Ari Weinstein

1. Key Themes

Computer use is crossing the "faster than a human" line, and expert-level speed is the next threshold

Ari Weinstein frames speed as the unlock that changes how everyone interacts with computers. Once the agent beats an expert user, the default flips: manual work becomes the exception you choose. He says: "We've gotten computer use now to a place where it's faster than the average human at using a computer, and I think we're getting really close to a place where it is superhuman, meaning it is faster than an expert computer user like probably one of us at this table, and I think that's going to actually be a sea change when that happens." [00:09:49 region; see 00:10:15 block] He continues: "anything that you could do manually, it could do faster. So, unless you really want to do it yourself for some reason... in the default case, you probably just want the agent to do it for you." 00:08:37

Universality is the structural advantage: computer use works with everything, with no integration required

Unlike APIs or connectors, computer use needs no developer adoption. Ari Weinstein: "It's unlike any other thing that we've built, where, you know, the developer has to do some work to adopt it, someone has to build a connector or MCP server, you know, computer use works with everything." 00:00:34 James Sun reframes it as the model's fallback layer: "computer use is not like a thing you pick up and just say, 'I want to do something with it.' It's really like the universal kind of API that the model has to do whatever it takes." 00:16:37 The implication is that the more ambitious the task, the more likely it eventually touches computer use.

Enterprise legacy software is the beachhead, because the capability was ready before the buyers were

James Sun explains why enterprise demand is "fairly obvious": "Inside enterprises, there's lots of legacy software that doesn't have any MCP tools, they are software that's been used for 15, 20 years..." 00:04:31 He adds that "enterprise is where the capabilities have been there, maybe for a little bit longer, and the companies are catching up." 00:04:31 Consumer, by contrast, is "where we try to push the frontier on what's possible." 00:04:31

The "boring" systems of record are the highest-confidence automation targets

James Sun lists the categories: ERPs, CRMs, procurement, finance reconciliation, sales prospecting, and market research. "Every single system of record." 00:21:13 On finance: "a lot of these recon- us financial software don't have APIs to be able to quickly pull things... also a very boring, but useful workflow." 00:21:13 The procurement example, where a 1,000-row spreadsheet is passed from teams to ops to procurement to vendors, is a multi-hop human relay that computer use can collapse.

Control, auditability, and deterministic guardrails are what unlock enterprise adoption as speed outruns human observation

James Sun argues that once agents are too fast to watch, governance becomes the product: "once you can get to a point where you can no longer observe and audit everything it's doing, um how do we give, especially enterprises and businesses, the ability to control um the amount of human-in-the-loop there is, and then the auditability, right?" 00:10:15 The mechanisms: a "Guardian" or auto-approve model that "is a really smart model looking at every single action that the computer use agent is taking to decide if it's safe," plus deterministic sandboxes: "it can use certain apps, it can't use other apps, it can use certain websites and not certain other websites." 00:18:03 00:18:03

Proactive, background agents are the next product form factor

James Sun describes Dot flagging an external invite missing from his calendar and monitoring Slack for critical feedback with a high-level goal rather than a fixed script. "I think it's really good at proactively looking for tasks where the shape of it is not super, let's say, concrete every single time." 00:06:34 Ari Weinstein: "I'm really excited about a world where the agent can sort of adapt to you instead of you adapting to it." 00:40:14

The web's business model is a friction point for agents, and cooperative standards like WebMCP are emerging

James Sun: "You think about advertisers, publishers whose business models entirely on advertisement. If the agent's using the website for them, then where's that revenue coming from? And I think most places, a agent's ad impression does not actually count towards real uh revenue." 00:12:09 The cooperative path is WebMCP, which lets "the agent is elevating the experience of using the site, but is not entirely replacing it." 00:12:09

Cheap software raises the bar for craft, and personalization becomes the new frontier

Both guests expect legacy software and new software to coexist, with agents-under-the-hood plus real UIs on top. James Sun: "the pressure of cheap software will make it so that software has to be better." 00:27:29 Ari Weinstein: "any piece of software could be designed for 100 people or one person." 00:29:57

The long tail of non-tech industries is under-diffused, which is the startup opening

James Sun on restaurant groups: "those restaurant groups are probably not like tech companies that are um super you know, familiar with all the tools that are available." 00:35:04 His diffusion chain: "even understanding what the tools are capable is one, connecting tools in a way that makes sense is two, deploying it is three, servicing is four." 00:35:04

Growth is inflecting, and mobile/cloud is the next unlock

James Sun: "it's like YC level of growth week over week." 00:42:16 Next step: "how do you give everybody who's just using their phone for their day-to-day life, um a powerful computer that they don't even need to understand." 00:42:16

2. Contrarian Perspectives

Agents that are too fast to watch make "human-in-the-loop" the scarce, productized feature, not the default

Conventional wisdom says speed is the headline. James Sun argues the opposite dynamic: slow agents let you supervise, but that caps the tasks you can delegate. "When the model is slow, you can kind of watch it do stuff, but then you're limited to the tasks that you can do, because you can't just sit here watching the thing doing it, like it's actually less effective." 00:10:15 So the value migrates to configurable oversight: "a consumer might be like, 'I just want it to do everything, and I don't want to be asked for approval,' whereas an enterprise is able to kind of decide which type of actions actually go through." 00:10:15

Computer use may be the fallback that makes the MCP/API integration world less central, not more

The industry consensus is to build connectors for everything. Ari Weinstein points out that computer use inverts the burden: "someone has to build a connector or MCP server, you know, computer use works with everything." 00:00:34 He goes further: computer use "may put more pressure on companies to make software work better with agents, because all the software is going to work with agents by default." 00:26:38

The AI-native-software thesis is overstated: legacy software persists, and UI isn't going away

Against the "agents replace interfaces" narrative, James Sun says: "I don't think UI is going away." 00:27:29 Ari Weinstein: "humans are so hardwired for direct manipulation." 00:29:57 Ari adds that "a lot of [software in production] was written 10, 20, 30, 40 years ago, and I don't think that that stuff's going to go away overnight." 00:26:38

Even OpenAI says the labs won't absorb everything: execution depth in verticals is a durable moat

James Sun, an OpenAI employee: "even though we're quite big now, thousands of people, we are still feeling very stretched thin internally... speed and pain is kind of the mode that exists today." 00:35:04 He insists the "labs can do everything" assumption in San Francisco doesn't hold for the average SMB.

Atlas, the AI browser, underdelivered on agentic use because people treated it as a normal browser

James Sun, candidly: "in the Atlas product, the agentic experience, I think for a few reasons, wasn't actually like a big selling point. Not many people used it, partly because I think people mostly treated it like a normal browser." 00:42:16 Agentic usage took off only when embedded into existing products.

3. Companies Identified

OpenAI

Frontier AI lab; maker of ChatGPT, Codex, and the computer-use stack. Mentioned as the source of "the fastest, cheapest, and most reliable computer use product." Ari Weinstein: "I think we have the fastest, cheapest, and most reliable computer use product. And, I think there's actually so much further that we can take that." [00:09:49 region; see 00:10:15 block's preceding statement at 00:08:37]

Dot (OpenAI agent product)

OpenAI product giving each user an agent with its own cloud computer. Mentioned for proactivity and dual local/cloud control. Ari Weinstein: "each Dot gets its own computer, and you can ask it to do things for you in the cloud, it gets a full web browser, it can buy things for you, it can book haircut appointments." 00:05:44 James Sun: "Dot is really cool... One is really proactive." 00:06:34

Codex (OpenAI)

OpenAI's coding agent; used in James Sun's workflow and for its computer-use cursor. Fabrizio Serafini: "The codex cursor is so cute, I love that... watching it use the computer gives you like confidence." 00:20:18

ChatGPT / ChatGPT work

OpenAI's consumer/enterprise app, which includes a cloud browser. James Sun: "ChatGPT work also has a browser in the cloud." 00:38:40

Atlas (OpenAI browser)

OpenAI's AI browser product; cited as a learning on agentic adoption. James Sun: "I joined to work on the Atlas product, and love love Atlas, rip Atlas." 00:42:16

WebMCP

Emerging standard letting sites expose structured interfaces to agents while retaining the human UI. James Sun: "WebMCP is a standard that we're really excited about. We added support for it in the in-app browser." 00:12:09

Sentry

Error-monitoring tool; part of James Sun's automated feedback-triage workflow. "log into Sentry, pull the logs, analyze it, give me feedback." 00:06:34

Slack

Internal comms tool; target of James Sun's agent monitoring for critical feedback. "we use Slack internally, and there's a ton of messages that's going on at all times." 00:06:34

Gmail / Google Calendar

Used in the example of Dot surfacing missed external invites. "Do you want me to like accept it in your Gmail, so it shows up?" 00:06:34

Apple Numbers

Obscure spreadsheet app; Ari Weinstein's evidence that the model handles niche software. "I use Apple Numbers, which is like a obscure spreadsheet program, but the model still knocks it out of the park." 00:24:25

Microsoft Paint

Demo environment for Astra's creative computer use. Ari Weinstein: "it's drawing things on Microsoft Paint. It's way better at that sort of thing." 00:02:27

Fallout (game)

Cited as a benchmark-style win for the model. James Sun: "Astra's able to beat Fallout in 30 hours, or something like that." 00:03:23

Restaurant-ops startup (unnamed)

A company James Sun spoke with that photographs back storage, forecasts needs, uses computer use to order groceries via procurement software, and adjusts staffing in the ERP. "it takes that to then use computer use to go to like a Shopee or some kind of like procurement software to then order your groceries for the next day." 00:35:04

Pickleball/squash/tennis video-annotation software (unnamed)

Competitor products James Sun had the agent sign up for and critique while building his badminton tool. "I just had computer use go log in, create an account on these softwares, try it, take screenshots, give me feedback." 00:21:13

ERP / CRM / procurement and finance systems (category)

Legacy systems of record flagged as top targets. James Sun: "all the acronyms you can think of: ERPs, CRMs... Every single system of record." 00:21:13

4. People Identified

OpenAI product leader on computer use and the Codex cursor design; consumer-focused. Mentioned for his product craft and ownership of trust-building UX. James Sun: "Ari builds a beautiful product with his team. The cursor was all him." 00:27:29

James Sun

OpenAI product lead working on enterprise computer use; former startup founder; joined OpenAI in January to work on Atlas. "you know, both Ari and I were startup founders before we joined OpenAI." 00:35:04

Seema Amble

a16z partner/host who co-authored the August piece on computer-using agents. "In August, we wrote about computer-using agents and how they're starting to do the real work inside companies." 00:01:52 She also supplies a key product framing: "Why should we settle for shitty software?" 00:30:33

Fabrizio Serafini

a16z colleague and co-author of the piece; probes startup strategy and human-AI collaboration. "it's so underappreciated the importance of like human AI interaction." 00:32:52

James Sun's wife (unnamed)

Learning the board game Omok; the motivating user for a personalized tactical trainer James built, illustrating one-person software. "I was able to like, you know, craft a software to be something like feels really good for her." 00:27:29

Ari Weinstein's friend, a personal trainer (unnamed)

Using the model to make customized plans for clients. "I have a friend who's a personal trainer who's making customized things for his clients." 00:24:25

5. Operating Insights

Make "agent first" a forced habit, not an option

James Sun runs a deliberate exercise: "whenever I have a task, the first thing I think about is just, 'Kind of just have the agent do it.'" 00:14:31 Ari Weinstein admits the failure mode is inertia: "it's so easy to get locked into the old way of doing things." 00:15:48 Operators should institutionalize the reflex, since the limiting factor is habit rather than capability.

Give agents goals and judgment criteria, not scripts

James Sun replaced a fixed hourly Slack-digest automation with a goal-based instruction: "if anybody in my reporting chain says anything, flag it to me... use your judgment about certain products, go for it." 00:06:34 He had previously "ended up tuning out a lot of the notifications, because it's just sometimes useful, sometimes it's not." Judgment-based filters beat polling digests.

Design for notification fatigue: summarize instead of alerting

Ari Weinstein on collaboration design: "Is something summarizing that, so that you don't get bombarded with notifications saying, 'Hey, this happened, this happened, this happened,' which actually you're going to ignore?" 00:33:45

Use agents as competitive-research and product-feedback loops

James Sun had the agent sign up for competing tools, screenshot them, critique them, and feed findings back into his build. This closes the loop from market research to product iteration inside a single agent run. 00:21:13

Build trust incrementally with visible cursors, permission gates, and graduated autonomy

Ari Weinstein: "you can always see the cursor... and it sort of animates in a fun way, so it's not boring to watch... You kind of want to have this trust-building mode moment where you figure out what it feels like... and then you sort of feel more comfortable over time." 00:19:27

6. Overlooked Insights

The browser/identity fragmentation problem is the quiet blocker to consumer-scale agents

In passing, James Sun lists four disconnected environments: "you can install the Chrome extension, and that uses your like password manager and everything in Chrome, and then we have an in-app browser in the desktop app which has a different set of cookies and different passwords, your computer in the cloud also has different things." 00:38:40 Cookies, credentials, and tab state living in separate silos is a real moat opportunity for whoever unifies identity and session state across local, cloud, and mobile agents, and it is a likely gap for startups.

"Harness" and a system-level view of speed matter as much as model improvements

The announcer's framing says the harness around a model can nearly double its speed, while Ari Weinstein notes Astra's speed gain comes partly from "fewer thinking tokens to do the same task." 00:02:27 Speed is a stack problem (tokens, scaffolding, environment), which suggests infrastructure and tooling around the model, not just the model, is where investable differentiation lives. The small reliability gap Ari describes ("between something being 99% reliable and 97% reliable") 00:02:27 is exactly the margin that decides whether enterprises hand over end-to-end workflows.