At HumanX this past week, I watched eight different tech leaders – including the CEO of AWS, the co-founder of Cursor, and the head of Anthropic Labs – describe essentially the same phenomenon without any of them naming it. A Novo Nordisk scientist asks a Databricks agent about adverse effects in an obesity study and gets referenced, percentiled answers in minutes; work that previously required routing through a data team over months. Anthropic’s finance department builds its own internal tools with Cursor despite having no coding background. Replit’s CEO watches board members and VPs walk into their companies with vibe-coded prototypes, telling their engineering teams “this is my vision” or, more pointedly, “why is this taking so long?”

The handoff machine

The narrative we’ve been hearing about AI and work is almost entirely about headcount: will jobs be created or destroyed? But what these leaders are actually describing is something structurally different. AI isn’t primarily replacing workers: It’s eliminating handoffs. And in so doing, it’s collapsing the division of labour that has defined how companies operate for the better part of a century.

The modern corporation is, if we’re being honest about it, largely a handoff machine. An idea moves from product to design to engineering to marketing, each transition consuming time, introducing opportunities for miscommunication, and requiring a layer of management to coordinate. A huge share of corporate headcount exists not to create value, but to manage the movement of information between people who do. A Medium analysis of AI-related layoffs in 2025 found that cuts clustered exactly in these coordination-heavy roles – “where work exists to move information, manage handoffs, and absorb friction.” Gartner’s prediction that 20% of organizations will use AI to flatten their structures by 2026, eliminating more than half of middle management positions, points at the same thing. The roles most at risk are those that exist because humans are slow at context-switching themselves, and bad at transferring context to each other, too.

Pulling the work back in

What I saw at HumanX was the other side of that coin – the reabsorption. People are pulling work back into their own hands that they’d previously just been forced to delegate. Matt Garman, the CEO of AWS, framed it through his sales organisation: reps currently spend maybe 20% of their time actually talking to customers, with the rest consumed by pipeline administration and preparation that used to require support staff. If AI flips that ratio, he argued, you don’t fire salespeople – you hire more of them, because each one can now support four times the customer relationships. The expansion is in the generalist role, not the support layer.

Bret Taylor, running Sierra and chairing OpenAI’s board, named something I think is genuinely important and underexplored. In the early days of a company, he noted, the most valuable person is often a great generalist – someone who’s “sort of an engineer, sort of a product manager, talks to customers,” who has real empathy for the end user and enough range to ship something end-to-end. Then the company grows, everyone has to specialize, and Taylor wonders aloud whether that forced specialization is correlated with “the enshittification of products,” the term coined and popularized by Cory Doctorow that describes a gradual decline in online products and services over time due to pressures on popular products in terms of maximizing profits. His real question: can that generalist now scale with the company, because AI handles the execution complexity that previously required breaking roles apart?

There’s already evidence this is happening. Michael Truell at Cursor reported that designers and product people – “technically light personas” – are now making changes in production codebases at enterprise customers. Mike Krieger at Anthropic mentioned his finance team building tools with no coding background. Eric Glyman at Ramp described a risk analyst who got frustrated with a 20-hour monthly underwriting process, screen-recorded his own workflow, and turned it into a 45-minute exercise.

I’ve felt this shift in my own work, too. I help run a PR and communications agency – SBS Comms – and over the past half-year I’ve been using AI to take entire campaign projects from strategy, through research and all the way to deliverable, solo, without pulling in the rest of the team. Not because I don’t trust them or because I’m trying to cut costs, but because the handoff itself was the bottleneck. By the time I’d briefed someone on the client context, explained the angle, reviewed a first draft, and iterated, I could have just done the thing – without signal loss. AI closed the gap between what I knew and what I could execute on my own. The result isn’t fewer people at the agency, though: it’s that the rest of the team can stay focused on their own priority work instead of getting pulled into mine.

These aren’t stories about AI doing someone’s job. They’re stories about people reabsorbing work they’d been forced to outsource to specialists or intermediary teams.

Ali Ghodsi at Databricks had probably the sharpest example. He described easyJet building an agent that can answer questions like “how many seats are taken on this flight from Paris to London?” combined with competitive pricing and historical demand data – the kind of query that used to require a multi-month engagement with a data team. The data team doesn’t necessarily disappear, but its gatekeeping function evaporates. The person with the question can now get to the answer without routing through three layers of organisational intermediation. That’s not automation in the traditional sense of replacing human-driven value-creation. It’s the disintermediation of internal expertise.

What gets lost

The shadow side of this is real and shouldn’t be glossed over. Korn Ferry’s 2025 workforce survey found that 41% of employees say their companies have reduced management layers, with 37% saying it’s left them feeling directionless. Middle managers held institutional knowledge, mentored junior employees, and served as a translation layer between strategy and execution – functions that don’t just evaporate because you flatten the org chart. Amjad Masad at Replit told a story that captures the emotional texture of this perfectly: he walked to work one day, the elevator opened, and an employee was standing there “dazed and confused,” muttering that “Zerg is taking over” – their internal code name for the AI agent that went from writing 0% of their pull requests to 50% in two weeks. The guy ran away and Masad hasn’t seen him since. It was a funny anecdote, but there was something underneath it that wasn’t funny at all.

Range over depth

What I find most interesting about the reabsorption pattern is that it reframes the entire debate about AI and employment. The “will AI take my job?” question assumes a static division of labour where AI slots into existing roles. But what’s actually happening is that the division of labour itself is being renegotiated. When a CEO can prototype, the power dynamic between leadership and engineering shifts. When a scientist can query clinical data directly, the relationship between domain expertise and data infrastructure changes. When a salesperson can prepare their own customer briefings, the support apparatus that existed to serve them becomes optional rather than essential. And in an ideal world, that support structure then becomes an entirely additive new value creation layer.

I’ve experienced a version of this myself – a media briefing document that used to involve a researcher, a writer, and a round of internal review is now something I can produce end-to-end in a fraction of the time, with AI handling the research synthesis and first draft while I focus on the strategic framing and client voice. The quality hasn’t dropped. If anything, it’s improved, because there’s no context loss between the person who understands the client and the person who writes the document. They’re the same person now. The function of what was the support and implementation team is now to focus on deeply understanding the customer and their industry and operating substrate in order to be able to parallelize and augment that process across the organization.

The career implication is fairly clear, even if nobody wants to say it this bluntly: range is becoming more valuable than depth. The person who can ship end-to-end – from idea through prototype through deployment – is increasingly worth more than the person who excels at a single link in the chain. Nielsen Norman Group, not exactly a hype-prone organization, recently documented “the return of the UX generalist,” noting that economic pressures will increasingly favour AI-augmented generalists over narrow specialists because the results are “cheap, fast, and good enough.”

That last phrase is going to make a lot of specialists uncomfortable, and it should. But it’s also probably the most honest description of where this is heading. The great reabsorption isn’t a story about machines replacing humans. It’s a story about humans reclaiming scope – and the organizational architecture that was built to manage their limitations quietly becoming unnecessary.

Disclaimer: Cursor and Ramp are clients of SBS Comms, my employer.

When HumanX debuted in Las Vegas last March, DeepSeek was a name that maybe a few hundred researchers would have recognised. Reasoning models were a novelty. Cursor had a devoted following but hadn’t yet become a verb among developers. And the idea that AI agents would be autonomously merging pull requests, managing corporate expenses, or finding and patching security flaws without a human in the loop would have sounded like a pitch deck fantasy – ambitious, sure, but comfortably distant.

Thirteen months later, every single one of those things is real and shipping. HumanX 2026 opens at the Moscone Center in San Francisco on April 6, and the sheer volume of ground that has shifted beneath the conference’s feet since its inaugural edition is absurd. The event has nearly doubled in size, but even that growth feels inadequately proportional to the pace of the industry it’s convening to discuss.

The move to San Francisco is more than a venue upgrade. Placing an AI conference at the geographic epicentre of the thing itself was the obvious play, but it also signals something about the maturation of HumanX as a gathering. The Vegas debut had the energy of a well-funded first draft – impressive speakers, genuine buzz, solid editorial programming. Moscone is where you go when you’re no longer proving you belong; it’s where you go when you’re setting the calendar for an entire sector.

And the sector desperately needs a coherent calendar right now, because the last year has been a blur. Consider just the broad strokes: DeepSeek’s R1 arrived in January 2025 (yes, only a year ago, and yet if feels way longer than that) and sent shockwaves through both Silicon Valley and public markets, demonstrating that a relatively small Chinese lab could produce frontier-class reasoning for a fraction of what Western labs were spending. The “DeepSeek moment” became shorthand for a kind of efficient innovation that challenged the assumption that you needed tens of billions in compute to compete. OpenAI shipped o3, o4-mini, and eventually GPT-5 over the course of the year. Anthropic released Claude 4 and iterated through many point updates across Opus, Sonnet, and Haiku variants at a pace that made keeping track of version numbers a part-time job. Google DeepMind’s reasoning models were winning gold at the International Math Olympiad. And underneath all of it, the real story – agentic AI moving from demo to deployment – was quietly rewriting how software gets built, how companies operate, and what “work” actually means for a growing number of knowledge workers.

That last shift is the one that makes this year’s HumanX speaker lineup feel particularly well-timed. The show’s roster spans the full stack of the current AI moment – from the researchers building the models to the founders deploying them in production to the policymakers trying to figure out what guardrails make sense when the ground won’t stop moving.

Fei-Fei Li – co-founder of World Labs and the person who, through ImageNet, arguably did more than anyone to make modern deep learning possible – will speak on what comes next. Matt Garman, now running AWS, and Mike Krieger, leading Anthropic Labs, represent two very different but deeply consequential visions of how AI infrastructure and products should evolve. Sarah Guo of Conviction, whose No Priors podcast has become essential listening for anyone tracking the space, will be there alongside Bret Taylor – co-founder of Sierra and chairman of OpenAI’s board – and Srinivas Narayanan, OpenAI’s CTO of B2B Applications. Vinod Khosla and Scaled Cognition’s Dan Klein will tackle the tension between super reliability and super intelligence, which is – if we’re being honest – the core question the entire industry is dancing around.

Then there are the founders who are translating all of this foundational capability into products people actually use every day, and three of them are worth calling out specifically – with the disclosure that both of these two are clients of SBS Comms, my current employer.

Michael Truell, the co-founder and CEO of Cursor, will be on the main stage discussing how AI is reshaping the act of writing code. Cursor has had a year that borders on the surreal. The company surpassed $2 billion in annualized revenue and hit a $29.3 billion valuation, but the numbers are almost secondary to the product story. Truell recently described a “third era” of AI-assisted development – one where developers are no longer writing code keystroke by keystroke, or even directing agents through synchronous prompts, but instead managing fleets of autonomous agents that operate like teammates. More than a third of the pull requests Cursor merges internally are now created by agents running independently in cloud VMs. Agent usage in Cursor has grown over 15x in the past year, and the company now has twice as many agent users as tab-autocomplete users – a complete inversion of where things stood just twelve months ago.

Eric Glyman, co-founder and CEO of Ramp, will talk about how AI is rewriting the rules of finance – a session title that might sound like standard conference fare until you look at what Ramp has actually shipped. The company’s agentic AI for expense management is now more than 99% accurate at tasks like determining whether a travel expense adheres to company policy – more accurate than humans working alone, and serving over a million end users. Ramp crossed $1 billion in annualised revenue and reached a $32 billion valuation late last year, and Fast Company just named it one of the most innovative companies of 2026. Glyman’s framing around “zero-click” experiences – stripping dozen-click legacy workflows down to a few clicks, and then eventually to none as the AI gets smarter – is one of the clearest articulations I’ve heard of what agentic AI actually looks like when it’s deployed at scale in a domain that isn’t software development. Plus, he recently dropped an essay on the shift that’s already emerging favouring companies who are leaning into AI spend, and that’s bound to be a buzzy topic of conversation at the show.

Beyond these three, the agenda is dense with sessions that reflect the breadth of where AI is actually landing: Demi Guo of Pika and Dean Leitersdorf of Decart will discuss the distinction between viewing and experiencing in AI video. Arvind Jain of Glean and Eric Yuan of Zoom will do a joint Q&A. Snowflake’s new CEO Sridhar Ramaswamy, Thomas Wolf of Hugging Face, and Anton Osika of Lovable all make appearances. There’s a whole track called “Control Room” dedicated to how AI transforms enterprise operations, and another called “Builders” that gets into the infrastructure layer.

The thing that strikes me most about this year’s HumanX, though, is how different the ambient mood of an AI conference feels now compared to even twelve months ago. Last year in Las Vegas, there was still a palpable undercurrent of “is this real?” – not about the technology itself, but about whether the enterprise adoption wave would actually materialise, whether the valuations were sustainable, whether AI would remain a tool mostly used by developers and early adopters. That question has been answered, decisively, in the affirmative. According to McKinsey's most recent global AI survey, 88% of organisations now use AI in at least one business function – up from 72% just a year earlier. Worldwide AI spending (across infrastructure, software, services, etc.) crossed $1.5 trillion in 2025 according to Gartner. The conversation has shifted from “should we adopt AI?” to “how do we not get left behind?”

Which means the conversations that matter most at HumanX this year won’t be the ones about capability – we all know the models are getting better, faster, and cheaper at a rate that makes Moore’s Law look leisurely. The conversations that matter will be the ones about deployment, integration, trust, and the deeply human question of what all of this velocity actually means for the people and organisations trying to absorb it. The conference’s tagline has always leaned into the “human” part of HumanX, and in 2026, that framing feels less like branding and more like the central challenge of the moment.

San Francisco is the right city for that conversation. And next week is very much the right time. I’ll be there, so drop me a note if you will, too, and let’s continue the conversation.

Disclosure: Michael Truell/Cursor, Eric Glyman/Ramp, and Glean are clients of SBS Comms, where I work. I was not compensated for this piece and all opinions are my own.

"I'd wager that the most successful companies of the next decade were founded and run by missionaries who earned high rewards through hard work and learning."

Kipp Bodnar is talking about the AI talent wars, where companies are throwing $250 million pay packages at researchers, and his take is heretical: it won't work. Not in any way that matters. The HubSpot CMO thinks the mercenary approach might deliver numerical wins, but "it will not create this new 10x, better than anything outcome."

When I caught up with Bodnar recently ahead of HubSpot’s annual ‘Inbound’ conference and the launch of the company’s new Loop Marketing playbook, he was deep into experimenting with Perplexity's new Comet browser, dictating thoughts through Willow Voice, and building custom GPTs that his marketing team apparently would "die on a hill for." But what struck me wasn't his tooling obsession. It was his contrarian take on what actually matters as AI reshapes marketing.

"Benchmarks don't matter in real life," he told me, cutting through the industry's obsession with model performance metrics. "They don't correlate well to real life."

The Taste Problem

While everyone else is talking about prompt engineering and management skills for the AI age, Bodnar is focused on something decidedly more analog: taste.

"Taste is that unique point of view and unique understanding of your audience, what they need, the experiences that they're not getting," he explained. During the pandemic, bored and looking for something to do, he decided to learn about contemporary art by spending 20 minutes a day on Artsy, just looking at piece after piece. "Now I actually kind of understand. And if I was in New York City and I was talking to somebody running an art gallery, and I knew what they were focused on, I could have a good conversation with them."

This isn't just philosophical meandering. Bodnar believes the last era of marketing let people get away without taste because everything was "so metrics driven,” and focused on adjusting “knobs and dials." But that world is ending. The SEO apocalypse that marketers are lamenting on Reddit? It's real, but it's not the end of the story.

The Volume-to-Value Inversion

"Visits and revenue are going to get decoupled for the first time," Bodnar said, describing what might be the most important shift marketers need to understand. Traditional conversion rates used to be static; you just needed more traffic. Now it's flipped. "AI can really help your conversion rates improve and change dramatically. But it's much harder to move that visits number because you have established channels like Google going away."

The data backs this up. HubSpot is seeing customers who arrive via LLM-based search queries converting into customers twice as fast and at "12x the rate, not like 1 or 2x like over an order of magnitude more." The problem? There are far fewer of them.

This creates a paradox that's breaking marketers' brains. The old playbook (more traffic equals more revenue) is dead. The new reality requires something different: extreme personalization at scale.

"I don't know that we're that far away from me having like a dedicated part of any website just for you with custom content and everything," Bodnar mused. "Because I could programmatically do it and make it good."

Why Your CEO Shouldn't Hire That $10M AI Researcher

When I asked about the talent wars and those eye-watering compensation packages for AI researchers, Bodnar had a surprisingly measured take. Yes, he understands why companies are doing it: "the opportunity cost of not being great at artificial intelligence over the next 24 to 36 months is probably the highest opportunity cost, at least of this century."

But here's the kicker: "I'd wager that the most successful companies of the next decade were founded and run by missionaries who earned high rewards through hard work and learning." The mercenary approach of being top bidder for a particular researcher might work numerically, he argued, but "it will not create this new 10x better than anything outcome."

This extends to his philosophy on skills development. While others are preaching management skills for overseeing AI agents, Bodnar is skeptical. "Management has long been kind of overrated even in human culture. Professional managers will always get beat out by skilled craftspeople most of the time."

Instead, he champions curiosity, creativity, and ambition. "If you're willing to think about things from a slightly or very different perspective and from first principles and you're willing to have some ambition... it's 100x easier to make that vision come true now than it was five years ago."

The Browser Wars Nobody Saw Coming

Perhaps the most unexpected part of our conversation was Bodnar's enthusiasm for what's happening with browsers. Using Perplexity's Comet browser has him convinced that "the browser wars are kind of everything."

"I use Perplexity way more now because of Comet," he said, demonstrating how the browser can fill out forms, draft emails, and analyze YouTube videos in real-time with contextual understanding. It's not just a Chrome extension – it's full computer use that's actually fast.

This matters because it reveals something crucial about AI adoption: integration beats innovation. The best AI isn't the one with the highest benchmark scores; it's the one that slides seamlessly into your workflow.

The Diminishing Returns Nobody Wants to Talk About

Here's what the AI optimists won't tell you: "AI can't do everything." Bodnar was blunt about the reality of implementation. "If you use AI to solve customer support, you can resolve a lot of your customer support. But it's not like you're going to get 100% solved. You're going to get 50, 60, 70, 80, and then it might cost more money to solve the rest with AI than not."

This pragmatism extends to his advice for marketers navigating the chaos. When I asked about generative engine optimization (GEO, AEO, or whatever acronym wins), he was clear: pay attention, but understand the tradeoffs. "You get way less visits. And the attribution is way harder." But those visits? They're gold. HubSpot literally has recordings of customers saying, "I talked with ChatGPT for an hour about my business and they said that you were the best choice for me and I would like to buy your product."

What Apple Should Do (But Won't)

As our conversation wound down, I asked what he was most excited about. New OpenAI model releases, sure. Better video generation, definitely. But what really has him fascinated is watching Apple navigate this moment.

"Just give your whole market cap to Sam and have him run the company," he joked. "Do something crazy."

The company that once attracted mission-driven talent at a discount now finds itself desperately behind. It's a cautionary tale for any organization that thinks its current advantages are permanent.

The Real Transformation

Bodnar has seen this movie before. He watched the second generation of the internet create fortunes for early adopters. He's telling his team the same thing now: "My expectations around AI skills and AI adoption will likely be higher than anywhere else. And that's going to be great for you because... whenever you decide to take the next step of your career, you're going to go out to the market and you are going to have this premium skill set."

But unlike the social media revolution, this one requires something different. Not just early adoption, but taste. Not just management skills, but creativity. Not just more traffic, but better conversion.

"Creativity and ambition are now the limiting factors of successful people, successful companies," he said.

In other words, the robots aren't taking over. They're just raising the bar for what humans need to bring to the table.

Google has made some very interesting product announcements over the past week around AI-powered image editing. There are actually two separate announcements, and no overts link between the two, but it seems likely that they at least share some common DNA under the hood.

Today (Tuesday, August 26), Google announced that its DeepMind team developed an extremely sophisticated and capable image editing model nicknamed ‘nano banana,’ which, aside from calling to mind Nintendo’s excellent new Switch 2 exclusive Donkey Kong game, also brings very powerful editing tools to its Gemini AI-powered assistant.

And then last week, Google unveiled its new Pixel device lineup, including the Pixel 10, Pixel 10 Pro and Pixel 10 Pro Fold smartphones, plus new Pixel Buds hardware and capabilities, and a new Pixel Watch 4. One of the highlight features of the new Pixel 10 line is the ability to edit photos taken with the devices in Google Photos using simple, natural language editing commands - things as simple as ‘make this photo look better,’ as well as more specific commands like ‘fix the lighting in this picture’ or ‘frame the subject in a more interesting composition.’

Google — and any other smartphone maker or photo editing software provider — have long offered ‘auto’ adjustment features that make an algorithmically-determined set of optimizations and changes to the pictures people take. And Google has used generative AI-powered editing for things like ‘magically’ erasing elements from pictures or improving specific parts of photos, but these editing tools go beyond that to offer more comprehensive, powerful and flexible editing options with a user interface that’s accessible to even the most non-expert and non-technical user.

These features are not dissimilar from some of the more recent feature releases from companies like Adobe whose entire business is focused on image editing. Adobe offers generative erase/removal and canvas expansion features, for instance, that either automatically try to intuit what you want to accomplish, or that can take natural language instructions to steer their results.

Adobe’s tools are still aimed at and optimized for professionals who want to start to make use of advanced AI features to lighten the load of highly repetitive, painstaking manual work during edits. Google’s tools are aimed squarely at non-expert users, many of whom have probably never opened a dedicated image editing app in their entire lives. They’re very different ideal customer profiles, but they also meet in the middle, and I think the bigger question is whether Google’s features and services start pushing up into Adobe’s market before Adobe’s tools start pushing down into Google’s.

Even Google’s own tools are both aimed at slightly different use cases and segments: The Google Photos editor is focused more on naturalistic, small changes to existing photos that keep them looking mostly like the photos you actually took. Gemini’s enhanced editing capabilities are designed for more imaginative use cases, like changing your entire outfit or setting, or combining subjects from multiple photos into one generated image. But they’re both on a spectrum that runs the gamut of uses that act as entry points for many Photoshop-curious users who pick up Adobe’s suite for the first time and end up becoming lifelong users.

I wrote previously about how OpenAI’s latest image generation tools powered by GPT-4o represented a huge shift in AI-powered graphics and visuals, and Google’s newest tools represent a continuation and significant evolution of that shift. It’s actually much easier to anticipate a future in which ‘vibe photo editing’ is a going professional concern and paid opportunity than it is to envision a scenario where ‘vibe coder’ is a real and sustainable job. Rick Rubin on 60 Minutes famously talked about how he accomplished everything he did without any technical ability, ascribing his success instead to taste, and photo editing has more in common with music production than it does with coding.

What I think might result from continued improvements like natural language editing in Google Photos and Gemini’s new ‘nano banana’ features is a changed creative industry where taste and style, rather than technical capability, will determine who becomes most successful. I don’t actually think it represents a path to ‘democratization’ of creative skills and expression: instead, I think it means value will accrue better to those who actually deserve it, rather than to people who are able to decode the specific technical tools and advanced software incantations that separate ‘professionals’ from ‘amateurs’ today. An editor with poor or mediocre creative vision, but an adept-level grasp of Photoshop or Adobe Premiere is unlikely to get very far in this predicted world, but a true visionary who can barely operate their iPhone stands a much better chance, by contrast.

Even speaking as someone who has spent 20+ years developing proficiency with all kinds of ‘professional’ level creative software, personally, I’d love to see more of those visionaries get a real chance.

The iPhone’s ‘killer app’ didn’t launch with the original hardware – it came later, when Apple changed course from an original vision of allowing only first-party native applications to run on the device, to opening up access to third-party developers who wanted to build locally-installed, downloadable ‘apps.’ Apple’s App Store, combined with new technological capabilities ushered in by, and unique to the smartphone – like location services and cellular connectivity – resulted in an explosion of creativity and innovation that crowned some of the kings who still stand astride of the tech world today.

Apple has been aggressive in its defence of its iOS platform advantage (the straining at the seams of which is starting to show in things like its ongoing EU regulation troubles). Much of its defence has built on lessons learned from the platform winners and losers of version one of the consumer computing revolution. But like all incumbents, their eventual weaknesses don’t look the same as the ones they exploited to ascend the throne they now occupy.

Like most incumbents, Apple has spent a lot of time tilting at windmills when it comes to anticipating where the next big platform shift might come from. The Apple Vision Pro is maybe the best – and most expensive – example of a company betting big on a direction for the future that turned out to totally miss the actual disrupting transformation. The Vision Pro is an outstanding technical achievement, don’t get me wrong: but Apple’s AI game is severely lacking relative to the other serious players in the field, including some other incumbents like Google that placed better bets on what tech shift would have the most potential to reshape the world.

Apple’s still trying to effect a do-over in its AI product debut, with new reporting from Mark Gurman at Bloomberg suggesting it’s now looking to external partners to help shore up its deficiencies, at least temporarily. And yet one of those potential partners casually lobbed a grenade into its strongest moat – the aforementioned App Store – with a new feature launched last week.

Anthropic launched Claude Artifacts last week, providing a space within their interface to house a library of user-generated interactive applications created using their Claude AI. Claude has a great reputation for being able to generate sound code, and the company also recently released Claude Code to capitalize on this. Claude’s new Artifacts library also trades on this advantage, but does so in a way that’s much more accessible to their most non-technical users.

Claude Artifacts collects some of the applications that users are generating on the platform, making them available for other Claude users to either employ directly, or to remix however they want to suit it to their specific needs.

I want to take a second to shout out my former colleague and pal Kyle Russell, who recognized early that accessible, user-generated software was going to be a fundamental change in the way people interacted with the digital world. Kyle was operating on this assumption in a pre-LLM world, which was the ingredient that unlocked its broad applicability and flexibility, but he clearly foresaw what Claude Artifacts is now making a reality for millions of users.

Whether it’s more complex stuff made possible by tools like Replit, Lovable and Vercel; or more simple things like the current Artifacts Anthropic is curating in its collection, it feels inevitable that a massive portion of the application market that once paved the way for a generation of indie devs, is going to move up the abstraction ladder and become something users spin up on demand as easily as they do Studio Ghibli-fied versions of their selfies today.

Apple’s App Store dominance feels shakier than ever, and the foundation isn’t being shook by anti-trust cases or threats from peers – it’s coming from massive changes to how people solve problems they’re encountering. Once, we sought out comparisons and read through charts pitting the feature sets of one off-the-shelf application vs. another: Now, we’re creating apps that meet our immediate needs on-demand, with near-zero cost and little need for it to survive beyond our immediate use case. Snapchat also understood that ephemerality was a key ingredient for the future of tech – they just didn’t anticipate that that ephemerality extended to the utility and leisure applications we use every day, too.

Btw – if you want to hear more about Claude Artifacts, and lots more, including a recent Anthropic project to see if an LLM could effectively own and operate a vending machine, check out Artifactor, a brand new podcast hosted by myself and my friend and former TechCrunch colleague Greg Kumparak.

Last week, OpenAI announced a new feature relatively quietly, without the usual fanfare it assigns to product introductions or capability expansions: ChatGPT Record, a fairly innocuous new ability for its LLM-based chatbot that’s very limited in scope and availability at launch – but that seems symptomatic of a potentially significant issue for anyone building on top of OpenAI (or other model makers’) foundations.

ChatGPT Record (which I also talk about on my new weekly podcast Artifactor, with Greg Kumparak – check it out) is just what it sounds like: A recording function for working with ChatGPT. It’s Mac-only at launch, exclusively available via the desktop Mac app that OpenAI released last year. You also have to be on either ChatGPT Enterprise, Edu, Team or Pro, containing the bounding box of those who have access even further.

And at launch, in terms of functionality, it’s not doing all that much: It provides almost an async voice mode, where you record a voice note, memo or local meeting and it can then turn that into various outputs, including meeting summaries, to-do lists, brainstorm notes and more.

OpenAI really seems to bet slow-rolling this one, given how narrow its user group is at launch relative to the whole sum of ChatGPT’s user base. Plus, there’s no accompanying blog post and limited first-party hype, beyond a mention in an overall ChatGPT for Business video update from a couple of weeks ago, a help article and a LinkedIn update from a few days ago when it actually went live.

Increasingly, I think we’re going to see OpenAI play the tricky dance of trying to deploy useful and intuitive new products and extensions of its core features that work for both business and creative users – while also trying not to compete too directly (or at least, loudly) with the partners who build on top of its API and developer-facing tooling and infrastructure.

ChatGPT Record is a direct shot across the bow of startups and companies building in one of the buzziest and most active generative AI subcategories there is: meeting notetakers. They’re so prolific that talking about them and their omnipresence is replacing idle chitchat about the weather as the default small talk for when you’re waiting for everyone to join at the start of a call.

You can bet that the Granolas, Fireflies and Fathoms of the world sat up and took notice when Record was announced; likely, they were already on high alert before. ChatGPT’s potential to become a true productivity suite is reminiscent of when Gmail first arrived, and some startups focused on building things like a ‘calendar’ app specifically for Gmail users (check out the origin story of Justin.tv, which became Twitch, for more about how that ended up working out).

For those of you who follow the Apple blogosphere and news cycle, you’ll probably already be familiar with the term ‘Sherlocking’ – basically, the process of a platform provider expanding its core to include the features offered by a third-party building in its ecosystem. Astropad, which was itself Sherlocked, has a good practical primer for those looking to learn more. It’s not a new phenomenon in tech, nor is Apple the lone example (Shopify is another repeat offender), but in the case of OpenAI and the other foundational model makers who also have user-facing products, I think the speed and frequency of Sherlocking is going to increase rapidly.

Intuitively, it should already be clear that the current glut of similar products at the application layer in generative AI is unsustainable and bound for consolidation. Exits to foundational model companies will be part of that, as will mergers of startup competitors looking to survive and thrive against the big dogs. Meanwhile, the information gap works very much in favor of the platform providers: OpenAI, Anthropic, et al. know way more about what kind of products are working on their platform thanks to access to metrics like volume and frequency of API calls across customers than those building on top.

But I think that a surprising number of startups – including many with promising early traction – might just end up caught under the treads of the relentless metastasizing of the product machinery of OpenAI, Anthropic and others as they seek revenue and build on some of the fastest-growing direct user bases of any tech era to date.

When evaluating tools to adopt in an enterprise, business or product setting, there’s a long list of considerations – with modern LLM-based AI power tools, now there’s a thorny new one to contend with: Personality.

It’s been apparent for a while now that choosing among the various consumers AI chat-based answer engines is at least partially a matter of taste as well as an exercise in comparative capability. People I knew from an editorial background have been generally more willing to lean on Claude for feedback vs. ChatGPT, for instance, and technical-minded folks likewise have their own preferences. That changes with each model iteration, and sometimes within a model generation, as well, as tweaks are made. But the implications are even more profound when you consider what it means to incorporate these tools and their APIs as core elements of a product or products upon which you’re building a business.

Given my new gig, I’m starting to see this up close and personal – but I think Twitch co-founder Emmett Shear may have captured the spirit of this best in a recent tweet, pictured here:

This is something that dramatically changes the way in which tool selection is r should be done. I don’t want to overstate it, however; ‘vibes,’ called that or otherwise, has always been at least part of the equation when you pick what elements make up your stack. But the attitude or persona associated with a particular platform has been more a summary of what the various elements and design choices that make it up convey, in a soft and ambient sense. The UX of Slack vs. the UX of Teams, for instance, is very different, despite similar use cases and end results.

Differences like these are subtle, indirect and likely to have less impact on outputs than you might even expect. Same for a choice like using Google Cloud or AWS. But selecting between comparable technologies like LLMs that have the differing ‘personality’ traits that I think Shear has accurately identified above has a potentially massive impact on what you end up producing as an end product.

Think about those descriptors Shear uses to characterize these competing generative AI products: Basically, they represent a spectrum of appetite for risk, mental malleability and susceptibility to groupthink. All of these are core considerations when making key personnel decisions, particularly in the case of leadership roles.

Models that make up central elements of your technology stack effectively are leadership ‘hires’ in the current approach to AI-powered product building. The non-deterministic nature of their output means that they leave their fingerprint on whatever you’re having them do, to varying levels of visibility and effect based on how far from the customer-facing ‘surface’ of what you’re building they are, and to what extent you’re modifying or mixing the results with outputs from other models and tools.

I think we’re already seeing the model makers lean into the personalities and reputations of their products in their go-to-market strategies for how to sell them, and to who. Anthropic is increasingly interested in the enterprise, and ‘balanced, but timid’ as Shear puts it, is a good way to describe the average ideal enterprise hire, particularly in the middle management bucket.

As with discussions about their capabilities and prowess, one big issue remains that these personalities are also shifting sands, since new model releases seem to bring different characteristics. OpenAI had to specifically and intentionally dial back behavior from its latest model tweaks to make them less cloying and cheerleading, for instance. People, too, change who they are and how they show up over time, and they’re also core to the businesses we build.

I still think it will take a significant shift in mindset to properly adapt to a world in which building out your product and technology architecture also includes making essentially a hiring candidate assessment, alongside a technology capabilities and features evaluation. But that’s yet another exciting facet of just how fundamental the changes resulting from widespread AI adoption will be in terms of how we think about business building, and mostly discrete functional areas like people operations, engineering and product.

In a week absolutely stuffed to overflowing with AI-related news and announcements, one in particular stands out: Sam Altman and Jony Ive revealing ‘io,’ their collaborative effort to design and develop the first generation of truly mass-market consumer AI devices. We’d heard through earlier reports that the collab was on, but this week, the two dropped an absolutely wild trailer for their joint company – along with the news that OpenAI would be acquiring io for $6.5 billion in OpenAI shares.

All delivered in an extremely cinematic, over-the-top, ~9 minute video of two of the world’s most influential tech figures saying with words that they have inimitable connection, while somehow demonstrating absolutely zero actual rapport.

The fact that this odd love story is delivered with nary a hint of human warmth or true affinity probably isn’t due to a lack of connection between these two giants of consumer technology; it’s much more likely that it owes more to the fact that both of these two have achieved mythic status and wealth, and it’s hard for relatability to survive something like that. And regardless of what you think about their interpersonal relationship, the potential of a mashup between the man who designed the single most popular consumer hardware device in all of history and the guy who built the fastest-growing consumer application in history, is off-the-charts interesting.

While we do get some gorgeous shots of San Francisco in this video, and a lot of weird eye-line and reaction shot cut choices in the conversation between the two men that makes up the bulk of the clip’s running time, what we don't get is any actual look at what they’re building. Sam does reference a prototype he says he’s been able to actually use, and definitely doesn’t get hyperbolic with his assessment:

“I’ve been able to live with it and I think it is the coolest piece of technology the world will have ever seen.”

Ive, meanwhile, expresses an insight that I’ve had, and that so have many others, including members of his former employer Apple who went off and built the Humane AI pin:

“The products that we’re using to deliver and connect us to unimaginable technology, they’re decades old, and so it’s just common sense to at least think, surely there’s something beyond these legacy products.”

I don’t mean to criticize Ive for pointing out something obvious – one of his strengths as a designer and maker of iconic products is that he’s not afraid to dive into those ‘obvious’ truths and help express those insights in the things he makes with exactly the right approach to make them strike their users as the ‘obvious’ default, too. But it is maddeningly faint broth for a ‘big reveal’ that contains no real revelations.

One thing I will criticize is when Altman goes into informercial pitch person mode to describe in extremely overwrought terms the current state of the art of working with AI. From the clip:

If I wanted to ask ChatGPT something right now, about something we had talked about earlier, I would reach down, I would get out my laptop, I’d open it up, I’d launch a web browser, I’d start typing, and I’d have to, like, explain that thing. And I would hit enter, and I would wait, and I would get a response. And that is at the limit of what the current tool of a laptop can do.

Is that really what you would do Sam? Can you not think of a better way to ask ChatGPT a question mid-conversation with a friend in a café? No? Here’s a hint: The man you’re having the conversation with CREATED THE DEVICE that makes that so much simpler and less convoluted! Oh and also your company has a partnership with its maker that bakes ChatGPT directly into the system at the OS level??

Yes, I mean the iPhone.

Set aside that even if you do go with Altman’s method, it’s not actually that complicated a process, and the answer engine will still provide you what you’re looking for with remarkable speed and efficiency – he’s also purposefully choosing the most complicated path when a much easier one is right in front of him. It reminds me of those infomercial segments where people struggle with even the most basic tasks, or with Khaby Lame’s entire oeuvre, for a more contemporary example.

I still do fully agree that we haven’t yet figured out the right approach to take advantage of all that generative AI can offer with our current consumer hardware, as I’ve expressed many times. And I’m still super excited to actually see what these two come up with, to be very clear.

One final note is that it’s extremely entertaining to see the level of petty on display among the foundational model makers right now. This past week alone, we had Google basically fill everyone’s brain buffer in terms of how much AI news we can all process with their I/O developer conference, and Anthropic introduce the next generation of Claude with its Claude 4 Sonnet and Opus models.

Scrambling to one up one another is par for the course for AI model makers these days, but OpenAI pulled off a special coup by announcing one of the most interesting partnerships in technology period during Google’s I/O week while hijacking its own keyword: Not sure if that’s the reason Ive and Altman landed on ‘io’ as the company name for their hardware joint venture, but it certainly helped inform the timing of the drop.

Watch the full announcement in all its self-indulgent glory here:

Betaworks’ Camp program is evolving alongside the emerging AI economy, and its latest cohort is another iterative shift: It builds on the last group’s focus on AI-native product experiences, but now moves into AI-powered applications that appeal specifically to enterprise, SMBs and future of work. The class of nine startups (the first of two cohorts betaworks is running this year) includes a wide-ranging field of founders and concepts, as is typical for the firm’s wild and wonderful approach. All of these participated in the official Camp demo day last week, which I unfortunately couldn’t make, but luckily my great pal and betaworks partner Jordan Crook filled me in, so here’s a rundown of the class for the first half of 2025:

Decode (https://getdecode.dev)

What is the whiteboard-to-code pipeline was pretty much instant? That’s what Decode proposes, with an IDE-integrated white boarding app that takes an alternative approach to moving up a layer of abstraction from the more established set of ‘vibecoding’ products like Cursor and Replit.

It’s an interesting approach and could find a sweet spot in terms of customer appeal among the class of front-end developers who are also design-oriented and/or visual thinkers. Plus, an approach that could eventually scale to offer a platform that best represents an approach to AI-powered coding which makes maximal use of spacial relationships as a core concept sounds like a solid differentiator.

CEO and founder Francois Laberge was previously Principal AI engineer at popular AI-based podcast editing tool Descript, and founding engineer Sriraam Raja has a masters from Harvard in the intersection of AI and education technology.

Trampoline (https://trampoline.ai)

Responding to RFPs is a massive time- and effort-sink, as anyone who has ever done one will tell you. It’s also often highly formulaic and repetitive, especially in the case of government contracts, and others in highly regulated industries. Trampoline aims to automate the RFP process, and says it can save between 30 to 80% of a company’s time as a result of doing so. This can help unlock new potential markets and customers for smaller companies that previously couldn’t compete with large firms on RFPs strictly as a result of not being able to dedicate resources to responding.

The team has ample experience in RFPs across a range of industries, and the startup has support from some of Canada’s top AI institutions, including The Vector Institute and Creative Destruction Lab.

TabTabTab (https://tabtabtab.ai)

One of the areas where I think there still remains a ton of work to do, and also a ton of opportunity to exploit, is in interaction paradigms involving AI and its use across computing flows. TabTabTab is a startup aimed squarely at that challenge, and is building a layer for everyday computing that resides above the OS and connects across all your various applications and productivity software.

I’ve definitely seen other startups and companies taking this same approach (including Highlight, a portfolio company of my previous employer OMERS Ventures) but this one has a great founding team in Gyanendra Mishra and Vasin Wongrassamee, two engineers with experience across some high-impact information and intelligence delivery organizations, including Bloomberg, Palantir and Citadel. Plus, this is an area where I think we’re in very early days, and I expect novel solutions to come from multiple vectors and surprise us in terms of how much they differ from computing paradigms in past eras.

Superposition (https://www.superposition.ai)

Perhaps it was because one of the investment areas that OMERS Ventures focuses on, but I feel like I’ve seen a huge number of AI-powered recruitment startups, and Superposition is one of those. There are some obvious reasons why recruitment is a category that seems ripe for AI-based innovation: Similar to sales, another AI hotbed, it’s incredibly time-consuming, and huge parts of it are rote and repetitive.

Superposition’s unique differentiators vs. the Mercors and Moonhubs of the world include that it’s agentic first, and also that it inherits the taste, experience and perspective of its founder and CEO Edmund Cuthbert, a longtime recruiter who hired Brex’s first engineer. Superposition is also specifically narrowing its focus to startup hiring at launch, which could offer it a unique wedge in a sea of competitors mostly aiming at larger and more established enterprise buyers.

Hopper (https://hopper.dev)

AI coding is undeniably a massive success and directionally the future of software engineering. But many of the tools currently on offer actually start to break pretty quickly when they encounter the complexity and consistency requirements of large, multi-departmental organizations where product and codebase development occurs across and between teams and projects. Hopper is an AI planning companion that helps shape the process from ideation to detailed tickets, keeping in mind both organizational context and technical audits from the perspective of the organization as a whole.

The idea and the expertise to build it come from founders Nathan Ackerman and Austin Riedel’s experience at big tech companies including Amazon, Microsoft and IBM.

Afterimage (website forthcoming)

Seeking remedies after something goes wrong with your own health is an intimidating task in the best of circumstances, and generally if its come to that, the circumstances definitely aren’t ideal. Afterimage uses AI to comb through patient data and find things that may have been missed – sometimes amounting to malpractice – and then helping with the rest of the process, including drafting complains, and matching patients with legal representation suited to their case.

It’s a very interesting approach, given the current state of AI-powered diagnostics and medial imaging analytics. Many reports are suggesting that AI analysis of scans and medical images can at times be even more effective than human diagnostics alone, which does raise the question of what that means not only for lab analytics going forward – but also for uncovering past cases where care wasn’t adequate as a result of (sometimes avoidable) human error.

JigsawML (https://jigsawml.com)

Mapping and maintaining enterprise software architecture is a wildly expansive, never-ending exercise that can feel like firefighting combined with wrangling cats. It’s another one of those areas that are absolutely essential at a certain scale, but also not anyone’s favorite job. As such, it’s also a perfect area to explore in terms of automation, and how AI can help with that beast of a task.

Founder Pracheer Gupta was a lead engineer at AWS as it built-out the precursor to modern vector databases, and then was a founding member of the Pinecone team when they set out to productize vector databases for AI applications.

NetAssist (https://netassist.live/seniors)

Using the insight that AI represents a new layer of abstraction when it comes to computing and computer interaction, NetAssist is developing voice and text agents that make using web-based technologies easier for senior citizens. Increasingly, community services, resources and connections are exclusively or optimally available through online and digital services, but these can be a major barrier to entry for an aging population.

Founder Anthony Hagouel once developed AI technology solutions for the largest commercial shrimp farms in the world (yes, really) but was driven to build NetAssist after having personal exposure to the capability gap between seniors and the online services they require within his family.

Graze Social (https://graze.social/)

Social media is in a state of relative disarray and flux compared to the past several years, with major incumbents seeing their dominant positions begin to erode, and new protocols like ATProto and ActivityPub picking up in terms of both interest and audience. But this new, more disparate and distributed world is a very different one when it comes to building consumer products, and also when it comes to monetization and building sustainable businesses on top, or else incorporating them into existing businesses as net additive contributors.

Graze offers everything you need in one place to build custom feeds on Bluesky, and tools to help grow your audience, as well as to set up a sponsored content marketplace to help with creator revenue opportunities. Founder Devin Gaffney is an experienced machine learning researcher and builder at the intersection of ML and online social activity.

And that’s the class for this first Camp of 2025! Shoutout to Jordan, John Borthwick and Jon Chin for always picking a super interesting batch of startups.

I’ve been a little quiet here for a few weeks, and now I’m finally ready to share the reason: I’ve got a new gig, with GrowthX, an exciting early-stage startup that blends perfectly the areas in which I can have the most impact, and my beliefs about where the future is headed in terms of AI technology development and tool use. Today, GrowthX is making its $12 million Series A (led by Madrona) official, so I figured it was a great time to share a bit more about the company, some of the people involved, what it’s doing, and what I’ll be doing with them.

First, GrowthX’s focus is (unsurprisingly) growth – the growth of its clients specifically. Its approach to doing this blends a traditional services-based, agency-like model with AI-powered content creation and optimization, along with a flywheel of rapid iteration based on feedback and quick development and deployment of new technological tools.

GrowthX founder and CEO Marcel Santilli has described the approach as “service-as-software,” and it’s a concept that is gaining more purchase and traction particularly in the era of AI. Historically, venture capital has been more or less allergic to companies that resemble service businesses rather than ‘true’ tech or software startups, in large part because the revenue multiples on services businesses have typically not been able to take advantage of the inherent scaling benefits of software: Software is code once, deploy effectively infinitely – services businesses on the other hand, have historically required scaling spending on the capital expenditure side more or less in lock step with revenue growth, because new clients require new people, not just new copies of the same code.

One of the earliest companies to challenge the long-held maxim in VC that ‘services = bad’ was Palantir, which famously relied on a strategy of using ‘forward-deployed engineers’ (essentially embedded contractors) within clients to build alongside them, blending a services approach similar in style to traditional providers like KPMG or Deloitte, with a SaaS model for the resulting tech stack or tool build. Building somewhat inefficiently at first like this also nets you new product that you can adapt and resell to other clients, too, meaning that despite an initial approach that looks like it won’t scale on an exponential curve, you end up with revenue multiples that better fit what a venture investor expects to see.

The broad spread and rapid iteration of generative AI has made the ‘service-as-software’ approach both more attainable and more scalable from even early on in a startup’s lifecycle.

Knowledge work, and in particular work around content marketing and brand communications, seems especially well-positioned to benefit from generative AI in terms of building scalable, effective, and powerful tooling that can dramatically change how companies talk to – and serve – their customers. Even before the opportunity to work with GrowthX came along, I had been personally working on and experimenting with how my own output might be scaled using AI, and in fact, even ideating what it might look like to build an agency on modern AI tooling doing exactly that.

The other thing that really appealed to me about GrowthX is that it’s starting from a position where the product it's delivering is valuable enough to customers out the gate that it’s already profitable – another factor that has sometimes been greeted with skepticism from VC as a category. But because GrowthX offers a marketing stack that’s AI-enabled, but backed by world-class editorial and SEO talent, it’s already winning on value vs. traditional agency approaches. Plus, like Tesla’s efforts with autopilot, the great value delivered on a cost basis for clients helps fund ongoing tech development and continuous improvement – leaving everyone in the equation happier and better for it.

As for my role there, I’ll be working closely with my old boss (now also my new boss) Matthew Panzarino, the former Editor-in-Chief of TechCrunch, as well as fellow TC alum Megan Rose Dickey; ex-TripAdvisor/Red Ventures SEO expert Dave Capone; former Business Insider senior reporter Mara Leighton; ex-G2, ActiveCampaign and Scribe content pro Jakub Rudnik and ex-Semrush growth marketing leader Branko Kral. Together, we make up the team responsible for ensuring our client delivery teams are delivering best-in-class results that best represent our customers to their customers. Given the nature of the startup, I’ll also be working closely with our engineering teams on product ideation and development.

I’m super excited to be joining a company this early in its journey, and to be joining one where I have such strong faith in its vision, mission and direction. It’s definitely going to be a new challenge for me relative to what I’ve done before, but it’s one I welcome – and one where I think my skills and experience will be especially useful.

You’ll probably occasionally hear from me here about what’s going on at GrowthX, just because it’s a genuine thrill ride and I won’t be able to control my excitement. But know also that I’ll still be maintaining The Angle with its current mission of documenting the development of the AI economy, and the startups, investors and companies that factor therein.

Also, if you think GrowthX might be a fit for your company (or those of your portfolio, VCs reading this), please do reach out – happy to chat about what it is why do and what we’ve been able to do for our existing clients so far.

Redcar founding team, including (left to right) Cataleya Jiang, Ben Wu, Jeff Chen and Aidan Lynott

Sales enablement and go-to-market process augmentation has been one of the early beachheads for B2B AI tools, and Redcar, a two-year old startup based out of SF, fits squarely in that category. The company is building AI sales agents that take care of a lot of the repetitive, fundamental steps involved in the sales process so that the people doing the selling can actually focus on the close.

The company revealed $5.3 million in funding today, including a $4.3 million seed round led by Khosla Ventures and including participation from Greylock, and a $1 million pre-seed from Humbition, as well as individual checks from some heavyweights including Mitch Kapor, Siqi Chen, Nancy Xu and Elias Torres. It’s a remarkable list and a sizeable early financing, in a crowded category, but speaking to two of Redcar’s co-founders – CEO Jeff Chen and Head of Growth ‘CJ’ Cataleya Jiang, it’s clear why they were able to stand out from the competition.

Chen is a serial founder whose previous company Joyride built Android’s top voice assistant, Skyvi, and which exited to Google in 2015. He launched Google Assistant with the company, then eventually moved on to build another startup, Taste, which was a part of Y Combinator’s winter 2021 batch. It was at YC that Chen really dug into the sales process.

“One of the rites of passage for founders is how hardcore you are,” Chen said. “And you demonstrate this ‘hardcoreness’ by showing how many cold personalized emails you can send per day[…] I wrote 200 a day.”

After a while of doing that as part of a two-person team for Taste (with Redcar co-founder and Chief Software Architect Ben Wu), Chen decided it might be better to outsource it, and contracted someone to handle it for him for $5K per month. That ended up netting only one meeting with a potential customer, so he rolled up his sleeves and took over again himself, landing about 50 such conversations. It wasn’t sustainable, but something just over the horizon would make it manageable – as well as scalable.

That was the arrival of the large language model, which instantly struck Chen and Wu as something that would provide a much better and more transformative opportunity than what they were originally working on at YC. Redcar grew out of Chen’s experience with trying to stock the funnel for their prior company, combined with the efficiencies that LLM-based AI brought to the table in terms of handling a lot of the manual lift involved in that process.

Redcar helps B2B companies scale sales teams by providing building blocks to create tailored AI agents, letting human reps focus on closing deals. There are three core components to the product: custom research agents, prospecting, and personalized outreach. Part of the startup’s unique value prop is the degree to which these components can be customized to the customer, and to the sales rep, and combined in different ways to best match their workflow and particular sales motion.

Chen and CJ describe the approach as a ‘verizontal’ one, which describes the way in which packaging and coordinating a few vertical tools aimed at specific functions within the sales process (prospect research, for instance) ends up providing a much more broad, horizontal solution.

"If you combine a few vertical tools together, it starts becoming like a bundle,” Chen said. “That's what we think is a 'verizontal.' The better AI gets, the more we can make more software, and the more bundling happens.”

It’s an approach that could be (and is being) applied to a number of different problem areas, so I asked Chen why they started with sales as the immediate focus for Redcar. Chen explained that in addition to his own experience and frustration with the sale process, in part, it’s because of how the current state of the technology fits the shape of the challenge.

"Language models are probabilistic by nature,” he said. “You want to be in a space where you don't have to get things perfectly right – where 99.9% is good enough."

As for why CJ is interested in tackling sales and GTM specifically, she said that it’s surprisingly similar in some ways to her background studying theoretical mathematics in school.

“It's very logical, I really enjoy it,” she said. “It's almost like a puzzle piece you solve – you gather all the information and then it’s putting the blocks together.”

Creating and assembling this blocks is exactly what Redcar is all about, but Chen is clear that it isn’t about replacing the human elements of the sales process yet. He sees it as upleveling the profession in a way that has ample historical precedent.

“In the sixties, we had human computers [...] mostly women of color who did tedious mathematics by hand,” he noted. “When machine computers got invented, those same women became scientists and engineers – now we can do the same for our generation.”

Like many AI builders I’ve spoken to, Chen is partial to the analogy of his technology acting as an Iron Man suit for the human operator inside – it’s a technology that can take any ordinary person and put them on a level with a superhuman peer or adversary. Tony Stark without the suit isn’t going to be able to go toe-to-toe with Thanos, but the suit can make it so that anyone can stand a chance. Redcar, similarly, aims to make every sales person a top performer – all while letting them focus on the most rewarding part of the job.

"Closing is fun,” Chen said. “Who doesn't like to win a deal? But to get to that close... you have to persistently follow up, make your pitch deck, talk to the person. We're basically setting these people up for success so they can enjoy the fun parts."

People who I speak with often know I’m extremely bullish on this being the time for an exciting new period in the evolution of consumer hardware (and, at the same time, software and experiences). I’ve written before here about some of the recent entrants in the field focused specifically on building consumer devices from an AI-first perspective, including Humane (RIP) and Limitless – as well as rabbit, the company behind the r1.

I’ve also talked about how rabbit has generally improved the experience of r1 iteratively since its launch, despite stumbling out of the gate due to a mismatch between customer expectations around the product’s capabilities as described in pre-launch marketing materials and the reality once it was in their hands. Humane suffered from a similar affliction, but in its case the wounds were fatal – mostly because of the different approaches both companies took to getting their products to market.

Rabbit has also been very good about post-launch software iteration and development, as well as listening to, and engaging with their device owner and user community. They also landed some quick wins with sticky and charming features like ‘magic camera,’ which created pixel art-style renditions of photos taken with the device to be discovered and shared with a slight delay, almost like waiting for film to develop. And they also continued to work on their concept of the ‘Large Action Model’ or LAM, which was essentially an agentic AI system intended to perform tasks via the web on the user’s behalf before ‘agentic’ was the AI buzzword of the day.

In addition to the LAM, rabbit has continued to add new agentic experiences to its capabilities, including a way for users to teach it how to perform actions via ‘teach mode,’ and an Android agent that can effectively use Android apps on a user’s behalf. Now, it’s combining those capabilities and adding new ones to launch ‘rabbitOS intern,’ a new version of its operating system designed to capitalize on the clear appetite for consumer agentic products demonstrated by tools like OpenAI’s Operator and China’s Manus, while also charting a path forward for rabbit in terms of company evolution and revenue opportunities.

RabbitOS intern is aptly named for a couple of reasons, according to the company: First, it’s explicitly described as potentially “error-prone” and somewhat slow at launch. Second, it’s meant as a first step on an interactive path that should see it up-level its skills to eventually graduate to associate, mid-level, senior professional and beyond. It’s described as being able to perform tasks across “a variety of domains,” including doing this like creating a real estate research report for a specific area, making a fully functional (though simple) game in Three.js, creating a website for a TV show and coding a simple chiptune music sequencer as a web-based application.

Crucially, rabbitOS intern isn’t tied specifically to the company’s r1 hardware: At launch, it’ll be available as a free trial to all at rabbit’s ‘rabbithole’ website with some usage limits (with r1 owners getting more use). During the trial, the company will be evaluating pricing and plans for its eventual paid launch. The r1 launched as a product whose capabilities would specifically be available without an ongoing subscription, and those will remain with this as an optional upgrade once the trial period is over.

Rabbit says explicitly that rabbitOS intern is designed “to eventually power rabbit’s future products,” but also goes further to say it’s meant to also “integrate with any digital interface and other compatible devices,” which is why I suggested above that it feels like an Android-style approach to a future where mass market consumer experiences are primarily powered by AI under the hood.

I’ll caveat all of this by noting that I haven’t yet had the opportunity to try rabbitOS intern myself, and so can’t vouch for its state of ‘bakedness’ or capabilities. But what I like about it is what I’ve always liked about rabbit’s approach: It’s a novel and interesting, if very nascent, take on what the future could look like with entirely new device and software paradigms that use AI as the basis for their design and development, rather than just as a supplemental or additive feature.

I think that, as critics are extremely quick to suggest, it’s still much more likely that consumer adoption of AI will broadly happen via existing device and platform paradigms like iPhone, iOS and Android, used as the vehicle for AI software delivery both through OS integration and over-the-top through consumer apps. But while it’s less likely overall, I think it’s far more exciting – and potentially far more lucrative, from a market and investment perspective – to imagine and build towards a future where AI-first devices and software topple and supplant the relatively fixed and mature categories of devices like smartphones, laptops and computers.

As a company, rabbit is actively trying to create that future, and I think should be lauded for doing so, especially given that they launched their hardware with a relatively low selling point and no ongoing subscription commitment required, while iterating quickly and effectively post-launch. With rabbitOS intern, it’s debuting something that it admits is early, but that has the potential to become something big, and it seems to be doing so with even more care paid to not making its customers bear the brunt of the cost of testing and building something new in public.

If we get to a world where consumer AI devices and experiences are unshackled from our conventional, pre-AI devices and interaction models, it’ll be because of efforts like this to advance the state-of-the-art while also avoiding the alienation of people who don’t want to feel like they’re paying for the privilege of alpha testing the future.

AI image generators have already been impressive for quite a while, and iterative improvements in things like Stable Diffusion have added up to a fairly capable creative tool, but one with very clear, hard to escape limitations. This week, OpenAI launched a new image generation tool that’s built into its ChatGPT 4o model, which leverages technology developed for its Sora video generation tool. It’s a little bit slower than the DALL-E model it replaces – but in exchange for a slightly longer wait, you get vastly improved generation capabilities that seem to avoid a lot of the downsides of AI-based image generation that previously seemed unavoidable.

Social media was flooded with renders created with OpenAI’s new image generation tool after its official launch on Tuesday, and the majority of them were the result of so-called ‘style transfer’ – applying a specific aesthetic or set of associated visual characteristics to an existing image. During its launch stream for the new image generation features, OpenAI’s own staff, including Sam Altman, demonstrated a style transfer example live, using a ‘Studio Ghibli’-inspired aesthetic to transform a selfie shot during the presentation.

Studio Ghibli is the iconic Japanese animation studio behind classics including Spirited Away, Kiki’s Delivery Service, My Neighbor Totoro and Princess Mononoke to name just a few. Its most iconic director and co-founder is Hayao Miyazaki, whose thoughts on artificial intelligence and its use in creative endeavors went viral about a year ago, but are circulating again now given the irony that the signature look and feel of his studio’s visuals are now the hallmark of a new era of capability in AI image generation. In case you haven’t seen, Miyazaki is very much not generally in favor of using AI to generate creative works.

As a photographer and artist myself, and someone with a lot of close personal friends who are also working artists, illustrators and photographers, I find my reaction to products that can potentially replace huge swaths of their livelihoods viscerally repulsive. I’m also obviously, at least in part, still a professional writer by trade. Leaps in performance and capability like the one made by OpenAI this week with 4o image generation are a shocking, painful and existentially daunting occurrence which no amount of intellectualizing or silver-lining thinking can entirely ablate.

Conversely, there’s an essential type of joy that comes through when people are using these tools, especially when they themselves don’t possess the kind of artistic talent that would let them produce anything approaching this level of quality on their own. Exceptional talent and skill is obviously a core human trait and one to be celebrated – but so, too, is democratization of domains and experiences that once had a high threshold to entry, but that are made more accessible through technological advances.

The fact is that with its new image generation tools, OpenAI has crossed a threshold whereby AI-powered visual asset generation can now do, in minutes, a significant part of what it used to take a professional hours to accomplish working with expensive dedicated software like Illustrator, Canva and Photoshop. Past improvements have been more iterative, providing slow but steady improvements but keeping the event horizon of when the tooling could replace significant portions of the creative workflow pretty far out.

OpenAI’s 4o image generator will be transformative not so much because it replaces the need for a creative professional and attendant software in so many contexts, but because it will actually end up adding an entire new layer of abstraction on top of that process. Just like the modern visual operating system simplified the concepts of interacting with a computer vs. text-based system UI configurations like MS-DOS, so, too, will OpenAI’s current and subsequent iterations of visual generation tools radically change how we think about where that sits in the stack of modern user-facing applications.

There are a lot of signs that a technology or industry has ‘arrived,’ meaning that it’s reached a point where it impacts things in a way that can’t be reversed and that’s far-reaching. One is when it earns enough attention to make a large, lavish and well-crafted industry event viable. Human(X), a new conference happening in Vegas right now (which is where I am right now) is that for AI – loosely defined and broadly applied.

The show is put together by a group of people that includes creators of many other very successful trade shows, including Money20/20, HLTH and Shoptalk – and the benefit of that combined experience shows in everything from the quality of programming to the fit and finish, and also in how well-run it is. Having helped organize and run TechCrunch Disrupt for 10 years, I can safely say these folks have got it right in many ways.

The other key ingredient to a successful show is of course appetite from the market, and AI has that in abundance. What’s more impressive, and what makes a show like this so challenging to pull off, is that ‘AI’ as a concept is as loose a connective tissue as you can come up with when it comes to an organizing principle for doing a centralized thematic event. It touches every industry, and its adherents focus on everything from the physical hardware and infrastructure that makes it possible, to the UX of consumer-focused products, to the knock-on effects of workforce transformation that results from its uptake.

From my perspective, Human(X) is doing a good job of both delineating and connecting the various disparate streams that play across AI. It’s definitely a good sign that people I’ve talked to on the ground have repeatedly cited FOMO for cross-programmed events and discussions as one of their main complaints about the show so far – there’s clearly a lot of value in a cross-discipline, cross-industry approach.

Switching to my main takeaways from day one, they fall into two categories:

1. Human replacement in job functions is a hot topic, and seems to have crossed the Rubicon from a controversial topic into a clear desirable, at least for this crowd of mostly business and technical executive leaders

2. There’s a lot of buzz around the potential in latent data, and how to uncover existing stores of that, how to capture it, and how to assess and quantify its value

First, regarding human replacement, it’s come up as a topic in most of the sessions I’ve managed to tune in to. Yet there’s a general optimism about it as far as how it will impact the workforce and job opportunities. OpenAI Chief Product Officer Kevin Weil, for instance talked a lot about our historical tendency to adapt to the introduction of massively impactful new technologies that wipe out entire job categories or professions.

Even in conversations where the emphasis is on human augmentation rather than replacement, if you pull the thread it’s clear that the aim is to replace the productivity of ten people with the enhanced productivity of one. Past major technology shifts have of course been about value creation, but AI seems to be the one in which human headcount reduction (at the individual company level) is becoming unambiguously the goal.

Reading that back, it sounds like it contains an implied value judgment: it does not. I don’t think that in isolation job creation or elimination by AI is inherently bad or good. Humans are, collectively and individually, resilient and resourceful, and I think that as much as there are many potential negative unintended consequences of broad AI deployment, there are also many unseen or unpredictable positive outcomes, too, including the creation of entirely new job categories, modes of productivity and maybe even socio-economic organizational models.

On the second point, latent data is indeed an extremely exciting and interesting opportunity, especially as it pertains to the tech, VC and startup scene. The potential for existing, untapped data sets to create new value in businesses that may be stagnant or negative in terms of their primary growth trajectories is significant. Plus, the ways in which we might capture new and unexploited repositories of data that could enable new types of AI models and applications is very exciting, in terms of what it means for new venture formation and new data capture tech.

This was a major point of discussion in the panel I hosted on building AI infrastructure that’s suited for C-suite adoption, which included Honeycomb.io CEO and founder Christine Yen, Google Head of AI Developer Assistant/Agent Bin Ni, Roam Home founder (and August Home founder previously) Jason Johnson and McKinsey Senior Advisor Brian Goffman. Once the recorded version of that is available I’ll definitely share it here.

I also spoke to David Cox, VP, AI Models at IBM Research, and we had a great discussion about IBM’s latest Granite 3.2 enterprise-oriented foundation models, DeepSeek and its broader implications, open source and what that means for AI, and much more. Likewise, I’ll be sharing that when available.

More to come from day 2 and 3, and also stay tuned for thoughts on Cerebras and a few announcements they made at the show, along with takeaways from an illuminating chat with Cerebras founder and CEO Andrew Feldman.

Generative AI has done more than maybe any other technology to help solve the ‘blank page’ or ‘cold start’ problem when it comes to content production and idea generation. It’s the first time in history that we’ve developed a technology that can capably act in a co-creator role with human beings, as opposed to just an enhancement or accelerant to human origination. But it still operates primarily on a pull model – you have to prompt or query it in most of the dominant user experience paradigms we have for generative AI today.

On a recent episode of the always excellent No Priors podcast, Harvey CEO and co-founder Winston Weinberg articulated an under-appreciated point about how we learn from experts: We gain the most from smart people around us by just spending time literally with them around us, rather than by proactively going to them and asking them specific questions in isolated bursts, exclusively inquiring along a self-directed vector.

There are a number of limitations that arise naturally out of a dynamic where you have a non-expert or non-specialist person trying to learn from someone more knowledgeable in a specific area. One of those is commonly referred to as ‘not knowing what you don’t know’: Basically, you’re not going to necessarily ask the right questions or seek out the most relevant information from a source if you’re not already at least somewhat familiar with the lay of the land. Somewhat related is the problem of contextual gaps, or not having a fulsome picture of everything that surrounds a given answer to a specific question and therefore mis- or under interpreting the response.

Given the era in which generative AI emerged, and the paradigms that it was best set up to compete with, a query-based interface made the most sense for general use tools, including ChatGPT. It’s still also the best way to use AI for most people most of the time, for most tasks. But in the category of general enrichment and knowledge acquisition – a category I’d argue is growing – I think other models stand to offer considerable benefit vs. a request-based chat interface where most of the onus is on the person asking questions to know what’s important to focus on.

Some more recent AI-based products approximate what I think of as a more autonomous knowledge companion: OpenAI’s voice mode is good at prompting additional conversation through clarifying questions it appends to the end of its answers, for instance. Google also has its much-vaunted NotebookLM, which can generate topic-specific, naturalistic podcasts on-demand from whatever materials you choose to feed into it.

NotebookLM’s subject matter boundary (which is determined by the materials you put into it) is a feature not a bug for its intended use cases, which range from study guides to research and writing assistance. Yet it also makes it more prescribed than I think is strictly useful in a situation where your goal is to improve your general familiarity on a particular topic, in a way that is less about a specific directed outcome and more about building context.

I think a rich, under-explored product model for delivery of generative AI would be a kind of library of experts, which can perhaps be queryable much like ChatGPT or Claude are today, but with a very different primary delivery method that’s actually much more like traditional terrestrial radio or broadcast television.

Key to this working well would be that the AI expert is essentially self-directed in its knowledge sharing – perhaps responsive to prompts or queries in terms of very generally orienting its ‘thinking' out loud, but mostly meandering down a path guided by its own ‘interests’ as they pertain to its discipline. The experience should be more or less equivalent to asking, say, a professor of contemporary poetics what they’re interested in or paying attention to right now in their field, and then maybe asking them clarifying questions when they venture into jargon, or seeking expansion in case they hit on something you find particularly fascinating.

As I mentioned earlier, some tools out there now do resemble this very closely, but I don’t think it has received enough attention as a ground-up, intentional design exercise with the intent of building an AI interface that is, from first principles, about spontaneous and self-directed sharing of expertise.

If, however, I’m wrong and someone is already building this out there and you know about the project, please do point me in their direction because I’d love to see what they’ve found.

The launch event for the Humane AI pin was basically a pageant of hubris, a pantomime of an Apple event that proved almost all style and very little substance. To be clear, I’m not critical of Humane founders Imran Chaudri and Bethany Bongiorno’s ambitions: the concept behind Humane was truly trailblazing, with bold bets on emerging technologies and novel approaches to consumer devices, a category otherwise infected by a rictus of lack of innovation.

What I didn’t love about the Humane approach was the bombast – the slick presentations that cribbed from Apple’s shiny, perfect and fully realized marketing machine. Chaudri and Bongiorno cut their teeth at the iPhone maker, so it was somewhat inevitable that they’d borrow from that playbook. But just about every one of Apple’s consumer devices is a very mature product, from industrial design, to software experience, and everything in between.

Humane’s AI pin was basically the opposite.

Despite the veneer of professional hardware finishes, Humane’s pin was a lab prototype masquerading as a shipping consumer product. It was breaking a lot of new ground with respect to user interaction paradigms, and the fundamentals of what people expect their personal devices to be able to do for them. It was also doing all of this on top of a foundation made up of emerging, non-deterministic AI technology that was itself extremely rough around the edges and shifting in terms of both capabilities and performance on a daily basis.

Compare and contrast Humane’s AI pin launch with that of its closest contemporary, the Rabbit R1. The latter was also a brand new piece of consumer tech, which aimed to ride the wave of generative AI to produce real value for its users in a novel way. Like Humane, Rabbit definitely engaged in some stunt marketing around its launch, with founder Jesse Lyu delivering a vaguely Apple-like ‘keynote’ presentation at CES 2024 when it had its official launch.

The difference was that where Humane’s approach was self-serious and imbued with an unearned level of professional polish, Rabbit’s was playful, purposefully experimental and generally more resonant with early adopters who had at least some grasp on the trade-offs inherent in entirely new product categories.

Rabbit still overpromised and underdelivered. But it had a much lower cost of entry (not to mention a lack of ongoing subscription commitments), less bombastic claims about its functionality at launch, and what seems now like a better course of iteration following its official launch to help its real-world performance more closely resemble its pre-launch claims.

Like I said, I don’t begrudge Humane’s team their big swing: Especially in consumer tech, ambitious, wild and truly novel ideas are precious and few. But my friend and former TechCrunch colleague, Greg Kumparak texted me to express concern that Humane’s failure will dampen interest from others who might otherwise have tried similarly interesting things, and that captures perfectly my primary concern.

Generative AI has been a game-changer in many ways across the broad landscape of tech, but I think it still has massive unexplored potential in consumer devices. Yes, Android and iOS are incorporating it more and more, but it still feels like just another feature addition (and a sometimes very poorly executed one at that). Plaud’s AI note pendant and the similar Limitless AI wearable are perhaps more successful (if less ambitious) examples of AI-first consumer hardware, but there’s still some undiscovered paradigms out there that won’t be found without a healthy appetite for risk from more brave explorers.

At this point in the AI arms race, there are a few new things that are considered table stakes – basically while they also remain the bleeding edge of the state of the art. Chain-of-thought is one, and the other is ‘deep research,’ which comes under different branding depending on which model maker you go to, but which amounts to an undergraduate level introductory topic-based survey paper about whatever you want to learn more about.

xAI’s Grok 3, which is available now for Premium+ members, has been top of mind for Elon Musk – as much as anything is able to occupy his attention for long these days, while he’s busy gutting the US government and now apparently also trying to take on ‘woke’ gaming with his own studio.

Rolling a few products into one, Grok 3 now offers a “DeepSearch” option for queries, which resembles both a search engine-style tool for sourcing information from the web like OpenAI’s search feature and Perplexity, as well as the more thorough ‘deep’ search modes that both offer. In practice, Grok 3’s implementation of this tool is much faster than those of the two competitors mentioned above, but it also seems like it produces a much more high-level, casual overview vs. the other tools (though that might be an artifact of the time it takes and the way it articulates its process and what it’s currently working on more than the actual result).

Grok also now has a reasoning mode called ‘think’ that checks its answers as it goes, but it doesn’t articulate its thoughts like some competitors – an intentional decision that Musk says is to prevent ‘distillation’ by competitors like DeepSeek. One of the prevailing theories about how DeepSeek was able to achieve its incredible cost/performance efficiencies was that it distilled (using a large, complex model’s answers to train a much smaller one with relatively little quality loss) from OpenAI’s most powerful models.

Musk claims that Grok 3 is the best AI chatbot in the world, but of course whether that’s accurate or not depends on who you ask. It is definitely receiving some favorable reviews, and performing well on benchmarks thus far. Andrej Karpathy, one of OpenAI’s founding team and a former director of AI at Tesla, has a good rundown of how he found it performed relative to the other state-of-the-art models available right now. Consensus seems to be that OpenAI’s best models still outperform the Musk-backed’s best efforts, however.

I’m less interested in how Grok 3 stacks up to the other AI chatbots out there, and more in what this release tells us about the direction of the industry and what matters (at least in the eyes of the AI foundation model companies) when it comes to product iteration.

It’s definitely clear by now that all these companies will need to offer the ability for their tools to ‘think,’ ‘search’ and ‘research.’ Consider this relative to the original launch state of GPT-3-powered ChatGPT, which for the purposes of this article we can call the starting point of this product category. The original ChatGPT could do none of these things in the way we understand them today.

There’s a desire and a tendency to see these new implementations of the various generative AI models as simply ‘repackaging’ efforts that simply obscure scaling limits, or that minimize and mask but don’t eliminate fundamental challenges in the tech, like hallucinations. These takes underestimate the real utility value of thinking, researching behavior equipped with real-time access to current information in order to confirm a pre-existing bias that because AI can’t directly replicate human thought and intuition, it’s therefore useless.

There’s no question in using these new incarnations of AI that they’re actually massively useful, and that they’re far less susceptible to the failings of earlier versions. Most importantly, these gains were realized only in part on the back of scaling the number of resources expended in their development and during their use; recombination of the elements involved, and the addition of new and innovative pieces to complement the existing recipes, have made all the difference. And will continue to – next year’s capability set is probably just as difficult to predict as this year’s baseline was at the outset of 2024, which is incredibly exciting.

The financial markets are currently having a very broad and fairly deep reaction to the technically impressive debut of DeepSeek’s V3 and R1, and Janus-Pro-7B models – a lot of which is showing up as a significant sell-off of US tech stocks, all the way up and down the stack. Some are saying this undermines not only all of OpenAI’s value premise, but also the recent funds of the venture firms that put money into model companies altogether.

This is definitely an overreaction: OpenAI’s value won’t go to zero, and NVIDIA’s wedge as the anchor of the AI revolution isn’t dissolved to dust.

In my last post for this newsletter, I talked about how the need for AI compute is just going to keep scaling (possibly faster) because of the growing need for inference-time compute resources. I briefly addressed DeepSeek there, but it’s worth expanding why I don’t think R1 or its amazing efficiency is an extinction-level event for anyone else out there building at the foundational level.

What DeepSeek’s impressive launches do show is that 1) AI and the distribution of its value across the chain are far from a settled question, and that 2) a focus on fundamentals continues to be the key way for investors and builders to suss out and make bets on the allocation of said value.

In the era of highly attenuated attention spans, it’s easy to forget that ChatGPT launched in 2022. A little over two years just isn’t long enough for a nascent and particularly explosive technology like generative AI to achieve any kind of reliable stasis.

Because markets and funders need to try to seek out some kind of stability against which to place bets, and because there was initially an immense amount of upside in just continuing to iterate on what had already been achieved in a linear direction, it started to look like gen AI was settling into a set of comfortable ‘knowns.’ All along, however, some of the smartest people in the room were cautioning that this likely was not, and could not be the end state with respect to even the fundamentals of AI technology. Meta’s Yann LeCun, for instance, has long held that some of the key assumptions around LLMs and transformer-based technologies just won’t hold over time for how AI is developed and deployed.

DeepSeek is impressive because it’s a master class in escaping constraints through non-linear thinking and orthogonal approaches to problem solving. But the technical improvements and advantages that it incorporates are not locked to this one company or this one effort.

Companies like OpenAI, Anthropic and others (as well as their open source equivalents like Meta’s Llama) can and will incorporate these learnings into their own approaches, and the result will be massive resource savings along existing parts of the technical stack. But these savings won’t just drive costs for these services down to the point that well-capitalized companies collapse under their own weight and nimble ones offer the same services at cut-rate prices.

We already know that one way in which the AI landscape is shifting is that the importance of having a well-resourced training facility is being replaced with the importance of inference-time compute. DeepSeek-R1 and what it reveals to others in the space is just another way to shift resource allocation to that new and exciting part of the stack, and to pursuing other, complimentary technologies that are equally demanding at the data center but that can also fundamentally change the quality and nature of output at the level of the model supplier.

This reveals that for venture investors, the right approach remains to look for businesses that are focused what offers fundamental value to their customers, and what addresses their specific pain points. I still don’t think the foundational layer becomes ‘commodity’ in the way that bears would have you believe, but it reinforces that the best bets for most will be around companies and founders that understand the core problems they aim to solve, and that are nimble, smart and flexible enough to adopt the best of the base layer even as it shifts and transmutes underneath them.

OpenAI just revealed a new venture called ‘Stargate’ in partnership with SoftBank that will direct $500 billion over the course of the next four years to build out the data centre infrastructure to power its technology in the US, beginning with a facility in Texas that has already broken ground.

Stargate’s active partners are primarily OpenAI (operations) and SoftBank (finances), but it’s a collaborative effort that includes equity investment from Oracle and MGX as well, and technology support from Arm, Microsoft, and NVIDIA, alongside Oracle and OpenAI.

This isn’t a surprise move: Reports from the past couple of years had OpenAI’s Sam Altman traipsing around the world looking for capital partners to help fund infrastructure build-out to the tune of trillions, not just hundreds of billions. And the stakes continue to grow, with rivals like Elon Musk out there rough-hewing multi-billion dollar data centers virtually out of thin air in the outskirts of Memphis.

Musk’s Colossus supercomputer is massive in scope, and the billionaire technologist’s ambitions are even bigger: He’s said that it will eventually expand to house over 1 million GPUs, 10x the 100,000 that call it home today. Even so, the $500 billion figure that OpenAI dropped yesterday dwarfs the financial investment Musk and co. have sunk into xAI’s infrastructure to date, as well as its projected plans.

Already, Musk is attempting to throw cold water on the apparent disparity in capital expenditure between Stargate and his own ambitions. He posted on X on Tuesday following the announcement that the new consortium doesn’t “actually have the money,” with “well under $10B secured” by finance lead SoftBank as of the announcement according to the Tesla CEO.

That’s a bit rich coming from the same guy who once infamously tweeted ‘funding secured’ when actually that was super not the case. Still, he has a point here that the eye-watering number is at this point mostly aspirational rather than concrete.

Certainly, challenges remain in terms of getting to these lofty investment goals, but whether it’s $10, 20, 50 or 500 billion, the fact is we’re going to see increasing amounts of money shovelled in to these kinds of infrastructure projects.

There are reasons to doubt the sustainability of an AI-supporting infrastructure boom. I think there’s an argument to be made that we’re in the midst of an AI bubble that may pop to some degree, and people in the industry I respect to a great degree are also saying that there are diminishing returns in terms of applying compute and spend at the training part of the AI tech stack. Just this week we saw DeepSeek AI roll out a relatively lightweight, open source model that claims to be able to outperform some of OpenAI’s best and biggest, so it also seems increasingly like the foundational model game is going to be rapidly commoditized.

On top of that, unlike even the highly-commoditized cloud providers, AI models are actually almost trivially easy and cheap to swap in and out. With marginal switching costs and a demonstrated ability to leapfrog larger models with small ones in terms of performance in short order, what’s the value in massive underlying infrastructure spend?

I think the answer lies in inference time compute, and how that will increasingly be a valuable differentiator – particularly when it comes to running multiple inference tasks targeting a single result, and then using that array of results to achieve an optimal outcome through orchestration, combination and revision.

A lot of product layer companies are doing this kind of thing already, and delivering real value there. I think we’ll increasingly see foundation model providers like OpenAI explore what else they can do at this layer of the stack – and especially as they work out what kind of defensible advantage they can bring to the table by amassing an unmatchable pool of sheer hardware horsepower to drive those inference-time gains.

This past week, two isolated pieces of news about Perplexity came out – one that was very high-profile, riding high on the heightened drama surrounding the TikTok ban (and unban, just hours later). The other flew more under-the-radar, a passing report about the startup exploring product direction decidedly outside of its current wheelhouse. Combined, they paint a picture of a company looking to grab as much real estate as possible during a pivotal paradigm shift.

The news you likely saw was that Perplexity made a bid for TikTok in the waning hours leading up to the ban; the news you may have missed is that Perplexity made an acquisition, of a professional social networking tool called ‘Read.cv’ that resembled a LinkedIn re-conceptualized from the ground up for the TikTok age.

This follows recent product moves by Perplexity to extend their capabilities in shopping, enterprise data lakes and collaboration, web publishing, podcast creation and more.

Given their initial launch product and direction, the general characterization of Perplexity as a potential Google-killer makes a lot of sense. And if it weren’t for the incredible (unprecedented?) entrenched advantage they have in terms of usage and spread, I think Perplexity’s product chops would allow it to outcompete Google’s search for a variety of uses – including much of the problem space of what search is actually intended to solve in the first place.

The focus on whether or not Perplexity is able to replace a juggernaut like Google, a bastion of the early internet age and the bulwark of the modern tech monopolists, distracts from more interesting discussions about what else Perplexity could become.

Supplanting legacy leaders in tech generally does not take the shape of a 1:1 replacement arriving and becoming a direct swap-in replacement for the incumbents. Instead, the vectors that challengers exploit are slightly off-kilter to the giants they hope to dethrone – basically, it’s the Death Star trench run (the Rebels never built a mirror Death Star of their own to take on the Empire, they tried to exploit a minor design flaw instead).

In the early days, Perplexity looked a lot like an OpenAI or Anthropic feature that neither of those two foundation model companies had gotten around to building yet. They managed to differentiate themselves even in the fundamentals to the point where it’s clear they have a product advantage even when compared directly to the OpenAI web search features that eventually launched. But even focusing on that is probably missing the forest for the trees.

Perplexity’s moves this past week reveal an opportunistic and questing curiosity when it comes to product direction and problem space expansion – this isn’t a company looking to rival Google in its place of power, but one that wants to find its way into fissures and gaps in Mountain View’s armor and pry open its defenses that way instead.

It’s still unclear whether the rise of generative AI will be a platform shift at the level of transitions like the move to desktop computing, the rise of the web, and the dominance of mobile. But, if it does reach that scale, the industry’s topographical changes are going to be shaped by moves made now, including some that happen on the periphery or without much fanfare.