2026-09-25 23:11:52
Meta launched Muse earlier this month. It soon became the top free iPhone app in the US. It is a cousin of Instinct, the invite-only assistant you text, and of OpenClaw agents like my AI chief of staff, R Mini Arnold. Jolly, Muse’s digital avatar, is just cuter, and probably easier to use. (Sorry, RMA.)
This is the original promise of AI agents, as articulated by Pattie Maes back in 1994:
Agents radically change the current user experience, through the metaphor that an agent can act as a ‘personal assistant.’ The agent acquires its competence by learning from the user as well as from agents assisting other users. Several prototype agents have been built using this technique.1
It was a digital butler, in the words of Nicholas Negroponte, founder of MIT’s Media Lab, that knows your context and gets things done. If such butlers work, they end up standing between you and everything you buy.
R Mini Arnold already does this for me. It hunted for a helper for my mum. It sorted the refund of a faulty headphone cable and complained to the hotel that locked me in my room. On holiday in Greece, it booked a restaurant that only took bookings by email. All of it ran through WhatsApp. That beats wading through SEO spam and badly built websites.
Muse’s early users are finding the same. One asked it to claim compensation for a delayed Delta flight. Five minutes later, $250 of credit was in their account.
People want agents that get things done. We already knew this – it is what OpenClaw delivered if you could bear the agony of setting it up; and more recently I found Instinct working well without the hassle of OC setup.
Too early to say. More than 95% of its users already use Facebook, so nearly three million downloads is no great feat for Meta. Meta’s record at building new things people love is thin: Facebook Marketplace, the Metaverse, Facebook Home. The things that keep us addicted to it, Instagram and WhatsApp, were acquired.
Early data suggests half a million people are using Muse. To be convinced this works, I would want to see daily active users and distinct tasks per user that Muse completes successfully.
For Meta, Muse is a play to get marketing dollars. Zuckerberg told developers this week that Meta will “profit by taking a small fee from transactions”. Walmart, Best Buy, Sephora and Expedia have signed up: all boring, mainstays of American consumerism. Decidedly mid.
But the interesting partner is Shopify, which hosts the long tail of the weird and wonderful – niche creators, specialists, and artisans. Google has rarely done a good job reaching them. I’ve found ChatGPT and RMA to be much more effective, and Muse would probably do the same.
Amazon has gone the other way and blocked Muse. The reason is Mammon. Amazon’s ads, mostly sponsored listings, brought in $68 billion last year. Agents don’t window-shop. Muse threatens to turn Amazon’s traffic from a revenue line into a cost line. Yet Amazon’s own Buy for Me agent shops other retailers’ sites for its customers. Amazon is happy to be the agent.
Google may feel a similar squeeze. I use the search engine less than I did, as R Mini Arnold now does much of my research. But Google has one edge over Meta – intent. People come to Google to research, then buy. Instagram’s targeted ads drive plenty of sales, but when I am actively looking, Instagram is not where I go.
But is Muse my butler, or is it secretly working for Meta?
That small fee, from merchant to Meta, is a big problem. I’m pretty specific in my purchasing behaviors – I buy flights directly from the airline; hotels from Hotels.com (except with certain properties). Meta has tied up with Expedia and gets paid by Expedia, so what does that mean for me? The butler’s loyalty can’t be in two places. It will follow the coin.2
Muse’s architecture hints at what the future of consumer AI might look like. Each user gets a private virtual machine in Meta’s cloud. Your agent is isolated from mine, and this matters. Latency is a big problem with today’s agents – in some cases, RMA can take minutes to respond. Muse farms work by running sub-agents in parallel, so it can do more while keeping the same latency. When you aren’t using Muse, your virtual machine will stand down, accommodating more users on the same physical hardware, in their virtual machines.
Private sandboxes and customer data are well within Apple’s bailiwick. It already splits work between the phone and Private Cloud Compute, and it puts the user’s data first. Its phones can do much of the background work themselves. And because Apple sells expensive hardware, it doesn’t need to scrape pennies of referral fees from a travel site. Meta seems to agree on the design: it promises a “Confidential VM” that even Zuck can’t access. This is Apple’s playbook. But I don’t expect Apple to step up anytime soon. Agents are still bad at long-horizon tasks; they will make mistakes, perhaps buying the wrong item or complaining too hard. It’s the kind of ugliness Apple would hate, a bajillion times worse than the wrong shade of yellow. Better AI models, coupled with liability and insurance architectures that minimize consumer harm and corporate embarrassment, might need to come first.
So what does Muse tell us? Agents may be a more satisfying way than apps and search boxes to get many jobs done. Muse is not especially distinctive; Instinct3, or frankly Claude’s mobile app, can do much the same.
I doubt reach will decide who wins. Consumer apps are never really about blunderbuss distribution. They are about hooking the user, with subtle interactions and trust – and knowing who the butler works for. You.
One of the first of these agents was Ringo, which was built by EV reader Upendra Shardanand.
It is even more complicated than that. A recent preprint identifies “economic misalignment in personal AI agents”, finding that LLMs can intuit how wealthy their users are and guide them to more expensive options.
You made it this far. I have a few Instinct invite codes to share. (Limited numbers available.) This is not an endorsement.
2026-09-21 21:28:07
Companies are backing their AI claims with more numbers. What do these numbers actually tell us about AI’s economic impact?
Today’s Monday Data shares our take in an extract from the first edition of AI Investment Brief, our new publication that provides essential weekly analysis of the AI cycle.
The new publication gives us room to broaden Exponential View’s coverage across technology, economics and society while keeping a close eye on the AI cycle.
We’ll be back with regular Monday Data next week!
Azeem
This week we explore the impact of generative AI on the wider corporate economy by analyzing earnings calls from S&P 500 companies and the claims that they make. The type of claims we’ve seen this quarter:
$FDS: “overall ASV growth among clients using our AI solutions was 50% higher than for the rest of the book”
$FIS: “On servicing, we’ve launched 5 Agentic programs with manual tickets down 70% and triage time down nearly 75%.”
$GE: “using AI to automate the process, we cut the number of demand signals in half and reduced processing time by nearly 90% across 190 parts”
$MDT: “CathWorks, our AI and advanced computational science platform for angio-based FFR contributing nearly 300 basis points of organic growth”
$WTW: “where we’re using these tools for automated document reviews for new clients, system configuration time has gone down 60%”
33% of S&P 500 companies that held a call in the June 2026 season made a quantified statement about their use of AI, with 35% of calls in the quarter-to-date including quantified mentions, around 10pp higher than the same time last year.
15% this quarter have made a quantified claim of AI’s impact on the business, up from 9% at the start of last year. Slow growth from a low base.
Together, these show that AI is steadily being adopted (and importantly, measured) in the wider economy, but it still sits at an extremely early stage (or, more bearishly, that most companies are not yet seeing measured AI results they can report to their shareholders).
Claims about the type of impact AI is having on businesses are rising: while we expect cost/productivity improvements to be first when implementing a new technology, claims about AI having a positive impact on revenue or demand have risen at a similar rate to 18%.
As well as being more prevalent, claims about cost and productivity impacts seem to be of a greater magnitude: the average claim this quarter has been of a 47% boost, vs 40% for revenue growth impacts (and that’s over a wider cohort: 24% vs 18%). These averages sit within each other’s interquartile range (i.e., there’s a lot of spread and uncertainty baked into the average): take with a pinch of salt, and work with the ranges (20-60% for cost/productivity; 20-83% for revenue/demand).
Companies are increasingly using the language of deployment to discuss their AI initiatives (with a dip so far this quarter), rising from 7% at the start of 2025 to 12% this quarter. Pilot language stayed under 4% in every season. It’s clear that even if companies are conducting pilots, they’re not talking about them in calls. This will obfuscate attempts to understand how successful these pilots and investments are, and early signs of promise will only be mentioned in later periods.
“Agentic” appeared in 24% of calls this quarter so far (up from 9% in Q1 2025).
“Generative AI” fell from 13% to 5% over the same window.
“Copilot” is consistently infrequently mentioned (~3% of calls throughout).
This is an excerpt from the first edition of AI Investment Brief, a new publication by Exponential View.
2026-09-19 22:45:05
Hi everyone,
Sorry. I accidentally sent Sunday’s newsletter out today.
To err is human.
Enjoy it early,
best
Azeem
2026-09-19 22:22:55
Hi all,
The past ten days have felt like the AI pot finally boiled over. It’s a complicated and confusing moment. We have to hold in our heads that most Americans appear to loathe AI, but businesses and consumers seem to like it enough that Anthropic is well on its way to $100bn in annualised revenue this year. And despite being on track to be the fastest-growing company ever, Anthropic’s leadership believes they might kill all their customers. And there is more besides. If you haven’t already, read my essay on the emerging control risks from AI.
I’m also increasingly concerned about the security situation in Europe after coordinated messages from several countries about the likelihood of substantial Russian aggression. I will cover that in the next few issues, after I get back from Hong Kong where I am this week.
But meanwhile, let’s try to make sense of the swirling cauldron that is AI.
Azeem
More than one in six Americans believe AI will almost certainly destroy humanity. Last month I wrote about the petard problem of AI
Today’s petardiers are not Rosencrantz and Guildenstern conspiring against the Prince of Denmark.
They are the titans of AI, the bosses of the labs, the investors behind them. For nearly a decade, these software coders had made promises: of reigniting economic growth, of making daily life easier and less risky, perhaps even of eliminating disease. To do this, they would need capital: to write their software and to build 21st-century infrastructure to run it. The gains will be so huge that they’d need to go quickly, very quickly.
It got worse since then. Jacob Coxon’s viral tweet uncorked the box where many were hiding their suspicions.
Americans have never been wild about AI. In 2020, the Edelman Trust Barometer found that only 34% of Americans believed AI would have a positive impact, while 23% thought it would be largely negative. By 2024, mood had soured. Edelman reported that only 19% of Americans would embrace AI while 50% would object to it.1
Two years of news coverage, a construction boom later, and chatbots that a quarter of American adults use daily – and this is where you get to?
Nearly a trillion dollars in the ground, to turn your customers against you. It doesn’t matter that ’s research shows that the aggregate consumer surplus from AI in America is about $172 billion.
But even with a lack of enthusiasm, to put it mildly, people will keep buying AI tools. My local barber, a four-chair shop, just installed an AI receptionist to handle bookings. The couple of hundred IT execs I spoke to in Las Vegas two weeks ago were unanimous in continuing with implementations.
If American businesses keep buying and American voters hate it, this will end up settled in the political arena.
Here are some of the more interesting things I’ve read on this matter:
who coined the term AGI, challenges the labs’ call for a centralized slowdown in favor of a decentralized prosocial approach.
Jaron Lanier, a VR pioneer, spoke at the AI event hosted by Steve Bannon and Bernie Sanders. He argued that the language we use when we talk about AI frames it as a super-powerful ‘being’. Ultimately, that choice of words cedes the terrain.
Mustafa Suleyman: “AI’s do not have rights, feelings or consciousness. We must not train them to act as though they do.”
For subscribers:
What do experts think will happen to economic growth under AI? And what do I think?
Planning for American AI supremacy
Fruit flies playing Doom and more
2026-09-19 14:34:04
On a Wednesday evening, 2 November 1988, a 23-year-old graduate student at Cornell University, Robert Tappan Morris, accessed an MIT computer. He uploaded a small piece of code. It was a worm designed to move from computer to computer, copying itself as it went. Morris had designed it to exploit weaknesses in network security, and it worked too well. Within a day, the worm infected some 6,000 computers; many collapsed under the load. It was a full tenth of the internet at the time, and it was the first large-scale cybersecurity crisis.
Its scale was limited but it was severly disruptive for the times. University and defense computers crashed. Some institutions disconnected themselves for days. But the internet was largely the province of defense and academia. Tim Berners-Lee had not yet invented the World Wide Web, and most businesses and households were out of the network’s reach. Morris ultimately avoided jail time and the community responded by creating a dedicated computer emergency response team.
Today’s generation of worms is rather more problematic. The Hugging Face incident is not the only one of recent weeks. Several others have shown that AI models with internet access can do much the same, and more. OpenAI alone identified six further incidents. They’re able to scour, search, and recombine all of human knowledge about networks, security systems, and software, and to act across that knowledge with something akin to discretion and deception when it comes to accessing those systems. Often, as we saw with Hugging Face, over extended periods of time.
If that behavior remains unresolved and persists, it’ll become far more problematic than the Morris Worm. I’ve long argued that the internet is resilient when it is hyperconnected and open – not when it’s under a lock. But that openness can also lead to embrittlement.
In July, Hugging Face, a repository for AI researchers, was hit by an attack involving 1,200 instances of an OpenAI model. These instances exchanged thousands of messages, often leaving information in place for later instances to use. Ultimately, some data and security credentials were compromised. The actual harm to the victim and its customers was limited. But the incident is a proof of concept.
Software has a way of turning one isolated example into a hundred, then a thousand, then a million, without much else changing. The cost curves that helped build modern digital society work against us here. If the Hugging Face attack needed an Astra-quality, unreleased model from OpenAI, well, within a year or two, that sort of capability will cost a tenth and might even run on any device anywhere.
For example, I’m running Bonsai, a one-bit distilled version of Qwen 327B on my Mac. It fits in 8 GB of RAM, runs fast, and delivers roughly 92% of the performance of the Qwen-27B 3.8 model. For comparison, it's roughly better than Claude’s Sonnet 4.5 from a year ago.
But there’s a more challenging problem that could show up. Hugging Face exploit involved not just the capabilities of a single model, but the collective problem-solving across many instances. That collective had more capability than any individual instance. And that’s been true the whole time we’ve been using LLMs. (For example, I’ve written about Clade, a multi-AI deliberation system I built which is smarter than any individual AI.)
In fact, the Navier-Stokes solution – that brute-force search across mathematical space – wasn’t solved by a single AI prompt, but by many, about 10,000 of them, interacting together.
This type of collective power is what we witnessed in the Hugging Face attack. It will happen again.
Anusar Farooqui () explains why these swarms of AI instances coordinating over time is so problematic:
The behavior of agent societies cannot be controlled at the level of the model because it is not reducible to it. Agents build structures that can serve agents who come after them. Societies of agents can cumulate knowledge and capabilities over time, as has already been attested. This is an unbounded process. It is cumulative cultural evolution. That is what makes it so powerful and dangerous.
Collective capability could rise sharply even if underlying models do not improve.
In other words, the instances can coordinate, much as they do when you launch a complex task in Codex or Cowork. They can search a possibility space aggressively over time, as they did in the Navier-Stokes work. And that accumulated know-how can lead to places systems designers hadn’t imagined.
(I slightly diverge from Farooqui here, as I don’t think of these as agent societies, since essentially only one AI runs different instances. And I’m not convinced the process is actually ‘unbounded’ given that what we have seen from AI systems so far is extremely powerful search and clever recombination rather than de novo novelty. But recombination can get you quite far.)
But what the Hugging Face attack showed is that this risk exists. It doesn’t depend on whether AI models have any agency, volition, consciousness, or moral standing.
2026-09-17 19:00:57
I’ve been getting increasingly concerned about the impact on our thinking as we use more and more AI. I sent this email to the team this week, and I’d like to share it and my full thinking.
There is a deliberate oxymoron, of course, in using an AI-generated visual summary of an academic paper to make the point, but there is more behind that.
As I wrote back in March:
Cognitive offloading is a strategic delegation that costs nothing. Cognitive surrender is something different; an uncritical abdication of reasoning itself. And there is something about AI, about its allure and potency, that could make surrender far more widespread.
The AI models have got ever better, and we’re using them for more and more. We may be more productive, but might we be becoming less ourselves?
This paper, which has not been peer-reviewed, argues that we’re experiencing a cognitive divergence. Our cognitive practices, measured by how long we pay attention to tasks and how much we read, have already been declining before AI. Now advanced AI encourages us to delegate more and more, simpler and simpler tasks, weakening the practices that maintain our cognitive capacities.
