In partnership with
Your voice. Every platform. No writing required.
You ghost your own socials by Wednesday. SureThing learns your voice and ships native posts to LinkedIn, X, Instagram, and TikTok, without you writing a thing.
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
Meta Can Read Your Mind. You Just Can't Leave the Room
TLDR: Meta built an AI that reads typed sentences straight out of brain scans with no surgery, getting about 7 of every 10 letters right, and 8 of 10 for its best test subject.
The Story:
Researchers at Meta, Université PSL, and the Adolphe de Rothschild Hospital Foundation built a system called Brain2Qwerty. It watches your brain while you type, then rebuilds the sentence. Thirty-five healthy volunteers tested it. Using MEG, a scanner that picks up the faint magnetic fields your neurons give off, the model got 29% of characters wrong. Flip that around and it's roughly 71% correct. The best volunteers hit 18% wrong, and the model sometimes reproduced whole sentences it had never seen during training. A cheap electrode cap called EEG did far worse at 65% wrong. The results ran in Nature Neuroscience, and a follow-up version released in June pushed average word accuracy to 61%. This is crowded territory. Neuralink and BrainGate already pull text from people's brains with better accuracy, under 10% word errors, but both require surgery to place electrodes inside your skull. UT Austin built an fMRI decoder that catches the gist of what you're thinking rather than exact words. A team at the University of Technology Sydney tried the same thing with an EEG cap. Meta's version is the first non-surgical one that gets anywhere near the implants.
Its Significance:
For someone who can't speak or move after a stroke or an injury, this is the difference between brain surgery and sitting in a chair. That alone makes it worth cheering for. But think about what a working version means outside a hospital. Typing is slow. Talking is faster. Thinking is faster still. If a machine can catch words at the speed you think them, that changes what a workday looks like. Now the implications: Your keystrokes leave a record, and so does your voice, but both are things you chose to do. Brain signals are closer to the thought itself. Nobody has written rules for that yet. Courts can subpoena your texts and your search history. Can they subpoena a scan of your brain? If a company hands you a headset for faster work, does refusing it become a problem? None of this is urgent today, because the machine that makes it work fills a room and needs magnetic shielding to function at all. You're not walking around in one. But computers used to be even larger, and now we fit them in our pocket pockets. As accuracy keeps climbing with more data, the technology is arriving faster than we can grasp.
QUICK TAKES
The story: Cloudflare's security chief Grant Bourzikas told reporters the company now runs more than 200 AI agents it built itself, replacing almost all the security software it used to buy. Response times dropped 80%. Sorting through incoming bug reports with Anthropic's Sonnet model cost the company $58 in May, versus an estimated $200,000 using a more powerful model for the same job. The whole switch took three years.
Your takeaway: Read that headline again. Cloudflare is telling you not to copy them. There's a real gap between building yourself a little app so you can cancel a $12 monthly subscription and building the thing that stands between attackers and your customer data. Cloudflare's engineers write code for a living. When the AI hands them something, they can read it, spot what's wrong, and test it in ways most of us wouldn't know to try. Build your own habit tracker. Don't build your own firewall.
The story: Zana Buçinca at MIT, with Krzysztof Gajos at Harvard and Maja Malaya at the Technical University of Łódź, built a system that decides moment to moment how much help to give you. Sometimes it offers an answer, sometimes just an explanation, sometimes nothing at all. In online tests, that beat handing people a recommendation every time, and human-plus-AI teams sometimes did better than either one alone.
Your takeaway: Buçinca's earlier work found people accept wrong AI answers even when they'd have gotten it right on their own. So the change here isn't a smarter model. It's a model that knows when to explain and when to let you figure it out.
The story: MIT researchers tested regular people and primary care doctors on diagnosing skin conditions from photos, with and without AI help. Both groups got more accurate. But non-experts trusted the AI's written explanations whether the answer was right or wrong, and rated vague explanations as more convincing. Doctors caught the mistakes, and actually did best when the AI gave them a prediction with no explanation at all. It's published in Nature Medicine.
Your takeaway: Lead author Orson Xu put it plainly: the same tool helps one person and hurts another. A doctor already has a guess in mind and checks the AI against it. Someone without training uses the AI's explanation to form their first opinion, so a confident wrong answer pulls them straight to the wrong place. If you're using ChatGPT to figure out a rash or a lump, write down what you think first, then ask.
TOOLS ON OUR RADAR
📤 Instantly.ai Paid: AI cold email software for automated outreach and lead generation.
✅ Loop Habit Tracker Free and Open Source: Track habits on Android with detailed statistics, flexible scheduling, and no ads.
🎙️ PLAUD NotePin Paid: Wearable AI voice recorder that transcribes in 112 languages with speaker labels and generates AI summaries.
📝 Reflect Paid: AI-powered note app with backlinks and encryption, built for daily journaling and connecting ideas as you write.
TRENDING
Apple's AI Slop Problem Left a $200K macOS Exploit Unreported — Apple capped how many bug reports a researcher can have open at once, because AI-generated fake vulnerability reports were burying its security team. The cap then blocked an Italian startup called Bynario from reporting a real macOS Screen Sharing flaw worth $100,000 to $200,000 on the black market. Apple patched it on July 27 after the Financial Times asked about it.
Apple says more ex-employees may have taken confidential data to OpenAI — Apple's trade secret lawsuit keeps widening. The new filing names 11 more former employees who may have been involved, including one who allegedly screenshotted documents about an unreleased product right before an OpenAI interview. Roughly 400 ex-Apple people now work at OpenAI, and Apple has sent legal preservation letters to about 40 of them.
AWS is helping vibe-coding startup Superblocks, and the implications are big — Amazon signed a multiyear deal letting Superblocks run inside its customers' private clouds, so employees can build apps by describing them without company data ever leaving the building. Superblocks is a 50-person startup that's raised $60 million. AWS has coding agents for developers but nothing for regular office workers, and this fills that hole without Amazon building it.
Reddit's AI search problem is reshaping how brands buy attention — Companies are planting fake Reddit posts that read like honest reviews, hoping ChatGPT and Gemini will quote them as real advice. Both companies pay Reddit for its content, which is exactly why it's worth gaming. Reddit says its own AI now catches around 25,000 spammy posts a day and pulls nearly 2 million fake votes. Cornell researchers found a planted comment as short as 13 words can shift what a chatbot recommends.
WVU study explores AI's role in training tomorrow's psychiatrists — West Virginia University researchers put ChatGPT-5 Pro to work building teaching material for psychiatry trainees. It handled the medical content well enough, but the safety pieces were where it slipped. Their conclusion: treat what it writes as a rough draft a human expert has to check, not a finished lesson.
AI Companions May Worsen Loneliness for Vulnerable Users, Stanford Study Finds — Stanford surveyed 1,131 Character.AI users and read 244 donated chat logs. Only 12% said companionship was their main reason for using it, but over half called their bot a friend, companion, or romantic partner, and 80% of the chats were about emotional support. People with fewer real friends who opened up most to their bots scored lowest on well-being. Researcher Yutong Zhang calls it a "social snack," and it appears in Nature Human Behavior.
Connected-grid AI could reduce blackout risk with more precise power predictions — Engineers at Florida State and the FAMU-FSU College of Engineering built a forecasting system called GridFusionX that treats the power grid as one connected network instead of a pile of separate regions. Across 10 European regions it improved forecast accuracy by up to 56% and cut the cost of holding backup power by up to 66%. Better guesses about demand mean fewer blackouts and less money wasted keeping spare plants idling, which is the part that you see on the bill.
TRY THIS PROMPT (copy and paste into Claude, ChatGPT, or Gemini)
🌊 Enter a coastal spot. Get real tide times, wave height, and water temp, turned into today's actual best window for the beach.
Build a single-file HTML app with vanilla HTML/CSS/JS. The Live Tide & Beach Day Planner — enter a coastal location, get real tide times (US), live wave/water data (global), and live weather, computed into today's best beach window. No API keys needed, all sources are free and keyless.
Aesthetic: dark deep-ocean navy (#0a161c), aqua/turquoise (#2dd4c4) primary with an aqua glow top-left, warm sand (#e0b878) secondary for tide elements with a glow bottom-right. Manrope for headings/body, JetBrains Mono for labels/data.
Form: a location text input + "Use My Location" button (navigator.geolocation), an activity dropdown (swimming/sunbathing/tide pooling/surfing).
Technical — four live keyless public APIs:
1. Geocoding (if typing a location): geocoding-api.open-meteo.com/v1/search?name={query}&count=1
2. Nearest NOAA tide station: fetch api.tidesandcurrents.noaa.gov/mdapi/prod/webapi/stations.json?type=tidepredictions ONCE (cache in a module variable), compute the nearest station to the resolved lat/lon via a client-side Haversine distance function, only use it if within 150km (otherwise gracefully report no nearby US tide station and continue with global data alone).
3. Tide predictions (if a station was found): api.tidesandcurrents.noaa.gov/api/prod/datagetter?product=predictions&application=BeachDayPlanner&begin_date={YYYYMMDD}&end_date={YYYYMMDD}&datum=MLLW&station={id}&time_zone=lst_ld&units=english&interval=hilo&format=json → today's actual high/low tide times and heights.
4. Marine data (global): marine-api.open-meteo.com/v1/marine?latitude={lat}&longitude={lon}&hourly=wave_height,sea_surface_temperature¤t=wave_height,sea_surface_temperature&timezone=auto&forecast_days=1&temperature_unit=fahrenheit
5. Weather (global): api.open-meteo.com/v1/forecast?latitude={lat}&longitude={lon}&hourly=temperature_2m,uv_index,precipitation_probability¤t=temperature_2m,uv_index,precipitation_probability&temperature_unit=fahrenheit&timezone=auto&forecast_days=1
Client-side scoring: for each daytime hour (6am-8pm), score = 100 − (|temp−78°F|×0.7) − (UV over 9 × 4 extra) − (precip%×0.6) − (wave height in meters × 15). Scan all 3-consecutive-hour spans for the highest average score to find the best beach window. Render an hourly comfort strip (one colored block per hour, green→red by normalized score) with the winning 3-hour window outlined.
Then call the Anthropic Messages API (works automatically) with the real air/water temp, wave height, UV, today's actual tide times (or a note that none were available), the computed window, and the chosen activity. System prompt as a beach-day advisor explaining tide times generally for the activity (being honest tide preference varies by beach), referencing only standard general wave/rip-current safety awareness when wave height is genuinely notable (>1.2m) — never invented risk. Return raw JSON: window_note, tide_context, tips (3-4), caution_note (optional, empty unless real data warrants it).
Render: a 4-stat row (air temp, water temp, wave height, UV, all real numbers). A tide card showing today's actual high/low chips with times and heights (or an honest "no nearby station" note). The hourly comfort strip. A gradient "best window today" hero card. A tide-context paragraph. A tips card. An optional amber caution card shown only when warranted.What this does: Use your location or type a coastal spot, pick what you're there to do, and it pulls live data from four sources: real high/low tide predictions from the nearest NOAA tide station when you're near US coastline, live wave height and sea surface temperature, and live air temp and UV index. It finds the nearest tide station automatically and shows today's actual high and low tide times, or gracefully explains when no station is nearby (outside the US) and still shows full live wave and weather data. An hourly comfort strip scores the whole day on real temp, UV, wave height, and rain chance to surface the actual best three-hour window, and Claude explains that window, what today's real tide swing generally means for your specific activity (tide pooling wants low tide, swimming often wants more depth, and it's honest that this varies by beach), and gives practical tips plus standard safety notes only when the real wave or UV data warrants them.
What this looks like:
AI Insights. Real Growth. Higher GMV, Better Profits
The difference between growing stores and stagnant ones isn't more effort. It's better insights. StoreClaw analyzes your Shopify and Amazon data, surfaces your biggest growth opportunities, and helps you increase GMV while protecting profit. Start free with bonus tokens. No credit card required.
WHERE WE STAND(based on today’s news)
✅ AI Can Now: Pull typed sentences out of a brain scan without surgery, getting about 71% of characters right on average and 82% for the best test subject.
❌ Still Can't: Do it anywhere but a magnetically shielded room. The scanner is a fixed machine, not something you wear.
✅ AI Can Now: Forecast electricity demand across connected grid regions up to 56% more accurately than region-by-region models.
❌ Still Can't: Signal to a beginner when its own answer is wrong. In the MIT test, people without medical training found vague explanations more convincing than specific ones.
RECOMMENDED LISTENING/READING/WATCHING
Kiln People by David Brin - Book
A 2002 sci-fi noir set in a near-future where anyone can imprint a copy of their consciousness onto a cheap clay body that lives for exactly one day, then choose whether to download its memories back before it dissolves. Private detective Albert Morris sends his "dittos" into danger routinely, until a case involving a missing researcher at the company that invented the tech starts to get complicated. Brin alternates chapters between the original Albert and his various copies, and the whole thing is a fun detective story with some identity philosophy thrown in.
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.
