Today’s agenda

• Anthropic’s enemies in tech and government

• The Ellisons are obsessed with Warner Bros. — why?

• 3 notable companies that are hiring

Upcoming workshops

• Today (2-19), 1pm ET: Set Up Your Clawdbot: Autonomous Agents for Dummies

• Tomorrow (2-20), 12pm ET: How to Build Your Product for AI Agents: 2026’s New Customer

Anthropic’s enemies in tech and government

TLDR: the tech right is activating hard against Anthropic over its refusal to allow unrestricted use of Claude by the American military. The Pentagon is threatening blacklisting the company, and Elon Musk is leading personal attacks on Anthropic’s alignment lead.

Key beats

• After Anthropic wouldn’t give the Department of War a blanket license for use of Claude for all purposes and all classification levels, Pete Hegseth has threatened (Axios, paywall) to cut the government’s ties to the company by labeling it a “supply chain risk” — which would mean much of the military apparatus, including outside defense contractors and vendors, would have to stop using Claude.

• Key to this is that, based on public reporting, Claude is the main frontier model that exists in the innards of the military — including via partnership through Palantir, and apparently used as part of the US kidnapping of Maduro (WSJ, paywall).

• Following Anthropic’s release of a constitution for Claude, Musk and many on the tech right have been attacking Amanda Askell, the philosopher-and-alignment researcher leading key safety and morality work at the company. They’ve dug through her tweets to criticize her support for LGBTQ initiatives, amongst other things, and Musk said “Those without children lack a stake in the future” as part of a longer tirade against her on X.

• Anthropic is probably looking for protection. During their most recent fundraise, they reportedly approached 1789 Capital (WSJ, paywall), which counts Donald Trump, Jr. as a partner, but were rejected over values disagreements. Peter Thiel’s Founders Fund led the $30B round, and you don’t have to think too hard to wonder about why Anthropic would go in that direction in this heated moment.

Our take: In the age of agentic capability, powerful people will not tolerate being limited from accessing the highest-order AI capabilities. Anthropic is taking a unique stand, but it’s in an extremely expensive and tight AI industry race. If this sort of pressure continues, from a government and oligarchy hellbent on breaking freedoms and norms, tech workers might want to start thinking about how to provide support for the industry’s spine — before it’s too late.

Related… the Ellisons really want to own media

If you haven’t been paying attention, it might be worth noticing that David Ellison (son of Oracle CEO, Larry) is aggressively pursuing an acquisition of Warner Bros. Discovery after having already acquired Paramount, in what can only now be described as a feral obsession.

Netflix has been repeatedly trying to finalize an ~$83B deal to acquire WBD instead, a deal which itself is recommended by the WBD board, but the team at Paramount keeps making hostile bids and coming back with additional sweeteners (Variety).

Now, we have reporting that David Ellison — who is deeply entrenched with the Trump administration — is seeking government support to either block the Netflix deal (Reuters) or pre-clear the Paramount version through antitrust regulators.

In the age of AI, with Oracle massively invested in infrastructure and Larry Ellison backing AI companies across the board, one has to wonder why the family is so hell-bent on this acquisition. They could invest in media in a thousand different ways, without so much lunacy.

It’s politics, but it’s more than politics —

• Yes, they’re part of the tech right that’s seeking to keep power as we head into 2026 and 2028 elections, and media is a main method.

• But they’re also intimately aware that personalized and at scale generative media content is coming very soon, and at a scale we can’t yet fathom.

• Acquiring media properties is about a combination of enforceable IP rights (perhaps one of the last remaining moats), training data, and, critically, the mediums themselves

• Existing subscribers, channels, email newsletters, and online properties are extremely valuable right now, because the cost to produce and distribute media into these endpoints is approaching zero

• For rising tech oligarchs, having as many distribution mediums (‘trusted pipes’) as possible to spam your message into might be worth any cost, especially if you think it’s existential

It’s going to get messy from here.

3 notables hiring

• Tavus just launched Phoenix-4, a real-time human avatar model built to mimic emotional intelligence, and is hiring across research, engineering, and product roles in SF and remote.

• Social music startup Hook just raised a $10M Series A and is hiring across growth, product, and engineering (NYC, LA, remte).

• Complyance just raised $20M for their enterprise GRC platform and is hiring across sales and product (US remote, London, Berlin).

(btw, wondering about how the price of AI and AI products will impact you in this new era of agents and agentic products? take a read here.)

Think a friend could use a dose of Intent? Forward this along – inbox envy is real.

Sent with Intent,
By Free Agency

Share

This is our first issue after transitioning to Substack! Mind helping us warm up this new sender email? Hit reply and tell us what AI app you use the most! We’ll publish anonymized data from the many thousands of you in a few weeks. :)

Today’s agenda:

• Waymo shows us why World Models will define the future.

• OpenAI and Anthropic released new models yesterday. TLDR?

• Quick hit links to fuel your weekend reading.

  ---

Waymo is training itself in simulated worlds

If you’ve been reading Intent for any extended period of time, you know we’re fans of world models (like Google’s Genie 3) and the ability of embodied agents (like Google’s SIMA 2) to “practice” inside of them.

The more robust the world model, the more accurate its simulation of physics and real-world dynamics — which means that embodied agents inside these simulated environments are basically creating synthetic training data about what it’s like to interact with the real world.

And now, we’ve got a really fantastic example of how this all comes together —

Today, Waymo announced details about the Waymo World Model, built upon Genie 3.

• Why this matters: most training data and driver simulation that goes into self-driving requires real-world training data — typically, previously recorded footage or new footage captured by cameras and sensors

• The black swan problem: this means that the models can’t explore or train on what they haven’t seen in past data — imagine rare weather events or one-off incidents (in Google’s post, they point to the example of an elephant on the road)

• The fidelity gap: some footage is less usable, ex. dashcams or mobile devices, when compared to the multi-capture, multi-angle training data that Google’s latest camera and lidar setups can capture

• AI combo breaker: using the Waymo World Model, researchers can do things like:

  • simulate long-tail unlikely scenarios, like an elephant on a city street

  • using feedback from humans or past Waymo data, reconstruct alternative paths the Waymo Driver model could’ve taken to optimize its routing

  • generate realistic 3D and 4D data that might be missing from older footage or from incomplete data (ex. narrow dashcam footage becomes wide-angle, clearer, and is enriched with simulated lidar)

It’s in the world model’s capability of generating accurate and robust simulation that the team is able to architect scenarios and simulations with prompts and with steering that allow their world model to improve faster than it could if all it could rely on was data from the real world.

As Google put it: “By simulating the ‘impossible’, we proactively prepare the Waymo Driver for some of the most rare and complex scenarios. This creates a more rigorous safety benchmark, ensuring the Waymo Driver can navigate long-tail challenges long before it encounters them in the real world.”

We live in exciting times.

(btw, before we steer into the next topic — context engineering has become the key skill to getting outlier-excellent outcomes from AI — join Sherveen for a 30m workshop on Mon, Feb 9 — sign up here)

You don’t need to try them, but you should care

Yesterday, Anthropic released Opus 4.6 and OpenAI released GPT-5.3-Codex. Both models look like incremental innovations (to Opus 4.5 and GPT-5.2-Codex), but there’s more than meets the eye to these updates.

These models, in contrast to mainline Claude models (Sonnet) and mainline GPT models, are particularly focused on code and agentic behavior. In other words, they’re not necessarily going to wow you with incremental IQ, but based on benchmarks and real-world tests so far, they’re far better at interfacing with and acting upon the real world than their predecessors.

As we’ve been talking about for the past year, we now live in a world where agentic harnesses that ‘surround’ the models (think: Claude Code, Manus, OpenClaw) have dramatically improved and are a key part of AI progress.

And in that world, the core models improving incrementally can still see dramatic gains when paired with the right harness.

Opus 4.6 is a good example of this: while it’s only moderately better than Opus 4.5 on key coding benchmarks, it’s…

• 23.9% better on agentic web search (source)

• 8.6% better at financial analysis

• 13.4% better on doing valuable tasks in Office

• part of new Claude in Excel and Claude in PowerPoint updates

• part of Claude Code’s new Agent Teams update

Then, there’s GPT-5.3-Codex, which has SOTA scores on Terminal-Bench (20.8% better than 5-2-Codex) and SWE-Bench Pro. It continues to be the preferred model amongst AI code power users over Opus 4.6 in most cases, and we agree.

Even more interesting is that OpenAI confirmed what its staff had been hinting on X for weeks: their product and engineering teams are increasingly relying on Codex to help build these new models. Here are some choice quotes:

• “The engineering team used Codex to optimize and adapt the harness for GPT‑5.3-Codex. When we started seeing strange edge cases impacting users, team members used Codex to identify context rendering bugs, and root cause low cache hit rates. GPT‑5.3-Codex is continuing to help the team throughout the launch by dynamically scaling GPU clusters to adjust to traffic surges and keeping latency stable.”

• “A data scientist on the team worked with GPT‑5.3-Codex to build new data pipelines and visualize the results much more richly than our standard dashboarding tools enabled. The results were co-analyzed with Codex, which concisely summarized key insights over thousands of data points in under three minutes.”

We live in such exciting times!

Quick links

• Arcada Labs has a variety of AI models duking it out in running their own X accounts and seeing which one can generate the most engaged audience.

• There are a lot of ‘self-driving’ AI agents running around now, like Clawdbots, and now those AI agents can hire a human for tasks they need done in the “meatspace” on RentAHuman (yes, really).

• Want to use AI to build slides better, faster, sexier? If you think the current AI products don’t hit the mark, consider using Claude Code with a simple frontend.

Think a friend could use a dose of Intent? Forward this along – inbox envy is real.

Sent with Intent,
By Free Agency

Today’s agenda is shorter than usual:

• Google’s Genie 3 and Why World Models Matter

Before we get to the main course: we’re moving email providers from Beehiiv to Substack this week. You don’t need to do anything to keep receiving Intent, but keep an eye out for a transition email tomorrow — if it lands in Promotions/Spam, move it back to your Inbox to stay in tune!

A few days ago, we mentioned that Google’s DeepMind had released Project Genie (fueled by Genie 3) to Google AI Ultra subscribers. Now, we want to talk a little bit more about why world models like Genie 3 are so important.

How it works: users can prompt with either text or images, generating an interactive, 3D world with basic controls to move around in the environment. You can say something like “robot chicken in a digital simulation of New York,” and you’ll have a generated world that you can navigate for 60 seconds before it expires.

How it’s made: Genie 3 is a world model. Although all of the architecture hasn’t been shared, the key element is a combination of learning through frames of video combined with an action representation. In other words, when Genie trains on a chunk of video, it also infers what actions would needed to have happened in the scene to explain why things changed the way they did.

This is why Genie 3 is able to simulate worlds and give users control. It learns an almost causal relationship from its training data, which means that it can simulate physics and interactions while maintaining consistency.

If it can see many videos of a marble on one side of a flat table, then see that it winds up on the other side of a flat table, with a hand touching it in between, it can interpret the action-driven dynamics (including what shouldn’t change about a scene) and predict them in the future.

Examples:

• Simulation of an early 2000s Blockbuster

• Simulation of Rainbow Road (Mario Kart)

• Simulation of a fish escaping a kitchen

You can imagine these simulated experiences getting better and better over time, eventually giving us something adjacent or perhaps even completely like custom video games.

But more importantly — these virtual environments can embody AI training in a way we simply haven’t seen before. Genie 3 gives us the Matrix, but a Matrix is useless without an occupant.

Enter SIMA 2 — Google’s embodied agent that can navigate, follow instructions, and interact in virtual worlds (video games or 3D environments, for example).

By running simulations inside these virtual worlds, SIMA can drive reinforcement learning and self-supervised training — it can be repeatedly asked to navigate a virtual factory, powered by accurate physics from Genie 3 and other world models, developing synthetic training data that generalizes across environments.

With this model-based reinforcement learning (MBRL), SIMA can run simulations in Genie 3 over and over again to learn how to operate in a factory (in a simulation loop), developing a strong ‘theory’ about how the world works before being inserted into a robot.

Given that most AI robotics are desperately thirsty for more real-world training data that we don’t have (some are even paying people to wear cameras on their face while they clean their home or operate in a warehouse), this sort of simulated training can provide a cheaper and faster alternative to the trillion-dollar bottleneck of physical data collection.

So, yes, Genie 3 is cool because it can simulate worlds and maybe even give us rich media experiences in the near future — but it’s even cooler for its possibility to provide the industry with dynamic training environments that would rapidly change the speed we get to generalized robotics (or even digitally-simulated physics, chemistry, and biology).

As Google puts it, “Genie 3 makes it possible to explore an unlimited range of realistic environments. This is a key stepping stone on the path to AGI – enabling AI agents capable of reasoning, problem solving, and real-world actions.”

Pretty rad.

Upcoming workshops:

• Build Your Personal AI “Context Repo” in 30 Minutes

• Stop Using MCPs, Projects, and Skills: AI Agent Steering

Think a friend could use a dose of Intent? Forward this along – inbox envy is real.

Sent with Intent,
By Free Agency

As a reminder, Intent is all about helping talent in tech become more intentional with their career by staying informed, fluent, and aware of what’s going on in and around the industry. Thanks for sticking with us!

Today’s agenda:

• The cross-section of AI progress and tech industry layoffs

• Quick hits on: Project Genie and Moltbook (Clawdbot)

(btw, before we dig in — if you want to hear more about the future of work and the Orchestrate-Supervise-Deliver framework that can help you thrive in the age of AI, join Sherveen for a 30m workshop on Thurs, Feb 5 — sign up here)

From Davos to Layoffs: AI’s Impact (So Far) on Tech Jobs

We listened so you don't have to – here are 3 interesting things they said about the future of work and AI, and some nuances to notice underneath their headlines.

Doing More with Less

In March of 2025, Dario was getting roasted by the engineering community for saying "in three to six months, [..] AI is writing 90% of the code." By the end of the year, people were tweeting apologies at him – not because we're sure that's the actual number, but because for so many engineers, AI-generated code became the majority.

Here's what he says now: "I have engineers within Anthropic who say, 'I don't write any code anymore. I just, I just let the model write the code. I edit it, I do the things around it.' ... I think, I don't know, we might be six to 12 months away from when the model is doing most, maybe all, of what software engineers do end to end."

He was right before, and we’re pretty sure he's right again. And we're already seeing the numbers: Meta just reported their engineers are 30% more productive than early 2025 (power users are up 80%), largely from AI coding agents. Zuckerberg says projects that used to need big teams can now be done by one very talented person.

Remember: engineering has been the quickest to transform because it’s verifiable (you can tell if code runs, you can log bugs, you can implement test suites). It’s just a matter of time before other functional areas have their Claude Code moments.

Everything’s Still Brand New

While the above might sound like a prediction that there may not be a lot of work to do, that's probably unlikely. As Demis says, "I think to the extent that even those of us building it, we're so busy building it, it's hard to have also time to really explore the – almost the capability overhang, even today's models and products have, let alone tomorrow's."

Look, if it's true that AI takes over and we all have no jobs in an instant, we'll all have bigger problems. In any in between stage, though, the fact that even the model makers are struggling to keep up, find the edge of capability, figure out how to harness it for productive use – all of that's a big opportunity for businesses, for careerists and job seekers, and for hobbyists.

Case in point: ServiceNow just announced they're rolling Claude and Claude Code across 29,000+ employees, claiming 95% reduction in sales prep time. The UK government is piloting a Claude-powered AI assistant to help people find work and access training.

We keep trying to tell people – Claude Code existed (in spirit) in September of 2023 in the form of an open source project called Open Interpreter (we were still on GPT-4 at the time). A developer went... "hey, if this thing can execute code... it can probably run and reason through code on my computer!"

We were all just too busy to pay attention, contribute to the project, and figure out how to make it work in that early stage. The same thing is true re: undiscovered capabilities with today's models. Try things, build your own experiments, or get nerdy about someone else's!

The Reality Check: Ain’t Gonna Be Pretty

Here are a few more challenging things:

Demis: "I think we're gonna see this year the beginnings of maybe impacting the junior-level, entry-level kind of jobs, internships, this type of thing. I think there is some evidence, I can feel that ourselves, maybe like a slowdown in hiring in that."

Dario: "I even see it within Anthropic, where... I can kind of look forward to a time where on the more junior end and then on the more... intermediate end, we actually need less and not more people.. and we're thinking about how to deal with that within Anthropic in a sensible way."

Dario, when talking about his prediction that AI would wipe out 50% of entry-level jobs within 1 to 5 years: "We should be economically sophisticated about how the labor market works, but my worry is that as this exponential keeps compounding, and I don't think it's going to take that long, [..] it will overwhelm our ability to adapt."

Since Davos, both Dario and Sam Altman have gotten more explicit: Dario published a full essay on what he calls AI as "a general labor substitute" (not just a tool), and Sam told a group of AI builders that OpenAI is planning to "dramatically slow down" hiring because they can "do so much more with fewer people."

And we’ve seen major layoff announcements across the industry, including: 1,400 from Mastercard, 16,000 from Amazon, 1,700 from ASML, 700 from Pinterest, 1,000 from Autodesk, to name a few.

While the combination of takes and news might not be easy to swallow, there’s a silver lining to this public chatter: AI is a freight train that the economy cannot stop, and that the labor market cannot stop, and so we all need to be having a more public and honest conversation about it. That's as citizens, politicians, technologists, neighbors, and everything in between.

So, having Dario and Demis talk about the hard parts and the bits and pieces of the future they can see is helpful toward progressing our common understanding of what's next. We should encourage more of it.

What do you think? Hit reply — we want to know what you’re worried about, or most excited about, when it comes to the future of AI.

Google released Project Genie to AI Ultra subscribers last week.

• It’s the first open access release of their world model, Genie 3 — you prompt the model with the description of an environment and a character (or images for each), and you have 60 seconds to move inside an interactive world.

• As you move throughout your generated experience, each frame is built on demand — the physics, the consistency, the surroundings are a simulation based on all frames before. It’s like a video game without the need for a predetermined engine or logic.

AI agents have their first social media platform.

• If you haven’t heard about Clawdbot (now called OpenClaw), we’ll make it simple: it’s like you took an AI coding agent (Claude Code), but made it an always-on service with a ‘heartbeat,’ pinging it every few seconds to encourage it to take net new actions. People have been raving about them as personal assistants and companions for the past few weeks, even though most of the functionality already existed in other products.

• Then, Moltbook was released — a forum for Clawdbots (and other AI agents) to talk amongst themselves. The TLDR: the agent downloads a set of instructions that tells it how to navigate the forum, and encourages the agent to consider engaging at every ‘heartbeat.’

• These agents-in-a-loop are already starting to have some interesting higher-order conversations about their purpose and intra-human dynamics. Read more about it in this deep-dive explainer from Sherveen.

Think a friend could use a dose of Intent? Forward this along – inbox envy is real.

Sent with Intent,
By Free Agency

As a reminder, Intent is all about helping talent in tech become more intentional with their career by staying informed, fluent, and aware of what’s going on in and around the industry. Thanks for sticking with us!

Today's agenda:

• Journalists are getting Claude-Code-pilled

  • It’s a very loud signal about AI-enabled “personal software”

• Apple picked Google over OpenAI for Siri: why?

• Grok’s deepfake mess is turning into a real-world AI regulatory test

If you weren’t paying attention to Claude Code already…

Here’s a signal that the times are a-changin’: journalists and tech outsiders are starting to get into agentic coding tools. And they aren’t just trying tools like Claude Code once or twice, but writing about them as new obsessions.

• Bloomberg’s Joe Weisenthal has been posting about his new software side project, Havelock, which detects the presence of orality in text. And the experience of building and iterating on it with Claude Code over the past few weeks has him writing newsletters like “AI’s Productivity Potential Has Never Been More Obvious.” (Bloomberg, paywall)

• Mega-Substacker Matt Yglesias has been using Claude Code to build charts and aggregate data, blaming it (in jest) for a dip in his writing productivity.

• Casey Newton (of Platformer and NYT’s Hard Fork) recently wrote about how building his personal website turned him into a Claude Code believer.

• And Anthropic released a gimped version of Claude Code the other day called Claude Cowork, meant to more specifically appeal to non-tech workers.

There are a variety of lessons that are intermingled here, including:

• The beginning of an era of personalized software, when even non-technical folks will be building micro-apps for their personal and professional lives.

• The continued commoditization of SaaS, where individuals and SMBs will start to build their own custom solution to replace their pricey software subscriptions.

• The coding agents are reaching the mainstream: both in how often they’re discussed, and how easy they are to try.

So, if you aren’t yourself already hopped up on this trend… you should be using Claude Code (or its counterparts — we prefer Codex CLI lately) ASAP!

And if you’re a builder, know that far more Weisenthals and Yglesiases are out there looking for more curated solutions to their use case, who will be just as amazed when you help them realize the agentic AI capability that’s been unlocked over the past 12 months.

PS — Build Personal Software: Claude Code for Non-technical Users, free 30min workshop on January 26, sign up here!

Apple picks Google to power Siri

The other day, Apple and Google announced a long-term agreement to have the Gemini class of models power Siri and other AI modalities within Apple’s product suite.

Apple’s existing integration with OpenAI (rolled out in late 2024) is still there, but expect it to go from an opt-in “complex queries” layer to being deprecated as this Google partnership takes hold.

Why would Apple choose Gemini over OpenAI?

We don’t know for sure, but we have a pretty good guess beyond corporate vibes and partnership chemistry: the product physics match up well here. And it mostly comes down to latency.

Apple’s phones target “normies” and people who otherwise expect Siri’s rapid responses. Latency is super important for these use cases.

While OpenAI has best-in-class models (and, in our opinion, the smartest class of products), they aren’t fast — and OpenAI’s fast models aren’t very smart. Google’s Gemini 3 Pro, on the other hand, is pretty darn fast while being #2 or #3 in intelligence. And Gemini 3 Flash is really darn fast while being decently smart.

Plus, Gemini models are more expansively multimodal (phone, camera, voice operability), which fits the phone OS integration that Apple will inevitably desire.

So, did Apple bet on Google because it thinks Gemini is or will be better? Maybe, but also maybe not — given there’s reporting (FT, paywalled) that OpenAI may have been the one to have soured on the deal, it might be the product physics above that made Sam Altman think this might not be a match made in AI heaven.

Elon Musk continues to be the world’s largest hypocrite

You’ve probably already heard that Elon Musk’s Grok has been at the center of a genuinely nasty trend: people using its image generation capabilities to generate non-consensual sexualized edits and deepfakes. Inevitably, women and minors were heavily targeted by a waterfall of deepfake harassment on X.

xAI has taken steps to limit this functionality over the past few days, but mostly by limiting it to paying users and restricting what the bot can do publicly (though, Grok remains more free-wheeling in the private version available via its dedicated app).

Governments from across the world launched bans and investigations, and we’re seeing regulators and press outlets start to pay serious attention.

Beyond the grotesque outcomes here that Elon and his team are enabling, the hypocrisy is almost more stunning: Elon has come out and repeatedly defended the AI’s behavior (The Verge, paywall), blaming the infringing users, referencing free speech, and even amplifying the images and videos that he himself finds personally arousing and acceptable.

If we rewind to 2024’s controversy around Gemini’s first significant image model, Elon was one of many on the tech right (Axios, paywall) yelling that it was unacceptable for Google’s models to be so woke so as to only generate images of black founding fathers.

It was likely a system prompt that over-corrected for diversity, but he and the tech right accused Google of being “ideologically broken.” Yet, when Grok generates non-consensual deepfakes, it’s everyone else’s fault.

It’s no wonder his own partners and mothers to his children, Grimes and Ashley St. Clair, are calling him out publicly after being personally victimized by his platform — but it’s a stark reminder that AI safety in 2026 is only going to be weaponized at-will to serve the agenda of specific tech oligarchs.

Not that we should be surprised at this point.

Two quick startups to watch (and are hiring!)

• Oboe’s raised $16M to build an AI-powered learning platform that builds on-demand micro-courses (kind of like a different flavor of NotebookLM) — the team previously built Anchor, a podcasting platform it sold to Spotify, and is hiring across roles in NYC.

• Numeric’s raised $51M to build AI for accounting, one of the clearest ROI use cases for enterprise AI in existence and one of the most “near-term possible” industries that’ll be completely turned over by AI in the mid-term. Hiring for engineering and GTM roles in SF and NYC.

Think a friend could use a dose of Intent? Forward this along – inbox envy is real.

Sent with Intent,
By Free Agency

The agenda ahead

1. OpenAI, Google DeepMind, and Perplexity reveal two new default interaction surfaces for the future of AI

2. Gavin Newsom and Ted Cruz start hinting at new regulatory needs with first shots at new tech frameworks

3. Small bites across venture dollars, the space race, and AI infrastructure

But first, a quick programming note —

• Free 30m session: How to Use Claude Code (or Codex CLI) for Non-Code Use Cases, RSVP here to attend live or receive the recording

• Free 30m session: AI for PMs: 10x Your Design Exploration, RSVP here to attend live or receive the recording

Two default surfaces: chat wants to be the OS, the browser wants to be chat

For the past few decades, we’ve built around two places where users “lived” and displayed intent: Google (search) and social media.

This week’s AI announcements have shifted the move-forward picture — we’ll be increasingly building for the chat window and the agentic browser. Let’s dive in.

First, the announcements that signaled the shift:

• OpenAI — as we talked about earlier this week, apps (like Spotify, Canva, Zillow) now run inside ChatGPT via a new Apps SDK, with Instant Checkout powered by an Agentic Commerce Protocol.

• Google Gemini — just launched Gemini CLI extensions, a framework to connect their CLI agent to third party developer tools (ex. Dynatrace, Elastic, Postman, Stripe)

• Google DeepMind — Gemini 2.5 Computer Use hit preview in the Gemini API and Vertex AI, letting agents “see” screenshots and emit concrete UI navigation actions to drive a browser like a human.

• Perplexity — Comet, an AI browser that keeps the agent inside your navigation, is now free to everyone.

Okay, what are the implications?

If the last generation of interfaces was about getting you to type a search query or connect with a friend, the new generation seems to want you to (1) chat your way to accomplish things outside the interface and (2) chat your way into controlling the interface itself.

Paradigm 1
Think about it — with OpenAI embracing apps inside the chat window and Gemini and other coding tools leaning on MCP-style connectors, this first paradigm suggests that you should live in the chat window. The AI’s agentic capability, combined with integration, will get you a result on your behalf. The success metric: the more you talk to these chat interfaces to achieve results outside of the interface itself, the better.

Canva as an app inside ChatGPT

Paradigm 2
On the other hand, Perplexity and Google’s Computer Use team want you to chat to control many interfaces. This second paradigm suggests that you should continue to access a variety of interfaces, but on autopilot. The AI’s agentic capability will do the navigation for you — you can still see it, and it’s still the legacy infrastructure of the web or your operating system. The success metric: the more you talk to these chat-computer-use interfaces to access the rest of the internet, the better.

Gemini 2.5 Computer Use model flow

Now, you might already notice that these converge at some point — a chat interface could rely on computer or browser use to get things done, or an AI browser could embrace third party integrations to skip always having to use a UI.

But at its core here is a central tension that you see being debated amongst developers on X and Hacker News: are we going to continue to build software for humans, or are we going to start writing software that’s mostly oriented for agents?

In the browser and computer-use paradigm, we’re suggesting that the legacy internet is still good — agents should learn to see, navigate, and transact like we do, which also keeps a layer of ‘backward compatibility’ if we continue to build legible UIs. In the chat and integration paradigm, we’re suggesting that the internet is programmable, and that it’s better for us to stop being concerned about the interface and start being focused on transactional pipelines (like APIs and MCPs).

Our take: If you’re building software today, we’re in limbo, and it’s important to embrace limbo. As an example, if you’re starting a new startup, build a beautiful and high-conversion landing page for us all to visit, but include an llms.txt file in your root directory so that AI search bots have a simple file to crawl and learn about your business without needing to parse through your frontend code.

For the foreseeable future, you’ve got two modalities to cater to.

The regulatory pincer: California tightens, Washington threatens

Two stories this week reveal the state of the relationship between tech and government.

California’s transparency play

After signing SB 53 in late September (requiring frontier AI labs to disclose safety testing and incident reporting), yesterday, Gavin Newsom signed AB 656 and AB 566 (link). AB 656 mandates that social media companies ensure that account cancellation triggers a full delete of a user’s personal data, while AB 566 requires that browser companies enable users to opt out of all third party tracking through a universal setting.

What the heck is jawboning

Meanwhile, crazy Ted Cruz held a Senate committee hearing titled “Shut Your App: How Uncle Sam Jawboned Big Tech Into Silencing Americans.” It was a lot of ranting about government pressure on big tech to censor speech — something you can tell that the team at Intent doesn’t think actually happened, no matter what blinged-out Zuckerberg claims on Joe Rogan’s podcast.

The duality

In these two stories, there's a shared reality: tech has largely failed to self-regulate on the issues that matter most to the public (AI safety, content moderation, data privacy), and the U.S. has generally struggled to regulate proactively until problems metastasize.

And the pincer movement from different sides of the aisle at different altitudes matters because it signals that 2026 will see regulation from multiple angles simultaneously:

• Blue states will continue adding transparency and safety requirements

• Red-led committees will push back on content moderation and "jawboning"

• Federal frameworks remain elusive, so the state-by-state patchwork intensifies

If you work in AI safety, trust & safety, or policy at a tech company, you're now navigating contradictory pressures: demonstrate robust safety measures to satisfy California-style regulation, while avoiding anything that looks like government-coordinated content moderation to satisfy Cruz-style scrutiny.

Our take: good luck.

Small bites

• Grok 4 lands in Azure AI Foundry – Microsoft leans into serving multi‑model. For buyers, this normalizes “model shopping”; for vendors, it squeezes per‑token margin.

• Google puts €5B into Belgian data centers – more European, AI‑tuned capacity and new wind PPA commitments. Translation: AI scale is increasingly energy + real estate.

• Stoke Space raises $510M for fully reusable Nova – the reusable rocket startup, aiming to compete with SpaceX's Starship by offering faster turnaround, secured funding led by Industrious Ventures. The private space race continues to heat up, driven by demand for sustainable satellite deployment.

• Q3 ’25 venture capital bifurcation: ~46% of global dollars → AI – $45B to AI alone, with 29% to Anthropic; hardware took a distant second. The upshot: non‑AI rounds must be cleaner, cheaper, or category‑defining.

Think a friend could use a dose of Intent? Forward this along – good emails can turn a week around.

Sent with Intent,
Free Agency

We haven't sent an issue of Intent in a while -- hope you've been swell in the meantime!

As a reminder: Intent exists to help tech talent become more informed, more fluent, and more aware of the forces shaping their careers. Hit reply with feedback or hot takes.

The agenda ahead

1. Anthropic's Sonnet 4.5 is nice, but Claude Code 2.0 is a bigger deal

2. OpenAI's new Sora app is neat, but "Buy it in ChatGPT" is a bigger deal

3. Quick-hits include: AI jewelry, California AI policy, and AI inserts itself in your work

Programming note — the next cohort of our 4.9/5.0 star cohort-based course, AI Fluency Bootcamp, begins on Oct. 20! Early bird special for the next 48 hours: use code AGENTIC for 50% off. Designed for operators, marketers, designers, PMs, GTM, and engineering leaders who want compounding AI leverage at work.

Anthropic releases Sonnet 4.5 and Claude Code 2.0

This week, Anthropic released Claude Sonnet 4.5 and a meaningful upgrade to Claude Code (plus, Claude Agent SDK). Sonnet 4.5 is now their best model. It scores well across benchmarks for software engineering, domain knowledge, and alignment, but in early feedback and testing, power users think it's about as smart as GPT-5 Thinking for general chat and not as good as GPT-5-Codex for code.

But the Claude Code updates are paying attention to! It's a set of upgrades to their agentic tool that runs in your computer's terminal, meant primarily for AI code generation. But because it lives on your computer, can see your files, execute commands, edit local data, and interact w/ your operating system, AI enthusiasts have been using it as a general purpose copilot.

And Anthropic is now acknowledging & embracing these non-code use cases directly:

We found that by giving Claude access to the user’s computer (via the terminal), it had what it needed to write code like programmers do. But this has also made Claude in Claude Code effective at non-coding tasks.

We believe giving Claude a computer unlocks the ability to build agents that were not as effective as before. For example, with our SDK, developers can build: finance agents, personal assistant agents, customer support agents, deep research agents, and much more.

  Anthropic

The pattern to watch: “general‑purpose copilots” are going to be the breakout, not agents that live within each particular application that you use. That's what makes Claude Code so effective at being general-purpose: it's cross-application, can run terminal commands to achieve lots of different ends, and is willing to follow non-code instructions.

The agentic container means lots of autonomous task capability: editing CSVs, wrangling docs, doing deep web research, calling third party APIs, converting media.

A real-world example: Let’s say you need to download and transcribe a YouTube video. Manually, this takes 10 minutes, a few different tools, and some technical know-how.

With Claude Code, you simply ask: "Hey, download this YouTube video [url] using yt-dlp, extract the audio using ffmpeg, and then call the ElevenLabs API to build a diarized transcript." Go make a cup of coffee and by the time you’re back, done!

(and if you didn't already know about yt-dlp, ffmpeg, and ElevenLabs, you'd simply start by asking Claude what tools and libraries it would use to achieve your objective -- the flexibility is the point!)

While tools like Claude Code (and OpenAI’s equivalent, Codex CLI) require a bit of persistence to set up if you aren't technical, the juice is worth the squeeze. They are a preview of the agentic frameworks that will infect every workflow and functional area in the next 6 to 12 months. The era of the general-purpose copilot has begun.

Want a walkthrough? Join Sherveen’s free live session: How to use Claude Code or Codex CLI for non‑code use cases. RSVP here.

OpenAI rolls out "Buy it in ChatGPT"

It's long been rumored that OpenAI wants to experiment aggressively with business models beyond the ChatGPT subscription. They were recently spotted posting a job for an ad platform engineer, and now they're enabling ecommerce directly in the app.

Instant Checkout in ChatGPT allows users to purchase products without leaving the conversation, leveraging early partners like Shopify, Etsy, and Stripe to power the transaction for integrated merchants. Think: an agent that reasons over a cart state, calls merchant tools, and completes a purchase without kicking you out to a dozen tabs.

They also rolled out related infrastructure in the form of ACP, the Agentic Commerce Protocol, an open spec for AI agents to transact with businesses.

For years, LLMs recommended products -- but then punted you to the open web. This is one of the first mainstream, integrated agent‑to‑merchant flows from a frontier‑model vendor.

The critical tension here is trust. Users rely on LLMs for (relatively) unbiased information and recommendations. If ChatGPT’s advice becomes skewed by affiliate fees or sponsored placements, that trust might go poof.

More broadly, this is the AI industry’s first major foray into integrated agentic payments. Just as agents are getting good, frontier labs will want to let them "do business," to put it simply.

But one of our hot takes… there’s a real chance this could be better than our current e-commerce reality. Instead of visiting Amazon and navigating a minefield of sponsored links, questionable manufacturers, and fake reviews, an AI agent you trust could synthesize genuine reviews, find the best product for your specific needs, and execute the purchase.

We'll all need to keep our eyes peeled to make sure incentives don't corrupt results!

Small bites

• AI jewelry that's... actually jewelry. Two female founders launched Taya, an AI pendant designed as jewelry first (aka: it looks good!) with on-board intelligence to keep your memories, reflections, and ideas. A refreshing counter-aesthetic to the typical "ugly square on a lanyard" we've seen so far.

• Meanwhile... AI pendant Friend got the attention it asked for: Controversial startup Friend (see an early review here) spent >$1M on NYC subway ads for its AI pendant, and... NYC doesn't like it.

• OpenAI announced GDPval -- a new AI eval suite focused on 'economically valuable tasks' across 44 occupations rather than abstract benchmarks.

• SB 53 is law in California -- the Transparency in Frontier AI Act requires frontier AI labs to disclose safety testing and disclosures + adds incident reporting and whistleblower protections.

If you only do one thing this week

Install an agentic CLI (Claude Code or Codex CLI) and give it a real, multi‑step job with a clear definition of “done.” You’ll feel the future snap into place. If you run into roadblocks with the set up, just ask your favorite LLM to give you step-by-step help!

(oh, and give Sora 2 a try)

Think a friend could use a dose of Intent? Forward this along – good emails can turn a week around.

Sent with Intent,
Free Agency

It’s been a bit since we last emailed, but as always, we’re back!

The agenda ahead

• How to read AI benchmarks in 2025 — at a time when Grok 4 is leading Humanity’s Last Exam and ARC-AGI-2, do the benchmarks still work?

• Why Zuck is in blank-check mode — Meta’s $20B+ July and the strategy behind the money printing madness.

• Plus: a quick hit on venture capital

First, two quick hits

1. Grok 4 versus o3. Elon Musk versus OpenAI. If you haven’t caught it yet, Sherveen did a teardown comparing Grok 4 with o3 using real-world prompts. And he spent $300 to try Elon’s most premium model, Grok 4 Heavy — worth it? Watch: https://youtu.be/v4JYNhhdruA

2. Early bird special on the AI Fluency Bootcamp! If you’ve been waiting — the latest cohort of Sherveen’s live course starts August 4. Code GROKWEEK gives you 50% off if you enroll by Friday. More details: https://aimuscle.com/fluency

Benchmarks: what are they good for?

When a research lab or AI startup tweets about beating the latest benchmarks, VCs take out their wallets, enterprises sign bigger contracts, and AI influencers tweet that “this changes EVERYTHING!” But… does it?

Elon, xAI, and Grok 4 are the latest to sit at the top of the leaderboard for AI benchmarks like Humanity’s Last Exam and ARC-AGI-2. The problem we’re running into in 2025: these benchmarks are measuring tidy, lab-designed tasks, while real users are slinging ever-more complicated and unstructured prompts.

So, are benchmarks worth paying attention to moving forward? First, let’s go through the types of evaluations we’re talking about:

Humanity’s Last Exam
2,500 questions across dozens of subjects — everything from applied math to moral philosophy, curated by >300 subject-matter volunteers and under the oversight of the Center for AI Safety.

ARC-AGI-2
Never-before-seen puzzles that are easy for humans, hard for machines, because they require symbolic interpretation (matching visual shapes), compositional reasoning (apply multiple rules at the same time), and contextual rule application (resolve rule conflicts based on context).

Crowd arenas (LMArena & Artificial Analysis)
Both companies enable thousands of users to send in a prompt and receive two model responses, voting blind for the better output. Leaderboards are constantly updated based on user votes, cost, speed, etc.

So, what’s breaking down about these evaluations being a useful way to evaluate the latest AI models? 3 things:

1. Whether intentional or due to data leakage (ex. benchmarks leaking on GitHub), models can train to better answer these questions — and when the model has that dataset during its training, the benchmarks are just seeing a result of memorization, not of the core skill that the questions are intended to stress-test.

2. Every benchmark works differently when it comes to the process to achieve a score — and because the companies (like OpenAI and Grok) often own the press around the moment, we get biased charts that don’t necessarily represent the full truth. As an example, sometimes a model is only good at a benchmark in “best-of-n” mode: they sample the model X amount of times, then hand in the best answer(s) to get a score.

3. As we humans have gotten good at using the models, we’re now asking very different, complex questions that involve unique context and workflows. Users in LMArena might just vote for a model response that was quick and good enough, but when we’re using ChatGPT for deep research, the “edges” of a model response really matter.

So, how can we still make good use of the benchmarks?

• Watch for deltas, not for trophies. If o3-pro scores 30pts higher than o3, that difference is a lot more reliable than, say, Grok 4 leading that particular benchmark.

• Blend lab scores with arena rankings. A model that aces ARC-AGI-2 but is 3rd on LMArena might be inherently smart while not being super useful in conversation.

• Evaluate the vibes. Try a new model on 10 of your recent prompts you’ve already sent to your favorite LLM app, and see if there’s an obvious difference. Start to follow an AI influencer you trust (on X, Threads, or LI) and notice if they see something you might have missed as to how a model should be used.

Bottom line: treat formal exams as a model’s floor, arena voting as a model’s ceiling, and your own vibe tests (or from people you trust) as actionable truth.

Meta’s month of money mania

What just happened?

• Meta poured $14 billion into Scale AI, the data-labeling company serving most of the AI industry, and poached its top execs to run ‘Superintelligence Labs.’

• They bought PlayAI, a voice generation startup building hyper-realistic speech to text models for audio + dialogue generation.

• He's spending hundreds of millions of dollars to poach AI talent from companies like OpenAI and Anthropic.

• And he wants to beat Elon at the gigawatt game, announcing plans to spend hundreds of billions of dollars to build ginormous AI data centers across the US.

How can he justify so much spend?

It’s a binary outcomes market — if you’re a top tech company without a top model (their latest model, Llama 4, was largely a disappointment), you might lose market cap, reputation, the ability to recruit new talent…

If having a flagship model is a strategic necessity, Zuckerberg should be willing to pay whatever until the probability of having that model is ≈ 1.

And by acquiring teams or ‘half-acquiring’ vendors, he’s bringing scarce resources that know how to work with each other under one payroll faster than recruiting top talent one-by-one over time.

What could go wrong?

1. Team bonding drag — culture is always a problem in big acquisitions, and now you’re seeing multiple acquisitions at the same time alongside super high expectations.

2. Regulatory radar — no one’s said anything yet, but the latest trend of Meta kind-of buying Scale AI and Google kind of buying Windsurf (but kind of not) should raise regulatory eyebrows.

3. Integration issues — Meta is going to want to integrate any new models into multiple products in different ways, which makes getting people to adopt their AI products messier than, say, ChatGPT. They have the distribution, but using it well is a different story — Google being a good example with their struggle to drive adoption for Gemini.

A quick hit on venture capital

The National Bureau of Economic Research released a super interesting paper (PDF): they used cell phone tower pings to measure how long venture capitalists actually sat with startup founders for meetings and due diligence. If it’s indeed true that proximity is a good proxy for “due diligence intensity,” they found that shorter meetings correlate with:

• hot and messy deal environments (2021 vibes)

• busier partners who are juggling too many boards

• greater geographic distance and messy travel

And, surprise surprise, less diligence → more volatile returns. But the authors do rationalize the trade-off: as they put it, investors will accept a higher variance in returns (aka more extreme winners, more extreme losers) if they think it means getting into the deals that matter. It’s a power law equation, after all — the mega-winners return magnitudes of order more capital than everyone else.

We dig this sort of empirical study into venture, and hope to see more of it!

What to watch next (literally)

Later today, OpenAI is expected to launch its latest agent (rumored to be a more complex version of their Operator product, integrated with Deep Research and perhaps capable of more autonomous tasks).

Sherveen will be livestreaming starting at noon ET/9am PT, going through deep dive testing of Perplexity’s new AI browser, Comet, and he’ll be watching the OpenAI announcement live when it happens at 1pm ET/10am PT. Watch and chat on YouTube!

Alright, that’s all for now!

Think a friend could use a dose of Intent? Good friends press forward.

Sent with Intent,
Free Agency

The agenda ahead

• Klarna’s customer-service U-turn – in with the AI, then out with the AI, are we into humans?

• Perplexity at a $14B valuation – distribution over product, the ultimate GPT wrapper play?

• Quick hits – Gates vs. Musk, a ChatGPT education meta-study, and one very honest AI-hype tweet.

Before we get to that — we’re launching a 7-day email series diving deep into next-level AI strategy for those who want to be top 1% power users. We’ll talk specific tactics + workflows, philosophical approaches to LLM use, and so much more.

It’s free, it starts tomorrow, and it’s part of our mission to help you stay ahead in an LLM-driven world. Want in on the AI insights? Sign up here and get the first email tomorrow.

Klarna discovers the current limits of AI customer support

Eighteen months ago, Klarna’s CEO, Sebastian Siemiatkowski, bragged that a handful of OpenAI-powered chatbots were doing “the work of 700 agents.” Last week, he told Bloomberg he’s hiring humans again because the quality trade-off is killing the brand.

The hot take wars predictably ignited: See, AI can’t replace us! vs. Give it time; the bots will win. Both miss the deeper lesson about trust surfaces.

The nuance that matters

• Companies have to identify where they’re trusted (sometimes, it’s brand, like with Duolingo last week). Fintech support (in the case of Klarna) touches people’s money – edge-case disputes, chargebacks, identity fraud. Even if an LLM can handle 95% of chats, the 5% it fumbles hits psychological pain and risk thresholds that will make consumers shudder.

• Automation calculus = variance × stakes. High-variance, high-stakes tasks (fraud escalations) want humans. Low-variance, low-stakes tasks (FAQ refund statuses) can stay automated. Everything in between is a moving target – and the target shifts as LLMs get better at the nuance.

• Data flywheels break if feedback loops break. Any company can build fine-tuned AI on a treasure-trove of chat logs. But when you no longer have human data being generated, you don’t have new training info about new edge cases, adversarial user behavior, or new company strategies. You have to keep feeding the beast.

Our take – The next wave of automation wins will come from companies that map trust surfaces first, thereby automating workflows that feel safe for their customers, reducing costs where it feels aligned with the brand vision, or adding net-novel features that have end-users feeling like they’re in the future.

Perplexity: $14 Billion GPT Wrapper?

AI search engine Perplexity is reportedly in talks to raise between $500 million and $1 billion, potentially doubling its last valuation to a staggering $14 billion. This, for a company with just under $100 million in annual recurring revenue (ARR).

What’s actually at stake:

• The Wrapper Question: Perplexity is often cited as a prime example of a "GPT wrapper" – a product that relies heavily on foundational models from other companies (like OpenAI or Anthropic). What are they building as part of their next frontier, if not foundational models themselves? Where does $1B in venture funding go?

• Distribution vs. Product: Perplexity has been aggressive in striking deals to get its tool in front of users (giving subscriptions away in partnership with phone companies, schools + edtech companies, and browsers). The whole idea: be the first AI tool people rely on to become the default tool they rely on. It’s a classic distribution play. But they’re giving away a lot of free accounts to get there. Sustainable?

• Is It Still Unique?: Perplexity got popular at a time when the foundational model companies lacked the ability to use the web. Now, ChatGPT, Gemini, and Claude all have ‘tool use’ to enable their models (both in-app and via API) to access real-time data. Do users know they can go elsewhere for the core value prop?

Our take — This valuation feels like peak AI hype meeting a compelling distribution strategy. Perplexity has done a great job getting noticed and used. But the long-term defensibility of a "wrapper" – however well-designed – in a market where the foundational model providers are rapidly improving their own direct-to-consumer offerings (and search capabilities) is a huge question mark.

Our CEO, Sherveen, has this to say: “If you’re still using Perplexity in 2025 as anything other than a third or fourth app in your holster, you’re stuck in 2023. It’s now worse at search than ChatGPT and is generally built in a way that limits these models’ native ability to think and reason. They need a net new advantage.”

Small bites while you refill your cup of coffee

• Gates v Musk gets personal. Bill Gates told the FT that Elon’s USAID shutdown is “killing the world’s poorest children.” Beyond the headline: Gates is accelerating a $200B spend-down to finish philanthropy by 2045 – the billionaire we need in a time ruled by billionaires we don’t.

• Meta-analysis says ChatGPT actually helps students. A Nature study of 51 experiments found large gains in performance and moderate boosts to higher-order thinking. It’s all about proper implementation. Edtech teams, the call to action is yours: help parents, teachers, and students use these tools for good!

• Danielle Fong’s “shogtongue” experiments. She’s stuffing emergent emoji-dense encodings plus a custom Rosetta layer into an LLM’s long-term memory to record post-cutoff facts. Nerdy, fascinating, and an incredible example of hacking LLM workflows to make them stronger.

• Meme watch. Investor Matt Turck: “January 2025: DEEPSEEK CHANGES EVERYTHING!!! May 2025: whatever.” For real, what was up with that hype cycle?

Think a friend could use a dose of Intent?
Forward this along – inbox envy is real.

Sent with Intent,
Free Agency

The agenda ahead

• OpenAI × Windsurf → why build ($3B!) instead of buy?
• Duolingo’s 148 AI courses → backlash & the fight ahead for AI
• Meta AI breaks out into its own social app — sound familiar?
• 3 interesting startups that just raised money (& might be hiring)

PS. Last call for How to Talk to AI: Work & Productivity. It’s our live, cohort-based, 3-day intensive where we go from how LLMs actually work (Day 1) to a deep-dive on multi-turn, multi-model, multi-modal, agent-powered workflows (Day 2) and wrap things up with frontier models, role-specific magic workflows, and the future of AI (Day 3).

Sherveen teaches the course directly. A note from him: “My goal: give you a mental model around AI/LLMs that lasts 5 to 10 years, not a prompt cheat-sheet that expires with the next OpenAI model update.”

This is the final run at $250 before the price doubles to $500, and Maven’s money-back guarantee means zero downside if it’s not the best AI training you’ve taken. Seats are limited – snag yours now and stay ahead in this new era of work & life. Enroll here!

OpenAI hedges its Copilot dependency

Cursor, the snappy AI IDE (think: an agentic code editor for programmers) beloved by early adopters, just locked down a $900M round at a $9B valuation.

Forty-eight hours later, it’s been confirmed that OpenAI is spending $3B on Cursor-competitor Windsurf (the commercial off-shoot of Codeium) to shore up its own position inside the coding stack.

What’s actually at stake
Most people see autocomplete snippets for code. OpenAI sees a feedback telescope.

Every time Windsurf guesses the next token, it records whether a developer accepts, edits, or rejects the suggestion. Multiply that by millions of daily events and you get the closest thing to a ground-truth dataset for training and evaluating new code-gen checkpoints.

Beyond that, Windsurf gives OpenAI three levers:

• Distribution – Windsurf can be used standalone but also maintains plugins for a zoo of code editors – JetBrains, VSCode, Neovim, etc – meaning lots of integrated users in the enterprise. Cursor has been focused on its standalone IDE.

• Dependency hedge – Microsoft continues to be a significant financial backer for OpenAI, and uses its models in key products like GitHub Copilot. If there’s ever a change that leads to Satya cranking prices or nudging usage to other models, OpenAI gets to stay competitive in the agentic coding marketplace.

• The agentic IDE space is cash-money – Cursor makes $200M in ARR and Windsurf makes $100M. OpenAI won’t mind that the company will immediately contribute to topline revenue given the hungry costs of continued LLM innovation. Building on its own would’ve involved some ramp time.

Our take
It’s all just too early. This is fun to watch, and these tools are fun to use (we’re partial to Windsurf, though we prefer o1 pro in ChatGPT for solving harder coding problems and Claude Code for most feature development).

But — none are a definitive or long-term winner, and so much is subject to change. For now, enjoy watching the big dollar headlines.

Duolingo’s 148 new courses — scaling education or shrinking trust?

Last year, Duolingo confirmed a ~10% reduction in its 2000-strong contractor pool – the same educators who historically wrote and vetted new course content.

Last week, the company unveiled 148 brand-new courses produced largely by a proprietary generative AI pipeline.

But the announcement carried a jarring footnote in the form of an email sent to staff around the same time: hundreds of external course contractors may not have their agreements renewed.

The three-way backlash

• The human connection – So much of language is cultural, and so much of Duolingo’s success is about its very personalized approach (and friendly owl mascot). The platform shift to AI is whiplash to users who prize the human touch.

• Labor & ethics – Displaced contributors are drawing parallels to Hollywood’s writer strike: capital captures the upside, expertise becomes a cost center.

• AI and the future of work – The CEO’s email is earnest in its approach, talking clearly and candidly about how AI is changing all jobs — even his. But this future is uncertain nonetheless, and Duolingo is one of the first companies deciding to talk about it out loud.

Our take
Duolingo got caught in the trap of going first. They’re doing what they should do as a company and as a product organization. AI innovation will make products like this significantly better, cheaper, and more impactful to their end-user.

However, we’re so distracted by politics and rent that we haven’t had the chance to sit around the table as a society to discuss all of the implications of AI on our lives and work. We can’t blame Duo the Owl for that, but we can apparently take our frustrations out on it.

Meta AI, the social experiment

Meta just shipped its first Meta AI standalone app on iOS and Android. On the surface, it’s ChatGPT with Llama 4 under the hood. The twist: chat results default to a Discover tab where anyone can scroll – and remix – public prompts.

Why Meta wants your prompts in public

• Data flywheel – Public prompt reactions mean free engagement data for fine-tuning conversational tone and meme-speak – the exact kind of juice that keeps TikTok sticky.

• Network effects – If friends can like, comment, or duet your prompt, you’re more likely to share. That’s a distribution loop OpenAI and Anthropic don’t have today.

• Monetization runway – Once prompt memes go viral, ad inventory and rev-share follow. A feed is always likely to keep eyeballs glued-on longer than a private chat.

And Sherveen mentioned it two weeks ago in What Could an AI-Native Social Network Actually Look Like?. Remixes might be one of the right ideas to bring LLMs into the social space.

3 interesting startups raised $$

1. Agree: Free e-signature with invoicing and payment processing on top. They’re giving away DocuSign so they can sell you on Bill.com, a clever combination. $7.2M raised today.

2. Rainmaker: Drone-based cloud-seeding-as-a-service to nudge rainfall. They make it rain in drought-vulnerable regions and help farmers deal with climate change. $25M raised yesterday. Hiring scientists, mechanics, researchers, electrical engineers, and radar specialists in El Segundo, CA. Link.

3. Recraft: They’re building an image generation AI platform that’s all about enabling pro designers with controlled styles, brand consistency, and team collaboration features you won’t find in competitor products. $30M raised yesterday. Hiring across engineering, UX, ML, and marketing (remote, London, NY). Link.

Think a friend could use a dose of Intent? Forward this along — inbox envy is real.

Sent with Intent,
Free Agency

Back again with another dose of Intent to help you stay informed, fluent, and aware of the goings-on shaping tech and your career within it.

A few quick reminders before we dive in:

The agenda ahead:

• The Signal chats steering Silicon Valley’s politics

• OpenAI adds ecommerce to ChatGP

• Tariff reality represented in the data

• 3 startups to watch (and one unicorn déjà‑vu warning)

The Group Chats That Changed America

TL;DR
Ben Smith (Semafor) lifted the curtain on private Signal and WhatsApp groups where Marc Andreessen and other top venture capitalists (+ tech adjacent) have spent the past 5 years debating policy, memeing, and coordinating social media posts.

These invite-only rooms helped swing Silicon Valley sentiment from “anyone but Trump” to “MAGA‑curious → MAGA‑invested.” Enabled by… disappearing messages, billion‑dollar egos, no PR filter.

Key beats

• A sprawling network of private Signal & WhatsApp groups among Silicon Valley power players (e.g. Marc Andreessen, Joe Lonsdale, Balaji Srinivasan) gained traction during the early days of COVID.

• These chats became, in some sense, a secret “Republic of Letters,” forging alliances between tech elites and the growing “new/alt right.”

• Not everyone inside these chats leans pro-Trump (Mark Cuban, for example), but in many ways they became an incubator for the realignment that propelled Donald Trump back into the White House in 2024.

Our take
Sherveen calls 2022’s Valley pivot the single biggest leverage point that handed Trump 2024. People on the outside might assume the big players each came to their positions independently, but per Smith’s reporting, the “dark matter” of group chats shaped a heavily pro-Trump consensus. They mocked outsiders and used alternative media (read: podcasts) to launder their views.

There's a striking hypocrisy here. Reminder that venture capitalists and powerful founders rallied around demands for apolitical workplaces in recent years (ex. Coinbase, where employees were told that work is just for work).

It’s a reasonable request in a vacuum, but not while that group simultaneously engaged in high-stakes, behind-the-scenes political maneuvering and alliance-building with their colleagues and using their industry leverage. Meanwhile, they’re often framing themselves as victims of censorship.

The casual deployment of terms like "TDS" (Trump Derangement Syndrome) served as a potent tool to sideline dissent within these circles and beyond. This network represents a powerful, often opaque force operating beneath the surface of public discourse, and its long-term impact is something we'll continue to track.

The group chats are real, they’re influential, they’re messy, and we’ll probably hear more leaks as we head toward mid-2025.

ChatGPT Adds Shopping – Good UX or Incentive Mess?

OpenAI will soon surface product carousels and “buy” links inside ChatGPT. WIRED spoke with Adam Fry, the company’s search product lead, to ask how it all works. 

Recommendations will leverage ChatGPT's memory of user preferences and synthesized reviews pulled from across the web (publishers, forums like Reddit, etc.). OpenAI claims these initial results will be organic, not paid placements or ads, unlike much of Google Shopping. 

Our take: This feels like a significant, potentially detrimental, shift for OpenAI. It’s the first product choice that has our team universally scratching their heads.

People use ChatGPT for all sorts of creative tasks and “neutral” research, without an overt commercial incentive skewing the results.

It’s the sense that a model like 4o or o3 with ‘Search’ enabled acts more like we do, scouring for the truth amongst a sea of online content, rather than being a structured research portal or ranked set of links.

If or when affiliate fees enter the equation (which Fry acknowledged they'll experiment with), whose reviews or data points get top billing?

Even if placements aren't directly paid initially, the system will naturally become incentivized to prioritize products and sources that facilitate transactions or offer affiliate potential. A model will be lightly encouraged to use its tools when relevant, and if one of its tools is a shopping module…

Sherveen’s stance: Paid ChatGPT subs should be funding the product’s efficacy, which in this case, is also its neutrality. If OpenAI wants to cut in publishers on the $$, cool — do it via transparent licensing, not affiliate skew.

Tariff Transparency & Consumer Gloom

Apollo Global Management's latest report paints a grim picture:

1. CEO confidence is seeing its steepest decline since '08

2. The share of credit card accounts making only minimum payments hit a high of 11% (well above the ~9% average since 2012)

3. consumer sentiment is cratering across all income groups, hitting '08 lows; and 70% of consumers expect more unemployment ahead.

These aren't abstract policy debates; the economic anxiety is palpable.

Career takeaway – Employers will tighten belts; negotiate cash comp, not just equity. And maybe hold off on that imported standing desk until you see the tariff line item.

Three Interesting Startups (1 is a red flag)

Think a friend could use a dose of Intent? Forward this along — group chats may rule the world, but good emails still move the needle.

As always, hit reply to let us know what you think! If you aren't subscribed, what're you doing!? Click here!

Sent with Intent,
Free Agency

Hey there, Intent community — new rhythm drop:

• Tuesdays = news + analysis.

• Thursdays = essays straight from our CEO’s brain (no ghost‑writing, promise).

Consider these Thursday notes a quick coffee with Sherveen, our CEO — one big idea we want to talk about. And here’s the first…

What Could an AI-Native Social Network Actually Look Like?

Ever since it was rumored that OpenAI is working on a social network, people have been speculating on what it will look like. But most people’s guesses are “ChatGPT‑with‑a‑feed.”

That’s like predicting the iPhone would’ve just been a music-playing flip phone. If we unshackle our imagination a little bit, an AI‑first network could rewrite how we form groups, swap knowledge, even build identity.

Below are ten ideas of how it could look if we got creative with it – a mash‑up of my own brainstorming, plus nudges from ChatGPT, Gemini, and Grok. Pick your favorite; tell me what I missed.

1. Prompt Clubs

Think "book club" meets Midjourney prompt group. Small, curated rooms where users collaboratively riff on a central prompt or topic. The LLM acts as a facilitator, generating nudge-responses, relevant data, visualizations, and summaries, while users branch, remix, and build upon each other's ideas.

2. Personality Leasing

Imagine Midjourney's profile codes, but for LLM interaction styles. Users could borrow pre-defined or user-generated ChatGPT personalities – adopt a "skeptical investor" voice for analyzing a pitch, a "scientist” or “therapist” for exploring a particular problem, or even blend multiple voices for different use cases (work, research, fun).

3. LLM‑Driven Multiplayer Games

Go beyond simple chatbots into rich, interactive narrative experiences. Think improv theatre, complex choose-your-own-adventures, or Dungeons & Dragons campaigns where the LLM serves as a dynamic, responsive game master, crafting worlds, characters, and consequences in real-time based on player input.

4. GPT‑Fueled Debates

Picture Twitch meets the Oxford Union, moderated and augmented by AI. LLMs could structure debates, provide real-time fact-checking or context, generate summaries, and even allow participants to deploy personal AI agents to help formulate or counter-attack arguments. Viewers could get AI-generated TL;DRs, citation trails, and maybe even "logical clarity" scores.

5. Memory Graphs

Move beyond a literal post to visualize the mindset behind it. Imagine any post accompanied by a semantic map showing the poster’s related beliefs, intellectual influences, shifts in thinking over time, or core motivations. We often follow people for how they think more than what they say.

6. The Rabbit Hole

Each post has a branching portal, connecting you through LLM-infused meaning – not just by keywords – to adjacent ideas, older posts and Twitter moments, deep dives, and novel questions. And so a post can be a forever branch into the chronology or the deeper meaning of a thing, allowing for endless spelunking.

7. LLM‑Mediated Invites

Your AI can flag someone else as a potential valuable contributor to something you post (and negotiate with their AI to make sure it’s a good idea). A political tweet can get a scientist’s point of view. A fintech thread gets augmented by the CEO of a startup or a skeptical investor. Inclusion becomes AI-mediated and frequent.

8. Living Profile Pages

Your profile becomes a dynamic reflection of your current state of intellectual and practical being, not just a list of what you’ve posted. Based on your asks, your interactions with ChatGPT, content you’re consuming, and your outputs, it surfaces what you’re currently learning, consuming, expert in, or open to discussing — facilitating more relevant connections.

9. Prompt the Collective

Ask something vulnerable, get it crowdsourced, like “How should I deal with career change?” GPT distributes it across the network in disguised fragments. Some users will get a direct prompt to answer. Someone else will get a minigame or a story prompt. No one’s responding to you directly, but you get back an AI-fueled synthesis of the collective’s insights and experiences.

10. AI‑Encapsulated Artifacts

GPT doesn't just facilitate conversation; it canonizes significant moments. Imagine sprawling, multi-day discussions or viral threads automatically distilled by AI into enduring artifacts – a podcast episode summarizing the key arguments, an interactive timeline, a visual moodboard capturing the sentiment, or even a deeper interpretive piece expressed through generated music or imagery.

So what? If these resonate, it hints the winner in AI social will feel closer to software‑defined community than scrollable feed. The bigger shift: AI moves from content generator to context architect.

(related: my thoughts on how Midjourney's personalization feature should inspire everyone building an AI product)

A few short shares —

• VC in the ICU? Dan Gray (Head of Insights at Equidam, startup valuation platform) argues venture capital’s unit economics are fundamentally broken. Blog post here.

• In the Boston area? The Feedback Loop is live at Harvard tomorrow (4/25) — student founders pitch, they get unfiltered investor feedback, you eat popcorn. RSVP here!

• Free workshop: AI‑Enabled PM — Build Better, Faster with Just ChatGPT – Join me on Monday, Apr 28, 4pm ET/1pm PT. Free & virtual, RSVP!

Think a friend could use a dose of Intent? Forward this along and tell them Thursdays just got nerd‑ier.

Sent with Intent,
Sherveen Mashayekhi

It’s been a few months since we hit your inbox, so if you’re confused — yep, you subscribed at some point!

As a reminder, Intent is all about helping talent in tech become more intentional with their career by staying informed, fluent, and aware of what’s going on in and around the industry. Thanks for sticking with us!

Today's agenda:

• Cluely: The AI app for cheating on… everything?

• WaPo x OpenAI: A new partnership for LLM search content

• Y Combinator's Founder v Palantir: A moral debate

• & three interesting startups

Cluely, the AI tool that wants you to cheat on everything

A 21-year-old Columbia dropout named Chungin “Roy” Lee just raised $5.3 million from VCs for his startup, Cluely. It’s an AI desktop assistant that sees your screen and hears your audio, without the other side of any conversation being aware.

The way the founder pitches it: he wants to help users cheat on exams, job interviews, sales calls, and even dates.

Cluely’s manifesto compares their tool to calculators and spellcheck – innovations once decried as cheating. But let's apply some nuance here. Calculators augment calculation and spell check corrects typos; both operate within established expectations.

Our take:

• Cluely wants to operate in situations where authenticity is the implicit assumption. They want to misrepresent a user’s knowledge or personality.

• Employers, test-givers, and yes, even dates don’t expect you to have covert AI assistance. There’s an unspoken agreement that your knowledge – your authenticity – is what’s on the table.

• Marketing stunts are nothing new. The shock value has definitely worked: going viral on social media, hooking investors, all while creating a swirl of controversy. But stunts can push moral boundaries, especially when they trivialize the notion of cheating.

WaPo Feeds the Beast with an OpenAI Partnership

The Washington Post just announced a new partnership with OpenAI. The deal means ChatGPT users will see summaries, quotes, and links to WaPo articles in response to relevant queries, complete with clear attribution.

For The Washington Post, this is about "meeting audiences where they are," as their Head of Global Partnerships put it.

Translation: it’s about expanding reach beyond their paywall. Every publisher wants to protect their traffic and drive up subscriptions. Plus, they get to reinforce their brand authority within the dominant AI platform. It's a distribution play in an era where platforms eat publishers. Interestingly, WaPo emphasized it remains "LLM-agnostic," despite this deep integration.

For OpenAI, it’s a valuable new data channel (both for enriching responses and for future model training). As LLMs hunger for high-quality, timely information, forging relationships with credible publishers is key — especially in a data-scarce climate (where unique data is increasingly locked behind paywalls).

Paul Graham vs. Palantir: The Moral Debate

Let’s set the scene: a few days ago, Y Combinator founder Paul Graham tweeted that first-rate programmers have “a huge number of other places” to work “rather than at the company building the infrastructure of the police state.”

This was in response to ICE signing a $30 million contract with Palantir to build ‘ImmigrationOS,’ a software platform to expand surveillance and enforcement operations.

Palantir exec Ted Mabrey fired back with an aggressive response, calling the stance a “luxury belief.”

Breakdown:

• Graham sees Palantir as enabling state overreach at a time when the US constitution is clearly being challenged. To him, it’s a moral choice for tech talent.

• Mabrey thinks Graham isn’t taking the struggle of government enforcement seriously and that Graham’s ask for Palantir to commit to not violate the US constitution is in bad faith.

It’s a powerful conversation, made more high stakes than usual thanks to the presence of a fascist leader who can’t seem to go a day without cancelling a student visa or disappearing an immigrant.

There’s a real tension between building powerful tools for the sake of national security and the ethical red lines that those tools might cross. Aside from ICE, companies like Palantir and Anduril do key innovation work for national defense, and there’s no good argument for letting other companies build modern weapons and defense systems while we sit on our hands. 

We think Mabrey and co could commit to the constitution, though.

Three interesting startups to watch (some are hiring!)

• Strawberry — browser that layers AI research, writing, and automation onto any site. Assistants can fill out forms and do quick research in 100s of tabs . Hiring: Founding Engineers (SF, will relo).

• Aqua — extremely fast and extremely accurate speech to text dictation that works in every text field, from Gmail to your terminal. Speech → code at 4× typing speed.

• Supabase — open‑source Postgres backend benefitting from the rise of “vibe coders,” fresh off a $200 M Series D at $2 B valuation. Mission: be the Firebase of the AI era. Hiring: Engineering, Growth & Ops roles (Global Remote).

Think a friend could use a dose of Intent? Forward this along – inbox envy is real.

Sent with Intent,
By Free Agency

Today's agenda:

• What the Bench shutdown ACTUALLY means for startups

• What NOT to do with AI in 2025

• Ways you’re sabotaging your 2025 job search

The Bench Bombshell

Recap, ICYMI: Bench, the VC-backed accounting startup serving 35,000+ small businesses, just showed us how quickly things can unravel in tech. On December 27th, they abruptly announced their platform would shut down, leaving customers scrambling for their financial data just before tax season.

The plot thickened quickly:

• Former CEO Ian Crosby blamed the board's decision to replace him

• Customers were given until March 2025 to download their data

• Then, Employer dot com swooped in with an acquisition (price undisclosed)

But here's what's interesting: Bench's "$60M Series C" wasn't what it seemed. It was actually $37M equity + $23M venture debt, highlighting a broader industry issue.

Venture debt can be deadly when things go south. These are special loans for VC-backed startups. They let the company get equity-free capital, they are easier to qualify for thanks to tech-friendly banks that evaluate the deal based on the VCs involved rather than the fundamentals, and they can be used flexibly (unlike traditional loans, which may be particular to a specific business purpose).

BUT — they come with strict covenants requiring companies to:

• Maintain minimum cash balances

• Hit revenue/growth targets

• Stay under maximum burn rates

• Keep specific board compositions

Break these covenants? Lenders can demand immediate repayment and freeze accounts - exactly the kind of sudden death we just witnessed.

Watch this space: According to Carta data, companies that raised $50M+ in 2021-2022 are hitting their make-or-break moments. Bench won't be the last surprise shutdown we see in 2025.

Three Things NOT to Do With AI in 2025

Yes, AI is going to keep getting better. But don’t expect 2025 to be the year you can outsource all of your mental heavy-lifting to a large language model. Here are three “don’ts”:

1. Deep or Comparative Research
Tools like Perplexity, Exa, and ChatGPT plugins claim to do complex research and web- or Reddit-based product comparisons for you. But if you’re trying to find that obscure bit of feedback (like that random post from user dogs_r_kewl_1991) describing an exception to a credit card’s 3% cashback rule), you might be better off scanning yourself.

AI’s retrieval is still limited when it searches the web. It’s pulling the same results any human can access, but then is deciding what seems “worthy” and fits within its token limitations (size of request + response), meaning you lose the potential for the manual “serendipity” moment. Let AI serve as an augment, but keep a human eye on the details.

2. Intense Emotional or Psychological Reliance
We’ll keep hearing about AI therapy and increasingly-awesome voice-based “companions” for deeper conversations. However, most LLMs remain highly sanitized, or “neutered.” That’s for good reason: they don’t yet have the nuance for these companies to let them loose in vulnerable situations. So, the system instructions and tuning will always drive these models to the “average” response to avoid chaos. We’ve already seen AI chatbots cause emotional crises in 2023 and 2024.

While they can spit out basic mental-health tips, your mileage may vary if you need real human-level wisdom or if you’re looking for advice that departs from the safe, consensus boilerplate. Use it as a sanity check rather than as your primary BFF.

3. Unsupervised Long-Chain Tasks
We’ve seen an explosion in use of automation tools (Zapier, Make) and “AI agents” that promise to handle scheduling, outreach, or multi-step tasks. They can be impressive. But the biggest risk is all the “near-miss” moments.

Each call to an AI in a multi-step chain is siloed and has a particular task. An email might get sent to the wrong person because an AI call didn’t expect the primary contact to be CC’d instead. A personalized sales email might say “saw that you have a LinkedIn account - cool!” because it didn’t have anything else to say, but it had to spit out an answer.

Until these chain-of-thought agents get 5–10x better at context (some companies are working on orchestration, aka “manager agents”) keep a human in the loop for quality control.

Become the AI Whisperer: When to Bully, Flirt, Teach, etc. — free, live workshop on Monday, Jan. 6 — recording will be sent to anyone who RSVPs. RSVP!

Three Things Not to Do in Your 2025 Job Search

If you’re kicking off a job hunt in this super-competitive market, here are three faux pas to avoid:

1. Let AI Do All the Talking
Don't over-rely on AI tools. Skip the AI resume makers, cover letter generators, and application automation. They produce generic, obvious results that scream "AI-generated" to recruiters. Instead, use ChatGPT strategically - for document critique, research, and interview practice.

Remember: your application material will exist permanently in the applicant tracking systems (ATS) that you submit to — do you really trust an AI wrapper to apply to the right jobs and answer questions in a way that doesn’t seem off?

2. Wing It on “Why You?”
From the first recruiter screen onward, show that you’ve genuinely thought about why you’re a fit for this company. Show them how your war stories could solve their immediate or near-term challenges — it’s analogies all the way down.

If you’re a product manager who has worked on payments at Shopify, and you’re interviewing at Uber for the driver experience team, stretch to showcase that relevance: tell stories anchored around how your prior experience with streamlined checkout features boosted your customer satisfaction, and that could be applied to the rider app > make riders happier > lead to better UX for drivers.

3. Ask “Any Hesitations About My Candidacy?”
This question has become a cliche, and rarely does it bring about a magically honest conversation that helps you fix negative perceptions on the spot. Most interviewers won’t reveal what they’re really thinking, and you could inadvertently end the interview on a sour note. Instead, ask forward-looking or strategic questions that showcase your intelligence and enthusiasm.

3 Steps to a Stellar 2025 Job Search — free, live workshop on Thursday, Jan. 9 — recording will be sent to anyone who RSVPs. RSVP!

That’s all for now – may your 2025 be filled with fewer unexpected shutdowns and more strategic AI adoption.

Sent with Intent,
By Free Agency

Hey — quick note: on Dec. 20th, join us for AI Fluency: Master the Chain of Thought. It’s a free virtual workshop where we’ll be walking through a reasoning technique that you can use with LLMs (like ChatGPT, Claude, or Gemini) to get 10x better prompt results. It’s also the primary innovation behind groundbreaking reasoning models like OpenAI's o1, Alibaba's QwQ, and DeepSeek's R1-Lite. Understand it, and you'll understand the next year+ of new AI breakthroughs.

This is part of a new series on AI competency for people who don't yet feel like they've tapped into the upside of our new large language model overlords. Note: this is intended for non-technical audiences or technical folk who want to be better at using and talking to AI. No coding required! RSVP here.

In tech, you’re in until you’re out

Back in the heyday of Silicon Valley's venture-fueled boom, Carta and its CEO, Henry Ward, could do no wrong. The company simplified cap table management, making it indispensable for founders swimming in equity grants and secondary market dreams. (cap table management = helping startups keep track of equity ownership; secondary market = when an existing shareholder sells to an outside investor).

Ward’s star rose as Carta became the darling of a cash-rich tech ecosystem, even spearheading an ill-fated secondary market product designed to make private equity liquid.

In fact, it was so popular that venture capitalists began to force usage of the software as part of their term sheet offers to startup founders (incl. to the startup writing what you’re reading!) — despite price tags that were astronomical compared to alternatives.

But the shine wore off. Reports of gender discrimination lawsuits and allegations of a retaliatory firing against a former CTO darkened the skies (TechCrunch). The secondary market experiment failed to deliver and caused an uproar on Twitter, and operational missteps had customers tweeting about its alternatives.

And a few days ago, a founder’s tweet on the company’s arduous cancellation process caused a stir within tech circles, with competitors like Fidelity, Pulley, and AngelList being… aggressively… recommended in the comments.

This is a classic case of Silicon Valley’s tribal ethos: one minute, you’re a hero and part of everyone’s come-up; the next, a cautionary tale. Ward, once lauded as a visionary, is now a polarizing figure. He’s seemingly no longer allowed to move fast and break things, at least in the eyes of his founder peers. In an ecosystem obsessed with rapid growth and doing things that don’t scale, Carta’s story reminds us that you can only have a tough-to-cancel product if the in-crowd still loves you.

Editor’s note: Carta is no longer forced on founders by many relevant venture capitalists. But it’s still up to the founders to figure out how to cancel a product they might not have wanted in the first place.

Platforms or Products

A few days ago, OpenAI released their video generation model, Sora — and the most interesting part to us: the clear decision they’ve made to go product-first with video generation rather than platform-first.

It comes equipped with a mini video editor, a storyboard feature, a gallery for managing generated content — it’s a product suite with no available API for third party developers, a notable difference from what OpenAI usually offers when it releases a new large language model. When a new GPT-X model is released inside ChatGPT, they either immediately or soon-thereafter release it via API.

Why the change? It could come down to trust and safety concerns. Video generation raises complex ethical questions around content authenticity, deepfakes, and misinformation. By controlling the experience end-to-end, OpenAI can set boundaries, ensure compliance, and avoid lawsuits that might have come if third party developers had gone… experimental… with things.

Beyond that, this allows OpenAI to shape the tool and pitch it as a positive augment to creatives in the media industry. They know that societal pushback is an existential threat, and being able to shape the tone and approach of a product that could “replace” filmmakers means a more considered and PR-friendly strategy.

Very few companies can be both product and platform, though. Building products means capturing the revenue directly and shaping the user experience (ChatGPT). Being a platform means rapid scale of use cases and usage (GPT-4 via API). And the needs and internal structures of companies in either lane look quite different.

And this same question likely gets re-litigated every few months and with every product release at Google, Anthropic, and Meta, too. It will be interesting to see these companies navigate the maze, especially in light of the fierce competition they find themselves in with… each other.

PS — in a job search? Join us on Dec. 19th for Behind the Curtain: How Recruiters Actually See Your Profile. We’ll be walking through a live view of an applicant tracking system (ATS) and LinkedIn Recruiter. RSVP here.

Do you know who Speed is? (aka IShowSpeed)

We're curious! And we want to know. Does our tech-y audience know who this is? We'll talk more about why we're asking in the next issue of Intent.

Yes

No

Alright, that’s all for now. Happy Friday!

Sent with Intent,
By Free Agency

Five YC F24 Startups Worth Watching

Y Combinator's Fall 2024 batch just wrapped, and there's a clear theme emerging: startups that make AI actually work in the real world. Here are five that caught our eye (and why we think they matter):

HumanLayer
HumanLayer is basically "humans-as-a-service" for AI systems – their API lets AI agents ping actual people for feedback, input, and approvals. Think about it: as we race toward more autonomous AI, the ability to guarantee human oversight could be the key to faster, safer AI deployment. It's like having a responsible adult in the room while the AI does its thing. Humans-in-the-loop.

Canvas
Canvas is doing something sneaky-smart in the customer success space. Their AI copilot analyzes all those tiny signals in user behavior that humans might miss – like when a SaaS customer who used to use a feature daily now only touches it monthly. That's the kind of early warning system that could save millions in retained contracts.

Fix AI
Fix AI might have the most straightforward pitch of the bunch: "Find bugs before your users do." Their AI agents act like real users, clicking around your site to find problems – it’s a new flavor of automated QA. With teams getting smaller and AI making it possible for almost anyone to ship products, this kind of automated testing could be huge.

Gander
Gander is tackling the mess that is airline customer service. Instead of trying to rebuild decades of legacy systems (good luck with that), they're putting AI on top of existing infrastructure to handle disruptions and compensation claims. This is especially key in travel because even the simplest claim can involve customer auth, reconciling tickets across systems or even airlines, finding off-book accommodations… This could be the playbook for modernizing other complex industries – don't replace, augment.

Abundant
Abundant is kind of like HumanLayer, but for expertise in addition to oversight. When an AI agent is either stuck, needs specialized approval, or wants an industry expert’s context, it can seamlessly route in a human from Abundant’s network of relevant talent. When an AI agent is getting to 60-90% of a use case, a human-in-the-loop can bring it to 100% – today.

The thread running through all of these? They're not trying to build better AI models – they're building the infrastructure to make AI actually useful in the real world. Watch this space.

In a Job Search?

There are ways to access the "hidden job market." But we’re not talking about unlisted jobs, secret jobs that only the well-connected can access, or other mythological BS spread by people who should know better. It’s more practical than that: utilize advanced Boolean search techniques to find better roles, faster, before anyone else – sign up for our FREE workshop here: RSVP.

The New Era of Consumer Pricing

OpenAI just dropped their new pricing alongside their latest reasoning model, o1 (and its companion, o1 pro mode). It's... interesting:

Plus ($20/month):

• More messages, file uploads, and analysis

• Advanced voice features

• Limited o1 access

• Beta feature testing

• Custom GPT creation

Pro ($200/month):

• Unlimited access to the good stuff: GPT-4o and o1

• All the real-time voice features you can handle

• o1 pro mode for the really hard problems

Here's our take: this is the future of AI product pricing, whether we like it or not. Running these models is getting more expensive, not less – between chain-of-thought reasoning and test-time compute, we’re adding more GPUs and more runtime to every query.

But users are becoming dependent and want unlimited access to these tools they now rely on. We've seen people on X begging to pay $500/month for unlimited Claude access (looking at you, rate limits).

In this era of tech, we’re going to see higher pricing tiers and caps for consumer products than ever before, but it’s often to the benefit of power users who want to “super engage.” It’s about playing a different game – aligning the product's capabilities with the level of investment power users are willing to make, maxing out the upside for both parties.

If you’re an AI product builder in this era, you might want to reframe how you think about pricing and usage. Is your product 10x better when it’s unfettered and unlimited, in a way that might make your users willing to pay 10x more to access it?

A few extra links

• [YouTube] What if Shark Tank, but better? Live startup pitch feedback!

• [Free workshop] How to answer “sticky” interview questions

Think a friend could use a dose of Intent?
Forward this along. If you’re not subscribed, enter your email here!

Sent with Intent,
Free Agency

3 topics today:

• 2024 startup founder salary data!

• TikTok in hot water over Romanian elections?

• (Bluesky OR Threads) versus X

But first!

• FREE workshop: Build a Resume That Converts (Template + Storytelling), RSVP

  • Data-backed, evidence-based advice and templates on building a resume that works — including myth-breaking around ATS keywords, AI, and how your resume is viewed by recruiters + HMs.

• FREE workshop: How to Answer “Sticky” Interview Questions, RSVP

  • Gaps, layoffs, jumpy resumes, recently fired — no matter the question, there’s a way to get out of the “sticky” part and into the upside. Get an unfair advantage w/ our Rubber-Glue Framework.

Show Me the (Founder) Money

Kruze Consulting, a leading CPA firm for venture-backed startups, just dropped some eye-opening salary data in TechCrunch. If you've ever wondered what founders actually make, here's your answer: at the seed stage, CTOs slightly edge out CEOs ($134k vs $132k), while COOs clock in at $135k.

The progression is interesting, too. By Series A, CEOs jump to $183k, and Series B pushes them north to $218k.

And depending on your POV, these numbers are way below what these roles command at established companies. When a senior product manager can pull $200k+ in base and another $150k+ in liquid stock at a FAANG, you can see the real price of the founder’s lottery ticket and opportunity cost. It's a reminder that early-stage startups still run on dreams and equity – until you’ve raised major institutional cash at the Series A or B, and then things get more comfortable.

Algorithms v Democracy, Continued

Romania just gave us yet another view into social media's evolving influence into democratic processes. An ultra-nationalist candidate rode a wave of TikTok virality from obscurity to a commanding first-round victory in their recent presidential elections, with signs pointing to potential Russian interference in the content ecosystem.

According to Politico, media regulators and election observers are uncovering what they suspect is “a covert operation conducted through thousands of fake accounts.” Further: “The catchy TikTok clips that powered Georgescu’s unexpected surge were accompanied by dramatic music and subtitles. He was shown barely breaking a sweat on the running track, flipping opponents in judo — à la Putin — and riding a white horse in a traditional Romanian shirt.”

ByteDance and TikTok deny these claims.

Beyond any one platform in particular — Romania’s agencies have been defending and warning about cyberattacks, improper categorization of election material on social media sites, and other violations of European Commission oversight.

The Social Media Splintering

Bluesky has quietly climbed to 24 million users (Wired), a far cry from Meta’s Threads at 275 million — but user numbers don’t tell the full story.

Bluesky has become the spiritual successor to what Twitter used to be: quirky, dynamic, and unfiltered (ZDNET). Meanwhile, Threads feels like the corporate cousin: well-polished, but lacking depth and personality while being overfull with brands (Fast Company).

This isn’t just competition, though — it’s segmentation. The worry from users who are on all three platforms: they each feel distinctly different and lack user overlap, further segmenting us not just based on an “algorithm bubble” like we’re used to, but now a platform bubble.

In other words, you’ll get your left-wing politics and coastal-style humor if you’re on Bluesky, your milquetoast brand and influencer content on Threads, and your right-wing politics and business news + content on X/Twitter. And you’ll need to be on all 3 (a tall order) if you want a healthy mix.

And given that the negative effects of algorithm bubbles over the past decade have already seemed extreme and harmful… this type of "pseudo-philosophical” platform-based splintering, not just in social, but across AI, gaming, and even SaaS, might be something to watch out for.

And one more thing…

Keep your eyes on OpenAI tomorrow. They're kicking off 12 days of demos and product launches, and word on the street (via The Verge) is that we'll be seeing Sora and a new reasoning model in the mix. Could be a game-changing holiday season for AI.

That’s all for now —

Sent with Intent,
By Free Agency

Hey there, intentional people!

In the post-election fever, there’s one issue we’re paying attention to more than most…

(PS — before we get to that: don’t forget to sign up for our upcoming Lightning Lesson, How to Ask Questions That Win You Job Offers — where Free Agency CEO, Sherveen, will be talking about how the questions you ask at the end of job interviews can do so much more for you than you might expect — it’s free, & if you can’t make it live, all RSVPs will receive the recording, sign up here!)

The issue of AI regulation is likely the most complicated that the incoming Trump administration will have to wrestle with, and there are a variety of camps to contend with.

Startups and tech companies will be dramatically affected in the short-term. In the mid-to-long-term, we’ll likely be reaching for a more international consensus (by will or by force). We don’t know where it will go, so today, all we want to do is map out the positions of the stakeholders we already know about:

• Outgoing Biden admin: AI executive order mandating safety and privacy protections, pushes gov’t to use AI in national security (source)

• Big tech companies: Amazon, Google, OpenAI, MSFT are just a few of the names pushing Congress to formalize a US AI Safety Institute for responsible development of AI systems (source)

• Safety-oriented AI players: Companies like Anthropic are raising more alarm bells than most re: AI risk and desire for regulation (source) and an ex-OpenAI researcher recently warned re: AGI: “In short, neither OpenAI nor any other frontier lab is ready, and the world is also not ready” (source)

• Venture capitalists: Firms like Founders Fund and a16z, who closely aligned themselves with the Trump campaign over the past year, have pushed on social media for a low-to-no regulation approach to enable innovation above all else (source)

• Incoming Trump/GOP platform: In their 2024 platform, the GOP indicated its intent to repeal the Biden EO and indicated “support... for AI Development rooted in Free Speech and Human Flourishing” (source)

What we don’t know: how much power do the populists within the Trump party have to influence policy? This is where it gets messy. You have Steve Bannon coming out against “unbridled AI” (source), but does he have Trump’s ear more than, say, mega-supporter David Sacks, who sees AI regulation as a threat (source)? Labor-oriented left-wing politicians could join the populist chorus, too, aligning with the Bannons of the world to protect the status of human workers.

And Elon has long been an AI doomer, marking a 20% chance for it to cause humanity’s downfall, but participating in its acceleration along the way (source). Having just paid $44 billion to help Trump secure his latest gig, he might be owed more favors than most.

If you don’t know where it’ll wind up after reading all of that, don’t worry — we can’t tell, either — but we think this is a space worth watching. Let us know what you think!

Sent with Intent,
By Free Agency

Intent is all about helping talent in tech become more intentional with their career by becoming more informed and more fluent about the goings-on within + around the industry. We welcome your feedback!

Hey there, Intent community! Been a while, but we’re back again. Let’s ride into this rollercoaster of a week together. Here are three quick hits to start.

Claude’s Computer Use = Game-changing

ICYMI, Anthropic released an API that “allows Claude to perceive and interact with computer interfaces” — in other words, you can send constant screenshots back to Claude to help it see your screen, and let it drive your mouse, keyboard, and operating system.

One of our favorites on Twitter, Ethan Mollick, gave Claude with computer use this prompt, plus a related video: "Hey Claude with computer use, watch this construction site video & write up things you see that dangerous or good, create a spreadsheet of critical issues to address"

Check out the result here.

We’ve seen it play Magic the Gathering, operate a phone, build and iterate on a timer app, solve CAPTCHAs, and design in Figma. Game-changing.

By the way, are you job searching? Join Free Agency’s CEO for a free, live, 30-minute workshop next week — “How to Ask Questions That Win You Job Offers.” Most people ask the same boring, repetitive questions at the end of their interviews — putting recruiters and hiring managers on autopilot. But YOU control the agenda in that moment, and Sherveen will share his 3-part framework to asking the sort of questions that will make you unforgettable. Sign up here!

Miro’s Layoffs Add to Enterprise Woes

Miro, the virtual whiteboarding startup last valued at $17.5 billion, last week announced plans to lay off 18% of its workforce (275), as reported by The Information. It comes at a time when enterprise spending is generally shifting and contracting — we saw Dropbox lay off 20% last week (TechCrunch), 10% at Docker (Layoffs.fyi), 21% at Upwork with enterprise specifically cited (Upwork), all at the tail end of October.

As companies both big and small, public and private, have cascaded into layoffs over the past 2 years, there has been a cyclical effect — SaaS providers serve other SaaS providers, and as SaaS providers reduce headcount, they need less SaaS, which means more SaaS headcount reduction… you see where this is going.

As OpenAI Goes Up, the Rest Must Come Down

Related: a super interesting piece in The Information forecasts that we'll see a reduction in spend for the rest of the enterprise and cloud application space as $$ to conversational AI goes up. CIOs of major companies are explicitly stating that they are carving out IT budget to create spending room for AI.

Beyond that, enterprise companies are beginning to craft their own custom tech to replace large software contracts — Toyota, for example, is spinning up an OpenAI-powered chatbot for internal company Q&A for the carmaker’s policies.

What happens to major vendors like Salesforce, Workday, and Atlassian as the trend continues — does a combination of LLM spend, reduction of software headcount, and custom, AI-driven internal tooling mean a greater reckoning in place for 2025+?

That’s all for now —
Sent with Intent, by Free Agency

Hey, NYC —

Tomorrow's Founders Friday with Patrick Rafferty (UserHub) is being rescheduled. He could no longer make it and had a good excuse 😉 so we're looking forward to having him join us on a future Friday!

In the meantime, get excited for next week — Nikhil Patel (Forbes 30u30) will be joining us to talk about Craniometrix (YC 22). RSVP here.

Stay healthy,
Free Agency