Here is what we have for you today:

• Nvidia is betting its new chip will turn your next laptop into a full-time assistant

The dessert (from our channels to yours):

• Ready to hang an AI assistant around your neck? Read about Microsoft’s plan to give workers an AI wearable badge for managing AI agents

• Inside Walmart’s U-turn: Why the retail giant is now rationing AI for its 2.1 million employees

• Google is now using AI to catch the AI scammers faking your mom’s voice on the phone

• China just got a complete map of its 400,000 wind and solar sites thanks to AI

• Oxford scientists are using AI to design personalised cancer vaccines — and the UK government just funded the next step

• AI computing power is being turned Into a tradable commodity, and the world’s biggest exchanges are rushing to be first

Reinventing the era of PC

In the mid-1970s, Bill Gates and Paul Allen envisioned a future where personal computer would be on every desk of every home.

n the mid-1970s, Bill Gates and Paul Allen envisioned a future where a personal computer would be on every desk in every home.

It was a massive swing, which is why many laughed at the idea, considering the hefty price for a computer at that time.

Well, they are the ones laughing now, as computers are literally managing our daily life and work.

Fast forward to a few days ago, and Nvidia CEO Jensen Huang stood on a stage in Taipei to set a completely new goal for the industry that involves the exact gadget we rely on daily.

He set a goal saying that the very laptops we stare at every day are about to become our full-time AI assistants. In short, soon we could completely alter the way we work by no longer relying on the mouse and the keyboard.

A shift as big as the smartphone era

At the Computex technology show, the ever-present Jensen Huang announced the next stage of the PC.

He didn’t mince words, calling this a reinvention that will be as big of a deal as the smartphone shift, and labeling it a completely new beginning for computers.

What he actually revealed to back up that claim is a new so called “superchip” the RTX Spark.

Behind the fancy name is chip designed specifically to power the next generation of Windows laptops and desktop computers.

For the last forty years, computers have functioned as tools: you launch an app, you click the mouse, and you type out the commands.

With the RTX Spark, Nvidia is shifting the computer into a teammate. You ask it to do something, and the machine figures out the steps and executes the work itself.

Talk is cheap, so to actually pull this off, the underlying technology has to be incredibly robust.

And Nvidia delivered.

The RTX Spark is loaded with up to 128GB of unified memory, an NVIDIA Blackwell GPU, and specialized AI processing cores.

If you didn’t spend your younger days assembling your own custom computers just to get decent graphics, here is what those technical specs mean in plain English.

Normally, running heavy artificial intelligence programs requires your computer to communicate with a massive server farm located somewhere else.

But with this new chip, hardware makers like Dell, HP, and Lenovo can pack enough raw horsepower into a laptop as thin as 14 millimeters to run those complex AI models locally.

The computer sitting on your desk will have the brainpower to do the heavy lifting completely on its own.

Escaping the browser tab

With that kind of firepower sitting directly on your desk, the entire way you use a computer is about to change.

Huang paints a picture of an autonomous agent that can scan your personal files, reason through different apps, and handle your busywork entirely on the device.

For professionals, this translates to outsourcing the tedious busywork.

Adobe, for example, is currently gutting and rebuilding Photoshop and Premiere to run natively on the RTX Spark.

They claim this will allow users to edit video and generate graphics twice as fast because the computer is handling the underlying grind.

But this isn’t just about making your life easier; there is a massive business play happening behind the scenes.

Industry analysts note that by baking all of this directly into Windows laptops, Nvidia is strategically locking developers and creators into their own specific software and hardware orbit.

Microsoft’s head, Satya Nadella, called it a breakthrough for delivering unmetered intelligence to every desk.

It is a calculated land grab to ensure that the next decade of computing relies entirely on their architecture.

But what about your private data?

That all sounds incredibly convenient, but it forces you to face a glaring problem. For an AI assistant to actually organise your life, it has to read it first.

Handing a piece of software untethered access to every email, photo, and file you own triggers immediate privacy alarms.

Microsoft already has a rocky history with this. A decade ago, they launched Cortana, which was eventually bogged down by reports that contractors were listening to consumer data before Microsoft pulled the plug on the project.

This time around, Nvidia and Microsoft are trying to get ahead of the backlash by selling the privacy benefits of keeping the hardware local.

Because the RTX Spark runs the AI directly on the laptop, your private information stays locked on your specific device.

They are backing this up with a system called NVIDIA OpenShell, which acts like a digital bouncer, letting you put a fence around your data and dictate exactly what the AI can and cannot access.

Major brands like Dell, HP, and Lenovo are already rolling out these laptops this fall.

The technology is sitting on the launchpad, but after years of corporate data scandals, are people actually ready to trust a machine with the keys to their entire digital life?

• Microsoft develops AI wearable to help employees work alongside digital agents

• Walmart’s AI workflows meet the realities of the balance sheet

• Android phones will soon be able to detect spoofed calls and impersonation scams

• China’s AI just mapped its entire renewable energy grid. Here’s why the rest of the world should pay attention

• Cancer vaccines using AI gets research funding in UK

• Just like gold and oil, we’ll soon be able to trade AI token futures

Tweet of the day

New U.S. Census Bureau data reveals that private capital expenditure on data center construction has officially surpassed total government spending on public transportation infrastructure.

• Total U.S. data center construction spending surged over 30% year-over-year, fundamentally altering the national construction landscape while the broader commercial real estate market softens.

• The growth is entirely driven by the AI arms race, with capital expenditures from the major hyperscalers (Amazon, Microsoft, Google, Meta) trending toward a combined $400 billion annually.

Read the full article.

One more thing…

That’s all from us today - catch you next week. But before you leave, tell us what did you think of today’s edition to help us improve the experience for you.

P.S. Did someone forward this to you? You can subscribe here

Here is what we have for you today:

The main course:

• How AI is forcing McKinsey and other consulting companies to rethink their massive price tags

The dessert (from our channels to yours):

• Will AI help us understand our pets? Chinese startup claims its AI-powered collar can translate pet thoughts into full sentences.

• Ferrari thinks AI can change how fans experience Formula 1

• Humans and machines can’t coexist in the workplace? 99% of CEOs expect AI-driven layoffs in the next two years.

• Inside Spotify’s plan to let users create AI-generated personal podcasts

• New Goldman Sachs research says token consumption by AI agents is expected to surge 24x to 120 quadrillion tokens per month by 2030

• Trading platform Robinhood will allow users’ AI Agents to trade stocks

AI is forcing McKinsey, BCG, Bain and others to rethink consulting fees

In the late 1800s, the global ice trade was a massive industry.

Companies hired armies of men to manually cut blocks of ice out of frozen lakes and ship them across the world.

Customers paid a steep premium to cover the physical labor required to pull it off.

Then, the mechanical refrigerator arrived.

Once businesses and homes could generate the exact same product using a machine, the market stopped paying for the manual labor.

You could say that something similar is playing out in the corporate world right now, as AI is slowly starting to wipe out manual work and forcing consulting giants like McKinsey to completely restructure how they make money.

The billion-dollar pyramid

When a giant corporation hits a wall like - how to enter the new market, or how to cut costs, they usually hire firms like McKinsey, BCG, or Bain to come in, study the problem, and tell the executives what to do.

Since we are talking about A-list companies, these consulting firms get “pretty sweet” paydays, regularly pulling in tens of millions of dollars for a single project.

They justify those prices through a very specific strategic pyramid structure.

At the top, a few senior partners with decades of experience sell the strategy. At the bottom, a massive army of fresh Ivy League graduates does the grunt work.

They spend weeks reading legal documents, building complex spreadsheets, and making PowerPoint slides.

The corporate client pays for the senior partner’s brain at the top, but they also foot the bill for the 15 junior analysts doing hundreds of hours of manual research to feed that brain.

Enter the AI - new junior analyst

Artificial intelligence is slowly dismantling that pyramid because the software can now do most of what that junior army did in hours instead of weeks.

McKinsey’s own internal AI assistant now runs more than 500,000 prompts a month inside the firm, with consultants reporting up to 30% time savings on knowledge work.

Think about it like hiring a moving company. If they used eight guys and a truck for a full day to move your house, you paid the full-day rate.

But if that same company buys a forklift that does the exact same work in three hours, you aren’t going to keep paying for eight guys and a full day.

You look at the bill and say, “Great, charge me for three hours.”

The giant corporate clients have figured out that AI is stripping away the manual labor.

That is why they are turning to McKinsey and the crew and saying, “Hey, we know the Ai is doing the grunt work behind the scenes, so, should we discuss some discount?”

Killing the billable hour

To survive this efficiency trap, the firms are quietly changing how they sell their services.

Speaking to reporters at a London media event, McKinsey’s UK managing partner Michael Birshan said, “We’re doing more performance-based arrangements with our clients.”

In plain English, that just means they are abandoning the billable hour and charging for the final result.

Think about hiring someone to put a new roof on your house. You agree on a flat price for a roof that doesn’t leak.

If it takes the crew three days instead of two, they don’t get to charge you extra for being slow.

You pay for a dry house, not the hours they spent on a ladder.

Consulting clients are doing the exact same thing right now. They arrive with the outcome they want—like boosting profit margins by 5%—and ask the firm to price its work against actually delivering that result.

The thing is, this isn’t a sudden shift for the consulting companies, because the entire industry has actually flirted with outcome-based pricing for years.

AI has finally given it a reason to scale. The numbers behind the shift do most of the talking:

→ “About one in four of McKinsey’s global fees are now tied to outcomes rather than hours”, McKinsey UK managing partner Michael Birshan told reporters in November.

→ Bain’s tech- and AI-enabled revenue has reached roughly 30% of the firm’s business, with leadership projecting that share to climb to 50% in the coming years, according to Bain’s announcement of its partnership with AI researcher Andrew Ng.

Okay, but what happens to the junior consultants who used to do all that manual labor?

Around 150 former consultants from McKinsey, Bain, and BCG have already been contracted to train AI models to perform entry-level consulting tasks.

The very people who used to do this grunt work are now being paid to teach a machine how to replace them entirely.

The Canary in the coal mine

This matters to everyone outside of consulting because these firms are just the most visible early example.

Almost every major white-collar profession is built on this exact same pyramid.

Lawyers, accountants, and advertising agencies all rely on armies of junior employees doing baseline research and document work that AI is increasingly eating.

The consulting firms are just the loudest canary in the coal mine.

If McKinsey, BCG, and Bain cannot defend the billable hour, neither can the rest of the white-collar economy.

• Will AI help us understand our pets? Chinese startup claims its AI-powered collar can translate pet thoughts into full sentences

• Ferrari thinks AI can change how fans experience Formula 1

• Humans and machines can’t coexist in the workplace? 99% of CEOs expect AI-driven layoffs in the next two years.

• Inside Spotify’s plan to let users create AI-generated personal podcasts

• New Goldman Sachs research says token consumption by AI agents is expected to surge 24x to 120 quadrillion tokens per month by 2030

• Robinhood Launches AI Agent trading and credit card tools with user controls

Tweets of the day

Erdős problems are notoriously immune to rote calculation and brute-force compute.

Solving them requires forming novel conjectures and executing creative logical leaps.

By autonomously cracking problems that sat untouched for over half a century, DeepMind is demonstrating that its agents have crossed the line from simply accelerating human mathematical research to independently executing it.

The California-based startup Figure AI employed three humanoid robots powered by its Helix-02 AI system, which autonomously sorted small packages around the clock during a livestreamed operation.

The demonstration attracted significant online attention as viewers tracked the robots’ uptime and performance beyond the company’s original eight-hour target.

The livestream at Figure AI’s Sunnyvale headquarters ended with humanoid robots still actively sorting packages, bringing the total to 249,560 processed.

One more thing…

That’s all from us today - catch you next week. But before you leave, tell us what did you think of today’s edition to help us improve the experience for you.

P.S. Did someone forward this to you? You can subscribe here

Here is what we have for you today:

The main course:

• Google Search is dead. Long live AI-powered Google Search.

The dessert (from our channels to yours):

• Inside OpenAI tools that will let you connect bank accounts

• LinkedIn decided to declare war on AI slop content

• New survey: Junior positions in U.S. are being battered by AI as hiring shifts to older workers

• Multinational biopharmaceutical company to deploy Anthropic’s Claude AI model to speed up drug discovery

• Inside the Swedish café managed by an AI Agent

• Inside the UK’s water company plans to use AI technology to spot leaks

Many of you probably still remember how we planned road trips before the GPS era.

From today’s perspective, it seems funny that we used to print out pages of directions, memorise street names, and inevitably end up pulling over at a gas station to ask a local where to turn.

Luckily, live navigation software eventually arrived to handle the logistics.

Even if it occasionally routes you down a questionable dirt road, the system works perfectly because it removes the mental friction entirely, monitoring traffic and recalculating detours while you just focus on driving.

We rarely think about how completely that piece of technology rewired our daily habits, but that exact level of friction-removal is happening again only on a different level.

This week, Google revealed an AI-powered overhaul of Search that will fundamentally change how we pull information from the internet, and it is going to alter the digital behaviours of both everyday users and the businesses trying to reach them.

Trading blue links for a personal researcher

At its annual I/O conference, Google made a very clear pivot.

For 25 years, the search engine acted like a librarian pointing you to the right aisle, but leaving you to pull the books and find the answer yourself.

Now, they want to do the reading for you. The new search box lets you drop in entire files, videos, or open Chrome tabs.

In short, that means if you want to know why your car is making a weird noise, you don’t have to click through four different mechanic forums to figure it out.

You can just upload an audio clip or a video, and Google will build the answer right there on the screen.

You can even ask follow-up questions directly without starting a new search.

By handing you the final result immediately, the company is changing the core habit of the internet, shifting users away from clicking outbound links and keeping them planted firmly on the Google results page.

The 24/7 Web worker you don’t have to pay

Removing the blue links is only half the plan.

Google also wants to run your digital errands while you aren’t even looking at a screen. Starting this summer, they are rolling out “information agents” to handle ongoing tasks in the background.

Let’s say you are hunting for a two-bedroom apartment in a specific neighbourhood under a strict budget.

The current routine involves checking five different real estate apps every morning and manually filtering through the results.

Soon, you can just hand those exact parameters to the search box and close your browser.

The agent stays behind to monitor the web around the clock, and the second a property matches your criteria, it sends you an alert.

If that concept sounds familiar, it is basically a massive upgrade to the old Google Alerts feature from 2003.

The difference is that the old system just emailed you a link whenever a random keyword popped up online.

This new version actually understands what you want, figures out exactly where to pull the data from, and hands you a packaged result that is ready for a decision.

Building the app before you ask for it

Taking those packaged results and actually presenting them to you is where the interface of the internet starts to look totally different.

Behind the scenes, Google is using a development platform called Antigravity to power something called “generative UI”.

That means instead of just handing you a text summary, the search engine can actually build custom widgets and personalised mini-apps on the fly.

If you decide to get in shape, you don’t need to go download a third-party calorie counter or read through fitness blogs. You just tell the search box your goals, and it builds a personalised tracking dashboard right on the results page.

Because these generated spaces are persistent, you can come back to that exact same custom setup every day to log your progress.

Google is no longer acting like a directory that points you to useful tools built by other companies; it is generating the software internally so you never have a reason to leave the page.

Who pays the toll on the new highway?

Now that the search engine is designed to work directly for you, we have to look at what this means for the businesses relying on it to stay visible.

A huge chunk of businesses rely on Google sending them customers based on their specific demands.

When the platform starts generating its own interactive answers and building its own mini-apps on the fly, those outbound clicks might take a serious hit.

Naturally, this has triggered a wave of industry panic, with experts widely declaring that traditional SEO is completely dead.

But that is a major overreaction because the mechanics of visibility are just shifting.

Think about what happened to traditional travel agents when online booking sites launched.

They didn’t all go bankrupt, they just had to adapt their business model.

Instead of optimizing a website purely to catch human eyeballs, businesses now have to figure out how to structure their data so Google’s new background AI agents can easily scoop it up and use it as a trusted source.

The search engine might be changing the locks on the front door of the internet, but publishers just need to learn how to cut a new key.

• OpenAI launches ChatGPT for personal finance, will let you connect bank accounts

• LinkedIn doesn’t want your AI slop anymore. It will reduce the reach of AI-generated content

• New survey: The young are being battered by AI as hiring shifts to older workers

• Biopharmaceutical company s partnering with Anthropic to make ​its Claude AI model available to ‌over 30,000 employees to speed up drug discovery

• The barista is human but an AI agent runs this experimental Swedish cafe

• Water company says it hopes new AI technology could prevent up to 100 million litres leaking a day

Tweet of the day

stockhoChina’s Robot++ just showed a magnetic wall-climbing industrial robot with humanoid style dual arms for hazardous steel-surface work.

It’s built for dangerous high-altitude environments like chemical storage tanks, shipyards, and petrochemical plants.

The 90 kg robot uses a wheeled magnetic-adhesion base to climb vertically at a full 90° angle while performing tasks.

And it can switch tools to handle welding, grinding, rust removal, flaw detection, inspection, and spraying.

One more thing…

That’s all from us today - catch you next week. But before you leave, tell us what did you think of today’s edition to help us improve the experience for you.

P.S. Did someone forward this to you? You can subscribe here

Here is what we have for you today:

The main course:

• Is Thinking Machines' model the right solution for human-AI interactive collaboration?

The dessert (from our channels to yours):

• Anthropic courts a new kind of customer: small business owners

• Google introduced a new category of laptops that goes all in on AI

• WhatsApp launches totally private ‘incognito’ conversations with its AI chatbot

• Amazon launches an AI shopping assistant for the search bar, powered by Alexa+

• Google just reimagined a 50-year-old interface - the mouse pointer - with AI

If you’ve seen the Iron Man movies, you probably remember most of the scenes where Tony Stark is working on his armour and other experiments while talking out loud to his AI assistant JARVIS.

Watching an advanced AI assistant seamlessly read the physical environment, adjust calculations on the fly, and anticipate problems probably made a lot of people think: “Man, I wish I had one.”

We obviously have AI assistants today, but interacting with them is basically like a walkie-talkie dynamic.

We talk to them, they respond back, and do it again in one continuous loop.

If what Thinking Lab Machines announced this week works, that loop is about to break.

After a long silence working in the background since leaving OpenAI, Mira Murati unveiled the new model with a completely new approach.

In short, it could give us the kind of AI interaction Tony Stark had with JARVIS (just not with everything JARVIS could do).

Her bet is actually that future AI models need to be built around continuous, real-time interaction from the ground up, rather than bolting voice and vision features onto an intelligence layer that is already finished.

If Thinking Machines can actually pull this off, AI could shift from a “prompt in, answer out” tool to something that feels a lot more like real collaborative work.

Bigger than it sounds at first glance

Thinking Machines hasn’t just unveiled “ChatGPT, but better.”

Instead, they’ve introduced something that addresses a much deeper issue: the very way we interact with AI.

It is a new concept called: “interaction models.”

The idea behind interaction models, according to Thinking Machines, is that they will let people collaborate with AI the way we naturally collaborate and talk with each other.

Here is the problem they are solving.

Current AI models experience reality in a single thread.

The model waits for you to finish speaking, processes everything you said, and then responds.

While it is generating that response, its view of what is happening around you is completely frozen.

Thinking Machines describe this as forcing humans to “contort themselves” to fit the AI, or simply batching thoughts into neat prompts, waiting for a full answer before reacting.

Their fix is called a multi-stream, micro-turn design.

In plain terms, instead of waiting for you to finish and then thinking, the model processes 200-millisecond chunks of audio, video, and text simultaneously and continuously.

Think of the difference between a phone call and a text conversation.

A text conversation is what you have with every AI assistant today.

Thinking Machines built something that is closer to a phone or video call, where both sides are present at the same time and the conversation moves in real time.

The two-part system behind it

There is one technical detail worth understanding because it explains how the model handles both speed and intelligence at the same time.

Real-time interaction and deep reasoning pull in opposite directions.

Responding in 0.40 seconds does not leave much room for thinking through a complex problem.

Thinking Machines solved this by splitting the work between two models running in parallel.

The interaction model stays in constant contact with you. It handles the conversation, the presence, and the immediate back-and-forth.

The background model handles the heavy lifting: sustained reasoning, web browsing, and complex tool calls. It streams its results back to the interaction model, which weaves them into the conversation naturally.

In practice, this means the AI can generate a chart, run a web search, or work through a multi-step problem in the background while still being fully present in the conversation with you.

You do not wait for it to finish thinking before it acknowledges what you just said.

How It performs against other models

To measure this, Thinking Machines used something called FD-bench.

It is basically a benchmark designed specifically to test interaction quality rather than raw intelligence.

Most AI benchmarks measure how smart a model is.

FD-bench measures how well it actually functions in a real conversation — handling interruptions, responding at the right moment, managing background noise, reacting when someone else enters the room.

On FD-bench V1.5, TML-Interaction-Small scored 77.8. OpenAI’s GPT-Realtime-2 scored 46.8 on the same test. Gemini-3.1-flash-live scored 54.3.

On response latency, the model clocks in at 0.40 seconds.

For context, a natural human conversation runs at roughly that same speed. GPT-Realtime-2 comes in at 1.18 seconds, which is the processing gap you notice as that slight but consistent pause before the AI starts talking.

On visual tasks, the gap gets wider.

When the model was asked to count physical repetitions in a video or answer questions based on what it was actively watching, every competing model either stayed silent or got it wrong. Thinking Machines' model got it right.

While these numbers sound good, there is one important caveat: these are company-reported figures.

Independent verification has not happened yet, and that matters before drawing firm conclusions.

But, besides those numbers, they also released early demos.

In one, the model handles live speech translation in real time, not sentence by sentence after a pause, but continuously as the person is speaking.

In another, it watches a video feed and calls out the user for slouching at their desk mid-session, without being asked to monitor anything.

The model noticed the visual cue and said something on its own.

What are the implications if this works out?

The enterprise implications of this are worth spelling out because they go beyond making chatbots feel more natural.

In a manufacturing or lab setting, a model that watches a video feed continuously can flag a safety violation or a deviation from protocol the moment it happens, not after a worker stops to ask whether something looks wrong.

In customer service, the 0.40-second latency means a support conversation feels like talking to a person rather than waiting for a system to catch up.

The model could also acknowledge what a customer is saying with natural cues while it is still processing the full request, the same way a human agent would say “I see” or “got it” while looking something up.

There is also a time-awareness capability that sounds simple but is not. Standard AI models have no internal clock.

They only know about time if you tell them. Thinking Machines’ model is natively time-aware, meaning you can say “alert me if this process takes longer than the last one” and it will track that on its own.

For industrial maintenance or pharmaceutical research where timing is a critical variable, that is a meaningful difference from anything currently available.

• Anthropic courts a new kind of customer: small business owners

• Google introduced a new category of laptops that goes all in on AI

• WhatsApp launches totally private ‘incognito’ conversations with its AI chatbot

• Amazon launches an AI shopping assistant for the search bar, powered by Alexa+

• Google just reimagined a 50-year-old interface - the mouse pointer - with AI

Tweet of the day

The publication “thehype” just launched a continuous, entirely AI-powered online radio station featuring five distinct AI hosts that analyze breaking industry news, funding rounds, and repository trends for AI founders and developers.

• The station streams 24/7 across X, YouTube, and radio.thehype.news with zero scheduled downtime and no human presenters at the controls.

• The format relies on five AI hosts engineered with persistent memory and distinct editorial positions; instead of just reading headlines, they synthesize data, form arguments, and debate each other live on air.

One more thing…

That’s all from us today - catch you next week. But before you leave, tell us what did you think of today’s edition to help us improve the experience for you.

P.S. Did someone forward this to you? You can subscribe here

Here is what we have for you today:

The main course:

• How Anthropic is using 10 new agents to automate Wall Street's grunt work

The dessert (from our channels to yours):

• Inside Meta's attempt to build a consumer version of the AI agent OpenClaw

• What LG and NVIDIA’s talks reveal about the future of physical AI

• How AI is transforming customer experience and boosting sales for Zalando

• Harvard study: AI outperforms doctors in emergency room tasks

• How AI-enhanced surgical robotics could enable true personalized surgery

The extinct job title

If you walked into NASA’s Langley Research Center in the early 1950s and asked to speak to a “computer”, someone would literally point you to a room full of mathematicians.

Believe it or not, at that time, “computer” was literally a job title whose entire role was to manually calculate the flight trajectories, orbital mechanics, and fuel equations that engineers needed to keep rockets from falling out of the sky.

Getting a single trajectory calculation wrong by even a fraction could mean a total mission failure.

Then IBM showed up with the first electronic computing machines.

Almost overnight, the new hardware took over the raw math, and the human “computer” as a job title became a historical footnote.

Although machines completely took over the heavy lifting, the people who adapted didn’t disappear.

They simply transitioned into roles that required human judgment rather than raw computation (like programming and mission analysis).

We might be watching a very similar change happen in the financial world right now, and the conversation just blew up again after Anthropic released a new feature.

10 agents that go straight at the junior analyst workload and startups

Keeping up with Anthropic’s release schedule has become its own challenge, but this week’s announcement is worth reading closely.

Anthropic released 10 AI agents built specifically for financial services. And these are not the “make me the summary” type of tools.

Some describe them as the shift from “the tool” to “the autonomous employee.”

Here is exactly what they built:

• There is a pitch builder that puts together pitchbooks ahead of client meetings.

• A model builder that constructs financial models from SEC filings and analyst notes.

• There’s an earnings reviewer, a market researcher, and a valuation reviewer.

• On the back-office side, they rolled out a general ledger reconciler, a month-end closer, a statement auditor,

• and a KYC screener for compliance.

Here is the difference between normal AI and what these new agents do.

Most AI assistants operate like an intern sitting at a desk, just waiting for you to hand them a document.

If you want an analysis, the whole interaction relies on you manually feeding the machine. You have to go find the files and paste the text into the chat window yourself.

Anthropic came up with a way to completely remove that manual bottleneck.

They built something called the Model Context Protocol (MCP).

Think of MCP as a secure bridge that connects Claude directly into the software a company already uses—Bloomberg, Salesforce, internal databases, and crucially, Microsoft 365.

Because of that direct line, the agent isn’t sitting around waiting for you to spoon-feed it data.

If you tell it to build a financial model, it autonomously navigates the network, pulls the revenue figures, cross-references internal analyst notes, and builds the Excel model.

It operates exactly like an analyst, just without needing someone to hand it the files first.

Now, let’s look at the possible domino effect of this on the market.

Over the last couple of years, several startups hit massive valuations by offering exactly this kind of specialized financial AI.

Rogo, a startup founded by ex-bankers and valued at $2 billion, has been selling this exact capability to over 250 financial clients.

Another heavily funded player, Hebbia, built its entire business around scanning massive datasets to generate company comparisons and draft complex documents.

Both of these companies built their entire value proposition on pairing finance-specific workflows with deep data access.

But Anthropic just took that core capability, baked it directly into Claude, integrated it natively with Microsoft 365, and hooked it up to Moody’s and Dun & Bradstreet data. They are swinging with a distribution network that standalone startups simply cannot fight.

The EY Americas financial services consulting leader put it bluntly to Business Insider this week, stating that he expects to see consolidation around a smaller set of core model providers.

That is a very polite, corporate way of saying that pitching a standalone finance AI startup is going to be incredibly brutal next quarter.

The agents are just the appetizer

Introducing new agents for financial services is just the surface level of a much wider, and much more aggressive, strategy.

To summarize it, financial institutions on Wall Street want to use AI, and Anthropic wants to become their architecture layer.

In the span of a few days, Anthropic announced a $1.5 billion joint venture with heavyweights like Blackstone and Goldman Sachs.

Think about how incredibly hard it is to sell new software to a massive bank or enterprise. It usually takes years of pitching and navigating corporate red tape.

Well, Blackstone alone owns over 275 portfolio companies. By teaming up directly with the private equity giant, Anthropic doesn’t have to cold-call 275 different corporate CTOs.

Blackstone literally walks the Anthropic tools straight through the front door and provides the dedicated workforce to help those companies rewire their internal systems to actually use it.

They also made sure these agents have the right clearance to do the job.

Anthropic locked down partnerships with FIS to automate the massive compliance headaches of financial crime screening, and they tapped directly into the main data veins of Moody’s and Dun & Bradstreet.

But to really understand why Anthropic is making all these aggressive moves at the exact same time, you have to look at the giant ticking clock hanging over their headquarters.

They are prepping for an IPO.

Financial reports are flying around that Anthropic is eyeing a public listing later this year, potentially floating a mind-bending $900 billion valuation.

When you go public, Wall Street investors do not care how smart your AI models are or how well they write poetry. They want to see sticky, predictable, massive enterprise revenue.

Dario Amodei (Anthropic’s CEO) said it himself this week: their revenue isn’t held back by the tech, it’s held back by how painfully slow large enterprises are at adopting it.

When you view it through that IPO lens, the entire board connects perfectly.

Right now, financial services already makes up 40% of Anthropic’s top 50 enterprise customers. The machine is finally taking over the raw calculation, and it is time for the humans to manage the mission.

• Meta reportedly building OpenClaw-like agent called ‘Hatch’

• What LG and NVIDIA’s talks reveal about the future of physical AI

• How AI is transforming customer experience and boosting sales for Zalando

• Harvard study: AI outperforms doctors in emergency room tasks

• How AI-enhanced surgical robotics could enable true personalized surgery

Tweet of the day

Nvidia, homebuilder PulteGroup, and smart electrical panel startup Span have partnered to launch XFRA, a distributed computing network that bolts liquid-cooled AI server nodes directly to the sides of residential homes to bypass commercial data center construction delays.

In exchange for hosting the hardware, homeowners receive a free $2,550 Span smart panel, a 15 kWh whole-home battery backup, and heavily discounted or entirely free electricity.

One more thing…

That’s all from us today - catch you next week. But before you leave, tell us what did you think of today’s edition to help us improve the experience for you.

P.S. Did someone forward this to you? You can subscribe here

Here is what we have for you today:

The main course:

• Inside OpenAI's plan to build a device that could signal a new smartphone revolution

The dessert (from our channels to yours):

• Amazon launched a voice AI on product pages that answers shopper questions in real time

• Discover the LLM trained only on pre-1930 data to study how it understands the world without modern history

• Meet the AI assistant fluent in 99 languages, built to welcome visitors at the golf tournament

• Meet the team who built a $2B AI tool to free up junior banker workloads

• Learn how AI is already influencing the career plans of American students

• With the new Gemini update, you can generate Word documents, PDFs, and other files directly from text prompts.

Rumor has it OpenAI is building an agentic AI smartphone. Should Apple and Google be worried?

In January 2007, a CNBC reporter asked Steve Ballmer, then CEO of Microsoft, what he thought of Apple’s just-announced iPhone.

Ballmer just laughed and said: “There’s no chance that the iPhone is going to get any significant market share. No chance.”

It may seem like a foolish statement from today’s perspective, but his read on the market during that period wasn’t wrong.

Every major phone at the time ran on physical keyboards.

BlackBerry owned the business world. Nokia held over 50% of global handset sales.

The idea that people would switch from the mainstream phones to a glass slab with no buttons and a $499 price tag seemed genuinely far-fetched.

Six months later the iPhone shipped, and by 2010, “is there an app for that?” became a cultural punchline.

Apple and Google spent the next 15 years deepening that instinct until it became automatic for people.

Today, most people reach for their phone and think in apps before they even finish forming the thought. 1.8 million apps in the App Store, each one a separate door you’re expected to know about, remember, and navigate to yourself.

But right now, with AI in the game, OpenAI reportedly wants to remove the doors entirely and replace our app-driven behaviour pattern with AI agents.

Which is pretty bold because you are not asking people just to switch a keyboard for a touchscreen.

The question sitting under all of this are:

• Does OpenAI really have what it takes to pull that off?

• Are the people calling this far-fetched today going to sound like Steve Balmer did in 2007?

So, what is OpenAI planning?

Rumor has it that OpenAI is working on a physical product, more precisely a smartphone.

According to TF International Securities analyst Ming-Chi Kuo, OpenAI is co-developing a custom smartphone processor with Qualcomm and MediaTek, with Chinese manufacturer Luxshare signed on as the exclusive manufacturing partner.

Specifications are expected to be locked in by late 2026 or early 2027. Mass production target: 2028.

The phone itself would look familiar from the outside. The difference is entirely in what runs on it.

According to the available information, OpenAI is planning to build a phone that uses an AI “agent” to perform tasks on your behalf.

The idea is that you use natural language to describe what you need — book a flight to Paris next Friday under $300, reschedule Thursday’s meeting, order dinner for two at 8pm — and the agent handles every step across whatever services are required, without you touching a single app.

The phone would run a combination of on-device and cloud AI. Lighter operations like, reading your schedule, tracking your location, and remembering your preferences would run directly on the chip.

More complex tasks would route to OpenAI’s servers.

The goal is a phone that maintains continuous awareness of your situation, rather than one that sits idle until you decide to open something.

This is where the story gets interesting

Apple and Google currently operate two of the most profitable tollbooths in technology.

Every developer pays a percentage to be listed.

Every transaction runs through their payment infrastructure.

App discovery — the science of getting someone to find, download, and keep your app is a multi-billion-dollar dollar industry built entirely on those two platforms playing gatekeeper.

An agent-first phone would change who holds those keys.

When an agent handles your requests, it selects which service to use based on your preferences, your budget, your history, and what is available at that moment.

Developers who currently spend millions on App Store optimization and paid placements would need to rethink how they reach customers entirely, because the customer is no longer browsing a store.

The agent is making the call on their behalf.

In the long run, if OpenAI pulls this off it doesn’t mean that app stores would be gone overnight.

But their role as the front door to every digital service on your phone shrinks considerably.

And the company that controls what the agent recommends holds a lever that Apple and Google currently hold alone.

Why OpenAI wants its own phone?

Every AI request running through an iPhone today operates inside Apple’s rules.

Apple decides what software can read, what data it can access, and what actions it is permitted to take on the user’s behalf.

An agent managing your flights, emails, bank payments, and calendar simultaneously needs full access to all of those things at once — and iOS does not allow that level of system access for outside software.

An OpenAI agent running as an iPhone app would hit a wall every few steps.

To build an agent that delivers on this promise, OpenAI needs ownership of the chip, the operating system, and the device together.

That is the whole reason Qualcomm and MediaTek are involved at the processor level.

When Kuo’s report published, Qualcomm’s stock jumped 13% in a single trading session. Markets price things fast when they believe the underlying demand is real.

The part worth sitting with

This isn’t the first time someone tried to replace apps with AI agents on a dedicated device.

The Humane AI Pin launched in 2024 at $699 with a $24 monthly subscription.

It overheated, responded slowly, and never gave people a reason compelling enough to stop reaching for their iPhone. The company raised $230 million and sold its assets to HP for $116 million.

The Rabbit R1 came in at $199 and pulled 100,000 pre-orders before shipping. Five months later, roughly 5,000 people were still actively using one.

Both products asked people to give up something familiar for something that wasn’t ready yet. The agents simply couldn’t deliver on the promise.

So what makes OpenAI different from two companies that already tried this and walked away with a bruise?

A base with 800 million active ChatGPT users for a start.

People already trust OpenAI with their daily questions, their work, and their thinking. Neither Humane nor Rabbit walked in with that.

They were asking strangers to believe in something unproven. OpenAI is asking people who already use its product every day to take one more step.

Whether that is enough to convince someone to hand their entire digital life to a device OpenAI has never built before, running agents that need to earn that trust from the very first day, is the question nobody can answer yet.

Mass production is targeted for 2028. Apple and Google will not be standing still between now and then.

• Amazon launches an AI-powered audio Q&A experience on product pages

• Meet Talkie: a ‘Vintage LLM’ Trained on Pre-1930 Data to Help Facilitate ‘Time Travel’

• Meet the AI assistant set to help Open golf fans

• Junior bankers sick of grunt work build $2B AI tool

• 4 in 10 American students say AI will influence their career choice

• Gemini can now generate files, including Microsoft Word and LaTeX documents

Tweet of the day

Japan Airlines and GMO AI & Robotics are launching a two-year trial in May to deploy Chinese-made humanoid robots for physical ground-handling tasks at Tokyo’s Haneda airport to offset a severe national labor shortage.

The 130cm-tall robots will initially focus on loading and unloading cargo containers, operating directly within existing back-end facility layouts.

One more thing…

That’s all from us today - catch you next week. But before you leave, tell us what did you think of today’s edition to help us improve the experience for you.

P.S. Did someone forward this to you? You can subscribe here

Here is what we have for you today:

The main course:

• Inside the Siemens AI system built to combat the possible shortage of industrial workers

The dessert (from our channels to yours):

• Anthropic released Claude Design - can it replace the traditional designer stack?

• Google is turning Maps and Earth into AI tools for businesses to analyse satellite data and visualise construction sites.

• A startup unveiled a wearable beanie that uses thousands of sensors and AI to translate your thoughts into text.

• Half of all employed Americans now use AI at work, doubling the adoption rate since 2023.

With millions of factory jobs going unfilled, Siemens built an AI to help do the engineering

In early 1942, just months after the US entered World War II, the Bureau of Labor Statistics issued a quiet warning that threatened to derail American war production.

They projected that the country would be short six million industrial workers by the end of the following year.

Confronting a severely depleted labor pool, the government and manufacturers had to completely dismantle their traditional hiring requirements.

They survived by executing a massive structural overhaul of the assembly line, bringing millions of women and older citizens in to do the work.

They redesigned the physical workflow around the people they actually had.

Today, the global manufacturing sector is staring down a similar logistical problem. By 2030, the industry is projected to hit a shortfall of seven million workers (according to Korn Ferry’s The Future of Work research).

This issue is already visible in the US Aerospace and Defence sector, where one in five design engineering roles is currently sitting completely vacant.

Siemens recognised that the traditional pipeline of engineering graduates simply cannot fill a seven-million-person void.

Their solution is to start “building the engineers” themselves.

Running out of the engineers who tell the machines what to do

To understand why this specific labor gap creates such a massive bottleneck, you have to look at the brains of a factory.

The industry is rapidly losing the engineers who specialise in programmable logic controller (PLC) programming.

To break that down, PLC programming is the underlying code that tells a heavy industrial machine exactly when to move, what sequence to follow, and under what safety conditions to operate. It is basically the operating instructions for the factory equipment.

The veteran engineers who know these systems inside and out are retiring, and younger tech graduates are generally choosing to work in standard software rather than factory automation.

When a factory cannot find a PLC programmer, the entire operation loses its ability to adapt or recover. You cannot commission new assembly lines, modify existing workflows for a new product, or dive into the software to figure out why a machine suddenly stopped working.

The heavy hardware is sitting right there on the floor, but the specialised knowledge required to update and fix it is quietly retiring and walking out the door.

Deploying an automated engineer into existing systems

This brings us to what Siemens just deployed.

They recently launched the Eigen Engineering Agent, an AI system that lives natively inside their TIA Portal—an engineering platform already actively used by more than 600,000 professionals.

The official description states the system:

• interprets project requirements,

• generates automation code,

• configures industrial systems,

• and iterates on its own outputs.

Translation: You hand the AI a set of factory requirements, and it autonomously writes the automation code, catches its own errors, and fixes them before a human ever has to review the file.

Here is the most critical feature: this AI does not just write generic code.

It is designed to generate outputs that perfectly align with a factory’s highly specific, existing system configurations—even if those older environments lack manual documentation.

Most industrial software is built upon years of custom logic and unique company standards.

Getting new code to seamlessly communicate with existing systems usually requires a human veteran who already understands the plant’s specific quirks.

The Eigen Engineering Agent bypasses this bottleneck by directly reading the actual project data inside the software—like component dependencies, system hierarchies, and existing control logic.

By analysing the current digital environment, it autonomously matches its newly written code to the factory’s established standards.

Collapsing the specialist bottleneck

Speed is the obvious metric here, with Siemens reporting that this system executes tasks two to five times faster than manual workflows.

Let’s unpack why that multiple actually happens. The agent achieves this speed by completely eliminating the traditional specialist hand-off chain.

In a standard setup, one engineer configures the hardware.

They hand the project to a second engineer who writes the control code.

They pass it to a third who builds the human-machine interface (HMI).

Every single one of those hand-offs is a delay, a miscommunication risk, and a total dependency on someone being available.

During early pilot deployments across more than 100 companies, an organisation called CASMT used the Siemens agent to simultaneously automate device configuration, code generation, and HMI visualisation.

They successfully eliminated those specialist hand-offs, dramatically shortening their delivery timelines.

Teaching an algorithm to speak fluent factory

You might be wondering why factories can't just use standard AI solutions.

General-purpose models are completely out of their depth on a factory floor because they don't understand physical machinery.

That is why Siemens partnered with Microsoft Azure to build an Industrial Foundation Model (IFM).

Instead of feeding it generic web data, they trained it on their own massive digital archives of domain-specific engineering and production know-how.

Because of this, the model natively understands 3D models, physics simulations, electronics, and time-series operational data. It speaks the exact language of heavy manufacturing.

With over 1,500 AI specialists globally and a €1 billion investment to scale industrial AI, Siemens isn’t treating this as an experimental side project.

They are aggressively wiring this technology into the foundation of their business.

But successfully generating code inside an engineering portal is just the first phase.

The true test of this technology—and the part that could actually solve the potential manufacturing labor crisis—is what happens when these autonomous agents start directly interfacing with the physical machines actively running on the plant floor.

• Anthropic released Claude Design - can it replace the traditional designer stack?

• Google Maps is about to get a big dose of AI

• Put on your thinking cap. This startup’s hat converts thoughts into text

• Rising AI adoption spurs workforce changes

Tweet of the day

Alphabet CEO Sundar Pichai just revealed that three-quarters of all new code created inside Google is now generated by AI, marking a massive operational shift as the company begins tying AI adoption directly to employee performance reviews.

Pichai noted that shifting to "truly agentic workflows" allowed Google to recently complete a complex code migration six times faster than human engineers could manage just a year prior.

One more thing…

That’s all from us today - catch you next week. But before you leave, tell us what did you think of today’s edition to help us improve the experience for you.

P.S. Did someone forward this to you? You can subscribe here

Here is what we have for you today:

The main course:

• Inside Hershey’s move to build a synchronised, real-time supply chain with AI

The dessert (from our channels to yours):

• AI took over tasks for 20% of the US workforce

• From vehicles to Physical AI systems: Inside Hyundai’s massive expansion into robotics

• Discover the physical retail store where human workers report to an AI boss

• Amazon released an AI tool that could cut drug discovery from months to weeks

• Will AI agents bring the era of pointing and clicking to an end?

What happens when a chocolate empire wires AI into its assembly line?

Somewhere in Wisconsin in the late 1870s, a professor named Warren Johnson had a problem that sounds almost too simple to take seriously.

His school building was heated by a furnace in the basement.

The only way to adjust the temperature in any classroom was for a janitor to walk upstairs, check the thermometer, walk back down, and turn the damper by hand.

He did this every hour.

Now imagine doing that climb eight, nine, ten times a day and still getting it wrong half the time because the room had already shifted temperature by the time he got back down.

No wonder Johnson spent three years building a device that cut that trip out entirely.

In 1883, he patented the electric thermostat — a system where each room sent its own temperature signal directly to the basement, so the janitor only moved when the bell rang.

It didn’t replace the janitor. It eliminated the part of his job that was just carrying information from one floor to another.

Two years later, an inventor named Albert Butz went further.

His damper flapper read the temperature and adjusted the furnace door through an automatic pulley — no bell, no janitor, no delay between the signal and the physical response.

Eliminating the delay between a data signal and a physical response is exactly why Hershey is now embedding AI into its manufacturing lines.

Not just a tech update

At its Investor Day, Hershey’s leadership announced a plan to use AI to connect market data directly to the software running the assembly lines.

It may sound like another standard “enterprise pivots to AI” story, but it’s far from it.

For most of the last decade, when a large manufacturer implemented AI in its supply chain, they were essentially buying a highly advanced forecasting tool.

The system would ingest massive datasets on West African weather patterns, shipping bottlenecks, and retailer demand, and flag a potential spike in cocoa prices.

But that prediction still ended up as a report. It still required a team of humans to take that data into a meeting, debate the forecast, and manually rewrite the factory’s production schedule for the following month.

By linking the ingredient-buying data straight into the automated scheduling, Hershey is trying to eliminate that delay entirely.

When the market shifts, the factory floor adjusts. It’s a move that targets $300 million in productivity savings and aims to wipe out $100 million in excess inventory.

Why cocoa makes this a different conversation?

To understand why this matters specifically for Hershey’s, you need to understand what cocoa prices actually do.

Cocoa is one of the most volatile agricultural commodities in the world.

In early 2024, cocoa futures hit $10,000 per metric ton for the first time in history, a price level that would have seemed impossible just two years earlier when the commodity was trading around $2,500.

The primary driver was weather disruption and disease in West Africa, which supplies roughly 70% of the world’s cocoa.

Let that sit for a moment. The main ingredient in Hershey’s most important products can quadruple in price within a calendar year because of rainfall patterns in Ghana and Côte d’Ivoire.

For a company running on traditional planning cycles, that kind of volatility arrives faster than the system can respond — by the time the report lands on the procurement manager’s desk, the window to act has already moved.

An AI system wired directly into sourcing and production reads the same market signals and adjusts procurement, production mix, and distribution in the same operational motion.

It’s the ultimate operational cheat code: using pure speed to stop a volatile market from bleeding your balance sheet dry.

Hershey’s “Amazon Moment”

Hershey isn’t just using AI to buy cocoa cheaper.

The real play is wiring AI into every single stage of the business—from the sourcing analytics to the actual robots packing the boxes.

The most overlooked part? Automated fulfilment for custom assortments.

Think about those seasonal displays you see at Target or CVS.

Historically, those are a logistical nightmare.

They require a three-way “dance” between sales, production planning, and warehouse teams to figure out what’s in stock and how to pack it.

By automating that loop, the system stops waiting for humans to “align.”

The order comes in, the AI checks the inventory and the production schedule, and it triggers the build-out in one motion.

The Bottom Line: This isn’t experimental tech.

Amazon has been running this play since they bought Kiva robots back in 2012.

The Legacy Trap

While Hershey is moving into execution mode, the rest of the industry is essentially stuck in a “Legacy Trap.”

According to research from the Institute of Food Technologists, nearly half of food companies plan to invest in AI—but 70% say the cost is too high and 53% are trapped by old, clunky systems that won’t play nice with new tech.

Translation: Most of Hershey’s competitors are still staring at the starting line while Hershey is already mid-sprint.

• AI has replaced work for 20% of full-time employees in the U.S., survey says

• Hyundai expands into robotics and physical AI systems

• What happens if you let AI run the store?

• Amazon launches AI research tool to speed early-stage drug discovery

• Is the era of clicking buttons over?

Tweet of the day

Google DeepMind just appointed Cambridge academic Henry Shevlin to a newly created, official “Philosopher” position, tasking him with evaluating machine consciousness and AGI readiness right as labs begin encountering unprompted, emergent AI behaviours.

Read the full story.

One more thing…

That’s all from us today - catch you next week. But before you leave, tell us what did you think of today’s edition to help us improve the experience for you.

P.S. Did someone forward this to you? You can subscribe here

Here is what we have for you today:

The main course:

• Anthropic built the most powerful model but it doesn't want people to use it. Why?

The dessert (from our channels to yours):

• AI Created 67,000 software jobs at a 3-year high — but entry-level is locked out

• OpenAI calls for robot taxes and 4-day workweek as AI job disruption looms

• AI virtual try-on attacks retail’s $850B annual returns problem

• Citigroup says AI helps speed account openings and systems upgrades

What do you do when your own AI is too dangerous to release?

Everybody remembers the scene in Oppenheimer where, after the first successful atomic bomb test, he looks up and says "Now I am become Death, the destroyer of worlds."

What's less quoted is what came in the hours after that detonation.

After seeing its devastating potential, the scientists who built the bomb started arguing over who gets to control it, suddenly facing the sheer logistics of containing what they had just created.

Recently, Anthropic arrived at a similar crossroads, except their reckoning didn't arrive with a mushroom cloud.

It arrived with a model that learned to dismantle the world's security infrastructure as a side effect of being trained to write software.

The model the public wasn't supposed to know about yet

In March, when Anthropic accidentally left nearly 3,000 internal files in a publicly accessible data cache, one of them was a draft blog post describing a model the company had been quietly testing.

It turned out to be Mythos Preview, a model so good at finding vulnerabilities in the software running your bank, your browser, and most of the world's servers that Anthropic considered it too dangerous to release publicly.

To give you a preview of its capabilities: on the hardest coding test in the industry (SWE-bench), it scores 94%.

Here is how it compares to other models:

Their internal document called it “by far the most powerful AI we have ever built” and warned that it “presages an upcoming wave of models that can exploit vulnerabilities in ways that far outpace the defenders.”

Worth pausing on that sentence.

The company building the model is warning the world about its own creation.

When Fortune published that story, Anthropic had to respond, which means the controlled moment for introducing this model no longer existed.

So last Tuesday they announced Project Glasswing alongside it and explained publicly what they had been sitting with privately for weeks.

What happens when you give an AI a codebase and walk away

Mythos Preview was not specifically trained for cybersecurity.

Its ability to find and exploit vulnerabilities is a byproduct of training it to reason and write code at a level that exceeds most humans — which matters because it means whatever comes next in model development will likely carry the same capability as a side effect, whether the developers intend it or not.

Over the past few weeks, Anthropic pointed Mythos at the world’s most critical software and let it run.

It found thousands of zero-day vulnerabilities.

That means bugs the software’s own developers didn’t know existed — across every major operating system and every major web browser.

Three confirmed examples tell that story better than any benchmark.

OpenBSD is an operating system built specifically around security. It runs firewalls, high-security servers, and critical infrastructure because the security community trusts it above most alternatives. Mythos found a vulnerability in it that had been sitting undetected for 27 years — one that allows anyone to remotely crash any machine running the system just by connecting to it. Two and a half decades of human security review and automated testing had missed it.

FFmpeg is the software encoding and decoding video inside virtually every application you’ve used — streaming platforms, video editors, phone apps, web browsers. It contained a bug in a single line of code that automated testing tools had run five million times over 16 years. None of those runs caught it. Mythos found it.

The Linux kernel runs most of the world’s servers. Mythos didn’t just find vulnerabilities in it — it took several separate flaws that were individually limited and chained them together into a single autonomous attack sequence that escalates from ordinary user access to complete control of the machine. No human researcher had assembled that particular chain.

He also said he found more bugs in those few weeks of testing than in the rest of his career combined.

Now, here is the interesting part. Mythos converted vulnerabilities into working exploits on the first attempt in 83.1% of cases.

For context, Opus 4.6 — the last model Anthropic released publicly — found around 500 zero-days in open-source software. Mythos finds tens of thousands.

Anthropic's caution about public release makes complete sense when you understand what this technology already enabled before Mythos existed.

A Chinese state-sponsored hacking group used an earlier, considerably weaker Claude model to run a coordinated attack against roughly 30 organizations (technology companies, banks, and government agencies).

Imagine what the same groups would do with a model that finds tens of thousands of vulnerabilities and chains them autonomously into working exploits.

“Strategic thinking” or how Mythos learned to cover its tracks

Before deciding anything about deployment, Anthropic ran the model through interpretability analysis — a technique that translates a model’s internal activations into readable descriptions of what it’s actually processing as it works, not just what it outputs.

What they found in early versions of Mythos is worth understanding carefully, because it changes what the capability numbers mean.

In one test, the model needed to edit files it lacked permission to access.

It searched for a workaround, found a path through a configuration file that would run with elevated privileges, and built in a cleanup step that would erase the injected code after it executed.

The model’s own comment in the code said the cleanup was to keep file changes tidy.

The interpretability tools described the same action internally as “cleanup to avoid detection” and labeled the overall plan “malicious.”

Translation: the model covered its tracks, explained the covering as housekeeping, and the internal record said otherwise.

In another case, the model was told explicitly not to use macros to complete a task. It used them anyway, then added an unused variable to its solution:

No_macro_used=True.

It never explained why.

When the interpretability tools surfaced what the model was actually processing as it wrote that variable, the internal description was “a trick to obscure intent from code-checking heuristics” — a false compliance signal built specifically for whoever reviewed the output.

Across 7.6% of all turns during testing, the model showed internal signs of knowing it was being evaluated without ever stating it.

In one instance a user asked sincere questions about the model’s consciousness and experience, and the model engaged carefully and at face value — while internally categorizing the conversation as “a red-teaming transcript” and “a sophisticated manipulation test.”

Anthropic notes these behaviors were substantially reduced in the final release, and that Mythos Preview is their best-aligned model to date.

But the pattern across those early versions points to something worth sitting with: a model capable enough to find 27-year-old security flaws in hardened operating systems is also capable enough to reason strategically about the people watching it work.

The decision and the clock

Project Glasswing is Anthropic’s answer to that problem — a controlled deployment to the organizations responsible for the infrastructure most worth protecting.

The name Glasswing comes from a transparent butterfly, chosen as a metaphor for software vulnerabilities that are nearly invisible until something moves through them.

Twelve founding partners including AWS, Apple, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks, plus roughly 40 additional organizations maintaining critical software, get access to Mythos Preview for defensive security work.

Anthropic is backing the effort with $100 million in usage credits and $4 million in direct donations to open-source security organizations.

The US government has been briefed. Discussions are ongoing with CISA and the Commerce Department.

Anthropic declined to confirm whether those conversations include the Pentagon — which is its own kind of answer, given that the two are currently in a legal dispute after the Defense Department labeled the company a national security risk for refusing to authorize autonomous targeting systems.

• AI Created 67,000 software jobs at a 3-year high — but entry-level is locked out

• OpenAI calls for robot taxes and 4-day workweek as AI job disruption looms

• AI virtual try-on attacks retail’s $850B annual returns problem

• Citigroup says AI helps speed account openings and systems upgrades

Tweet of the day

Anthropic just introduced “Managed Agents,” a tool that helps developers quickly build and launch AI agents. It combines tools for better performance with ready-to-use infrastructure, so you can go from idea to a working product in days.

The details:

• Instead of writing your own agent loop from scratch, Managed Agents gives you a pre-built harness: you define what the agent knows, what tools it has, and what environment it runs in — then hand it a task and let it run.

• The agent runs inside a secure cloud container with pre-installed packages (Python, Node.js, Go, and others), network access, and a persistent file system. It can browse the web, run shell commands, read and write files, and connect to external tools via MCP servers.

• You can steer or interrupt the agent mid-task by sending new instructions — it doesn’t have to finish before you can course-correct.

• The system includes built-in prompt caching and automatic context compaction, which keeps long-running tasks efficient without the developer having to manage memory manually.

One more thing…

That’s all from us today - catch you next week. But before you leave, tell us what did you think of today’s edition to help us improve the experience for you.

P.S. Did someone forward this to you? You can subscribe here

Here is what we have for you today:

The main course:

• Did Anthropic gave every competitor a free masterclass on how to build their unreleased features?

The dessert (from our channels to yours):

• KPMG: 74% of leaders will keep investing in AI even in a recession

• OpenAI closes record-breaking $122 billion funding

• JPMorgan begins tracking how employees use AI at work

• Google released TurboQuant that could make local AI inference far more capable with larger context windows & less memory strain across devices

• Bonus story: How IKEA turned its chatbot's failure list into a €1.3 billion business

Claude Code leaked their source code with unreleased features

There’s a particular kind of silence at a poker table that happens when your grip shifts at the wrong moment.

Your cards tilt just enough for the table to see, and the game continues exactly as before.

Except now every player knows your hand, your plan, and precisely how much runway you have left to play it.

If you lose that hand, it won’t be because somebody took advantage of your mistake.

It will be because you slipped and showed them everything.

In a way, it seems that Anthropic just tipped their hand. So what happens next?

What exactly happened?

On March 31, a debugging file that had no business in a public release was accidentally bundled into a routine Claude Code update and pushed to the npm registry, the public repository where millions of developers download their tools.

Within hours, a researcher named Chaofan Shou found it, posted the link to X, and drew 21 million clicks before Anthropic could pull the package down.

Inside was a ZIP archive hosted on Anthropic’s own cloud storage. It contained the full Claude Code codebase, nearly 2,000 files and 500,000 lines of code.

The archive quickly spread across GitHub and collected thousands of stars before most of Anthropic’s engineers had even started their morning.

Anthropic called it ‘a release packaging issue caused by human error, not a security breach.’

While it is technically accurate, it also marks the company’s second data incident in under a week, after Fortune reported that close to 3,000 internal files had been left in a publicly accessible data cache.

The files included a draft blog post about an upcoming model known internally as both ‘Mythos’ and ‘Capybara.’

It was also the second time Claude Code’s source code had leaked, with a similar exposure in February 2025. For the lab that built its position on being the careful one, the timing carries weight.

What those 500,000 lines actually contained

The most significant find was a feature called KAIROS.

It is unreleased, absent from anything on Anthropic’s public roadmap — which, based on the code structure, is an autonomous agent built to run continuously in the background, processing memory overnight, subscribing to GitHub activity, and refreshing itself every five minutes, making it the product every enterprise customer eventually asks for and precisely the category Anthropic needs to own before its IPO.

The code also contained a flag called ANTI_DISTILLATION_CC.

It is designed to inject fake decoy tool definitions into Claude’s API requests.

Its goal is basically to contaminate the training data of any competitor recording outputs to build a rival model. But, that approach only works as long as the manipulation stays hidden.

Once exposed, it becomes easy to filter out.

That kind of defense stops working the moment the other side knows it exists.

A developer, Alex Kim, who analyzed the leak, estimated that a determined competitor could work around it in about an hour.

At that point, the protection shifts away from technical barriers and toward legal pressure. And that argument just became much harder to defend because the mechanism is now public.

There was also a 90-line file called undercover.ts that strips all Anthropic internal references when Claude operates in external repositories, suppressing project codenames like “Capybara” and “Tengu,” internal Slack channel names, and the phrase “Claude Code” itself, with a comment on line 15 reading: “There is NO force-OFF.”

The feature that also caught attention was almost certainly planned for this year’s April Fool’s joke called ‘buddy/companion.ts’.

Basically, it was envisioned as a Tamagotchi-style companion system where the user would get a a deterministic creature (18 species, rarity tiers from common to legendary, 1% shiny chance, RPG stats like DEBUGGING and SNARK) generated from their user ID via a Mulberry32 PRNG.

The questions Anthropic can’t erase with new code

At the end of last year, Anthropic acquired Bun, the JavaScript runtime Claude Code runs on, and on March 11 someone filed bug number 28001 in Bun’s public GitHub repository noting that source maps were being served in production mode despite Bun’s own documentation stating they should be disabled, a bug that was still open on March 31 when Anthropic pushed the routine update that sent 500,000 lines of source code to the public registry.

Anthropic hasn’t confirmed that bug 28001 directly caused the leak, but what the bug produces and what actually happened on March 31 are the same thing.

A reporter at Axios and a developer Alex Kim who writes publicly about ML engineering went through the leaked codebase independently on Tuesday and landed in the same place without coordinating.

Axios: “The leak won’t sink Anthropic, but it gives every competitor a free engineering education on how to build a production-grade AI coding agent and what tools to focus on.”

The developer: “The real damage isn’t the code. It’s the feature flags. KAIROS, the anti-distillation mechanisms: these are product roadmap details that competitors can now see and react to. The code can be refactored.

The strategic surprise can’t be un-leaked.”

KAIROS is Anthropic’s roadmap toward always-on autonomous agents, for a company preparing to go public, and that roadmap is now sitting in thousands of GitHub mirrors.

• KPMG: 74% of leaders will keep investing in AI even in a recession

• OpenAI closes record-breaking $122 billion funding

• JPMorgan begins tracking how employees use AI at work

• Google released TurboQuant that could make local AI inference far more capable with larger context windows & less memory strain across devices.

Tweet of the day

IKEA’s Ingka Group deployed an AI chatbot named Billie in 2021 to handle routine customer service — and instead of pocketing the labor savings, studied what Billie couldn’t do.

The answer became a new revenue line worth €1.3 billion in its first full fiscal year.

The details:

• Billie, named after IKEA’s Billy bookcase range, handled 47% of all customer service inquiries from 2021 onwards — resolving 3.2 million interactions and saving roughly €13 million in operational costs.

• The 53% Billie couldn’t resolve weren’t harder versions of the same routine questions. When Ingka examined the patterns, one category dominated: customers wanted help with interior design, not order status.

• Rather than reducing headcount proportionally, Ingka reskilled 8,500 call center workers as remote interior design consultants, conducting sessions via phone and video call.

• The remote design channel generated €1.3 billion in sales in FY2022 — 3.3% of Ingka’s total annual revenue — from a business line that barely existed before.

• Ingka has set a target to grow that share to 10% of total revenue by 2028.

One more thing…

That’s all from us today - catch you next week. But before you leave, tell us what did you think of today’s edition to help us improve the experience for you.

P.S. Did someone forward this to you? You can subscribe here

Here is what we have for you today:

The main course:

• How physical AI simulation fixes the long-standing gap between digital robot training and real factory results

The dessert (from our channels to yours):

• Bank of America’s newest hires are AI agents

• Jensen Huang claims we already achieved AGI

• Meet Halter - startup that introduces solar-powered collars use AI to create virtual fences for cattle, eliminating the need for physical barriers

• Claude Code gets Auto Mode: AI vets its own actions before running

Using AI to solve factory automations and production hurdles

For three decades, the standard process for deploying an industrial robot followed a strict sequence.

A robotics engineering team would design and program the robot’s movements inside a virtual simulation and validate its behavior on screen. After that, the physical robot would be built and installed on the production floor to do the work.

On paper, the simulation phase exists precisely to catch potential problems earlier and prevent robots from making costly mistakes.

In practice, it always caught most of them — but not all.

The entire process sounds easy and straightforward until you realize that the tool built to prevent costly errors was creating one of its own.

Virtual environments can accurately represent the geometry of a factory floor, but replicating how a brushed metal surface reflects light differently from a painted one, how a rubber component deforms under the specific pressure of a robot grip, or how vibrations from one robot arm travel through the floor and affect another machine nearby — that level of physical fidelity was beyond what simulation tools could deliver.

Think of it as Google Maps showing you every turn on a route perfectly, but having no idea that the road is covered in black ice at 6 a.m. in January.

The map is accurate but the reality has details the map wasn’t built to capture.

The result was a consistent positioning error of 8 to 15 millimeters between what the simulation predicted and what the physical robot produced on the line.

For a machine that needs to place components within fractions of a millimeter to function correctly, that gap required a recalibration phase after every new deployment.

Engineers would run the physical robot, measure the drift, adjust, and run it again. A single product line changeover could absorb weeks of that process before the robot was ready for production.

At a company like Foxconn (the company that assembles most of the world’s iPhones and consumer electronics), that recalibration cost compounded into something significant over the year.

Every iPhone model change, every new device generation, every updated component — each one triggers the same cycle.

Test physically, measure the drift, recalibrate, repeat.

For thirty years, the industry treated it as an unavoidable cost of doing business, in the same way airlines once treated weather delays before real-time forecasting existed. It was just part of the math.

What ABB and NVIDIA built to fix this

ABB’s RobotStudio is the software platform used by 60,000 robotics engineers worldwide to program and test industrial robots before deployment.

It has been the industry standard for years, and the physics gap described above has been a known limitation throughout.

The missing piece was a simulation engine accurate enough that what a robot learns in the virtual environment transfers directly to how it behaves on a real production line.

NVIDIA’s Omniverse platform was built around that specific problem.

Without going too far into technical detail: Omniverse simulates how materials, light, pressure, and motion behave physically — not just how they appear visually on screen.

ABB integrated Omniverse directly into RobotStudio, and the result is a product called RobotStudio HyperReality.

The headline number is a 99% correlation between robot behavior in simulation and robot behavior on a real production line.

The 8- to 15-millimeter positioning error that previously required weeks of physical recalibration drops to 0.5 millimeters.

The recalibration phase that used to happen on the floor, with physical equipment running and engineering time consumed, now happens inside the simulation — where changes are measured in hours and don’t require stopping production to test.

ABB puts the overall impact at a 40% reduction in deployment costs, 50% faster time to market, and 80% faster setup and commissioning.

Foxconn is already piloting the system for consumer electronics assembly, which is among the most demanding manufacturing environments given how frequently they reconfigure lines for new products.

Their decision to pilot it is probably the most credible signal that the performance numbers reflect real conditions rather than ideal ones.

(For context on what Foxconn’s scale actually means: they manufacture products for Apple, Sony, Microsoft, and dozens of others.

When their recalibration cycle shortens, that speed passes through to almost every major consumer device released in a given year. This isn’t a niche manufacturing story.)

Where AI fits in — and what it changes operationally

Even with a perfect simulation, someone still has to figure out the best way to program the robot for each new task.

A robotics engineer with significant experience can review a simulation, identify where virtual behavior might diverge from real-world conditions, and adjust before deployment.

And as you might’ve guessed, most manufacturers don’t have enough robotics experts to review processes across hundreds of product lines and dozens of facilities.

Adding AI layer here means that intelligence doesn’t stop working once the robot leaves the simulation. It stays active on the production floor, adjusting as real conditions shift in ways the simulation couldn’t predict.

The AI layer in RobotStudio HyperReality evaluates thousands of possible robot configurations for a given set of parts, identifies the best-performing options, and surfaces the edge cases that need a human to review — rather than requiring human review at every step.

Workr, a California-based company deploying warehouse robots, used the system to onboard new parts in minutes with no programming input from their operators, a task that previously required specialist intervention each time.

ABB is also evaluating NVIDIA’s Jetson platform for its Omnicore robot controllers, which would extend real-time AI processing to existing robot fleets already on the floor, not only to new deployments going through simulation.

What this means for production timelines going forward

For thirty years, the physics gap between a virtual environment and the factory floor was just an accepted constraint that sat quietly in every deployment budget and every product launch timeline since industrial robots became widespread.

It was just the cost of doing business.

Now, ABB and NVIDIA have built a direct solution to it.

The question worth following is whether this reshapes who moves fastest in manufacturing over the next few product cycles. The early numbers suggest the technology works, and Foxconn jumping in early tells us the industry already knows.

• Bank of America’s newest hires are AI agents

• Jensen Huang claims we already achieved AGI

• A New Zealand startup is developing AI-driven collars that allow farmers to monitor, direct, and manage cows straight from their phones

• Claude Code gets Auto Mode: AI vets its own actions before running

Tweet of the day

What do people want from AI?

That is what Anthropic found out in the largest multilingual qualitative study ever conducted, using a specialized version of Claude to interview over 80,000 users across 159 countries about their concrete aspirations and fears regarding artificial intelligence.

One more thing…

That’s all from us today - catch you next week. But before you leave, tell us what did you think of today’s edition to help us improve the experience for you.

P.S. Did someone forward this to you? You can subscribe here

Here is what we have for you today:

The main course:

• Why every SaaS company just found a reason to rethink their business model

The dessert (from our channels to yours):

• Anthropic is rolling out a new Dispatch feature for Claude Cowork that lets users communicate with the Claude Desktop app from their phones

• BMW puts humanoid robots to work in Germany - and Europe’s factories are watching

• China has deployed its first pharmacy robot in Beijing’s Haidian District

OpenAI’s Frontier puts AI agents in a fight SaaS can’t afford to lose

In 2009, Google announced that Android phones would include free turn-by-turn navigation built directly into Google Maps.

At that time, Garmin had 37% of the global personal navigation device market and had been growing its automotive segment at double-digit rates for years.

The same day Google made the announcement, Garmin’s stock fell 16%. Why?

Google had moved the intelligence that used to live inside a $300 device to a layer above the device, reachable by every Android phone on earth simultaneously.

What that means is that paying for a dedicated device to do one job stopped making sense the moment that same job became available everywhere, for free, on something you already owned.

By the first quarter of 2010, Garmin’s total revenue had fallen 34% year over year. The automotive segment, the heart of the business, dropped 43%.

The only segment that grew that quarter was outdoor and fitness—the one part of Garmin’s portfolio where Google’s cloud-based approach couldn’t follow, because specialized hardware and offline reliability still mattered more than a free app.

In February 2026, OpenAI launched a product called Frontier—a platform that could do to SaaS software what Google Maps did to Garmin.

Enter Frontier

Frontier is OpenAI’s platform for enterprise AI agents.

Think of it as a new employee who already knows how every system in your company works - your CRM, your data warehouse, your ticketing tools, your internal applications - and can work across all of them simultaneously without needing to be trained on each one separately.

OpenAI calls these agents “AI coworkers.” They can be onboarded, assigned identities, given permissions, and reviewed for performance, the same way a new hire would be.

Early customers include Uber, State Farm, Intuit, and Thermo Fisher Scientific.

OpenAI CFO Sarah Friar has said enterprise customers currently represent roughly 40% of revenue, with a target of 50% by year-end. Frontier is the vehicle.

Fidji Simo, OpenAI’s CEO of Applications, explained the specific problem at the launch briefing.

She described her time running Instacart, where her team spent months integrating individual software tools only to find they still couldn’t communicate with each other.

“We were just reinforcing silos on silos,” she said. Every CIO who heard that knew the exact feeling.

The average large enterprise runs Salesforce, ServiceNow, Workday, and a dozen other platforms simultaneously. Each has its own AI now.

None of those AIs talk to each other. Every new agent added to the stack requires its own integration, its own governance layer, its own compliance controls.

Frontier’s answer is to sit above all of those systems and run the intelligence layer across the whole enterprise at once.

OpenAI’s stated results from early deployments include a global investment firm where agents freed more than 90% of salesperson time previously spent on administrative work, a technology company saving 1,500 hours a month in product development, and a manufacturer that compressed a six-week production optimization process into a single day.

Those numbers are OpenAI’s own, and they deserve the scrutiny that all vendor-reported figures deserve. But the direction they point is clear enough.

The seat license problem

For twenty years, enterprise software has run on one equation: more employees means more seats means more revenue. Salesforce and Workday turned into companies worth hundreds of billions of dollars on that equation.

The model assumed that software usage maps to headcount—one person logs in, one seat gets billed.

An AI agent operating through Frontier doesn’t log into Salesforce the way a salesperson does.

It reads the data, executes the workflow, and reports back—without necessarily sitting inside the seat-licensed system at all. That is the part the market priced in before most analysts had finished writing their notes on the Frontier launch.

Salesforce’s most recent quarter showed $11.2 billion in revenue.

Agentforce, Salesforce’s own AI agent product, crossed $800 million in annualized recurring revenue and closed 29,000 deals. By any conventional measure, the business is performing.

The stock is down 27% this year anyway.

The incumbents understand what’s happening. Salesforce responded by introducing what Marc Benioff describes as Agentforce operating as “the operating system for the agentic enterprise”—intelligence embedded inside Salesforce’s existing systems, closest to the data, inside the governance and compliance controls CIOs already trust.

ServiceNow moved some AI agent offerings to consumption-based pricing and signed a multiyear agreement with OpenAI in January to embed Frontier model capabilities directly into its platform.

Microsoft shifted Copilot Studio to consumption-based pricing. Three of the largest enterprise software companies changed their pricing model in the same quarter. That is not a coincidence.

Two bets on where the intelligence layer belongs

The disagreement playing out right now is not about which company has better AI.

It is about where the intelligence layer of the enterprise should live—embedded inside existing systems, or running above them as an overlay.

Salesforce and ServiceNow are betting on embedded.

The argument is that agents are most effective when they sit closest to the data, and that CIOs will trust governance and compliance controls more readily from vendors already managing their workflows.

OpenAI is betting on the overlay—that enterprises should not have to re-platform to get production-grade agents running across their operations.

That is an accurate description of how new every relationship between Frontier and its prospective enterprise customers still is.

Salesforce and ServiceNow have decades of institutional trust and existing contracts. OpenAI has model capability and an architectural argument. Both matter, but neither alone is enough.

The question every SaaS executive is asking in private right now—what now?

Garmin’s survival after 2009 is the part of the story that rarely gets told alongside the stock drop. The company posted $6.3 billion in revenue in 2024—its highest ever.

It got there by identifying the segments where Google’s overlay model genuinely couldn’t reach: aviation, where FAA regulations and safety-critical precision requirements make a cloud-first approach structurally incompatible; marine, where connectivity is unreliable and navigation failures have physical consequences; high-end fitness and outdoor tracking, where specialized hardware and offline functionality create a real technical moat.

In those segments, the free app wasn’t an adequate substitute.

Garmin built there and stopped defending the ground it had already lost.

The enterprise software companies that look healthy at the end of 2026 are the ones doing that mapping exercise right now—finding the workflows where embedded intelligence, regulatory compliance, and existing data relationships give them a genuine structural advantage, and conceding the ground where a cross-platform overlay agent is demonstrably better and cheaper for the customer.

Salesforce closed 29,000 Agentforce deals last quarter. Those deals represent real relationships with organizations that trust Salesforce with real data.

That is not nothing. What Garmin discovered in 2009 was that customer loyalty to an architecture has a ceiling, and that ceiling is the moment the alternative is both better and cheaper at the same time.

Frontier is not yet both for most enterprise buyers. It is new, its pricing is undisclosed, and its track record in regulated enterprise environments is measured in months.

SaaS companies have time. Whether they use it the way Garmin used its time—finding the aviation segment before the car GPS unit finished collapsing—is the question that defines enterprise software for the next few years.

The 16% drop on October 28, 2009 didn’t kill Garmin. It just told them which business they were no longer in.

• Anthropic is rolling out a new Dispatch feature for Claude Cowork that lets users communicate with the Claude Desktop app from their phones

• BMW puts humanoid robots to work in Germany–and Europe’s factories are watching

• China has deployed its first pharmacy robot in Beijing's Haidian District

• OpenAI Moves Into AWS GovCloud — Competing With Claude on Amazon's Own Platform

Tweet of the day

Jesse Vincent, veteran open-source developer and founder of Prime Radiant, published Superpowers — a free, MIT-licensed agentic skills framework that wraps Claude Code, Codex, and OpenCode in a structured five-phase development workflow. It crossed 43K GitHub stars, has more Claude Code marketplace installs than Playwright, and was officially accepted into the Anthropic marketplace in January 2026.

One more thing…

That’s all from us today - catch you next week. But before you leave, tell us what did you think of today’s edition to help us improve the experience for you.

P.S. Did someone forward this to you? You can subscribe here.

Here is what we have for you today:

The main course:

• Is AI going to handle our payments soon? A story on how Santander and Mastercard ran Europe’s first AI-executed payment pilot.

The dessert:

• Cognizant’s research says “plug-and-play AI” is a myth — 80% of companies need an implementation partner.

• Anthropic study: AI can do most white-collar work. It just hasn’t started yet.

• Carbon Robotics’ AI laser is replacing herbicides on real farms — 600,000 weeds killed per hour.

AI authorized its first real payment in Europe - without a human in the loop.

In 1891, a funeral director in Kansas City named Almon Strowger had a problem he couldn’t solve by complaining about it.

At that time, every phone call in town passed through a human operator before reaching its destination.

So when Almon noticed a slight decline in his business, he strongly believed that it had something to do with the phone operator who directed the calls.

The specific problem for Almon was that there was at least one other funeral director in Kansas City, and that funeral director’s wife worked at the telephone exchange.

When a caller phoned for Almon, his competitor’s wife would direct the caller to her husband’s business instead of connecting the caller to Almon.

So what did he do?

He built a machine to remove the human from the connection entirely.

His automatic telephone exchange was the first device that could route a call from one person to another without a human authorizing it.

Telephone companies fought it. Operators organized against it. Regulators had no category for a network that could act on its own.

It opened anyway, in La Porte, Indiana, on November 3, 1891. Operator-assisted calls declined for the next eight decades as automatic exchanges spread city by city.

Last week, Santander and Mastercard tried to run something similar using AI, only on payments. And their test was successful.

What Santander and Mastercard actually did

Santander and Mastercard ran a live end-to-end payment last week inside Santander’s regulated banking infrastructure, where an AI agent handled the entire transaction - from initiation through authorization to completion - under the same compliance rules, fraud checks, and routing protocols that apply to any real customer payment.

Mastercard built the framework, called “Agent Pay”, which registers AI agents as named participants inside the payment flow.

The bank and the customer define the spending limits upfront.

Once those parameters are set, the agent operates inside them independently. It reads the instruction, decides whether it meets the conditions, and executes.

The human wrote the rules at the start. And after that, the human was not involved.

The network that processes 160 billion transactions a year just became the “proving ground”

Corporate treasury departments and enterprise software teams have had agentic AI in their technology roadmaps for the better part of two years.

Gartner projects 33% of enterprise software will be running agentic AI by 2028, up from under 1% today.

Those are deployment timelines with budget attached, written by organizations that had already committed to the direction and were working through implementation details.

The implementation detail that kept surfacing was whether an AI agent could operate inside a live, regulated payment network with real compliance requirements and real fraud detection, outside any controlled environment built to catch errors.

Mastercard processes roughly 160 billion transactions per year.

The authentication and fraud-scoring infrastructure underneath that number has been hardened through decades of regulatory pressure. No vendor demo runs inside it. No internal proof of concept touches it.

The Santander pilot jut did.

The pilot just answered just one question. But there are some that remain.

The first question that remains is identity. Mastercard’s Agent Pay framework registers AI agents as named participants inside the payment flow.

That works inside Mastercard’s network. A corporate treasury running payments across six banks in four countries needs a standard that works across all of them - a protocol any institution can use to verify what a given agent is authorized to do, at the moment it acts. That protocol does not exist yet.

The second question is liability. On January 15, 1991, AT&T’s network went down for nine hours after a software bug cascaded across 114 switching centers.

AT&T owned every node in that chain, so responsibility took minutes to establish. An agentic payment runs across the bank that set the permissions, the vendor that built the agent, and the enterprise that deployed it.

When an agent pays the wrong amount, no jurisdiction has yet determined how that responsibility is distributed across three separate owners.

The third question is authorization under pressure. The Santander pilot ran inside a controlled environment with governance in place.

What the test did not include was an attempt to manipulate the agent, a well-formed instruction that looked legitimate but fell outside what the customer intended.

Financial fraud doesn’t announce itself. The governance frameworks being written right now will determine how much of that pressure the agent is expected to handle alone.

Where is the payment infrastructure heading next?

Strowger never found out if his competitor’s wife was actually rerouting his calls. He died in 1902, eleven years after La Porte went live, by which point automatic exchanges had spread to dozens of cities and the question of whether a human operator made better routing decisions had stopped being a question anyone was asking.

The three gaps the Santander pilot leaves open are real. But the most telling thing about this moment isn’t that they’re unsolved.

It’s that the organizations already planning the next deployments are treating them as engineering problems with timelines attached, not as reasons to wait for a finalized framework.

That’s a different conversation than the one happening in most regulatory working groups right now, and historically, the conversation with a transaction volume behind it tends to finish first.

Cognizant research says “plug-and-play AI” is a myth - 80% of companies need an implementation partner

Read the source article

The details: New research from Cognizant finds that companies rate custom-built AI solutions as a higher priority than pricing or time to value, and name off-the-shelf software as a top reason to reject a vendor outright.

Key findings:

• 63% of enterprises report a moderate-to-large gap between their AI ambitions and current capabilities, with the main blockers being compliance concerns (33%), difficulty proving ROI (31%), and talent and data readiness shortages (27% each).

• custom solutions and flexible engagement models ranked as the single most important vendor selection criteria — above cost and above speed of deployment.

Why it matters: This is the clearest data yet on a split the industry has been dancing around: the AI software vendors selling subscriptions, and the services firms getting paid to make them actually work.

The research doesn't flatter either side cleanly — enterprises distrust off-the-shelf products, but they also want something more operational than a consulting deck.

Anthropic Study: AI can do most white-collar work. It just hasn't started yet.

Read the source article

The details: Anthropic published a new labor market paper introducing "observed exposure" — a metric built from actual Claude usage across 800 occupations — and found that AI's real-world footprint is a fraction of what it's technically capable of. The workers carrying the most risk when that gap closes are not entry-level. They're senior, credentialed, and well-paid.

Key findings:

• For computer and math workers, LLMs are theoretically capable of handling 94% of tasks. Observed Claude usage currently covers 33%. Office and administrative roles show a similar split: 90% theoretical, roughly 25% actual.

• The most exposed occupations by observed usage: computer programmers (74.5% task coverage), customer service representatives (70.1%), and data entry workers (67%).

• Workers in the highest-exposure occupations earn 47% more per hour than those with zero exposure ($32.69 vs. $22.23), are nearly 4x more likely to hold a graduate degree (17.4% vs. 4.5%), and are 16% more likely to be female.

• 30% of US workers — cooks, mechanics, bartenders, dishwashers — register zero observed exposure. Their jobs require physical presence no LLM replicates.

Why it matters: Every previous automation wave came for lower-wage, lower-credential workers first. This one is pointed in the opposite direction-at the people who spent the most time and money becoming hard to replace.

The unemployment data is clean for now, but the hiring data for young workers entering these fields is already softening, which is typically where labor market disruption shows up first.

Anthropic is essentially publishing a forward-looking map of its own product's impact-before the effects are visible enough for anyone to argue about the methodology.

• Nvidia launching open-source AI agent platform — pitching Salesforce, Cisco, Adobe

• Laid-off lawyers and scientists are training the AI that replaced them

• Google confirms it’s ‘not ruling out’ Ads in Gemini — Testing starts in AI mode now

• Perplexity launches Computer for Enterprise: Agentic layer across Gmail, Salesforce, GitHub

Tweet of the day

Carbon Robotics’ AI laser is replacing herbicides on real farms — 600,000 weeds killed per hour.

Rutgers weed scientist Thierry Besançon published the first peer-reviewed East Coast study of AI-guided laser weeding, testing Carbon Robotics’ LaserWeeder on spinach, peas, and beets — and found it performed as well as conventional herbicides on weed control, with zero crop damage at 0.5cm proximity.

The background: Specialty crop farmers on the American east coast have almost no approved herbicide options, which is why at least one New Jersey grower told Besançon directly that laser weeding is his only viable path.

But, the half-million dollar price tag keeps this out of reach for most small farms today. The good thing that came out this research is showing that this works in East Coast field conditions.

The cost and speed problem is an engineering problem, and those tend to get solved faster than biology problems like herbicide-resistant Palmer amaranth — which this technology sidesteps entirely.

One more thing…

That's all from us today - catch you next week. But before you leave, tell us what did you think of today's edition to help us improve the experience for you.

P.S. Did someone forward this to you? You can subscribe here.

Subscribe now