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
Elon Musk Promises an AI Feature Film of the Odyssey This Year
TLDR: Elon Musk says Grok Imagine will make a full-length, "historically accurate" version of The Odyssey before the end of 2026, less than a year after the tool couldn't make video at all.
The Story:
On July 22, Musk posted on X that Grok Imagine would produce a feature-length Odyssey before the year runs out. He attached a three-minute clip an X user had made showing Odysseus and Calypso. The post cleared 8 million views inside a day. Grok Imagine started making video in late July 2025, and back then it capped out at six seconds of 480p. By February it reached 10 seconds at 720p with much better sound. March added a way to chain clips together into longer runs. By May a new version took first place on the Image-to-Video Arena leaderboard, and xAI now says its fast model spits out a six-second 720p clip in about 25 seconds, down from 40 or more.
Its Significance:
Nothing in AI has improved faster than video. A year ago the output looked like a weird fever dream. Now it holds a scene together. That speed of development is the real impressive thing here. If a tool climbs from six seconds to something watchable in twelve months, small teams get choices they never had before. An old book nobody owns the rights to. An original script. A family story. Two people and a laptop can now take a swing at things that used to need a crew and a budget. The limits are still easy to spot, though. Characters move stiffly, faces go slack between lines, and mouths often don't match the words. Three minutes can hide that well with decent editing. Ninety minutes can't. Musk misses his own deadlines constantly, but in keeping a blistering pace, he also achieves more in short spans of time than many of his competitors over much longer timeframes.
QUICK TAKES
The story: Black Forest Labs, the German lab behind the Flux image generators, released Flux 3 on Thursday. It makes clips up to 20 seconds long with dialogue and sound built at the same moment as the picture, and the same core system also runs robot hands.
Your takeaway: The robot half is the part to watch. Audi is testing it on a real production line, fitting bendy rubber door seals, a job normal factory machines keep fumbling. The company says the full loop reacts in about 101 milliseconds, roughly as fast as your eyes tell your hands to move.
The story: Google now lets you record a short video of your face, turning your head a few times so it catches you from several angles, and save it to your account. Get locked out later and you record a fresh one, which Google matches against the saved copy to let you back in.
Your takeaway: Google says the head turns prove a live person is there and that the system throws out deepfakes. Still fair to sit with the timing: you're handing over video of your face right as fake video gets good. It's opt-in, you can delete it any time, and it's blocked for Workspace accounts, kids' accounts, and anyone on Advanced Protection.
The story: Researchers at Korea's ETRI, working with POSTECH and Sungkyunkwan University, built a system called MemEIC that parks new facts in a separate memory bank outside the model instead of rewriting the model's own wiring. Across a 1,278-question test it answered correctly about 70% of the time, up from the 36% to 52% older methods managed.
Your takeaway: Memory has been one of AI's stubborn weak spots for years. Teach a model something new and it tends to smother what it already knew, a problem researchers call catastrophic forgetting. MemEIC gets around it by keeping picture knowledge and word knowledge in two separate drawers, then pulling from both only when a question needs both. Show it a photo of a dessert and tell it that dessert is popular in Korea, and it can now connect those two facts instead of inventing a chocolate truffle from Europe. Fix this properly and AI can finally keep pace with things that change constantly, like prices, laws, and product specs.
TOOLS ON OUR RADAR
🔀 OpenRouter.ai Freemium: Unified access to multiple LLM providers through a single API.
📧 Instantly.ai Paid: AI cold email software for automated outreach and lead generation.
🔲 Whimsical AI Freemium: Generate flowcharts, mind maps, and diagrams with AI assistance.
✍️ Easy-Peasy.AI Freemium: AI content assistant for writing blog posts, social media, and marketing copy.
The first way to trade directly inside Claude and ChatGPT
Superintelligence used to be locked inside billion-dollar quant firms whose algorithms quietly took advantage of everyone else.
Co-Invest puts it right in your chat window. Analyze markets, manage risk, and execute trades, all inside Claude and ChatGPT.
The institutions built the game, Co-Invest gives you a way to beat them.
TRENDING
Think Through Hard Problems in Voice Mode - Claude's voice mode now runs on the bigger Sonnet and Opus models instead of only the small fast one, so it can chew on harder questions and reach tools like Gmail and Slack while you're still talking. It takes turns, listening then pausing to think, and works in 11 languages.
ElevenLabs Adds References and Sound Control to Music v2 - You can now upload a short audio clip and have the model match its style, mood, and production sound instead of guessing from your typed description. A separate Vocals feature sings the track in your own voice or one from a library, and every upload gets checked for copyright before it runs.
Health in ChatGPT - Every US user 18 and older on any plan can now connect Apple Health plus records from hospital systems like Epic and Oracle Health, so ChatGPT can read new results against old ones. OpenAI says health data stays out of training and ads, and that the tool can still get things wrong.
AI Uncovers How Materials Behave at Large Scale From Microscopic Data - Researchers at the National University of Singapore built AI that works out how a whole material behaves by watching only a small slice of its atoms instead of tracking every one. They tested it on simulated alloys holding more than 500,000 atoms, and it can predict systems far larger than anything it trained on, which skips simulations that would otherwise tie up a supercomputer for weeks.
Playful AI Robots Mirror Toddlers' Language-Learning Mistakes and Recover - Robots given a small reward just for being curious learned language in about half the time as robots rewarded only for finishing the job. Partway through training they started knocking things over and trying moves nobody asked for, then made the same grammar slip-ups toddlers make before fixing them.
Quantum Randomness Solves a Troublesome Problem for Neural Networks - Researchers at the Joint Quantum Institute ran a small neural network across several kinds of quantum computers and found that the randomness baked into quantum measurements actually helped it read handwritten digits better. They also saw hints that the errors quantum machines are famous for might help as well, so long as there aren't too many of them.
Closed-Loop AI Framework Could Shorten Clean Energy Materials Development by Years - A setup published in Digital Discovery links AI models, automatic reading of research papers, and robotic lab experiments into one loop, feeding every round of results back in to pick the next test. The point is cutting years off the search for better batteries and solar materials.
TRY THIS PROMPT (copy and paste into Claude, ChatGPT, or Gemini)
🪴 Snap a photo of the sad leaves. Find out what's actually wrong before you panic and repot everything.
Build a single-file HTML app with vanilla HTML/CSS/JS. The Plant Parent Diagnostic — upload a photo of a houseplant (or describe it), get a likely diagnosis and care fix using the vision API. Persist to localStorage key 'plant_diagnostic_v1' (diagnosis text only, not the photo).
Aesthetic: dark forest (#0e1512), leaf green (#7cb668) primary with green glow top-left, terracotta (#c97b52) secondary with glow bottom-right. Cardo serif (italic for prompts/explanation) for headings, Karla sans for body, JetBrains Mono for labels. A dashed-border upload zone that becomes solid-border with an image preview (and a small ✕ remove button) once a photo is added.
Form: photo upload zone (tap to open file picker, accept="image/*", shows a live preview with a remove button), plant-type text input (optional, "or not sure"), symptoms textarea (encourage detail especially if no photo), watering-habits input, light/recent-changes input. Require either a photo OR symptoms text before submitting.
Technical: on file select, use FileReader to read the image as a data URL, extract the base64 payload and media type. When calling the API, if a photo was provided, send messages: [{role:'user', content: [{type:'image', source:{type:'base64', media_type: <file type>, data: <base64>}}, {type:'text', text: <symptoms/care text>}]}] — otherwise send the text content alone as a plain string.
System instructions: knowledgeable warm houseplant diagnostician. If a photo is given, look carefully at leaf color/spots/texture/wilting/soil as PRIMARY evidence; if not, work from described symptoms and note in confidence_note that a photo would sharpen the diagnosis. Consider overwatering, underwatering, light issues, pests, nutrient deficiency, low humidity, transplant shock, root rot, and normal lower-leaf aging (often not a problem — say so plainly if that's the case rather than inventing an issue). Confident where evidence is clear, honest where it isn't. No em dashes. Return raw JSON: diagnosis (plain statement), confidence_note (1 sentence), explanation (2-3 sentences referencing specific photo/symptom details), other_possibilities (1-2 short phrases), fix (3-4 concrete steps: step + detail), timeline (1-2 sentences), prevent (1-2 sentences), reassurance (1 short warm line).
Render: gradient diagnosis hero card (name + confidence note + explanation). "Also worth ruling out" list. Checklist-style fix-steps card. Two-column split: green "recovery timeline" and terracotta "prevent it next time." A centered italic reassurance quote. Copy diagnosis + archive keyed by plant type + diagnosis (photos are not persisted to localStorage to keep storage light).What this does: Upload a photo of the affected plant (optional, but sharpens the diagnosis a lot), note the plant type if you know it, describe the symptoms, and add your watering and light habits. It looks directly at the photo, leaf color, spots, texture, wilting, soil if visible, as primary evidence, combines that with what you described, and returns the most likely diagnosis with an honest confidence note, a clear explanation, and one or two other causes worth ruling out. You get concrete fix steps, a realistic recovery timeline, how to prevent it next time, and a warm one-line reassurance, since plant parenting has a learning curve. If nothing's actually wrong (normal lower-leaf aging, for instance), it says so instead of inventing a problem. Diagnoses save to localStorage.
What this looks like:
WHERE WE STAND(based on today’s news)
✅ AI Can Now: Build 20 seconds of video with dialogue, sound effects, and music generated in the same pass as the picture, not stitched on afterward.
❌ Still Can't: Hold a face together across a long scene. Lips drift out of sync with speech and expressions go flat, which is why the good clips are still counted in seconds.
✅ AI Can Now: Turn a video-prediction model into robot movement fast enough to react in about 101 milliseconds, close to human reflex speed.
❌ Still Can't: Tell a real face from a generated one using video alone. That's why Google's new sign-in makes you turn your head instead of just holding still.
FROM THE WEB
RECOMMENDED LISTENING/READING/WATCHING
Automata (2014) - Movie
A gritty European sci-fi film starring Antonio Banderas as an insurance investigator for a robot manufacturer, working in a near-future desert city where 99% of the world's population has died from solar storms and humans co-exist with cheap manual-labor humanoid robots, all governed by hard-coded "protocols" that prevent them from harming humans or self-modifying. When Banderas's character encounters a robot that appears to be violating those protocols, he starts asking questions the manufacturer would rather he didn't. Melanie Griffith and Dylan McDermott co-star, and the whole thing looks a hundred times more expensive than it actually was.
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.
