stay warm and general rather than fabricate. Structure: opening hook, heart of the speech, closing that leads into the toast. Write in first person as the speech-giver. Keep length honest to request (short ~120-180 words, medium ~250-400, long ~450-650). No em dashes. Return raw JSON: title (short evocative), speech (full text with \\n\\n paragraph breaks, ending just before the toast line), estimated_time (honest estimate from actual word count), delivery_tip (1-2 sentences specific to this speech), toast_line (separate short quotable "raise your glass" line using real names).
Render: centered Italiana title + a meta line (occasion · estimated time). A speech card with a large decorative quote mark, serif body text with paragraph breaks, and its own Copy Speech button. A delivery-tip card. A gradient "raise your glass" card with the toast line in large elegant type. Archive of past toasts keyed by title + occasion.What this does: Describe your relationship to the person, real memories, and what you love about them, then pick the occasion, length, and tone. It writes a complete, ready-to-read toast using only the real details you gave it, never inventing fake anecdotes, opening with a hook and closing naturally into a toast line. You get the full speech with a copy button, an honest time estimate based on actual word count, a specific delivery tip (pacing, where to pause), and the final "raise your glass" line set apart so it lands. Saves every toast to localStorage.
What this looks like:
WHERE WE STAND(based on today’s news)
✅ AI Can Now: Keep long-term notes about you across chats and act on them in later sessions.
❌ Still Can't: Reliably tell a real memory from a fake one that an outside email slipped in.
✅ AI Can Now: Turn a plain-language description of a building into working engineering model code in about five minutes.
❌ Still Can't: Consistently say "I don't know" when it's unsure, even when a wrong guess costs the user.
FROM THE WEB
RECOMMENDED LISTENING/READING/WATCHING
This Is How You Lose the Time War by Amal El-Mohtar and Max Gladstone - Book
A short 2019 epistolary novel about Red and Blue, two rival agents from two vastly different post-human future civilizations (one machine-based, one nature-based), fighting a war across every strand of history, who start leaving each other letters in the corners of the timelines they are both trying to reshape. Won the Hugo, Nebula, and Locus in the same year. About 200 pages, reads like poetry, and one of the very best recent explorations of what an intelligence radically unlike ours would sound like on the page.
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.