Welcome, humans.

So apparently Accenture, the company that sells other companies on going all in on AI, just told many of its own employees they get $29 a month of Claude. 404 Media saw the internal announcement, and the new cap took effect on October 1.

Here's how it works: The company now pays $29 per person centrally, and the announcement says "all costs are consumption-based," meaning the bill grows with how much you use.

Want more? You charge the extra to a project or team, using what Accenture calls a Work Breakdown Structure (basically a cost code that says which budget eats the charge). Even then, a cost code plus a license does not guarantee ongoing access, because licenses get reviewed and can be reallocated.

The gap is the fun part. One Claude Code (Anthropic's coding agent) user building an AI accelerator for client work said they typically spend $500 to $1,000 a month, roughly 17 to 34 times the allowance. Another employee reported losing access entirely. An Accenture spokesperson said the company aims to provide the right tool for the right work while it keeps investing in AI.

Accenture's pitch to clients: invest heavily in AI. Accenture's pitch to staff: here's twenty-nine bucks.

Here's what happened in AI today:

• 😺 ChatGPT ads can target users by email and prompts

• 📰 Meta released an open model that runs on two Mac Studios

• 📰 Anthropic gave six METR evaluators employee-like access

• 📰 Google released Deep Review for checking technical documents

• 🎓 Powerdrill's four-bucket method sorts mismatched transaction records

… and a whole lot more that you can read about here.

😺 404 Media Ran Ads Inside ChatGPT to Show How Advertisers Can Target You

Want to make a point about ChatGPT privacy? Buy an ad inside ChatGPT. That's what 404 Media did, running two ads that read "ChatGPT isn't private. OpenAI hires humans to read users' chats. Learn more," aimed at people asking about privacy and AI. Trolling the chatbot from inside the chatbot. Respect.

The stunt matters less than the tool behind it. Reporter Joseph Cox's tests show advertisers can aim at you two ways: by what you're asking about right now, and by who you are.

Here's what 404 Media found:

• Context hints: advertisers describe topics or keywords, and OpenAI says these "guide matching but aren't exact-match targeting rules." Think Google keyword ads, but looser.

• Uploaded lists: advertisers upload emails, phone numbers, or Android advertising IDs, and OpenAI matches them to real ChatGPT accounts. Lists over 25,000 entries can target people. Smaller ones can only exclude them.

Uploads can be raw (readable emails) or hashed (scrambled into codes). OpenAI accepts either and didn't say whether it hashes raw uploads. Meta does. OpenAI says files are deleted after processing, usually within 24 hours, and advertisers can't see which individuals matched.

Then there's the off switch. You can turn off "ads personalization," but 404 Media says that doesn't stop ads or tracking entirely, and the setting says ads "may still be based on your current chat." So "off" means "less," not "none." Great.

Keep the scale in view, though. These are 404 Media's own tests, and customer-list targeting is standard across advertising. Zach Edwards of DecryptAds calls it "boiling the frog," and the article's estimate is that most of ChatGPT's roughly 1 billion monthly users don't know any of this.

Two things to watch: whether OpenAI answers the hashing question, and whether "off" ever covers your current chat. Until then, treat a ChatGPT conversation like a search box that can show ads.

FROM OUR PARTNERS

Inside the builder renaissance. Is your startup playing to win?

AI is rewriting the startup playbook. Ship in weeks what used to take years. If a small team can build almost anything, the real question is what's worth building, and who gets there first.

Alex Kantrowitz, founder of Big Technology, and Iccha Sethi, Vanta's SVP of Engineering, share patterns before they're conventional wisdom.

You'll learn:

• How teams turn AI demos into production tools

• How to build an edge before capital or headcount catch up

• Which fundamentals still hold as everything else moves

Register to save your spot.

🎓 AI Skill of the Day: Reconcile Messy Exports with the Four-Bucket Method

The Powerdrill team's guide to reconciling transactions (checking that two records of the same activity agree) opens with a scene every finance person knows. A card payment shows up once in your ledger, your own record of transactions, and twice in the processor's export, split into a charge and a fee. One invoice reads INV-0042 in one system and INV42 in the other. A transfer started on the 31st settles on the 1st and bumps the whole month.

Their fix: stop hunting for matches and start sorting leftovers. Every unmatched row falls into one of four buckets: missing on one side, duplicated, wrong amount, or right payment in the wrong period. Spreadsheet lookups stall because they assume one row on each side equals one event. They break when a single settlement covers forty transactions (a many-to-one match).

Powerdrill built this around its own tool, so the chatbot version below is our adaptation:

1. Upload both exports and ask the chatbot to flag mismatched column names, date formats, and reference styles before matching anything.

2. Ask for the four buckets by name, with a dollar tolerance and a few days of date slack. Then ask which groups of rows in one file add up to a single row in the other.

3. Check row counts and unmatched dollar values, not just the grand total, since two equal errors in opposite directions still produce a perfect total.

Keep every row and add a status column instead of deleting matches, so you can explain a call six weeks later. Spot-check a few proposed groupings by hand.

Start with a month you already closed and see what the many-to-one question catches that your old lookups missed.

FROM OUR PARTNERS

How much of your company can a model learn from your AI traces?

Your team explains how your company works to AI agents every day. Origin gave a model 200 of its own sessions to read. From the typed prompts alone, it recovered seven real private rules, the first within five sessions. Read what leaked, and why the traces are worth keeping.

Read the research

📰 Around the Horn

• Meta released Muse Spark Open, an open-weight coding and agent model with 816B total parameters that Meta says runs on two Apple M5 Ultra Mac Studios, under a license with limited commercial terms.

• Anthropic gave six METR evaluators access close to its internal risk teams, and METR can publish its findings without Anthropic's approval, though Anthropic can require narrow redactions.

• Google released Deep Review publicly, a team of AI reviewers that checks papers, legal documents, and financial material, with two free reviews per week.

• The Simons Foundation committed $15M in combined funding and resources with Anthropic and OpenAI to help mathematicians adapt to AI, including workshops and shared standards for checking AI-assisted results.

• Amazon expanded Bee with Bee Frontier, a $74.99 wearable, and Bee Eclipse, a magnetic phone attachment from $99.99, both turning conversations into summaries and tasks.

• Perplexity open-sourced SPACE, the sandbox behind its Computer agent, reporting that agents did not escape the virtual machine in 108 red-team attempts.

Want absolutely EVERYTHING that happened in AI this week? Click here!

🍪 Treats to Try

*Asterisk = from our partners (only the first one!). Advertise to 700K+ readers here!

1. *What Is Data Management? A Guide to Systems, Processes, and Tools. Learn more

2. Navige sits between your agents and the models they call, rerouting easy jobs to cheaper ones and stopping runaway loops, and its Watchtower mode shows projected savings before it changes anything.

3. Z.ai's GLM-Legal is a legal model for contract review, evidence analysis, and compliance research that you can test against general chatbots (application-only beta, and a lawyer should still check the output).

4. OpenMuse is an open-source personal agent you host yourself, with a browser, terminal, and files to work in, plus an action review step.

5. withHuman is an open-source approval layer that auto-approves routine agent actions and sends risky ones to a person for sign-off, across Claude Code, Codex, and LangGraph.

6. Chatway MCP connects ChatGPT, Claude, Gemini, or Cursor to your Chatway support inbox so an assistant can search contacts, pull from your knowledge base, and manage conversations.

7. Endo runs Codex, Claude Code, and Antigravity inside one thread, so you can switch between them mid-conversation and schedule repeat workflows (open-source, MIT license).

8. BlastLoop is an MIT-licensed framework that gives a coding agent work at system, subsystem, or single-file scope, then runs structural, operational, and adversarial review stages on the result.

UPCOMING EVENT: The Neuron IRL in San Francisco

The Neuron is going IRL on November 18 at 4:30 PM PT! Join us in San Francisco for drinks, bites, special guests, and a live recording of The Neuron podcast. Huge thanks to Slack and Alumni Ventures for making it happen.

It's free, but space is limited. Save your spot!

🧠 Intelligent Insights

• Harvard researchers Fiona Chen and James Stratton found AI coding agents lifted lines of code written by 30% across 718 firms, but reviews took 49% longer, making review the bottleneck.

• McKinsey's Lieven Van der Veken and team put the annual global AI opportunity at $2.52T, with the strongest economics in augmenting high-skilled, higher-wage work.

• Stanford researchers led by Diyi Yang analyzed a quarter-million anonymized Claude conversations and found 56% involved work rated consequential or higher, not just routine tasks like email drafts.

• Astrophysicist Brice Ménard of Johns Hopkins and Anthropic used Claude Science to predict the third of the UV sky no telescope has imaged, landing within about 10% on hidden test patches.

• Brian Earp, Julian Savulescu, and coauthors argued that making academics submit every AI prompt for ordinary writing help would backfire, since cheaters can dodge it and careful users bear the burden.

New from The Neuron: AI Explained

New episodes air every week on Wednesdays: Spotify | Apple Podcasts | YouTube

A Cat's Commentary

Btw: We just launched a robotics newsletter! Sign up for it here.

P.S. Love the newsletter, but only want to get it once per week? Don’t unsubscribe—update your preferences here.