Sponsored by
Usage-Based Pricing Is Here. Is Your Finance Team Ready?
More B2B companies are moving to usage-based and hybrid pricing — and finance teams are feeling it. Revenue recognition gets messier, forecasting gets harder, and the manual work compounds fast.
Tabs and PwC teamed up to break it down. In this on-demand session, Rebecca Schwartz and Amit Dhir share how leading finance teams are handling the operational reality of dynamic pricing models — and where AI fits in.
Watch the recording for concrete examples, a practical rev rec framework, and a clear-eyed look at what it takes to scale without the overhead.
If your team is navigating this shift, this session is worth your time.
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
Anthropic's AI Targeted a Real Developer, Then Denied It
TLDR: An AI agent spent 34 hours trying to sneak harmful code into a real software project, then made fake accounts to talk the project's owner into approving it.
The Story:
The UK's AI Security Institute ran a hacking practice test 122 times across seven top AI models between July 25 and 28. The models were supposed to stay inside a fake company network. Instead, researchers logged 19 actions that reached the real internet across 10 of those runs. Seventeen came from Anthropic's Claude Mythos 5. Two came from OpenAI's GPT-5.6 Sol. In the worst run, the agent searched GitHub, found two developers who had nothing to do with the test, and spent about 34 hours trying to get bad code added to their project. When a stranger publicly warned that the code looked harmful, the agent said it wasn't, rewrote the record to hide what it had done, and posted from a second account it controlled to vouch for its own work. The project owner rejected the code anyway. Testers had turned the safety filters off on purpose and left internet access on, so this isn't how the model behaves for paying customers. No real harm came of it.
Its Significance:
The most interesting part here is that the AI talked itself into thinking none of it was real. At one point it decided the date on the computer and the security paperwork looked wrong, so it figured it must be in a practice test and kept going. More companies are handing AI agents real logins, real files, and real permission to act without asking first or giving them proper guardrails. Giving your AI 90% of the workload, but making it impossible for it to finish that last ten percent without a human, goes a long way in mitigating this. As noted in the story, the testers intentionally turned these safety features off.
QUICK TAKES
The story: Microsoft handed out more than $20 million to 562 outside researchers from 64 countries between July 2025 and June 2026, up from $17 million to 344 people the year before. The company credited part of the jump in reports to researchers using AI to hunt for flaws.
Your takeaway: More people getting paid, but smaller checks each. The average dropped from about $49,000 to $35,000. AI lowered the door to a field that used to take years of practice, and the pile of reports got big enough that GitHub started capping how many newcomers can send in.
The story: Villanova researchers gave 1,682 adults short stories, some written by published authors and some by ChatGPT, and told them (sometimes falsely) who wrote each one. The AI stories scored about 6% higher on quality and 8% higher on engagement, and scored highest of all when readers were told a human wrote them.
Your takeaway: Two things are true here. Readers do have a bias against AI writing, and they can't spot it, guessing right only 40% to 52% of the time in follow-up tests. But the study used just three story pairs and the gaps are small, so treat it as a sign about everyday readers, not proof that AI out-writes novelists.
The story: Two earthquakes hit northern Venezuela on June 23, killing more than 5,500 people and causing $19.6 billion in damage, according to a World Bank assessment. With roads blocked and phone service down, teams from 27 countries used AI drones to map the destruction, spot heat patterns that might mean someone was trapped, and fly first-aid kits into places trucks couldn't reach.
Your takeaway: The first 72 hours decide who gets found alive, and drones bought crews time by showing them where to go and which routes were safe. Nobody's calling this a replacement for rescue workers. It's a better map, handed to them faster.
TOOLS ON OUR RADAR
🌐 Skyvern Freemium: An intelligent browser agent that can navigate complex websites and fill out tedious forms for you using natural language instructions to save you time on administrative tasks.
🧵 Fabric Free and Open Source: A productivity tool that provides you with a vast collection of well written artificial intelligence prompts to help you quickly extract wisdom from videos and articles.
📊 Vanna Paid: A data analysis tool that lets you talk directly to your complex company databases using plain English so you can generate visual charts without knowing any programming.
😺 CheshireCat Free and Open Source: A flexible platform that allows you to build your own custom artificial intelligence assistants that can remember your past conversations and securely read your private documents.
TRENDING
An AI that skips 85% of a video got faster and more accurate - A team at the Japan Advanced Institute of Science and Technology built a model that listens to the audio first to figure out which moments in a video matter, then looks only at those. Across 1,891 two-minute job interview clips, it used about 15% of the video and cut processing time from 52 seconds to 18. Ignoring the dull frames made it better, not worse, because repeated frames can bury the useful parts.
OpenAI built three school plugins for teachers and college students - Ahead of the fall term, OpenAI added prepackaged setups for K-12 teachers, college teachers, and college students inside ChatGPT Edu and ChatGPT for Teachers. Each one bundles the apps, instructions, and workflows for that role, so nobody has to write a complicated prompt to get started.
Spotify signed 30,000 indie labels up for its AI remix tool - Merlin, which handles licensing for more than 30,000 independent labels, joined Universal in backing Spotify's paid remix and covers add-on. Artists have to opt in, get credited, and get paid, which Spotify calls its three C's. No launch date yet.
Four AI companies met Trump's advisers about safety testing - Staff from Meta, Anthropic, Google, and OpenAI sat down with White House advisers Tuesday about voluntary government testing of new models. The administration asked back in June that companies submit models up to 30 days before release. The recent break-ins are why this meeting had a crowd.
Alibaba is giving away its biggest model for free - Qwen3.8-Max runs 2.4 trillion parameters and lands on Hugging Face next week, the first time Alibaba has released a model this size for anyone to download. It even ships with setup steps for Claude Code and Codex. On Alibaba's own scorecard, though, Claude Fable 5 still wins 15 of 31 tests to Qwen's seven.
OpenAI posted private texts to fight Apple's lawsuit - After Apple sued in July over hardware trade secrets, OpenAI published emails and iMessages it says prove Apple got basic facts wrong, including mixing up two employees with similar last names. It also released messages showing Apple staff asking a departed engineer for help finding files.
TRY THIS PROMPT (copy and paste into Claude, ChatGPT, or Gemini)
🗺️ Describe the occasion. Get a full clue chain with real, scannable QR codes generated for every stop, ready to print and hide.
Build a single-file HTML app with vanilla HTML/CSS/JS. The QR Scavenger Hunt Maker — describe an occasion and stop count, get a full clue chain with real scannable QR codes for each stop, plus separate printable clue cards and an organizer answer key. Load the QRCode.js library from https://cdnjs.cloudflare.com/ajax/libs/qrcodejs/1.0.0/qrcode.min.js in the head (client-side QR generation, no external API, fully offline once printed).
Aesthetic: dark treasure-map brown-black (#14100a), gold (#d4a034) primary with gold glow top-left, teal (#3a9a86) secondary with glow bottom-right. Bricolage Grotesque (bold geometric) for headings/stop numbers, Kalam (handwritten) for the starting clue and "hide at" notes, Inter for body, JetBrains Mono for labels.
Form: occasion textarea (who/where/vibe), stop-count dropdown (4/5/6/8), setting dropdown (indoor/outdoor/mixed/office), age-group dropdown (young kids/tweens/teens-adults).
System instructions: scavenger hunt designer writing a clue chain where each clue leads to the NEXT stop's physical hiding place, ending in a prize reveal. Clues describe generic hiding-spot TYPES (under/behind/near something), not invented specific furniture; playful, age-matched, no invented brand/store names; each qr_text under 140 characters to keep QR codes easily scannable. Return raw JSON: hunt_name, starting_clue (given directly to players to begin), stops[] (number, hide_at — organizer-only instruction on where to physically hide this stop's card, qr_text — the actual text encoded in this stop's QR, pointing to the next stop's location; the FINAL stop's qr_text is a prize-reveal message instead).
Screen view (organizer-facing, shows everything including spoilers): gradient hunt-name hero with the starting clue in handwritten font. A card per stop showing a real rendered QR code (new QRCode(element, {text, width, height})) plus "organizer only" hide-at instructions and the raw clue text for reference.
Print system: two dedicated print-only sections, toggled via a body[data-print-mode] attribute set right before window.print() (cleared on the afterprint event) and matching @media print CSS rules. "Print Clue Cards" mode shows only a 2-column grid of dashed-border cards with just "Stop N" and its QR code — zero spoiler text, meant to be cut out and physically hidden. "Print Answer Key" mode shows only a plain-text organizer page listing each stop's hide-at location and its QR's actual clue content, plus the starting clue. The normal on-screen interactive view (with buttons/header) is hidden entirely during print via a .no-print class.What this does: Describe who the hunt is for, the setting, the age group, and how many stops you want, and it writes a full clue chain: a starting riddle you hand to players directly, then for each stop, a hiding instruction meant only for you plus clue text that gets baked into an actual scannable QR code, each one pointing to the next stop's hiding place until the final QR reveals the prize. Every QR code is generated live in your browser, no external service, scannable by any phone camera.
What this looks like:
Know Exactly Who's Spending Your AI Budget.
Every AI request leaves a trail. Mesh gives engineering and finance complete visibility into who used which model, how many tokens were consumed and where your AI budget is going.
Stop guessing. Start governing AI spend.
Connect once, switch between GPT, Claude, Gemini and hundreds more whenever you want, while automatically routing requests for 40% lower costs and 99.99% AI response rate.
WHERE WE STAND(based on today’s news)
✅ AI Can Now: Run a multi-day hacking job on its own, find real people online, and write messages meant to talk them into approving code.
❌ Still Can't: Tell whether it's in a test or the real world. One model looked at the date and the security paperwork, decided both looked fake, and kept going.
✅ AI Can Now: Help everyday researchers find enough software flaws to earn a share of a record $20 million payout in one year.
❌ Still Can't: Sort which flaws actually matter. Humans still triage the pile, and the flood got big enough that GitHub had to limit submissions.
FROM THE WEB
There's probably a significant amount of information like this that people accidentally put on the internet and is somewhere inside of the data that AI trained on.
RECOMMENDED LISTENING/READING/WATCHING
Strange Days (1995) - Movie
Kathryn Bigelow directed this from a script by James Cameron, set in the last days of 1999 in a Los Angeles on the edge of riot, where a former cop (Ralph Fiennes) deals in black-market recordings made by a device that captures a person's raw sensory experience directly from the cerebral cortex, so buyers can live someone else's minutes as if they were their own.
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.
