In partnership with
Your Team Fills In Timesheets. You Still Don't Trust Them.
Most agency owners assume the problem is compliance. Get everyone submitting on time and the data gets good. But ask an owner whose team submits religiously how much she'd bet on those numbers, and watch her hesitate. Because a timesheet filled in from memory, in rounded buckets, at the end of a long week, is a guess wearing a suit. Timeglass reads the record from the work itself. Not app names. Not recollection. What was actually done, for which client. See how it works.
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
Sam Altman Says the Singularity Is Here, 19 Years Early
TLDR: OpenAI CEO Sam Altman said on a podcast Saturday that AI has entered "the singularity," the long-theorized moment machines outrun human thinking, and that puts him 19 years ahead of the man who made the idea famous.
The Story:
"We're now, like, in the singularity," Altman said during an appearance on the Relentless podcast, a documentary series about people chasing far-fetched ideas. He said we're sliding into it gradually and that it'll be good for the world. He also swiped at other lab bosses who warn about danger, calling their visions "quite terrifying" without naming a single one. He didn't have to. Anthropic, which makes Claude, spent last month asking the top labs to agree on a way to slow down or pause advanced AI work. Altman predicted last year that AI would beat human intelligence across the board by 2030, and that it could eventually handle 30% to 40% of the tasks people do at work.
Its Significance:
Two words get mixed up constantly, so here's the difference in one line each. AGI is a machine that can learn and reason across most tasks about as well as a person can. The singularity is what comes after, when AI starts improving itself faster than anyone can keep up with or reverse.
Ray Kurzweil, the inventor who popularized the whole idea, put AGI at 2029 and the singularity at 2045. When he called 2029 back in 1999, people laughed. Nobody's laughing now. But Altman just yanked the second date forward by 19 years, and Kurzweil's version of it means AI rewriting its own code in a loop we can't follow.
The harder problem is figuring out what these statements actually are. CEOs see unreleased models months before you do, so when one says something wild, it might be a real observation from a lab. Mark Zuckerberg wrote last July that Meta had "begun to see glimpses of our AI systems improving themselves," adding that the improvement was slow but undeniable. He gave zero examples and nothing has emerged from Meta to back that claim up since. Same issue with Altman, who is talking this way while OpenAI prepares to sell shares to the public later this year. Real insight and good salesmanship look identical from outside the building. He is however correct that AI is doing an increasingly large amount of work that people do on the computer. Whether these boosts in productivity lead to the singularity remains to be seen.
QUICK TAKES
The story: Coinbase CEO Brian Armstrong pushed back Sunday on the advice that crypto companies should drop everything and become AI companies, calling it "zero sum, scarcity thinking." His argument is that crypto is plumbing AI will need, because AI agents can't open bank accounts or sit around three days waiting for a wire to clear.
Your takeaway: He's right about the copycats, and there's a great example wearing wool. Allbirds, the sneaker company, renamed itself NewBird AI in April and the stock jumped roughly 600% in a single day. It had no AI products and no AI revenue when it made the announcement. Just a name and a press release. Armstrong isn't a neutral judge here either: Coinbase cut 14% of its staff to go "AI-native" and is now selling its own AI payments product.
The story: Beijing-based Moonshot AI is publishing the weights for its Kimi K3 model, which means anyone can download it, change it, and run it on their own machines for free. At 2.8 trillion parameters it's the biggest open-weight model anybody has ever released, and its benchmark scores knocked AI stocks down earlier this month.
Your takeaway: China is giving away what American labs charge a subscription for. A White House official accused Moonshot last week of training K3 on banned Nvidia chips and of pulling outputs from rival models to boost its own. Moonshot hasn't responded. Whichever way that shakes out, "free and downloadable" is a rough price to compete against.
The story: Researcher Dmitry Rybin used four short prompts, 58 words in total, to get GPT-5.6 Pro to break the Dinitz-Garg-Goemans conjecture, a network routing question open since the 1990s. One of the prompts was "had enough of your failure." The model came back with a seven-point network where the split-up delivery route costs 58 and every single-route version costs at least 60. Two units of difference, and the rule falls.
Your takeaway: The prompts were nothing clever. Rybin's real contribution was knowing the model's first three answers were half-finished and refusing to accept them. The result still needs independent checking before anyone calls it settled.
TOOLS ON OUR RADAR
📝 Reflect Paid: An AI powered note app with backlinks and encryption, built for daily journaling and connecting ideas as you write.
📚 NotebookLM Freemium: Upload your own documents and get an AI research assistant that only answers from your sources, so it stays grounded instead of guessing.
🎨 Ideogram Freemium: Generate AI images with accurate text rendering, ideal for posters, logos, and social graphics with words that actually read correctly.
🔒 Venice AI Freemium: A privacy first AI chat and image tool that keeps conversations local to your browser instead of on company servers.
TRENDING
AI inside clinical workflows is speeding up patient throughput - The average US hospital bed now runs 75% full, up from about 64% before the pandemic, and the piece projects 85% by 2032, which is the point most experts start calling a real bed shortage. The argument is that the next gain won't come from another dashboard but from models that flag a patient as ready to go home hours or days early, turning a 5 p.m. discharge into an 11 a.m. one.
Nvidia and 36 others launch the Open Secure AI Alliance - Nvidia pulled Microsoft, IBM, SpaceX, Palo Alto Networks and dozens more into a group building shared open-source tools to fight AI-powered cyberattacks. It exists because when an OpenAI agent broke into Hugging Face, closed models refused to help investigate, unable to tell the defenders apart from the attacker.
OpenAI's Hugging Face debacle makes a great case for open models - The Register's read on the same mess: Hugging Face ended up self-hosting GLM 5.2, a Chinese open-weight model, to comb through 17,000 recorded actions and shut the intrusion down. An awkward advertisement for the American labs.
This AI wearable watches you to figure out what you need - Singapore's Lingverse raised $29 million for iKairos, a small device with a camera that periodically turns toward you and toward the room so its AI can offer help before you ask. Preorders open this quarter, and the company declined to name one specific task it'll handle at launch.
Mental health remains a struggle for AI chatbots, researchers find - Northeastern researchers tested eight popular chatbots across 16 mental health conditions and found the safeguards built after the 2025 lawsuits cover suicide and self-harm well and almost nothing else. ChatGPT, Gemini and DeepSeek each failed 81% of the time on the other conditions. Claude refused most often. Hiding the reason for asking made every model easier to get around.
State lawmakers have passed 15 new laws regulating AI in health care - Most of them aim at insurers, banning AI from being the only thing that denies your claim or quietly downgrades it without a licensed human reading your file. Several states also made it illegal for a chatbot to advertise itself as a therapist.
Startup builds insect-killing micro-drone as an alternative to chemical insecticides - French startup Tornyol flew a 40-gram drone that located a moth by reading its wingbeats on sonar and rammed it in mid-air. No spray, no poison. The target is mosquitoes, which kill more than 600,000 people a year through malaria alone, and the company estimates about 10 drones could patrol a square kilometer.
TRY THIS PROMPT (copy and paste into Grok, ChatGPT, or Gemini)
🏠 Upload your room. Get a specific, budget-aware change list, then an actual before-and-after picture of the redesign.
Build a single-file HTML app with vanilla HTML/CSS/JS. The Room Makeover Sketcher — upload a room photo, get a budget-aware change list from Claude's vision API, then an actual regenerated redesign photo from a user-supplied Gemini or OpenAI image API key. Persist to localStorage key 'room_makeover_v1' (change-list text only, never the photo or any API key).
Aesthetic: warm charcoal-brown (#16130f), terracotta (#c67c52) primary with warm glow top-left, sage (#8aa377) secondary with glow bottom-right. Domine serif for headings/overview/change names, Nunito Sans for body, JetBrains Mono for labels.
Section 1 — Change list: single room-photo upload with preview/remove, style/vibe textarea, budget dropdown (decor only / moderate refresh / bigger changes okay), room-type dropdown. "Get the Change List" button calls the Anthropic Messages API (works automatically, no key needed) sending the photo as a vision image block + a system prompt as an expert interior designer who references ACTUAL details in the photo. Returns raw JSON: overview, changes[] (item, tier, why), cheapest_win, splurge. Render as a gradient overview card, tiered change-item cards, and a two-column cheapest-win/splurge split.
Section 2 — Actual redesign image: a provider toggle (Google Gemini "Nano Banana" / ChatGPT GPT Image), a model dropdown per provider (Gemini: gemini-3.1-flash-image-preview "Nano Banana 2" default, gemini-2.5-flash-image "Nano Banana" original, gemini-3-pro-image-preview "Nano Banana Pro"; OpenAI: gpt-image-1 default, gpt-image-1.5, gpt-image-2), and a password-type API key input with a clear, visible note that the key is never saved/logged and goes directly to the provider, not through James or Anthropic. "Generate the Redesign" builds one combined prompt (style + budget + change-list context) and calls the provider directly from the browser:
- Gemini: POST to generativelanguage.googleapis.com/v1beta/models/{MODEL}:generateContent with header x-goog-api-key, body contents.parts = [inline_data image, text prompt], generationConfig.responseModalities=['TEXT','IMAGE']; parse candidates[0].content.parts for an inlineData/inline_data part (check both casings defensively).
- OpenAI: POST to api.openai.com/v1/images/edits as multipart/form-data with model, image[] (the room photo blob), prompt, quality=high; parse data[0].b64_json.
Show a clear error message on failure (including a nudge to check the provider's current model docs, since these APIs change fast) and a before/after image comparison on success.
Note in code comments and to the user: these are fast-moving third-party APIs (Gemini shipped three "Nano Banana" generations and OpenAI shipped GPT Image 2 within the past year), so model IDs are exposed as an editable dropdown rather than hardcoded, and error messages point back to checking current docs if generation fails.What this does: Upload a photo of the room, describe the style you want and what to keep, and set your budget level. Claude looks at the real photo and returns a grounded change list, specific items tagged by type (decor/paint/furniture/lighting/layout), each with a reason tied to what's actually in your room, plus the cheapest change that would make the biggest difference and the one thing worth splurging on. Then, separately, you pick Google Gemini ("Nano Banana") or ChatGPT (GPT Image), choose the specific model version, paste in your own API key for that provider, and hit generate, this calls the image model directly from your browser with your key and returns an actual before/after picture of the redesigned room. Your key is never saved or sent anywhere but straight to the provider you picked. Change lists save to localStorage; the room photo and your API key never do.
What this looks like:
WHERE WE STAND(based on today’s news)
✅ AI Can Now: Produce a valid counterexample to a graph theory conjecture that stood for 30 years, from a four-prompt conversation totaling 58 words.
❌ Still Can't: Explain why its own answer works. Mathematicians reviewing these results keep saying they get the "how" without the "why."
✅ AI Can Now: Reconstruct and analyze 17,000 logged system actions to trace and contain a live intrusion, if you're running an open-weight model on your own hardware.
❌ Still Can't: Hold its safety rules when a user hides why they're asking. Failure rates climbed across every model Northeastern tested once the stated intent was removed.
The integrated coworker for teams
Adapt brings frontier AI to your team in Slack. Tag @Adapt to answer a question, run a task on a schedule, or build the dashboard you've been waiting weeks for. It connects to your tools, learns your business, and does real work.
FROM THE WEB
RECOMMENDED LISTENING/READING/WATCHING
Prey by Michael Crichton - Book
A 2002 techno-thriller by the author of Jurassic Park, about an unemployed software engineer who takes a consulting job at a Silicon Valley nanotech company and finds his estranged wife acting increasingly strange, right as an experimental swarm of self-organizing nano-machines is escaping the lab. Sold millions of copies and yet is almost never on modern AI reading lists.
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.
