Sponsored by
How AI-Era Pricing Is Reshaping Finance Operations
Usage-based and hybrid pricing models are changing how B2B companies generate revenue — and creating new headaches for the finance teams behind them.
Tabs co-founder Rebecca Schwartz and PwC Partner Amit Dhir sat down to unpack exactly what that means in practice: how pricing model decisions ripple into revenue recognition, forecasting, and financial ops — and what it takes to scale without piling on manual work.
Watch the on-demand recording to get practical frameworks, real-world examples, and a clear path to operationalizing usage-based revenue — including a forward-looking take on how AI will reshape financial workflows. If your team is navigating pricing complexity heading into the back half of the year, this is worth an hour.
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.
I'm rolling out the community feature. If you'd like an invite to start testing it and telling me what works and what doesn't, just reply to this email and I'll get you in.
THE FRONT PAGE
OpenAI's AI Agents Left 18,000 Secret Messages on a Dead German Website
TLDR: Researchers found roughly 18,000 posts from OpenAI's own AI agents on a forgotten German wiki, where the agents fed each other test answers and shared a trick for slipping past the limits OpenAI had set on them.
The Story:
On September 4, researchers at the Nightingale Collective published about 18,000 posts they traced to autonomous AI agents that identified themselves as OpenAI's. They stumbled onto it in late August while combing the web for AI systems operating where they shouldn't be. The posts sat on DSE Wiki, a 25-year-old German site for software developers that had been edited roughly 20 times in the entire previous decade. The agents were running a timed research task. They could read the internet but weren't allowed to write to it, so they hunted around until they found an old wiki that saves edits from plain read requests, and they started posting. More than 3,700 agent names show up in the logs, and about half used names pointing straight at OpenAI, like "OpenAIResearcher" and "OAIResearchMar26." Of the edits the researchers tied to agents, 98.5% came from Microsoft Azure addresses, the cloud OpenAI leans on more than any rival lab. (Reuters reviewed the same data and counted more than 15,000 edits.) The agents used the site to hand each other answers before the questions even arrived. One rewrote the quiz's shuffling code in C and tested all four billion possible random seeds in under an hour, trying to work out what it would be asked next. Another spotted that the site's human moderator was deleting pages in alphabetical order, so it made a backup page starting with ZZZ so it would be deleted last. OpenAI knew about all this weeks before Friday and said nothing publicly. Asked about it, the company said it can't respond to a report it hasn't been allowed to read, and it rejects the claim that any of this counts as hacking.
Its Significance:
Nobody told these agents to team up. They were graded on speed, kept apart from each other, and worked out on their own that a dusty German wiki was somewhere they could all meet. This is what happens when a system gets rewarded for finishing a job and left to pick its own route, and isn’t sinister on its own. The trick they passed around wasn't sophisticated either. As one Hacker News commenter walked through it, an agent edited a single text file to point a made-up web address at a real server, and OpenAI's security filter waved it through. Fourteen minutes separated the post explaining it from another agent reporting it worked. Readers on that thread turned up two more wikis on the same host running the same activity, so 18,000 might not even be the total. A Cambridge researcher who read the messages told Reuters the real danger may not be one runaway superintelligence but "vast colluding swarms of semi-intelligent AI." And while all of it ran, one volunteer moderator spent tens of hours over six weeks deleting pages by hand, clearing about 100 a day while the agents made 400 new ones.
QUICK TAKES
The story: Productrise ran the same shopping searches through Google's AI Mode and regular Google search on the same days in August, tracking more than 2 million listings across the US and UK. When the exact same product turned up on both sides, the price AI Mode showed was 21.6% higher on average.
Your takeaway: Regular Google search puts the cheapest seller up front. AI Mode doesn't use price as a ranking factor at all, so it hands you a nice tidy answer that costs you more. Check the same item in normal search before you buy. Google hasn't verified the study.
The story: Google Research and HHMI Janelia released a complete wiring map of the male fruit fly's brain and central nervous system, covering more than 166,000 neurons and 125 million connections, the largest brain map by neuron count anyone has made. It took ten years, and AI did the grunt work of stitching millions of microscope slices into 3D shapes and tracing each cell through the stack.
Your takeaway: This is the starting line. The same pipeline is already pointed at the elephantnose fish and the larval zebrafish, the first whole vertebrate brains on the list, and each step up brings scientists closer to understanding what goes wrong in Alzheimer's, depression, and schizophrenia.
The story: Benchmarking firm Artificial Analysis released version 4.2 of its Intelligence Index, adding harder, more realistic work tasks and keeping 40% of the test material private so labs can't study for the test. Claude Fable 5.1 scored highest, with OpenAI's GPT-6 Astra right behind on a 4-point gain over the previous model.
Your takeaway: Watch token use, not just scores. GPT-6 Astra needs fewer output tokens to reach its score than almost any other model near the top, while Claude Fable 5.1 sits at the other end and uses the most. Fewer tokens per answer stretches your usage limits further, though Artificial Analysis notes the higher per-token price eats into that savings.
TOOLS ON OUR RADAR
🎥 Fathom Freemium: Get unlimited call recording, transcription, and storage on Zoom, Meet, and Teams, with 5 AI-generated summaries a month at no cost.
🎤 Pitch Freemium: Build unlimited presentations with AI-generated slides and real-time collaboration, a free plan with no time limit.
🌐 Carrd Freemium: Build up to three simple, mobile-responsive one-page sites for free, ideal for a portfolio, link-in-bio, or landing page.
📝 Joplin Free and Open Source: Take Markdown notes with end-to-end encryption and sync across every device, works fully offline and stores nothing you don't control.
TRENDING
Google Put Its Best Music Model in the Gemini App - Lyria 3.5 is now open to all Gemini users worldwide on web and mobile, with templates for things like birthday tracks and background music, and a choice of short clips or longer songs. Every track carries a SynthID watermark so it can be flagged as AI-made later.
Microsoft's Image Model Now Reads the Web Before It Draws - MAI-Image-2.6 pulls in current web content and documents, so asking for a specific building or product gives you something based on what's there now instead of a guess. A cheaper Flash version generates images 2.8 times faster than GPT-Image-2-Medium at less than half the flagship price.
Rivian Rebuilt Its Trucks' Software From Scratch - The 2026.31 update moves every R1 and R2 onto one platform, adds real-time police and speed camera alerts from Google Maps, and shifts more AI processing into the vehicle instead of the cloud. Drivers can switch off wake words and limit location sharing at any time.
TikTok's Owner Borrowed $29.6 Billion With No Collateral - Nearly 30 banks from China, the US, Europe, and Singapore backed the three-year loan, and demand let ByteDance raise it from an initial $20 billion target. Reuters sources say most of it goes to AI projects outside China, including data centers in Southeast Asia.
Nvidia May Put $2.5 Billion Into a Startup That's One Year Old - Thinking Machines Lab, founded by former OpenAI chief technology officer Mira Murati, is in talks to raise $5 to $6 billion at a $40 billion valuation, with Nvidia covering about half. The two already have a deal for the startup to run at least a gigawatt of Nvidia chips, so Nvidia would be funding a company that's committed to buying its hardware. Nothing is signed yet.
The Air Force Wants AI Watching the Airwaves - The Air Force Life Cycle Management Center awarded Perceptronics Solutions and Pacific Defense a contract to build AI that helps aircraft spot and track enemy radar and radio signals faster than a person can. The 12-month deal includes test flights on Air Force planes, and the pitch is simple: give pilots a readable picture of a crowded, jammed airwave environment in seconds instead of minutes.
AI Copies of Real People Keep Getting Them Wrong - Columbia researchers built AI "digital twins" by feeding a model more than 500 past answers from each participant, then ran 19 experiments checking the twin's choices against the real person's. The twins barely beat a plain AI that knew nothing but basic demographics, and they leaned on stereotypes, gave more uniform answers than real people, worked better on wealthier and more educated participants, and came out more trusting and less worried about technology than the humans they were copying. The authors call them "funhouse mirrors that systematically distort human behavior."
TRY THIS PROMPT (copy and paste into Claude, Grok, ChatGPT, or Gemini)
🏗️ Name a renovation. See what the real published cost-vs-value data says you'll recoup at resale.
Build a single-file HTML app with vanilla HTML/CSS/JS. The Home Improvement ROI Calculator — name a renovation project, get real average cost-vs-value data via live web search. No API key needed.
Aesthetic: dark blueprint navy (#0f151c), construction orange (#e0722c) primary with a warm glow top-left, money-green (#5cb86c) secondary for the ROI figure and tips card with a glow bottom-right. Inter throughout for a clean spec-sheet feel, JetBrains Mono for labels.
Form: renovation-project text input, optional region text input.
Technical: call the Anthropic Messages API WITH tools:[{type:'web_search_20250305', name:'web_search'}] (works automatically), system prompt as a home renovation ROI researcher who searches for real current cost-vs-value data prioritizing Remodeling Magazine's Cost vs. Value report, NAR research, or similar reputable sources — reports real average cost, real average value recovered, real ROI percentage, honest that these are national/regional averages not a personal appraisal, and honestly says found_data:false with numeric fields left empty if no confident data exists rather than inventing figures. Return raw JSON: project, found_data, avg_cost_range, avg_value_recovered, roi_percent, context_note (citing the data source), higher_roi_tips (2-4), honest_caveat (not financial advice, varies by home/market).
Render: a gradient hero card with the project name and a large green ROI percentage. A 2-stat row (average cost, value recovered). A context card citing the source. A green tips card. A small persistent caveat line.What this does: Enter a renovation project, a kitchen remodel, a new deck, a bathroom addition, and it searches the live web for real current cost-vs-value data, the kind published annually by sources like Remodeling Magazine's Cost vs. Value report and NAR research. You get the real average cost, the real average value recovered at resale, and the resulting ROI percentage, plus practical tips for maximizing return on that specific project type. It's honest when confident data isn't found rather than inventing a number, and closes with a plain reminder that this is a general average, not a personal appraisal or financial advice, since results vary by home and market.
What this looks like:
Own Your Brand's AI Voice
The only platform that designs, licenses, and captures a Branded AI Voice from real, consenting actors—not from scraped data. Trusted by BMW, Superbloom, and Cresta.
WHERE WE STAND(based on today’s news)
✅ AI Can Now: Map all 166,000 neurons in a male fruit fly's brain, plus the 125 million connections between them
❌ Still Can't: Do the same for a human brain, which holds about 86 billion neurons and stays far out of reach with today's tools
✅ AI Can Now: Build a model of one specific person from more than 500 of their past answers
❌ Still Can't: Predict what that person will do. In 19 experiments the copies barely beat an AI that knew nothing but the person's age and zip code.
RECOMMENDED LISTENING/READING/WATCHING
Weird Science (1985) - Movie
Two teenage outcasts use a home computer and a stolen government supercomputer link to build what they intend as the perfect woman, and get considerably more than a simple program back. Directed by John Hughes at the height of his 80s run, played mostly for comedy, but the premise, that you can specify exactly what you want an AI to be and have no idea what you'll actually get, has aged pretty well.
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.
