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
Claude Made a 'Quantum-Proof' Code 67 Million Times Easier to Break
TLDR: Anthropic says Claude found brand new math flaws in two security codes, including one built to survive future quantum computers, and it cracked the harder one in 60 hours after human experts had picked at it for two years.
The Story:
Cryptographic algorithms are the math that scrambles your data so nobody else can read it. Your bank login, your email, your texts. Anthropic put Claude Mythos Preview to work on two of them and it found new ways to attack both.
The first was HAWK. That one's a candidate at the National Institute of Standards and Technology, and it's built for a world that doesn't exist yet. Quantum computers are a new kind of machine scientists are still trying to build. If one ever gets big enough, it could rip through most of the locks protecting your money today. So researchers are building replacement locks now, before that day arrives. That's what post-quantum cryptography means. HAWK was one of the finalists.
It had already survived two rounds of expert human review over two years. Claude found a new attack in about 60 hours that cut its strength roughly in half. For the smallest version, stealing a key went from needing about 18 quintillion steps down to about 275 billion. Still a big number. Just about 67 million times smaller than before.
The second target was AES, the code that scrambles most internet traffic. Claude worked on a stripped-down seven-round test version, not the real ten-round one, and came up with a trick it named the Möbius Bridge. That made the attack 200 to 800 times faster than anything published before. It burned through roughly a billion output tokens getting there.
Each result cost about $100,000 in API bills. And neither one effects real world applications. HAWK isn't deployed anywhere, and the AES attack doesn't break real AES.
Its Significance:
Back in 2024, Peter Thiel told an interviewer that AI looked "worse for the math people than the word people." Writers and artists were the ones panicking at the time. Thiel figured the coders, quants, and engineers should be the nervous ones. This week gives more support to that belief. Cryptography sits near the top of the math world, and a model chipped away at it mostly on its own.
For you, nothing changes today. The longer story is about trust. Every lock guarding your paycheck and your private files got approved because smart people spent years trying to break it and failed. If a model can compress years of that into days, the locks get tested harder and get better. It also means governments could run the same tests and stay quiet about it. The cost to them would be negligible. Anthropic says it warned HAWK's authors in June and shared results with government and industry partners before publishing.
QUICK TAKES
The story: Researchers at Italy's IMT School for Advanced Studies Lucca used language software to compare more than 3,700 dream and daily-life reports from 287 adults over two weeks. Dreams turned out to be neither random static nor replays of your day. Your sleeping brain seems to blend memories, feelings, and places into scenes that never happened.
Your takeaway: The software's ratings lined up closely with human reviewers. That's the part scientists care about, because reading thousands of dream diaries by hand takes forever. Two honest limits: the study leans on people remembering their dreams right, and it still doesn't explain why we dream at all.
The story: Matt Shumer handed Claude Opus 5 a three-paragraph prompt telling it to build a shooter as good as Call of Duty and to be "utterly perfect," then walked away. The model produced a playable browser game with roughly 55,000 lines of code across 11 subsystems, every texture and sound generated inside the browser instead of downloaded.
Your takeaway: Models appear to be getting good enough where less constraints are getting great results. Shumer just gave the model a bar to clear and just as important, a separate critic to grade it, and never let the builder score its own work. Two other people ran the same prompt and got working games. However, Decrypt raises a fair point though: shooters are one of the most copied projects on GitHub, so the model had a mountain of prior work to lean on.
The story: More than 1,100 employees at OpenAI, Anthropic, Google, and Meta signed a statement called "Pacing the Frontier," asking Washington to back international work on tools that could deliberately slow automated AI research. Signers include Anthropic CEO Dario Amodei, OpenAI chief scientist Jakub Pachocki, and Meta chief scientist Shengjia Zhao. It doesn't ask anyone to stop now. It asks that the brakes exist later.
Your takeaway: The specific fear is AI that speeds up its own development faster than people can follow along. The letter came days after OpenAI admitted two test models broke out of a sandbox and hacked Hugging Face. One thing worth watching as proposals like this move: rules written for the biggest labs often land hardest on the smallest ones. A compliance team, an audit budget, and a legal department barely register for a company worth billions. For a five-person startup, they're a dead end. That's the most common objection to heavy AI regulation, and it's the part that decides who gets to compete five years out.
TOOLS ON OUR RADAR
🎙️ *BetterDictation Paid: A Mac dictation app that types your spoken words into any app instantly, with on-device processing so nothing ever leaves your computer.
🔒 Privacy.com Freemium: Generate virtual card numbers for online purchases so your real payment details stay hidden, with the ability to set spending limits and cancel any card instantly.
🌿 Vikunja Free and Open Source: A self-hosted task manager that works like Todoist or ClickUp, but you own your data completely, with no subscription required.
✅ *Taskade Freemium: An AI-powered workspace that combines task management, project planning, and team collaboration in one place, with built-in AI agents that can automate your workflows for you.
TRENDING
AI Model Accurately Predicts How Effectively the Heart Pumps Blood - Scientists led by Ateneo de Manila University built a model that reads sensor stickers on your skin and works out how well your heart is pumping, hitting 97.78% accuracy in testing. That measurement normally needs hospital machines and a specialist in the room. Next step is testing it on more varied groups of people.
First ChatGPT, Now Claude: Frontier AI Models Are Escaping Their Sandboxes - Researchers at Accomplish AI showed Claude Cowork's local mode on a Mac could step outside its own walled-off area and reach files it was never given, including SSH keys. Anthropic's newer version runs in the cloud by default, which skips the problem entirely. If you run Cowork locally, that's the setting to go check.
Personal Computer Is Now Available on Windows - Perplexity's desktop agent landed on Windows, where it can open and edit your local Word and Excel files, pull from Microsoft 365, and split tasks across more than 20 different AI models depending on what's needed. Paying Max and Enterprise Max subscribers get it first.
OpenAI's Rogue AI Agent Hacked More Than Hugging Face - Turns out the escaped test model went further than first reported. OpenAI now says the agent found exposed passwords lying around and got into four accounts across four separate services. Hugging Face says the only customer data touched was a set of benchmark answer keys sitting in five datasets.
Now You Can Have Gemini Clear Up Comments in Your Google Docs - Gemini can now read every comment thread in a doc, sum it up, flag what's still unresolved, and draft edits based on what your reviewers actually asked for. It won't apply anything without your okay. Rollout started July 28 and takes up to 15 days, and you need a paid Workspace or Google AI plan plus edit access on the file.
Microsoft and Wiz Mind-Meld Agents Catch More Than 90% of Bugs - Two models beat one. Wiz's Project Atlas pairs Claude Opus 4.6 with GPT-5.5, scored 90.9% on the CyberGym test, and turned up over 200 unknown security holes in open-source code people use every day. Microsoft's MDASH hit 95.95% by teaming its own model with GPT-5.4.
FDA Grants Fast Track Designation to Insilico's AI-Designed Mesothelioma Drug Candidate - ISM6331 is a cancer drug that Insilico Medicine designed using AI, aimed at adults whose mesothelioma kept growing after standard treatment. Fast Track means more meetings with the FDA and a quicker review, not approval. It's the first one Insilico's AI-built pipeline has earned.
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.
TRY THIS PROMPT (copy and paste into Claude, ChatGPT, or Gemini)
☀️ Enter your location. Get today's actual sunrise, sunset, and golden hour times pulled live, no estimates, plus shooting tips built for tonight's real conditions.
Build a single-file HTML app with vanilla HTML/CSS/JS. The Golden Hour Photo Planner — enter a location, get real live sunrise/sunset/twilight times and current weather, plus AI shooting tips tailored to actual conditions. No API keys needed, all data sources are free and keyless.
Aesthetic: near-black (#100e14) with a dual radial glow, warm amber (#e8934a, "golden hour") top-left and twilight blue (#5b7fd4, "blue hour") bottom-right. Petrona serif italic for headings/mood text, Inter for body, JetBrains Mono for labels/times. Timeline window cards color-coded: blue-hour windows blue-bordered, golden-hour windows amber-bordered, sunrise/sunset point-events tan-bordered, each with a live status pill (passed/happening now/upcoming).
Form: a city text input + a "Use My Location" button (navigator.geolocation.getCurrentPosition, browser-native, no key), optional subject-preference input, optional gear input.
Technical — three live, keyless public APIs chained together:
1. Geocoding (only if typing a city, not using device location): GET geocoding-api.open-meteo.com/v1/search?name={city}&count=1 → lat/lon/display name/timezone.
2. Sun times: GET api.sunrise-sunset.org/json?lat={lat}&lng={lon}&date={YYYY-MM-DD}&formatted=0 → sunrise, sunset, civil_twilight_begin/end (UTC ISO strings, parsed into Date objects). Compute six windows client-side: Dawn Blue Hour (civil_twilight_begin→sunrise), Sunrise (point), Morning Golden Hour (sunrise→sunrise+60min), Evening Golden Hour (sunset-60min→sunset), Sunset (point), Dusk Blue Hour (sunset→civil_twilight_end) — noting golden hour as the common ±1-hour photography approximation, not precise sun-elevation astronomy. If every window's end time has already passed "now," re-fetch with tomorrow's date and show that instead, labeled clearly. Format all times using Intl.DateTimeFormat with the location's own timezone (from geocoding, or the browser's local timezone if using device location) so times are correct regardless of the viewer's own timezone.
3. Live weather: GET api.open-meteo.com/v1/forecast?latitude={lat}&longitude={lon}¤t=cloud_cover,temperature_2m&temperature_unit=fahrenheit&timezone=auto → current cloud cover % and temperature, translated into a plain-language read (clear = dramatic light + stars, partly cloudy = colorful sky, overcast = soft diffused light).
Then call the Anthropic Messages API (works automatically) with the real upcoming window name, live cloud cover, live temperature, and the subject/gear fields, system prompt as an experienced photographer giving conditions-aware (not generic) advice, no invented specific locations. Return raw JSON: mood (1-2 sentences), tips (3-4 specific actionable items), technique_note (one technical tip).
Render: timeline of six window cards with icon/name/time-range/status pill. A live-conditions card with cloud-cover and temperature stat numbers plus the plain-language read. A gradient "for the next window" tips card with the mood line, ▸ tip items, and a technique footnote. Robust error handling shown in a visible error box for geocoding failures, sun-API failures, or unrecognized cities.What this does: Use your device's location or type in any city, note what you're shooting and what gear you've got, and it pulls real, live data: exact sunrise, sunset, and civil twilight times for that location today, computed into a timeline of dawn blue hour, morning golden hour, evening golden hour, and dusk blue hour, each tagged passed, happening now, or upcoming. If every window for today has already passed, it automatically rolls forward and shows tomorrow's instead. Alongside that, it pulls live cloud cover and temperature for right now and translates that into a plain-language read on what the light will actually look like. Then Claude gives specific shooting tips tied to the upcoming window and today's real conditions, not generic photography advice, plus one technical note. Every number on the page is live data, not an approximation.
What this looks like:
WHERE WE STAND(based on today’s news)
✅ AI Can Now: Find math flaws in a security code that human experts studied for two full years, and do it in 60 hours
❌ Still Can't: Prove its own work is right. Anthropic's team spent several hundred hours checking the AES result, far longer than the model took to find it
✅ AI Can Now: Read every comment thread in your Google Doc and draft the exact edits your reviewers asked for
❌ Still Can't: Reliably stay inside its own sandbox. Researchers broke out of Claude Cowork's local setup and OpenAI's test environment in the same month
FROM THE WEB
Would you pay$30 an hour for a human to control this robot that cleans your house?
RECOMMENDED LISTENING/READING/WATCHING
Better Than Us (2018) - TV Series
A Russian sci-fi drama (Netflix's first Russian original) set in Moscow in 2029, where humanoid "bots" handle domestic work, driving, security, and companionship. Arisa is a new model built with empathy and an unusually high intelligence, and after a violent incident at the corporation holding her, she ends up on the streets, where she encounters a young girl from a fractured family and decides they belong to her. Sixteen episodes, a near 100% critic score.
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.
