It shouldn’t come as a surprise that Amazon collects data about you based on what you buy. If you buy a lot of cat litter, Amazon will probably infer that you have cats, or if you buy a lot of anti-acne skin products, then Amazon might guess you’re prone to breakouts. Even if we know that Amazon customizes our shopping experiences based on our purchase history, it still comes as a surprise to see all of these personal details laid out in a convenient list. Especially when those details get uncomfortably intimate.

One Threads user came across this section of her Amazon settings, where you can see who the shopping giant thinks you are, and alongside more benign assumptions like “Shops from women’s departments” and “Probably owns a Shark robot vacuum,” she found that Amazon assumes she “has flat buttocks.”

“I stumbled upon a page of assumptions that Amazon has made about me based on my purchases and I’m literally speechless,” an Amazon customer, @fangirlinmegan, wrote on Threads. “I mean it ain’t wrong but damn did you have to call me out like that?”

In less than a day, the post garnered over a million views, including many from users who were curious what their own Amazon profiles might reveal. Here’s how you find out.

On desktop, you can find this page by hovering over where Amazon says “Hello, [your name]” in the upper right corner. Then click “Account” under the “Your Account” section. Under the “Ordering and shopping preferences” section, click “Your Shopping preferences.” From here, you can add in information of your own, like your shoe size, interests, and dietary preferences. But we’re not here to tell Amazon that it should stop selling you beef jerky since you’re a vegetarian. We’re here to see if Amazon thinks you have “flat buttocks.” When you scroll to the bottom of the page, you’ll see a blue hyperlinked option that says “Manage your information.” Once you click that, you’ll see what dirt Amazon has on you.

On the mobile app, you can get here by clicking the hamburger icon from the bottom menu bar, then navigating to “Account” > “Shopping preferences” > “About you.”

For me, Amazon has gleaned that I am someone who “practices photography,” “plays collectible card games,” and is “invested in the Apple ecosystem.” This is all true. Amazon has even flattered me by noting that I “read diverse non-fiction,” to which I say — why yes, I am indeed an open-minded intellectual with a thirst for knowledge. Thank you for noticing.

Other colleagues of mine found that Amazon correctly noted their interests in vinyl records and ceramics, as well as a penchant for “natural materials” in home decor and clothing that “prioritizes comfort.” None of us had anything as affronting as “has flat buttocks,” but the Threads user noted that this was likely because she had purchased “butt scrunch leggings.”

Tech companies know too much about us, and it’s jarring to realize just how closely they’re paying attention to how we use their platforms. We live in a world of alarming surveillance, but at least I know that Amazon doesn’t think my butt is flat.

In Brief

Apple is working on a new ecosystem of smart home devices via a partnership with LG, Bloomberg has reported.

The new devices — which include a smart lock, a thermostat, and a doorbell — are designed to integrate with the company’s smart home hub, another upcoming product that will reportedly launch on Oct. 13th. That hub is expected to include a 6-inch square display that can be placed on a countertop or mounted on a wall.

The two companies are also working on a slate of security cameras, including models for indoor, outdoor and a floodlight. The LG devices are expected to be announced, if not formally released, next week, too. TechCrunch reached out to Apple for more information.

Interestingly, the devices will reportedly carry the LG brand — despite being developed by both companies.

The new products could aptly be read as a swipe at Amazon’s smart home offerings, which currently offers a variety of similar products, including its Ring home security cameras.

Apple has sought to push into new territory with its recent hardware releases. One of those new territories is an expanded smart home presence.

Melius, an AI platform for generating ad campaigns, images, and videos, announced on Tuesday that it raised a total of $25 million in funding, including a $20 million Series A led by CRV and a $5 million seed round led by General Catalyst.

While the startup claims to have hit more than $1 million in annualized revenue within two months of coming out of stealth in July, co-founder Joowon Kim (pictured on right) admits that Melius didn’t hit it out of the park initially.

When the New York-based company first launched over a year ago, Kim and his co-founders, Young Kim (pictured on left) and Arnav Ramu (pictured center) set out to build an AI-powered performance marketing tool. The three knew each other from working as engineers at Ramp, the corporate spending and finance software company.

More than six months after building the product, the co-founders decided their original idea didn’t “have legs,” Joowon Kim told TechCrunch.

Instead of helping marketers manage and optimize ad spend, the team chose to focus on what it saw as a much larger opportunity: building the tools that generate the creative assets and campaigns themselves.

“We scrapped the entire codebase; we burned all of it,” Kim said. Melius quickly went to work on an entirely new product. Nearly a year after its initial launch, it revealed a platform that it describes as an “agents lab for creative work.”

Melius is far from the only startup helping ad agencies, marketers, and brands generate creative assets and campaigns with AI. Competitors include rapidly growing Higgsfield, which was valued at $5.4 billion in August, as well as other small startups, including Krea and Flora AI.

Kim isn’t concerned about the competition.

He acknowledged that Higgsfield, a three-year-old that has hit over $700 million in annualized revenue, “is growing like mad,” but according to Kim, the size of the market is so large that it can support multiple competitors.

“There are a lot of players, which is pretty exciting,” he told TechCrunch. “That means there are customers to be won and there is demand in this space.”

As for why former Ramp engineers are building an ad generation product, Kim, who describes himself as a social media influencer, says he has been passionate about making short videos since elementary school and once dreamed of becoming a famous YouTuber.

“I tried my best to make vlogs,” he said. “My dad was great at using FinalCut Pro, but I wasn’t.”

Those early vlogs never quite took off, but his fascination with making digital media remained.

Now, with Melius, he’s building the tool he always wanted: a platform where anyone, including seasoned creative directors, can turn their ideas into reality using plain language.

When self-taught coder Sigil Wen was 17, he moved to Silicon Valley and lived in an AI hacker house with famed AI researcher Andrej Karpathy.

While there, he hacked and coded alongside other people who would become the biggest names in AI, like Perplexity founder Aravind Srinivas and OpenAI researcher Noam Brown. He tested early versions of AI tools that would later become well known, including a chatbot shared by Anthropic co-founder Ben Mann that would become Claude, an image generator from David Holz that would become Midjourney, and what would become OpenAI’s GPT-3 and the image generator Stable Diffusion. Prominent investor and entrepreneur Naval Ravikant hired him for Airchat, Ravikant’s now-defunct rival to the Clubhouse social network.

For fun, he figured out how to get GPT-2 running on his Apple Watch. “It was a magical time,” Wen told TechCrunch.

Now a Thiel Fellow — the program from investor Peter Thiel that invites young founders to pursue projects instead of college — Wen on Monday launched an invite-only beta of Underdog, one of the most private AI assistants Silicon Valley has yet to offer. The model runs wholly on-device, meaning the user’s data remains on devices they already own, currently Macs and Windows PCs, with Linux, iPhone, and Android versions coming soon.

Underdog is powered by Husky, the inference engine Wen built to run AI models quickly on a user’s own hardware. Wen says it moves less data between a computer’s main chip and its graphics chip than rival on-device engines.

Underdog has other security features baked in, too, like encrypting the keys to the email and other accounts that users authorize Underdog to access.

Underdog is, however, using much smaller models than today’s state-of-the-art ones hosted in data centers. It currently uses a 27-billion parameter reasoning model fine-tuned from Qwen3.8-27B.

Wen argues that this model compares favorably with Claude Opus 4.6 in some benchmarks, or what was considered top performance six months ago. He says that means it can handle the everyday tasks that people want an AI assistant to do, like shopping research or answering math homework questions.

“You don’t need to sacrifice your privacy for the capability because they’re just as capable,” Wen says. He adds that small on-device models will continue to grow more capable over time.

Perhaps the most interesting thing about Underdog is its early business model. The app will be free at first and never ad-supported. Since the AI runs on users’ machines, Underdog doesn’t have the giant overhead of paying a provider for inference. “I don’t have to charge you a subscription to run this because my costs are so super low,” he said.

Instead, with Stripe co-founder Patrick Collison as one of his angel investors, he’s borrowing a play from the fintech era. Underdog will take a tiny percentage of payment transactions that the AI assistant makes using Stripe’s secure payment rails, something like an interchange fee. In this way, the AI assistant never has to mine your data. It is as aligned with you as your bank or credit card providers.

This runs contrary to the business motivations of many of the other players in AI assistants, whose privacy policies allow them to collect data on users that they may sell to advertisers or other third parties, or use to train other models.

That kind of data collection could be a particularly treacherous trade-off for users of an AI assistant, which may need access to the most intimate details about you in order to be useful, from medical conditions to financial data to data on your kids.

As Wen wrote in what he calls his AI manifesto, “Why should using AI require surrendering your private information?”

He tells TechCrunch: “I honestly want to build Underdog for myself. I’m building a product that I would be proud for my future children to use.”

The startup behind Underdog is named Conway Research, and Collison isn’t its only big-name investor. Conway is backed by Andreessen Horowitz via partner Chris Dixon, as well as Khosla Ventures, Hummingbird, SV Angel, the Anthology Fund (the partnership fund between Menlo Ventures and Anthropic), and a prominent list of angel investors that includes Vercel founder Guillermo Rauch, Noam Brown, and Deedy Das, among others.

As decision models spread across the industry, a company called Musubi has a new idea for how to put them to work: moderating content. On Tuesday, Musubi announced a lightweight decision model made for real-time moderation called PolicyLM-1.7B, released with open weights.

The idea is to take a content policy written in plain English and apply it to messages in under 50 milliseconds. Musubi’s model is designed to be similar in cost and speed to the AI classifier systems that power moderation on most social platforms — but because it has the flexibility of a modern LLM, it can apply complex policies without special training. Even more important, the model won’t need new training when the policy changes, allowing for human policy-setters to iterate as much as they need.

As Musubi co-founder and chief AI officer Filip Jankovic sees it, it gives platform managers a way to label content proactively.

“Product teams just want a better understanding of what’s happening on their platform, especially as the amount of content is exponentially increasing,” Jankovic says. “Being able to label all of that in a very scalable, customizable way is extremely useful.”

Decision models have become a hot topic in the AI world since the release of TypeSafe AI’s Jev in September, which was shortly followed by competing decision models from OpenAI and Amazon. Instead of outputting text, a decision model outputs outcome probabilities, though in this case the model outputs a binary judgement: Either the content is in the category or it isn’t. By limiting the model’s output to a set of predetermined choices, decision models are able to run faster and cheaper than large language models, while still maintaining the flexibility of the transformer architecture.

One early use case is reining in misbehavior by AI agents — so it’s only natural to apply the same technology to human misbehavior.

Notably, Jankovic says his interest in decision models predates Jev, tracing it back to a 2024 project called GLiNER (Generalist Model for Named Entity Recognition) that deployed many of the same techniques.

Still, Musubi isn’t wary of the comparison. If anything, the company is eager to use the new interest in decision models to shine a light on content moderation. “If Jev caught your eye, PolicyLM-1.7B is the same kind of model, trained specifically for content moderation, that you can run yourself,” the product announcement reads.

In Brief

Cloud provider Lambda is raising up to $4 billion at a $14.5 billion pre-money valuation, marking what could be its last private round before a planned 2027 IPO, according to The Wall Street Journal. Coatue Management and Blackstone are leading the round.

A letter to investors reviewed by the Journal shows Lambda’s backlog grew from $15 billion in June to $50 billion in September. While that might look like a hearty increase in demand, much of that increase appears to be driven by a $35 billion commitment from one company: Anthropic, which signed a deal with Lambda in late August.

That means Lambda’s valuation, which has climbed significantly since its 2025 funding round, could be leaning heavily on Anthropic’s ability to keep paying. Still, with reliable GPU capacity so scarce, investors are clearly still willing to bet on companies that provide it, especially ones with large contracts with a major AI lab.

For neoclouds like Lambda, demand isn’t so much the problem as is the cost of meeting it. Data center buildouts are largely funded by debt — of which Lambda just raised an additional $1 billion last week — and lenders are getting choosier about who they offer cash to and under what circumstances. Lambda’s decision to raise more now not only sets the tone for its IPO pricing, but also gives it access to more capital before the scrutiny of public markets arrives.

If and when Lambda does IPO — the company was reportedly meant to debut this year, but has pushed that back amid market uncertainty — it will join other Nvidia-backed neoclouds, like CoreWeave and Nebius, that now depend on the health of their stock to fund their data center buildouts. British neocloud Nscale filed for an IPO last month and is expected to begin trading soon.

Lambda, Coatue, and Blackstone did not immediately respond to a request for comment.

Personal AI agents, like Meta’s Muse, Instinct, ChatGPT’s Dots, and others, are kicking off a new wave of consumer AI that involves more than just responding to queries. These AI agents can actually get things done on users’ behalf, like booking flights, making dinner reservations, ordering groceries, and more, all without having to be technical enough to set up OpenClaw or some other agent on your own computer.

But consumers adopting these agents are often running into issues when the website on the other end of their request blocks the AI from completing the job.

Most recently, Amazon began blocking Meta’s Muse AI agent from its retail site, meaning that the agent could no longer browse or make purchases from its product catalog.

However, the problem is not limited to Amazon or Muse. User complaints across social media indicate their agents are often running into similar blocks. Some of these are intentional, like Amazon’s, while others are the result of websites’ traditional anti-bot measures designed to protect against spam and malicious activity.

What’s not clear to the end user is whether or not the block is accidental on the website owner’s part, which leaves them frustrated with both the service provider or retailer and the agent trying to facilitate the transaction.

For instance, TechCrunch has seen a handful of complaints on social media claiming that Muse couldn’t complete a purchase on Walmart’s retail website. (NBC also reported that this was a problem Walmart had faced in its tests of the AI agent.)

But a spokesperson for Walmart told us these weren’t intentional. In fact, Walmart is partnered with Muse, as was announced at Meta’s Connect developer conference in September. The company said it wants to be where its customers are, and that now includes being available through AI agent experiences.

The issue that Walmart customers often appear to be running into (see below) is that the website presents a button that has to be clicked to verify the visitor is human. If that experience is interrupted, the check could fail, and the agent gets booted out.

It’s an indication that the current technology we use to verify humans and keep websites safe from spam and bots may not be the best fit for the agent-first era of the web we’re now moving into.

To address this problem, industry partners including Meta, Walmart, Stripe, Sierra, Genesys, Rocket, NiCE, and Decagon have begun working on an open standard that would dictate how AI agents can communicate with businesses, helping to separate the good bots acting on behalf of the user and the bad. This protocol will be specifically focused on agent-to-agent communication for the purposes of online commerce, Meta said in a blog post.

This is frustrating. @Walmart in their attempt to stop agent from making purchases has gone too far and is essentially blocking actual human from making purchase on their website. I tried to purchase something directly and I can't even sign in. pic.twitter.com/CRAgkrCCa0 — Kevin Wang (@buzaza) October 5, 2026

In the meantime, consumer confusion abounds.

Although Walmart is at least trying to be accessible via AI agents, other companies aren’t so keen on the idea or haven’t fully embraced the technology.

For example, flight booking is currently one of the main use cases being positioned for agents, but many people have reported issues with airlines rejecting their agents’ requests.

Delta, for instance, told us that the company is taking steps to protect customers from unauthorized automated activity, and that if it were to open up to AI agents, it would have to be done with “security and the customer experience in mind.”

“While Delta does not currently have a partnership or integration that enables a third-party AI agent to shop for or book Delta flights on a customer’s behalf on our digital platforms, we’re continuing to evaluate how new technology, including external AI agents, could enhance the way customers engage with Delta,” said Delta’s general manager of Global Communications, Heena Chavda, in response to our questions about whether or not it was blocking agents, as some users complained.

United was not quite as direct as Delta, but a spokesperson pointed us to the part of its Terms of Use policy, which says that the user agrees “not to use any robot, spider, other automatic device, or manual process to monitor or copy this website or the content contained therein or for any other unauthorized purpose without United’s prior written permission.” (United didn’t respond to a follow-up question about whether it was blocking specific personal AI agents, like Muse.)

Early adopters’ complaints across social media have referenced a number of brands that are reportedly blocking AI agents, including Yelp, eBay, Zillow, Pizza Hut, Adidas, and various airlines. Some believe that the blocking activity seems to have recently increased.

A couple of people even said eBay suspended their accounts because of their use of agentic AI. Another complained that Google kicked out Instinct while it was cleaning up their Gmail inbox for them.

Looks like @adidas won't get my business tonight 😪 pic.twitter.com/DXcPYE5S6O — Martin Kessler (@kesslerIO) September 28, 2026

Despite the push to have agents book tables at restaurants and other tedious tasks, Yelp told us that it doesn’t permit non-human traffic on its platform unless the agent has paid to access Yelp’s content and data through its data licensing program. That means agents trying to use its site to get a quote, add a user to a waitlist, or book them a table will likely fail if they don’t have a formal partnership with Yelp.

eBay told TechCrunch that its approach is not to prohibit all third-party shopping agents from completing purchases. Instead, its policies restrict unauthorized agents and actions, like automated scraping and model training.

Adidas, Zillow, and Pizza Hut either didn’t respond or declined to provide a comment by the time of publication.

One major company involved in this space is Cloudflare, which offers sites protection from AI agents that could be training on its data and managing other automated activity. Some suspect that Cloudflare and other content delivery networks (CDNs) like it are currently disrupting Muse traffic. This is attributed to a change to crawler defaults on September 15, which allows site owners to block AI training while allowing other types of bots. Existing sites that were previously set to block AI bots for security reasons were then shifted to block AI agents on any pages that had ads.

However, when we asked if Cloudflare had any insight into what was happening with the blocking of these personal AI agents, like Muse and others, the company said it didn’t have any specific data to share at this time. It pointed us instead to its free public data hub, Cloudflare Radar. The site shows the increase in bot activity (which Cloudflare has also commented on previously), but it doesn’t help differentiate between the good bots and the bad ones.

Cloudflare also has a vested interest in mitigating AI traffic to websites, as it now runs a marketplace where AI bots have to pay for the data they access. It recently began testing a new web browser built specifically for AI agents.

Given the rapidly evolving space, Meta’s decision to form specific brand partnerships for Muse access seems to be the right way to go, as is its decision to focus on new standards. In doing so, it assures its customers that its agent, at least, is welcome at these partners’ sites — so any issues with access would be reportable bugs.

Meta also maintains a list of “connectors” (a.k.a. partners) inside the Muse app, which continues to grow. Many of Meta’s newly announced partners haven’t yet shown up in this list. But given Muse’s rapid expansion, which now includes partnered tools for small businesses, a bigger reorganization of Muse’s connector list is likely soon to be underway.

“Turning away a personal agent means turning away the customer behind it. We’d rather work with businesses to address the underlying concerns than see them lose that customer,” Meta said in its announcement about the newly announced plans to develop a standard. “There is a path that is good for users and good for businesses. It starts with clear norms, gives businesses more predictable control, and evolves toward a more efficient way for agents to work together.”

From Muse to Instinct to Dots, tech firms are hunting for the user experience that will make an AI personal assistant a must-have product.

Hark, a startup founded less than a year ago by serial founder Brett Adcock, is throwing its hat in the ring today with the wide release of Hark Pro. Users can access the service for free, with a subscription tier for heavy users.

The company sees its mission as building a user interface for AI, not creating AGI or curing cancer. The foundation is a model trained specifically for computer use, and that’s what Hark Pro does — it uses your computer for you.

That’s perhaps a less ambitious claim than some frontier labs are making, but a little humility goes a long way in winning over users. And rather than an app, Hark’s goal is to create an operating system for the AI computers of our future.

Hark Pro is a full-screen experience with a kind of homepage, featuring a central chat dialog and a feed populated by prompts for actions you might need it to take and visualizations the company calls “panels” that offer mini-dashboards for things you’re interested in. The background is a piece of custom imagery tailored to the weather and time where you are.

Like other AI assistants, users configure it to access their digital life — email, calendar, hard drive, credit cards, etc. Then, it can take on digital tasks as well as fielding classic chatbot questions.

Abidur Chowdhury, a former Apple staffer who leads design at Hark, demoed the product for me. It proactively asked him about expenses he needed to approve, emails he needed to respond to, a meeting with this reporter, and, of course, pointed out that he needed to book a flight for a trip and offered to take care of it. (“I know the flight thing is put to death,” he admits, grinning.)

That does get at the challenge with these apps: One of the harder things for users is actually coming up with useful tasks they are capable of accomplishing. Chowdhury says he was impressed when the AI assistant recently recognized that one of his colleagues needed to renew their car registration at the California DMV, and handled it through the agency’s website.

One notable UX choice that Hark makes is showing the user how the agent navigates the web in a small window, an action most AI assistants perform without it surfacing in the app, which Chowdhury says is intended to build trust with users that the agent is doing the right thing.

That matters because part of Hark’s pitch is that they are not Meta or OpenAI — “we’re not here to like sell you ads and steal your data,” Chowdhury says. To truly leverage these models, users will need to share their entire digital lives. Beyond very real cybersecurity concerns, will users be comfortable with a company like Meta having its products create detailed profiles of their friends and family?

Asked about that kind of profiling, Chowdhury says that “we are doing work to understand you, and be helpful for what you need, but it’s really like task-related — it’s really about how Hark can help take something off your plate and be useful.”

What sets Hark Pro apart from its competitors, Chowdhury says, is its effort to fully subsitute for your computer, but also his belief that it just works better.

“Please compare it to the others, be critical,” he says. “It might not be faster than Muse, but it will give you a better result, and it will work end-to-end instead of asking you to take over at a point.”

One open question is what kind of model best suits the AI personal assistant. Hark’s is explicitly trained for computer use, which makes it faster and cheaper at those tasks, but it’s not clear how much it is trading away the broader knowledge and capability of a frontier LLM.

This software is only half the promise of Hark; Chowdhury made his name designing devices at Apple and Hark promises to release an AI-native device to support its software in 2027.

They’re still cagey about the form it will take, with Chowdhury bemoaning the “creep factor” of wearable cameras. “I will give you this: The first product that we’re building is a product that’s never ever existed before,” he says. We’ll have to wait and see if this is mere hype and Hark puts out some retread tech instead of a truly novel product.

“It’s so hilarious to me that you have AI now, you have a chatbot, and you type stuff in with a keyboard,” Chowdhury says. “Fifty years later and the future of computing is still a terminal.”

JioHotstar, the streaming platform with more than 500 million monthly active users backed by Indian conglomerate Reliance Industries, is expanding into the Middle East and North Africa (MENA) through Starzplay, one of the region’s largest streaming services. The new overseas push comes after its debut in the U.K., Canada, and Singapore.

Under the partnership announced Tuesday, JioHotstar will have a dedicated section in Starzplay’s app and on its website, offering Indian films, TV shows, originals, and reality franchises across multiple languages. This is different from JioHotstar’s September launch in the U.K., Canada, and Singapore, where it replaced Hotstar and debuted as a stand-alone streaming service.

The expansion is part of Reliance’s push to build JioHotstar into a global streaming platform after establishing a massive audience in India. MENA has one of the world’s largest South Asian diaspora communities, giving JioHotstar a sizable potential audience for its Indian films, television shows, and original programming.

As with its earlier debut in the U.K., Canada, and Singapore, JioHotstar’s own sports offering in India will not be available in MENA. However, a plan bundling JioHotstar with Starzplay’s sports offering, including cricket, is priced at AED 49.99 (about $13.60) a month. JioHotstar is also available through Starzplay’s Plus monthly plan at AED 34.99 (around $9.50).

With MENA’s large Indian diaspora in focus, new JioHotstar originals will premiere in the region simultaneously with their release in India, with audio available in up to 10 languages. The catalog spans Hindi, Malayalam, Tamil, Telugu, and Bangla, among other languages, and includes shows such as “Bigg Boss” and “Koffee with Karan.” Additionally, JioHotstar is putting particular emphasis on Malayalam programming, given the large Malayali population in the region.

JioHotstar was created after Reliance and Disney merged their Indian media businesses into an $8.5 billion joint venture in 2024. The service, which combined JioCinema and Disney+ Hotstar, launched in India in February last year.

Before JioHotstar’s launch, Hotstar was available in MENA. Disney Star expanded its distribution deal with evision, e&’s media and entertainment arm, in 2024, making it the regional home of Hotstar Specials. Evision is also a major shareholder in Starzplay, where JioHotstar will now have its first dedicated streaming presence in the region.

Starzplay counts more than 2.3 million subscribers and has built a sizable South Asian audience in part through its cricket offering.

Financial terms of the partnership were not disclosed. Starzplay CEO Maaz Sheikh described bringing JioHotstar to the platform as a “strategic investment.”

Startups that promise to predict how humans will behave are having a moment. Over the past year, Simile raised $200 million at a $2 billion valuation; Aaru raised $88 million at a $1 billion valuation; and Humans&, an AI startup that announced a massive $480 million seed round in January at a $4.48 billion valuation, launched Persimmon to model human behavior.

The status quo for human behavior prediction today relies heavily on large language models (LLMs) that are prompted or fine-tuned to role-play as a target demographic. But two-year-old, San Francisco-based Mirror Particle thinks that approach is fundamentally broken.

“It’s like bringing a super soaker to Niagara Falls,” says Abhivyakti Ahuja, co-founder and CEO of Mirror Particle, which provides brands with an AI engine that predicts consumer behavior and the reasons behind it. “LLMs have been trained on hundreds of billions of data points. How much can you influence its behavior by [fine-tuning] with such a small amount of data? It’s still stuck in the past.”

Ahuja doesn’t think LLMs see the world the way a human does. “LLMs are modeling written language, but humans are made of visual perception, spatial reasoning, social intelligence.” Relying on them, she says, means getting insights based on what humans don’t notice, which is beside the point when trying to predict human behavior.

Mirror Particle is taking another approach: building a foundation model, or as Ahuja describes it, a world model built from scratch that simulates why humans do what they do and how human behavior changes over time.

“We don’t want to capture the static person,” Ahuja said. “We want to capture the changing person. That means capturing the longitudinal data on how people are changing, what triggers are changing them and to what degree.” If they aren’t changing, she added, “that’s also a signal.”

Mirror Particle has already raised an angel round and says it’s close to closing its first venture round. The company is also competing next week in Startup Battlefield 200, TechCrunch’s renowned startup competition taking place at TechCrunch Disrupt 2026 in San Francisco on October 13-15.

The startup relies on a proprietary combination of data that includes its clients’ customer data, current events, pop culture, social media, and more to model a demographic segment, thinking of it as a system that evolves over time and tracking how motivations shift as it moves through experiences. Much of the focus is on “revealed behavior” — what people actually do rather than self-reported survey answers.

Like its rivals, Mirror Particle’s initial go-to-market strategy focuses on where budgets already exist for these kinds of insights: market research and brand and product strategy. Mirror might, for instance, help a beauty brand not just write better ad copy for makeup that would appeal to Gen Z, but also determine if that demographic even wants that product.

“What if [the target demographic] doesn’t want eyeshadow palettes?” Ahuja said. “Maybe blush is a better option to go for if you want to sell a product to this market.”

Mirror Particle’s prediction engine also provides customers with the “why” behind current or future behavior — the motivations, constraints, and additional context that justify its recommendation, helping brands make smarter decisions.

In one early pilot, a well-known pet food brand wanted to know what imagery to put on the packaging to boost sales. Chicken? Beef? Vegetables? Mirror’s technology found that the brand was asking the wrong question. The imagery didn’t matter. The problem was that the brand was so recognizable that it was considered mass market and cheap, and sales would plateau until it addressed that perception issue.

“The way we see our model evolving is like how a baby learns about the world,” Ahuja said, noting that babies move from vision to language to body awareness to social intelligence.

That fundamental interest in modeling the human brain comes from Ahuja’s background studying neuroscience and computer science. Originally from India, she ended up studying at the University of Toronto, where she became inspired by AI pioneer Geoffrey Hinton’s contributions to neural networks.

After school, Ahuja ended up at Amazon Robotics building robots that build other robots. That’s where she met her co-founders, Will Song and Thomson Yen. Song has spent a chunk of his career building sales personalization engines, and Yen focused on using deep learning to learn about how AI agents understand human behavior.

The startup’s long-term vision is to be the “general layer for anticipating human behavior” and moving from broader population-level analyses to individual-level insights.

“We just need a better model of humans if we’re going to work alongside AI and with each other,” Ahuja said.

Check out Mirror Particle and many other innovative startups that have been vetted by TechCrunch’s editorial team next week at Disrupt in downtown San Francisco. The winner of this year’s Startup Battlefield will be decided by our slate of VC judges on the afternoon of Thursday, October 15.

Until recently, venture capitalists largely steered clear of defense startups. Investments in the sector were viewed as morally controversial, and the process of securing government contracts was too lengthy and uncertain to justify venture funding.

But that sentiment changed rapidly after conflicts in Ukraine and Iran began to deplete the U.S. military’s missile supply. Startups that prove they can build equipment fast and affordably now stand a real chance of landing a Pentagon contract on an expedited timeline.

Four years into the defense tech boom that arguably began when Russia invaded Ukraine, the storied Silicon Valley venture firm Benchmark is placing its first bet on a pure defense startup.

Benchmark is leading a $25 million seed round into Furientis, a startup on a mission to mass-produce interceptor missiles. Just one year old, the company—whose name means “of fury” in Latin—has already secured a funded Pentagon contract to build mid-range interceptors. The startup was valued at $125 million, according to a person familiar with the company.

Furientis’ seed round follows a $5 million pre-seed that was announced in May.

“What they accomplished with just a pre-seed round was nothing short of remarkable,” Benchmark general partner Chetan Puttagunta, told TechCrunch. “They manufactured prototypes with just $5 million. They had they had already successfully done [a dozen] launches by the time we invested.”

That efficiency is in stark contrast to the slow-moving, high-cost development models used by traditional defense primes. The startup’s velocity is driven by its co-founders, who bring years of aerospace and defense engineering experience. Brody Franzen served as deputy chief engineer at Virgin Galactic before joining Castelion, a hypersonic missile startup valued at $13 billion, while Aris Simsarian previously led rocket engine testing at Virgin Orbit.

The duo launched Furientis after realizing the U.S. military is facing a severe shortage of “exquisite” missiles, which can cost around $3 million each, to counter inexpensive adversary threats, Franzen told TechCrunch.

According to Franzen, the U.S. Navy receives a mere 300 to 500 interceptor missiles annually, compared to China, which claims a production rate of 3,000 anti-ship cruise missiles a month.

“We’re being outproduced by a factor of 100 plus,” he said. “That is why we started the company.”

Furientis is not aiming to match the performance of best-in-class missile interceptors; instead, it is designing and manufacturing its systems so they can be assembled quickly from readily available components. As a result, each unit costs a fraction of legacy interceptors, a compelling value proposition for a Pentagon desperate to scale inventory within strict budget limits.

Unlike traditional defense primes such as Raytheon and Lockheed Martin, which are bogged down by long design cycles, Furientis has been testing its missiles in the field on a bi-weekly basis. The startup’s ultimate goal is to manufacture 1,000 systems a year in each factory. But for now, the young company is working out of its facility in Los Angeles, and taking its prototypes for trial launches to White Sands, New Mexico.

Given the need, Furientis is not the only company working on helping the government restock its stockpiles. Franzen said that larger defense startups like Anduril, Castelion, and Shield AI are also racing to build mass-produced interceptors, while traditional primes scramble to modernize their own manufacturing cycles.

Benchmark’s Puttagunta is not fazed by the seemingly crowded space. “I think there’s opportunity for lots more companies in this sector,” he said.

There isn’t a single manual to read or prompt to give an LLM that can equip you with the skills and knowledge to build a company. But on November 4, TechCrunch Founder Summit gives founders the next-best thing: A packed day at Boston’s SoWa Power Station, where investors and operators share how they’ve navigated crucial decisions.



The Founder Summit’s agenda is built around core questions every founder eventually has to answer about capital, leadership, and what they’re actually building toward. And each session is designed to build upon each other. Like your startup’s founding team, these sessions work together to drive outcomes.

This is a breakdown on how the full event agenda fits together, and what you can expect to learn by attending. Keep in mind that tickets are at their lowest price point before they increase on October 16, so this is the ideal moment to grab your seat. You can save $200 on both Investor and Founder passes up until that deadline.

For a limited time, you can also get a second ticket for 50% off, and for even larger groups, there are bundle discounts available to groups of four or more.

Capital: How you raise, and who you raise from

Fundraising advice often stops at “How do I get the term sheet?” This track goes further, because getting a yes is only half of the equation. The other half is making sure it’s the right yes.

Underscore partner Brian Devaney opens with a candid read on today’s fundraising climate, including what investors are actually screening for right now, how founders can differentiate in a crowded market, and the specific moments where founders give up leverage without realizing it.

That’s the part most founders obsess over — as well as who ends up on the cap table once the round closes. This is where Vineet Edupuganti, co-founder and CEO of Cogent Security, picks up the thread. After raising an $11 million round, he learned firsthand that not every check is equal. He’ll walk through how to actually evaluate investors and navigate those trade-offs before you sign, not after.

This all frames fundraising as a two-part decision. First, win the leverage in the room, then spend it wisely on the right partners. Founders who only solve the first part often find themselves fighting their own cap table later.

Grab your pass and get all these fundraising insights. Bring your co-founder at 50% off.

Leadership: When the job changes, so does the team you need

No founder stays the same kind of leader forever. After all, the skills that get a company to its initial customers rarely match what’s needed when its customer list expands and the employee count grows. This track tackles that shift from two angles: the founder’s evolution, and the decisions on employees that either support or sabotage it.

HubSpot co-founder and Sequoia partner Brian Halligan has experienced this arc many times, first by building through HubSpot’s own stages of growth and now by advising founders doing the same from the investor’s seat. He’ll unpack how the CEO job transforms at each stage, the leadership habits that stop working once a company outgrows them, and how to keep reinventing yourself without losing what made you effective in the first place.

But a founder’s evolution doesn’t happen in isolation — it’s constrained by who’s actually in the building. NEA partner Melissa Taunton brings the hiring half of that equation, breaking down how early team decisions quietly set the ceiling on a company’s culture, speed, and ability to execute long before anyone notices the cracks. The founders who scale smoothly, in her experience, are the ones who treat early hires as architecture decisions, not just headcount.

Get your ticket, learn the skills every founder needs, and bring a teammate along.

Image Credits:Halo Creative

Building for the future: AI, market fit, and knowing a winner when you see one

This is the track where the agenda gets most forward-looking, and it’s really three questions in sequence: What are you building, do people actually want it, and will it be big enough to matter?

7AI co-founder and CEO Lior Div starts with the “what.” Adding AI features to an existing product is a tactical decision, and building a company where AI is the foundation from day one is a structural one, touching everything from team composition to how go-to-market actually works. He’ll explore what founders need to rethink when AI isn’t a feature, but the premise.

Next, Kent Bennett from Bessemer Venture Partners tackles the question, “Do people want it?” This is a necessary gut check right after a session on ambitious AI-native building: It’s easy to mistake early traction or investor enthusiasm for real product-market fit. He’ll unpack the signals that genuinely indicate fit and the metrics that mislead founders into scaling too early.

Finally, TDK Ventures investment director Tina Tosukhowong zooms out to the category level with an explanation of the firm’s “King of the Hill” framework, which is the discipline top investors use to judge whether a company isn’t just viable, but also positioned to define its entire category on economics, scalability, and timing. It’s the lens that explains why some well-built, well-funded products never break out, while others soar to success.

Get ahead of the race to make the most of AI tools to build and register today.

The priceless lessons from a founder who’s seen it all

It’s one thing to hear these ideas in the abstract. HqO co-founder and CEO Chase Garbarino closes the agenda as a working example of all three tracks at once. After all, he’s raised $200 million from the right investors, scaled a team across more than 30 countries, and built a category-defining company, all without ever relocating from Boston.

He’ll break down the real advantages and disadvantages of building outside a major hub, and what every founder outside Silicon Valley should know before assuming they need to be somewhere else.

The Summit doesn’t stop when the sessions end

The agenda is the backbone of Founder Summit, but the day is also built around forming connections that can kickstart a nascent venture or lead to a new investor. Roundtable discussions dig deeper into the Summit’s themes in smaller, more candid settings, where you can actively workshop with experienced startup community members. And the Braindate networking app pairs founders with investors and peers facing the same problems outlined in the agenda.

“Founder Summit Week” runs November 1-7, with Side Events happening across Boston for anyone who wants more than one day with the startup community.

By joining Founder Summit on November 4, you’ll leave with a lot more than helpful notes — you’ll leave with a working framework for the capital, leadership, and product decisions that would’ve otherwise taken months of trial and error to figure out on your own. Register now before prices increase on October 16. Bring your co-founder or partner with you at 50% off. Or, save even more by bringing your cohort of four more.

In Brief

On Tuesday, Anthropic announced an expansion of its Claude for Startups program, offering subsidized access to its AI models and a variety of other benefits to qualifying companies.

The new version of the program, which launched as part of Anthropic’s SF Tech Week event, provides a free year of Claude Team, Anthropic’s paid plan for groups, with up to five premium seats, as well as $1,000 in API credits for building with Claude. Companies will also get access to Claude Marketplace, which lets them build plug-ins for the service. Companies can also book virtual office hours with Anthropic’s Applied AI team.

“We created this program because we believe the benefits of AI will reach most people through the companies that build on top of models, rather than through the models alone,” Anthropic said in an announcement. “That makes partnering closely with founders and developers central to our mission.”

Companies are eligible if they were founded in the last five years or received funding in the last two years. Interested companies can apply through the Claude for Startups program page.

If you’re looking for a document editor that doesn’t use artificial intelligence, LibreOffice may be just the software for you.

The Document Foundation, the nonprofit software maker behind the open source document editor LibreOffice, said in a blog post this week that while it does not reject AI out of hand, it “will not add” AI to its software for the foreseeable future.

The blog post comes weeks after the release of the latest version of LibreOffice in late August, in which the nonprofit reiterated that the software “contains no generative AI features.”

LibreOffice said at the time that this was a “deliberate design position” to ensure that users’ documents are not uploaded to a server for processing and that no part of the software requires a network connection to function. The nonprofit added that its software is used by tens of millions of people and organizations, and that users should retain control over their data.

“Any organisation handling confidential, legally privileged or personal data needs to know where that data goes, and the only assurance that survives an audit is that it does not leave the machine,” the organization said.

LibreOffice says users can still choose to use AI tools by installing software extensions that connect the document editor to local AI models.

The decision to eschew including AI features bucks the trend of many software makers today that try to pack in as much AI as possible, even to the detriment of the user experience. The so-called “enshittification” of modern software by packing in unnecessary and bloated features, like chatbots and AI assistants, prompted some developers to create or include features that block or remove AI features altogether.

AI tools often require the users’ data to be uploaded and processed on a remote server, which the user has no control over how the data is stored or for how long.

LibreOffice said that there is currently “no integration” that meets all of its requirements, such as ensuring that the users’ data stays on their device and that there is no reliance on a single AI provider, among other considerations. The organization’s criteria for including AI into its software is not a “definitive rejection,” but the decision was an assessment of the technology today, rather than a “judgement on the value of the sector.”

“Until then, there will be no AI of any kind in the default installation; extensions will be available for those who wish to integrate AI features, and we will closely monitor where the technology is actually heading,” said LibreOffice’s blog post.

LibreOffice is one of the many kinds of open source and alternative software used to escape the clutches of Silicon Valley tech giants.

The personal-agent space has become closely competitive in recent weeks, with Meta’s Muse, OpenAI’s Dot, and Instinct taking the lead and more competitors being launched each week. But Vinod Khosla is betting on former Google and DeepMind engineer Shivani Poddar’s Wajo for one simple reason: trust.

“I think the most important thing in an agent, because it’s acting for you, is trust. You have to have trust in the agent, and I think compared to either Muse or Instinct, Wajo is the most trustworthy because it is architected to focus on trust and safety first, and some of the others are not,” Khosla told TechCrunch over a call.

“I mean, who wants to give all their data to Facebook [Meta]? I don’t. And if I give my data to Facebook, I’m the product, not the product I am buying, because they are in the ad business. So trust is important, and most people don’t trust a large company like Facebook,” he added.

In broad strokes, Wajo’s Fo agent works very much like Muse or Instinct. The agent can work across iMessage and WhatsApp, and has a web interface with cards of potential tasks. Errands are organized into categories like “scheduling,” “gifting,” and “life admin.” Its agent can also call businesses or other people on your behalf, booking or canceling services as needed.

In a few early tests, I asked it to call an airline to check for upgrades, and it successfully did so. However, when I asked it to call my partner and remind her to feed the cats, the call didn’t go smoothly, and disclosures were less clear. As rivals like Instinct and Muse are also rolling out calling ability slowly, we might see more discussions around how an AI calling on someone’s behalf should disclose itself. One of the unique things about the tool is that it can hire a human to complete a task when needed.

Apart from this, Fo can create a virtual credit card and pay on your behalf without disclosing your details. The agent also has its own email and can converse with people and businesses. When websites or services block AI agents, Fo can inform the user to take action or create a group with the user and the business for verification.

The startup has also secured backing from Jeff Dean and Gokul Rajaram, but didn’t specify the money raised. Wajo didn’t provide any numbers around users but said that it’s being used in 107 countries and has grown 10x since its launch in late September.

Founder story

Prior to founding Wajo, Poddar worked in various engineering roles at Meta, Google, and DeepMind and was invovled in making the first Facebook assistant. She said that she wanted to build products that were useful to people.

“The whole journey of bringing Gemini to the world and building safety for it made me realize the technological depth that you need to do this. Google is so good in technology that sometimes it can be left behind in product. I thought I needed to do something that is both technology and product at the same time. It has to be something different, which is very grounded in what the user needs,” Poddar told TechCrunch over a call.

Image Credits:Wajo

The former Google engineering lead mirrored Khosla’s thoughts around trust, and emphasized that it was one of the core principles for Wajo.

“Trust is something you have to build from the get-go because one of the technological things that has not been solved yet is teaching agents to unlearn. If they get access to your personal data on day zero, they will always have it for the future. I think trust needs to be built from day zero, not only in your architecture, but in product choices,” she said.

The company is doing work in a few areas in relation to trust and safety. It has applied for a patent for its virtual card system that hides users’ information from sites. For calls, the company said it will include a “trusted contact” list for users whose agents can call them. When you get a call, you can also tell the agent not to call you again, and Wajo records that preference.

In the meantime, wajo’s Fo has a tos complaint solution to agents on Amazon folks. Shop on :) https://t.co/yx2SoknZIy — Shivani Poddar (@shivanipod) October 6, 2026

The startup’s challenge is not to raise privacy concerns like other agents, especially as it is marketing safety as a core value. It also claims that despite tools like Muse being blocked by Amazon, its architecture follows the e-commerce service’s terms of service.

Wajo said that in the coming months it will explore how its Fo agent will work in social settings across multiple users, with the core idea of keeping privacy intact.

Bluesky CEO Toni Schneider teased long-term plans on Monday when he shared that the company has applied to make “.bsky” an internet domain suffix, like .com or .org. If approved, users will be able to turn their longer social handles into shorter, cleaner domain names.

For instance, instead of @sarahp.bsky.social, my handle could one day become just “sarahp.bsky.”

That could leave Bluesky room to innovate around user profiles, turning them from a simple page that houses someone’s bio and their posts into a more comprehensive personal home page that establishes someone’s online presence.

Bluesky’s application hasn’t been approved yet. Schneider said that ICANN (Internet Corporation for Assigned Names and Numbers), the nonprofit that oversees domain names for the internet, will publish the applications for the new domains on October 7, 2026, and Bluesky’s request will be among them. It’ll take 18 to 24 months for the process to be completed.

Reached for comment, Bluesky said it hasn’t finalized how .bsky domains would be requested, registered, or distributed, but the company did say it had no plans to either sell or monetize the domains.

That’s a notable decision given that domain registration can be a profitable business that could serve as a potential revenue stream for a startup that has raised significant funds but isn’t (yet?) generating ad revenue.

However, Bluesky says selling domains goes against its belief in an open web.

“Domains are a big part of the open web, as they allow users to control their own online identity,” said Rose Wang, Bluesky’s COO, in a comment shared with TechCrunch. “Bluesky was built on the belief that you should own your presence online, not rent it from a platform that can take it away.”

In Brief

Amazon announced on Tuesday that Prime Video will be the new home for the Emmy Awards, marking another major Hollywood awards show’s move from broadcast television to streaming. This will also be the first time that the ceremony has had a dedicated home, as the show rotated across broadcasters like ABC, CBS, Fox, and NBC over the past seven decades.

Prime Video will stream the awards show worldwide for six years for free, starting in 2027, where it will go live in over 240 countries and territories. Non-Amazon Prime subscribers will also be able to watch the awards show on the Prime Video platform.

The terms of the deal were not disclosed.

The move comes as the Emmys face potential viewership declines, as more of its audience moves to streaming services. This year’s ratings for the 78th primetime Emmy Awards on NBC were down nearly 10% from the year prior, after a couple of years of growth, the AP reported. However, the telecast’s 6.87 million viewers this year were higher than its all-time low of 4.3 million in January 2024.

The Emmys aren’t the only major awards show making the shift to streaming. Last year, the Academy of Motion Picture Arts and Sciences announced a deal giving YouTube the rights to stream the Oscars beginning in 2029, while Netflix has streamed the Screen Actors Guild Awards since 2024.

The race between open and closed AI is on, and Europe is still in the mix. On Tuesday, French AI lab Mistral AI released Mistral Large 4 (ML4), a new large multimodal model aiming to leapfrog both American and Chinese rivals — following what French president Macron described as “a third way in AI.”

Amid a growing divide between closed models that can be unplugged and open models that are often made in China, Mistral is positioning ML4 as an alternative to both. Nicknamed Le Chonk in reference to its 1 trillion parameters, ML4 is definitely not small; but is not an open-weight model yet. For the time being, it can only be accessed via a public guardrail endpoint, but Mistral plans to make its weights available in just three weeks, after safety testing is complete.

“In the meantime, we’ll work with trusted partners and governments to make sure that the open source weights can be used to defend, but not to [perform] malicious attacks,” Mistral VP Science Pierre Stock told TechCrunch.

Security concerns have been mounting in recent months, particularly among Mistral’s core audience — enterprises and institutions. At the same time, an open-weight model is easier to audit, Stock said.

Another important behind-the-scenes aspect is that ML4 was trained entirely on Mistral’s compute; using only 4,000 Nvidia GPUs “which is two to three times less than our Chinese competitors, and significantly less than the closed source competitors,” Stock said.

With benchmark results still pending, Mistral hopes ML4 will be best in class among open-weight models, especially outside of China, but not only, Stock said. Thanks to focused training, it could also outperform closed models in specific areas that are key to its customers, and where multimodal capabilities can add value.

According to Stock, ML4’s optimized use cases include cybersecurity and finance, but also chip design, which is core to two of Mistral’s main backers — Dutch giant ASML, which led its Series C, and Samsung, which led its Series D last month at a €21 billion valuation (about $24.39 billion).

At the time, the company tried to convey that its decision to host Chinese models wasn’t a pivot into becoming a mere inference provider. With Le Chonk in its corner, the company believes it should still be considered a frontier lab as well.