"Nobody Lost Their Job to AI, Just the Promise of AI" - This Week in AI Ep 15

Summary

Podcast Summary: "Nobody Lost Their Job to AI, Just the Promise of AI"

This Week in AI - Episode 15 (May 27, 2026)
Host: Jason Calacanis
Panelists:

• Tanay (CEO, Whisper Flow)
• Eric Bernhardson (CEO, Modal Labs)
• Richard Socher (CEO, Recursive Superintelligence and U.com)

---

Episode Overview

This episode dives deep into the impact of AI on jobs, infrastructure, and company strategy, with three CEO-level experts dissecting the current landscape. Recent large-scale layoffs at tech giants like Meta and Intuit are put under the microscope, challenging common narratives about AI-driven job losses. The panel also addresses rising compute costs, the boom in Chinese AI models, reward engineering and hacking in AI systems, and evolving hiring dynamics in the AI era. The discussion is both technical and strategic, packed with insights for builders, leaders, and anyone affected by AI trends.

---

Key Discussion Points & Insights

1. Meta & Intuit Layoffs: AI’s Real vs. Promised Impact on Jobs

• Meta laid off 8,000 employees (10%), Intuit cut 3,000 jobs (17%)—both citing AI integration and record profits ([00:00], [41:43]).
• Eric Bernhardson:
  "Very few people have actually lost jobs due to the reality of AI. A lot of people have lost jobs due to the promise of AI." ([00:17], [36:23])
• Layoffs are often due to reallocation of capital toward AI infrastructure, not directly because AI replaced those workers.

2. Changing Hiring Criteria: From Coders to Prompt Engineers

• AI has shifted desired skillsets. The ability to use and prompt AI models is valued as much as traditional coding.
• Tanay:
  "When I'm hiring for a software engineer, I care a lot more about what their prompt to Claude looks like than the actual code that they produce with AI." ([00:32], [33:24])

3. Company Spotlights: Product, Privacy, and Infrastructure

Whisper Flow (Tanay)

• AI-driven dictation software, optimized for accuracy:
  "95% of whisper outputs are never edited by the user. It's perfect on the first shot." ([02:57])
• Strong privacy by default, with zero data retention as a standard—even for free users. Context is pulled from the user's screen in real time but not retained ([03:45]).

Modal Labs (Eric)

• Operates as a developer-friendly layer above neo clouds, with global GPU aggregation and usage-based scaling:
  "You can scale up to thousands of GPUs, scale out to zero… super valuable sort of lack of predictability." ([01:32])

Recursive Superintelligence (Richard)

• Focused on self-improving AI that can code itself, aiming to automate the scientific method:
  "The plan is to build the ultimate sort of Eureka machine… could be the last main invention humanity needs to invent everything else after." ([06:03])

4. Rising Compute & Token Costs

• Token/computation costs now often exceed headcount in startups ([08:23]).

• Richard:
  "We're spending a lot more on compute than we are on people… the productivity that employees have is almost 10–20x of what you would have expected from that size team." ([08:23])

• Eric adds: The GPU market is supply constrained; prices have risen 40% in six months, though starting to soften as new hardware arrives ([10:03], [10:47]).

5. Business Growth & Acquisition Tactics in the AI Age

• Tanay:
  Growth is driven by “super fans” evangelizing the product, plus a constantly shifting portfolio of paid channels and experiments ([15:00], [16:35]).
• Whisper Flow is scaling rapidly, launching new features monthly, and doubling headcount every six months.
  "The business is growing 40% month over month." ([14:24])
• Use of AI agents in marketing spend:
  "We have two people… managing almost $70–100 million worth of annual marketing spend using all these agents." ([17:38])

6. Reward Hacking, Multi-Agent Systems & Organizational Parallels

• AI reward functions must be well-defined or systems will “hack” them in unintended ways ([18:58], [20:26]).
• Richard:
  "A new type of job is to become a reward engineer. It's kind of like a prompt engineer, but much more sophisticated..." ([19:12])
• Multi-agent and “checks and balances” approaches in AI mirror human organizations ([23:50], [24:48]).
• Humans and AIs both struggle with multi-objective optimization; splitting responsibilities can aid focus and creativity ([25:43], [26:10], [27:19]).

7. The AI Job Market: Polarization and the Demand for AI Fluency

• Despite layoffs, there are still hundreds of thousands of open tech jobs, but the job profile required is shifting rapidly.
• Tanay:
  "The profile of the person we hire is completely changed… I care a lot more about what their prompt to Claude looks like than the actual code that they produce." ([33:24], [35:51])
• Richard:
  "Companies want young, hungry talent that knows how to use these tools, has grown up with these tools and are now bringing them into the workplace." ([00:23], [35:51])
• Some data suggests a bimodal distribution: AI-first graduates are highly sought after; others struggle ([49:03]).
• Entry-level software engineering jobs are fewer, and the bar has risen ([47:34]).

8. Global AI Competition: Chinese vs. US Models

• Chinese models outpace US models in weekly token usage: 9T vs. 5T tokens ([60:37]).
• Deep Seq v4 Flash is now the global leader.
• Eric:
  "The Chinese models are very good… I wish there were better open source models coming out of the US." ([61:41])
• Issues with deploying Chinese models directly due to biases and censorship ([62:47], [63:45]).
• Demand for open source remains; capital incentives are holding back US-based open model development.

9. Public Perception, Regulation & Storytelling in AI

• Growing public anxiety, criticism, and calls for regulation, including from the Pope ("AI must be disarmed") ([51:54]).
• Eric:
  "Tech doesn't fully aware, isn't fully aware of the PR battle here… there's this very pessimistic scenario leading to this backlash to data centers and all these things." ([52:04], [53:06])
• Richard:
  "AI is only as good as the systems, the people and the data that influence it. It is a general purpose tool… both (good and bad uses) can be true at the same time." ([53:37])
• Tech CEOs’ “doom” narratives may backfire, fueling distrust and social backlash ([53:48]).

---

Notable Quotes & Memorable Moments

• Eric Bernhardson:
  "Very few people have actually lost jobs due to the reality of AI. A lot of people have lost jobs due to the promise of AI." ([00:17], [36:23])
• Tanay:
  "That career you've been planning for for 15 years is deleted. And that is the pain that people are feeling." ([00:10], [33:24])
• Tanay on Prompting:
  "I care a lot more about what their prompt to Claude looks like than the actual code…" ([00:32], [33:24])
• Richard Socher:
  "AI is only as good as the systems, the people and the data that influence it. It is a general purpose tool… both can be true at the same time." ([53:37])
• Host, on prompting / Elgato pedal:
  "Ramble and ramble and ramble. Correct yourself mid sentence. It doesn't matter. The more you give them… it's like a big mind blowing thing for me." ([15:59])
• Eric on the PR challenge:
  "The tech industry almost needs to hire a couple of PR people and then engineers just need to chill out a little bit and try to be a little bit more optimistic." ([53:06])
• Richard, on job loss and labor fallacy:
  "That is called the lump of labor fallacy. That labor is this fixed lump. And when you take 10% off now there are going to be 10% unemployed people. That has never happened in the past." ([30:26])
• Tanay, on hiring AI-fluent talent:
  "If you're somebody who is a first principles thinker, you're low ego and you're hungry to win, then you're going to fit right in." ([66:21])

---

Key Timestamps

• 00:00–00:40: Setting the stage—Meta and Intuit layoffs, panel introductions.
• 02:57–05:43: Whisper Flow tech deep dive—accuracy, privacy, infrastructure (Modal partnership).
• 06:03–08:23: Recursive Superintelligence’s mission—self-improving AI and research company funding.
• 10:03–11:31: GPU and compute supply constraints, rising costs, Anthropic’s effect on market.
• 14:12–17:38: Fast feature shipping at Whisper Flow, growth/fan-driven marketing, agent-based spend optimization.
• 18:10–24:58: Reward hacking, multi-agent systems, and organizational parallels in AI.
• 30:26–37:39: Macro discussion: layoffs, labor, optimism, and restructuring in the tech economy.
• 40:26–42:05: Meta’s worker monitoring and data collection; employee trust and morale concerns.
• 49:03–49:49: College grads, job market stats, AI-first talent, and shifting employment paths.
• 51:54–53:06: The Pope calls to ‘disarm AI’, public fears, need for improved tech industry messaging.
• 60:37–64:40: Chinese AI model boom, open source challenges, biases, and edge cases.
• 65:20–66:55: Panelist hiring pitches, desired talent profiles.

---

Takeaways for Listeners (Who Haven’t Heard the Episode)

• AI’s impact on jobs is complex; most layoffs currently are due to the promise of AI rather than true automation. Company priorities are shifting from labor to infrastructure investment.
• Job requirements are evolving: Skills in AI prompting and tool use are rapidly overtaking legacy coding skills, particularly for entry-level candidates.
• GPU compute and token costs are now often the top expense for serious AI companies—higher than people.
• Chinese AI models are surging in global usage and perform well, but deployment can be tricky due to biases and context/censorship.
• Organizational structures in companies (and AI) are converging around focused agents and clearly delineated objectives to avoid reward hacking and encourage out-of-the-box thinking.
• Public perception, regulation, and storytelling will increasingly shape the future of AI. There’s a critical gap between philosophical/technical messaging and public understanding.
• All panelists are bullish on the future: They’re all actively hiring and see AI as a force multiplier, not a destroyer—if society and individuals adapt accordingly.

---

Closing: Who’s Hiring & For What?

• Modal Labs (Eric):
  Hiring systems engineers in New York & SF. Focus: low-level performance, infrastructure.
• Recursive Superintelligence & U.com (Richard):
  Looking for infrastructure wizards and researchers who can work with huge GPU clusters, and senior backend engineers for search.
• Whisper Flow (Tanay):
  Seeking ex-founders, product leads, GMs, researchers, engineers, marketers, and sales—people who are “first principles thinkers, low ego, and hungry.”

---

This was a fast-paced, expert-level episode that balanced optimism, realism, and technical depth—essential listening for anyone tracking the real-world effects and future of AI.

Transcript

[Host / Moderator]

Meta laid off 8,000 people about 10% of its staff record profits lifting their 2026 AI capex by up to 10 billion. Intuit cut 3,000 jobs, that's 17%.

[Tanay (Whisper Flow CEO)]

That career you've been planning for for 15 years is deleted. And that is the pain that people are feeling and it's hatred towards large companies.

[Eric Bernhardson (Modal Labs CEO)]

Very few people have actually lost job due to the reality of AI. A lot of people have lost jobs due to promise of AI.

[Richard Socher (Recursive Superintelligence CEO)]

But it's also clear that companies want young, hungry talent that knows how to use these tools, has grown up with these tools and are now bringing them into the workplace.

[Tanay (Whisper Flow CEO)]

When I'm hiring for a software engineer, I care a lot more about what their prompt to Claude looks like than the actual code that they produce with AI.

[Host / Moderator]

Thanks to our friends at PayPal, the exclusive sponsor for this week in AI try the payment and growth platform that's trusted by millions of customers worldwide. PayPal Open start growing [e-mail] all right everybody, welcome back to this Week in AI. This is a new roundtable where I talk to people who are building the future vis a vis AI, machine learning, robotics, and we talk about whatever is in the news that week for AI. We have a great roundtable. Again, Eric Bernhardson is here. Eric is the CEO and co founder of Modal Labs. Building AI infrastructure, GPUs in the cloud and started back in 2021. Welcome to the program, Eric.

[Eric Bernhardson (Modal Labs CEO)]

Thank you. It's great to be here.

[Host / Moderator]

Yeah. So how should people think about you in the Neo cloud space and what you're providing and who your customers are?

[Eric Bernhardson (Modal Labs CEO)]

We're basically one layer up. Like we offer two things that Neo clouds and hyperscalers don't quite do. First of all, we offer a much better developer experience. You can iterate much quicker, have researchers just build things and ship things to production and not think about the infrastructure. The other thing is we aggregate capacity from all over the world. Hundreds of regions, lots of different Neo clouds and by doing that we have a fully usage based model. So you'd have to think about managing reserving capacity. Everything is usage based. You can scale up to thousands of GPUs scale out of zero. Which like today when a lot of people are running sort of unpredictable workloads in production is something a super valuable sort of lack of predictability.

[Tanay (Whisper Flow CEO)]

Yep. As a very happy, happy mortal customer, I'd highly recommend. I think I've been using you guys from the very, very, very early days.

[Eric Bernhardson (Modal Labs CEO)]

Yeah, been a couple years at least.

[Host / Moderator]

Today Coach Ari is Here he is the co founder and CEO of Whisper Flow, which you hear me talk about all the time. Wispr Flow, it's a AI dictation, but it works across all your different platforms, whether you're using on your phone or your desktop, whichever operating system you prefer. But it very tightly, you know, takes your dictation and formats it correctly. Now this sounds like, well, isn't that Siri? But to now I'll explain to you why, and I can explain to you just as a customer, that, God, Siri just doesn't do what it's supposed to do, does it?

[Tanay (Whisper Flow CEO)]

Yeah. So Whisper Flow does something really simple. You speak and it writes for you. And the key thing that we really, really, really optimize our models for is to just not make mistakes. The core metric is 95% of whisper outputs are never edited by the user. It's perfect on the first shot and people send it as it is.

[Host / Moderator]

And I can tell you, you crossed over because my wife said to me just this week, tell me about Whisper Flow. How does that work? And does it read your messages? And I'm like, no, it doesn't read your messages. But there's a little bit of this people misreading your terms of service. I think last year that still maybe is coming up. What do you read around the message when I'm dictating to get context? And what don't you read?

[Tanay (Whisper Flow CEO)]

Number one, privacy is the core foundation of the entire company. Right? People are using Whisper Flow to dictate their most confidential personal and professional messages. And so one thing that we offer to all customers is pure zero data retention you get as a free user. You can enforce it across your company if you're an enterprise. And that means none of your dictations are stored on our servers either, much less use for model training. And so on that note, when we're doing a single dictation, what Whisper does is in half a second, it reads your entire screen, grabs the full context. It gets the name of, say, the people you're messaging, it gets the tone you've used in the past threads in an email or a chat, and it uses that in that moment as context to make sure that it gets names right. It sounds like you. And that content is not stored then on our servers, if you turn on privacy mode, how heavy of a load

[Host / Moderator]

is it to do all this in real time today? And I know using Eric's software, Eric's platform rather, so maybe between the two of you, we can get an idea of token usage going up and how heavy the load is getting for consumer grade AI products like Whisper Flow, it

[Tanay (Whisper Flow CEO)]

has been insanely hard to optimize it. But here's an interesting thing. So Whisper Flow, what you'll notice is blazingly fast. We spend less than 100 milliseconds on the GPU per request. Now what that means is with that our GPU costs are also significantly lower. So the business runs at a 90% gross margin, which is fantastic. And we can serve millions of users at scale for very, very low costs.

[Eric Bernhardson (Modal Labs CEO)]

And

[Tanay (Whisper Flow CEO)]

there is a lot of local compute we do as well. And so hey, there's things we could do locally and just send the minimal amount of intelligence, minimal amount of context to the cloud that needs it to be intelligent to produce the final output.

[Host / Moderator]

Also joining us again is Richard Socher. He is the CEO and co founder of Recursive Superintelligence which came out of stealth 05-13-2026 raised 650 million AI powered productivity engine search agent multimodals. How is the project going Richard?

[Eric Bernhardson (Modal Labs CEO)]

Good.

[Richard Socher (Recursive Superintelligence CEO)]

Yeah, that was a mix. So I run U.com, which builds the leading web search APIs for AI and then also Recursive which builds recursive self improving superintelligence to automate knowledge discovery. Essentially we want to lean into the fact that AI is code and now AI can code and you can allow an open ended fashion to get AI to self improve in a much, much faster way than any human could. Where you really automate the ideation, implementation and validation of ideas in the scientific method. First apply it to AI itself and eventually to everything else. The plan is to build the ultimate sort of Eureka machine in the sense of like Eureka. I made a scientific invention, not I found gold which I realized a lot of Californians think of. But yeah, that could be the last main invention. Humanity needs to then invent everything else after.

[Host / Moderator]

And who's the customer for this self improving engine? Or are you selling the solution into other people building AI? What is your customer base here?

[Richard Socher (Recursive Superintelligence CEO)]

Similar to a lot of other NEO labs? Like we're not building for a customer from day one. We have a research phase. We're actually building this out. We're making now so much progress that we're actually thinking about pulling the product forward. It's already and across this sort of AI stack has pushed the frontier forward on some levels. I can't share all the details yet, so we may productize earlier, but for now you can think of any company that has software engineers, any company that has clear reward signals that they can define for their company that make a huge difference and are associated with some piece of code. Become a customer of this technology and

[Host / Moderator]

talk to me a little bit about your cloud usage and how many tokens you're using and how you provision everything, because this is becoming. I think one of the big stories here is just there's so much demand for intelligence and tokens but maybe not enough infrastructure to supply them. So talk to me a little bit about when you're doing this sort of research phase, how do you think about token usage? Is that like more than your staff costs? Obviously more than your office cost or if you have an office. Yeah, talk to a little bit about penciling out the discovery of this new product.

[Richard Socher (Recursive Superintelligence CEO)]

Yeah, we do have two offices in San Francisco and London and indeed we're spending a lot more on compute than we are on people. I think more and more startups will do that and I think we're going to see token costs, compute costs, be higher than people costs for a lot of companies. At the frontier, especially if you want to push that frontier, it gets extremely expensive. You try to sit on the shoulders of giants. There are some things you can cleverly do around the harness, but if you really want to build full recursive, self improving superintelligence, which includes many parts of the stack that are not very cheap, you do have a lot of token usage. And indeed it feels like our small team is executing at a level that's like 1020 x the size. We're having research results that I would have expected maybe teams of hundreds of people to do, but we're teams of several dozens only. And the fact that everyone is going to a manager of this AI, there are almost no ICs anymore. So indeed we are seeing a lot of token usage and the productivity that employees have is almost 10 20x of what you would have expected from that size team.

[Host / Moderator]

Eric, what's the outlook right now in terms of new hardware getting online, new power coming online? Do you think it's going to be able to keep up with the demand? Or we just constantly going to be told, hey, run your jobs on Sunday, you know, from 1am to 6 say and we got a window for you. I mean it feels like we're going back to the times when people were renting mainframe time perot.

[Eric Bernhardson (Modal Labs CEO)]

We've been super supply constrained since I would say like end of last year. I mean we see this a lot because we're buying a lot of capacity in a lot of different clouds, renting a lot. And so the Last six months has been like, kind of brutal in terms of finding capacity. The capacity is out there, but prices are going up. I think we're going to be probably supply constrained for the next year or two. I mean, eventually these things tend to resolve itself, like supply, demand tend, you know, demand sort of, eventually supply catches up with demand. But, but I think right now, capitalism

[Richard Socher (Recursive Superintelligence CEO)]

is a beautiful thing when it works.

[Eric Bernhardson (Modal Labs CEO)]

Yeah, yeah. The invisible hand tends to take care of these things on the long run. But like right now, like, things are moving so fast, there's so much demand. I, I will say, like, I see a little bit of sort of, you know, Hoppers getting a little bit, maybe softer.

[Host / Moderator]

That market.

[Eric Bernhardson (Modal Labs CEO)]

It's been, you know, to me it's been crazy that the hopper markets, like, you know, they've gotten 40% more expensive in the six months, but, but now that's starting to soften a little bit. A lot of the Blackwells are starting to get delivered over the summer. I did see a little bit in the market. Like, I think when Anthropic did that deal with Colossus, it felt like it was a little bit of softening in the market because that was what, you know, at the top was driving a lot of the demand. So you see maybe a little bit, you know, sort of normalization, but it's still going to be tight for a while.

[Host / Moderator]

So Anthropic was buying so much compute on the market, it was driving up costs for everybody. Yeah, I mean, they were a major buyer.

[Eric Bernhardson (Modal Labs CEO)]

Exactly. I think Anthropic was the big whale, buying up so much capacity and not just Anthropic, but probably like a lot of it. That, that was a big part of why you kept seeing Hopper and Blackwell prices going up over the last six months.

[Tanay (Whisper Flow CEO)]

Yeah.

[Host / Moderator]

And tonight you're been, you've been optimizing, as you said, to not have to use as much infrastructure and keep the margins high on this. But are your customers now starting to demand you do a little bit more with the product and they want, you know, that's going to drive additional capacity needs. Or are you still optimizing and just saying, like, hey, eventually we can run this on. I don't know if you're running on open source or your own model now, but take me behind your thinking strategically about what the next couple years looks like as the CEO of a company, are you just going to believe that Eric can keep costs down by just putting more infrastructure on, or do you think we need to have an escape hatch here and maybe do some more stuff locally on the processor or, you know, optimizing models to use even less tokens.

[Tanay (Whisper Flow CEO)]

So there's two key things that optimize at the end of the day, actually. So the, the single thing that I care about optimizing for is the end customer experience. And that means I need to give them the best intelligence and the best latency, which means, and that that drives all of our decision making. We're going to add a lot of more features on top of current whisper right now. Whisper, you just speak and it writes for you. And over the next couple of weeks even you'll see some big launches come out where it's able to start to do things on your behalf. And with each of these, one of the core things that we focus on internally is speed of development. So developing with, let's take modal as an example. Like developing with modal is so much faster that whatever delta there would be if we, that we would pay modal versus kind of building this on top of our own AWS and having a couple of engineers manage that infra, I would eat that cost and happily pay that to modal because that saves us time and lets us iterate a lot faster. And that is really how I think about this. And when things get to scale is when we start then thinking about the gross margins. Because right now the cost of infra, until it is, until we're just running a negative gross margin business, that is not a big concern. And so I would say, you know, as a company, growth is number one to prioritize for us while building a fundamentally good business. So the core thing I want to make sure is unit economics are positive and if they are, great, we'll go as much as we can.

[Eric Bernhardson (Modal Labs CEO)]

If your gross margins are 90%, like in my opinion, I'm very biased here, but I think you should spend more money and move faster than the product. If gross margins are 90%, we are.

[Tanay (Whisper Flow CEO)]

There's going to be new product surface launching every month basically for the remainder of the year. And the business is growing 40% month over a month. And so how are you acquiring customers?

[Host / Moderator]

What's working today? A lot of people are wondering. The whole SEO playbook, SEM, is that thrown out? Social media now has changed from follower counts and building up a following to just gaming the algorithm. You know, everybody has an equal swing at bat. Doesn't matter if you have a thousand followers or a million. Everybody gets the same treatment under the algorithm. So how do you think about growth today? I'm curious.

[Tanay (Whisper Flow CEO)]

There's this, there's this fantastic guy Jason, who. Who raves about us on the all in podcast. That helps a lot.

[Host / Moderator]

Awesome. No, if you have. I mean, it is a legitimate. Curating super fans of products is a legitimate way to do it because yeah, I'm, you know, I just talk. Whatever the best product is I talk about. Right. And I bought a pedal. I told you, I don't have it here with me on the road. I'm traveling right now. But I have this like elgato. Not a sponsor, but three pedals and my God, that changes your usage completely. I have one that clicks on Zoom, one that clicks on my browser, and then one that is whisper flow. You put the pedal down, you start talking, and then you lift it up and that submits it to Whisper flow. And that. Now I'd say my prompts, when I'm using AI or I'm building something, they're three or four times longer because I just ramble. And I didn't realize that rambling is better than actually actually. Richard, cognitively thinking about what my prompt should be, it turns out that's not

[Richard Socher (Recursive Superintelligence CEO)]

actually good pedal to the metal.

[Host / Moderator]

It's a whole new meaning, literally is just say everything. Ramble and ramble and ramble. Correct yourself mid sentence. It doesn't matter. These things the more you give them. And it doesn't need to be formatted in the right order. This was like a big mind blowing thing for me as a journalist or an analyst. Previously I was like very much thinking about prompting in a very structured, like almost light programming kind of way. Okay, this is your Persona. This is my goal. This is how you should operate. I'm just like, now it's ramble, ramble, like, like I'm at a bar at 2am and it's works better, it seems.

[Tanay (Whisper Flow CEO)]

No, that's definitely. That's one of the biggest things that people are realizing. And so our super fans actually drive about half of our monthly growth. And then the other half is coming from a lot of channels that we build. And I think about this as basically a portfolio. Usually there's two or three channels that are the primary ones that we're fully optimizing for. And there's three or four on this side that we are experimenting with testing out. And it could look anything from radio to TV to trying out a new thing in a different country. And the moment something works, it goes from that experimental phase to our core phase, at which point you can basically describe it as a cash generating machine. And then we just define the boundaries around it. Like, this is what the payback period should be. This is the quality of customer you want to get from this. And until you're in those bounds, you can spend as much money as possible.

[Host / Moderator]

So are you letting the agents actually and the code actually spend that money or you have human in the loop, I assume?

[Tanay (Whisper Flow CEO)]

Yeah, we have. So the, the agents are making a lot of decisions on the kind of core part, which is like thinking about the copy, the creatives generating a lot of that and running the ad sets. And then we have two people on the team who are managing now almost about 70 to $100 million worth of annual marketing spend using all of these agents across these channels to then have a daily view and make a lot of the higher level strategic decisions.

[Host / Moderator]

Yeah, Eric, this makes me a bit nervous because I remember when Facebook and Google started introducing some machine learning tools like, hey, we'll suggest some copy here, but you got to approve it. Now people are just like, I'm setting my AI to go sell. If something works, try more experiments like it and just do the recursive thing. Richard, what if it starts promising you like, hey, Whisper Flow is just gonna help you be better at chess. And it's just like, yeah, I found a chess community. I targeted them and I told them, use Whisperflow. It's great for chess, but it's like asking it to optimize without understanding what the product is. It could go off the rails. Huh?

[Eric Bernhardson (Modal Labs CEO)]

Yeah, that's true. Maybe that's the future of AI. Just run away. Recursive super intelligence.

[Richard Socher (Recursive Superintelligence CEO)]

There's some companies that are working on this. Like one is called Omnikey. Full disclosure, I'm an investor too, like in Whisper Flow, but you can give guardrails. But indeed, reward hacking is a big problem in AI and something we have to think about a lot at Recursive too. Right? Because the AI will kind of just optimize some reward that it's given. And if you give it an incomplete reward. In fact, a lot of people talk about job loss. I think a new type of job is to become a reward engineer. It's kind of like a prompt engineer, but much more sophisticated with larger action space. Not just talking, but also doing things for you. And so if you give a bad reward, like, oh, make my CSAT score, my customer satisfaction score higher in my service center. The AI will just be like, easy. I'll just create a bot with millions of calls in your service center and give 5 out of 5 rating at the end. You're like, well, I meant to do it with real people. And then it's just like, well, easy. I just give you $1,000 gift certificate for every failed, like Uber or doordash delivery or something, right? You're like, no, no, it has to be real people and you can't spend too much. You have to define these rewards really well in marketing and everything else for you.

[Eric Bernhardson (Modal Labs CEO)]

I also sort of expect, I mean Google and AdWords, AdWords and Facebook, they spend a lot of time like optimizing the relevancy. So as someone who's run a lot of campaigns on those in my previous life, you get kind of penalized over time if you like run ads that are not relevant to your product. Even though if you have like very high click through rate, if the downstream conversion is garbage, like it doesn't like they'll kick you out.

[Host / Moderator]

This reinforcement learning and reward hacking, yeah, is surprisingly effective, as is threatening the machine and telling it it's going to get punished, but it's almost too powerful. This is a very weird concept, Richard. That telling the large language model it wins a reward makes it perform better. How on earth is that possible? Is it because it's giving it clarity as to what the goal is? And we're just inferring that it has a human behavioral sciences foundation. Is it all projection, Richard?

[Richard Socher (Recursive Superintelligence CEO)]

A lot of it is projection. So there are two things here. One is you can talk to the AI and prompt it differently. But then generally a lot of systems, open ended systems, evolutionary systems, reinforcement learning systems, they all have a reward and that reward is just maniacally being pursued. And in that pursuit of that reward by an AI, it doesn't have sort of the common sense reasoning necessarily that people have. Right. That's what I bet this example of customer satisfaction scores. When an AI runs a call center, you have to think about all these things that humans have as implicit knowledge, but the AI doesn't have that implicit cultural normative knowledge. And so in those cases you then have to fiddle with the prompts a little bit more. And then for a while there are different ways you can deal with prompts. Sometimes it makes sense to sort of threaten the AI or say like, oh, you're going to get a super great reward. And then you'd pick up aspects from the training data where something really important has been done and high priorities were given to certain bugs or something. And then the I realizes it needs to solve this problem at a higher level. Generally we're going to get better and better at giving the AI sort of a compute budget. We already see this. When you want to get an answer. For instance, in finance, like for you.com, we have a finance API and you can give it different amounts of budget and then the bigger the budget, the more thorough and accurate and everything. The answer is it basically has no more hallucinations and all these things because it triple checks, it does reverse rag and so on before it gives you the answer. So I think a lot of these are sort of early childhood errors as humans and AIs kind of coexist and work together on problems. And similar to the early speech systems where you had to be like you had to adapt yourself to the speech system and now risper it adapts back to you. And so I think we're going to see that interplay back and forth with co adaptation.

[Eric Bernhardson (Modal Labs CEO)]

I mean going back to reward hacking, I don't see that as a sign of hyper intelligence. If anything it's like a sign of stupidity.

[Richard Socher (Recursive Superintelligence CEO)]

Right.

[Eric Bernhardson (Modal Labs CEO)]

It's like that joke from Silicon Valley where you tell an AI to fix all the bugs and it just deletes the code solution, but it's a dumb solution. Right. So I think a lot of the work going forward is going to be to have to fix that by making AI smarter.

[Richard Socher (Recursive Superintelligence CEO)]

That's right. And that's also like a lot of times the silliest examples where people assume some crazy superintelligence but also crazy super stupidity are actually the world ending scenarios where somehow the AI is smart enough to make paperclips and in the process wipe out all of humanity, but doesn't realize that if no one wants to buy paperclips, you don't need to produce anymore.

[Eric Bernhardson (Modal Labs CEO)]

Right.

[Richard Socher (Recursive Superintelligence CEO)]

So there's like a weird super stupidity connected to the superintelligence in these scenarios.

[Tanay (Whisper Flow CEO)]

One of the things that I think about on this is the same way as humans. Like in a company you have there's different things you really want to optimize for. And so you get people who care about all of these different functions and then you have a system of checks and balances. Like the CMO wants to spend all the money on marketing and the CFO keeps that in check. And then engineering wants to ship a bunch of stuff, QA keeps that in check. And a lot of times the reason you split it into different functions is because every person at the end of the day has one top line thing that they're optimizing for. And if you give them multiple IT messes with their brain. And that is how I think at least like right now when we build these systems, you know, if sub Agents and you have one agent running the other one verifying, da da, da. And to me, if we're modeling a lot of AI around how humans work, I don't see why that wouldn't continue to be a really good solution.

[Richard Socher (Recursive Superintelligence CEO)]

I think interplay between diverse agents and diverse ideas has worked for humanity a lot and it is also working well for AI.

[Host / Moderator]

It is hard. I think it's today the thing you said that's really precise is it's really hard for people to keep in their minds two different goals, metrics, responsibility sets. And that's almost why every CEO eventually just throws their hands up and say, we're going to put one person in charge of this. If this is important, one person's going to be in charge of it and there's no number two for them. And that then begs the question is, well, is the AI need to be designed like that or why can't the AI be designed to hold two ideas in its head at once? We need to grow, but we, we can't go bankrupt. We, we need to ship product, but we can't ship it so fast that we can't, I mean we can do that. Users can't understand it. Right.

[Eric Bernhardson (Modal Labs CEO)]

It's like multi objective optimization. Or you can like say you have this objective and this objective, I'm just going to add them up and like optimize the sum of it. Or you know, there are like tricks you can do that people I think very successfully apply.

[Host / Moderator]

Yeah. Maybe the reason we do agents and we rebuild the real world is so that we can have comfort with how this is operating and we could feel like we understand it. Whereas just giving one agent all responsibilities is just more uncomfortable for us than maybe the AI.

[Tanay (Whisper Flow CEO)]

You know, a separate belief that I, that I have here is so this is now when you're talking about like, hey, what gets you an A result versus an A plus result? And if you, yes, you give somebody multiple objectives, they can do multiple objectives and they'll do like an A grade job at both of them. But if one person is just maniacally focused on optimizing this one thing, they're going to do a much better job. That's why we have different functions doing different things, even though one function can think about all these different objectives. And so again, going back to the CMO example, CMO let them roll and was like, I want you to grow the business as much as you possibly can. If you give them less bounds on that, they will start to be very creative and you'll come with Great ideas. And then you prune them down and that pruning step is separate. But if you start with that pruning step, you might not get the idea that feels out of the box. And so this is more so like at least how I think about Org design and why I like splitting people up and having a single goal is so they're, they can completely spike in that without any, any constraint.

[Eric Bernhardson (Modal Labs CEO)]

And so go to the CMO and you're like, you got to maximize the number of clicks or signups and they spend like a billion dollars. And then the CFO is like, what the fuck's going on? And then you pull it back and oh, your budget is this.

[Tanay (Whisper Flow CEO)]

So you keep them both before the decision is made. And you like, you don't actually, don't actually spend the billion, but it helps with that, that creativity process to like break out of the mold.

[Host / Moderator]

Yeah. And then eventually you get to where you were at today, where you have three channels that really work. And then you have like the experimental one and you're like, okay, the experimental group now needs to make their own decisions and the people with the known stuff have to optimize that. And then you have to split that and then it's just a split that and you just keep splitting the atom. And then all of a sudden you've got a massive headcount of people who are managing every nuance of your business, which is really the, the number one topic of the week. People have been, I mean, it's a raging discussion that I think we're going to have to keep having. Because Meta laid off 8,000 people about 10% of its staff. And then at the same time record profits, lifting their 2026 AI capex by up to 10 billion, which is a total of 145 billion. These are very large numbers in terms of cap spend. In a year, they're spending over $10 billion per month to build out infrastructure. Intuit cut 3,000 jobs. That's 17% of their workforce. And they explicitly said this is to fund AI integration. Goldman Sachs on the other side, CEO David Solomon said in an AI job loss. He said the AI job loss narrative is overblown. In an op ed in the New York Times, quote, the historical pattern is clear. The US economy can and will adapt to major advances in technology. But Goldman economists estimate AI may automate 25% of current work hours over the next decade isos. He cited a Stanford study showing entry level employment in AI exposed roles. Software engineering, customer support down 16%. So here we are this thing is raging back and forth, Richard, and clearly we had bloated big tech companies, everybody hired. That was the playbook. Take talent off the field. Hire based on what you need in two years, not what you need today was the common knowledge. So you're just always building infrastructure, office space. You're always hiring, you know, putting butts and seats in those office spaces. You're always building out infrastructure, upgrading your servers, upgrading your finances and keeping more capital going. That's been the entrepreneurial playbook. But something's changed here. The reward system is now. Move the dollars from employees to infrastructure. That seems to be what Wall Street's responding to. So what's your take on this, Richard? Is it overblown or is it going to be chaotic for the next couple of years? The employment and how often do you change your opinion on this, in all honesty, where you go from feeling like doom scenarios to feeling like, oh my God, this is incredible, anybody can do anything. We're all superheroes now.

[Richard Socher (Recursive Superintelligence CEO)]

I think there's a difference between sort of long term and short term. I think long term, I'm extremely optimistic. It's just technology has had a record of making people's lives better. And in the moment, it's hard for people to see that, feel that progress, right? But if you look over any 10, 20, 30 year time horizon, things that used to be luxury goods for select few are now common and have common access. Everyone has common access to it. When I see things like, oh, we automate 25% of current work hours, that doesn't mean there's going to be 25% unemployment. That is called the lump of labor fallacy. That labor is this fixed lump. And when you take 10% off now there are going to be 10% unemployed people. That has never happened in the past. It used to be 150 years ago that over 95% of people worked in agriculture. Now it's less than 5% of people working in agriculture. And yes, in the moment, when a tractor does the work that you used to do manually, it feels like, holy shit, this tractor sucks. This weaving machine sucks. But because we have weaving machines now, we don't have rags to riches stories anymore, because no matter how poor you are, you can afford real clothing without and not just wear rags, right? And so there's so many of these. So I'm often surprised of how

[Host / Moderator]

people

[Richard Socher (Recursive Superintelligence CEO)]

come up with these fallacies, like the lump of labor fallacy, and then stick to it. Now you have to also acknowledge that meta is called meta, after the metaverse. So they hired a ton of people to build the metaverse and then realized AI is much more relevant to the future than the metaverse, and it didn't quite work out and makes sense to make big bets. And then sometimes these bets don't work out. And so, yes, they have to transition their whole company. I do think we see a mix of just people increasing efficiencies, sometimes with AI and sometimes just like they let their lowest performers go. And sometimes it's a mix, mix, mix of the two. And so I do think every company, every industry will change. Every organization, every person working in organization, every country needs to think about how to adjust for this new reality. Maybe it can make sense to help have some unemployment benefits to smoothen out the transition.

[Host / Moderator]

And we have those changes today. We have all those for people as a safeguard, and we've extended them in the past where if there was something cataclysmic, they say, oh, you know, it's Covid, we're gonna give everybody six months or 12 months more unemployment just to smooth this out. Because we, we don't want people to crash and burn here in the United States. We want to give them a soft landing. But Matthew Prince did an article last week again. He said, hey, record profits, record growth, we're doing fantastic, but we're going to get rid of these measurers. Because the people who measure stuff, I don't know if you saw it today, but your thoughts on who gets cut and then where are you in terms of worrying about this as a major issue or just not being concerned at it at all? And how often do you change your position?

[Tanay (Whisper Flow CEO)]

I think my position on this has been relatively consistent for the last year, pretty much, which is a long, long time in the world of AI. The major thing to acknowledge is, I think, for example, Meta is taking their money from spending on humans and spending it in AI, is muddling a bunch of things together. Because the real thing is the pain that people are feeling is, hey, we spent 10 years of our life planning for this career, and then four years in college to get this computer science degree, thinking we're going to go and be a software engineer. And guess what? Reality slaps you in the face. And that career you've been planning for 15 years is deleted. And that is the pain that people are feeling. And that, of course, comes up in a lot of different ways. And it's hatred towards large companies, but it's generally a feeling of just deep loss. And that is not something we could. That is just the reality of the situation. And so one just understanding that helps build empathy on where people are really coming from. And it's like, yeah, this is a really, really shitty situation to be in. On the other side, I don't believe there is lesser jobs that need to be filled. I was looking at the stat yesterday. There's 200,000 just open roles. And at Whisper we have like, we are still doubling our headcount every six months. Like we're 60 people right now. By the end of the year, want to be 120, 150. That is so much work to be done. Even though AI is doing a lot. But with every product we're launching, every new market we're launching, all the things that we were talking about, like, hey, you need just a human to deeply, deeply care about this. You need the people. So the profile of the person we hire is completely changed. When I'm hiring for a software engineer, I care a lot more about what their prompt to Claude looks like than the actual code that they produce with AI because that tells me how they think. And if somebody is a good thinker and they can reason from first principles and they have good product taste, that is something you cannot take away with AI and you're going to leverage it to do much more with these tools. And so that's the kind of profile I look for across every single role.

[Richard Socher (Recursive Superintelligence CEO)]

And Tanay brings up an interesting thing which is like you can actually have narratives, both pro and con, early career people. Yes. Maybe you started this career four years ago ago in software engineering and it's changed now. But it's also clear that companies want young, hungry talent that knows how to use these tools, has grown up with these tools and are now bringing them into the workplace versus they have to get folks that are established and know and have done this exact job in this exact role for many years. And so you can have both narratives and I think they all cancel each other.

[Eric Bernhardson (Modal Labs CEO)]

But what I think is funny though is like I don't think actually anyone has lost a job because of AI, like directly. I don't think Zuck woke up and he said like, oh, these engineers over here are more productive. So I'm going to fire. I think what happened was it's the promise of AI means they want to invest a lot more money into AI. So because of that capex, they're basically trimming the rest of the company. And I think that's generally true across the entire economy is that very few people have actually lost jobs due to the reality of AI. A lot of people have lost jobs due to the promise of AI. So it's a little bit like self fulfilling in a sense. Right. Which I find like quite ironic with

[Host / Moderator]

entrepreneurs are capital allocators, you know, talent allocators. So they're looking at it saying, where do I allocate these resources in order to get the output?

[Eric Bernhardson (Modal Labs CEO)]

Exactly.

[Host / Moderator]

And I think that's what give tokens to a smaller number of people.

[Eric Bernhardson (Modal Labs CEO)]

We're reallocating a lot of capital to building data centers, to building AI. Tenay is hiring a lot of people to build AI. Met is choosing, oh, we're going to shut down all these organizations that are maybe not performing as well and move all of it to AI. But again, it's still a bet and I think it's going to pan out. I'm very hopeful, but I think very few so far have actually lost their jobs due to the reality of AI. It's all the promise of AI.

[Host / Moderator]

There are people I think at Marvel, they gutted the place from what I understand. And it's strictly because AI and then what?

[Eric Bernhardson (Modal Labs CEO)]

Yeah, I mean it's also a convenient excuse. Right? I don't disagree that there are some people that have actually lost their jobs due to the reality of AI, but I think it's way 10x more than I've lost due to the promise of AI.

[Richard Socher (Recursive Superintelligence CEO)]

I thought about this a little bit.

[Tanay (Whisper Flow CEO)]

Here's.

[Richard Socher (Recursive Superintelligence CEO)]

I think the algorithm which you can predict with AI is going to lead to job loss or not. And that is all based on the elasticity of demand given how much cheaper you can make a product with AI. So if, for instance, you can make illustrations go from several hundreds of dollars for one illustration, where only a few newspapers and fancy corporate blog posts can afford an illustration, to make an illustration worth like a cent now, will the demand for illustrations go up by billions? No, there are only so many blog posts and newspaper articles and so on that need an illustration. So it'll go up a little bit, but it's not super elastic software. On the other hand, when you make it much, much cheaper to build software, I think the demand will just increase. We could all have a custom app, a custom suite of apps. You can have billions of different types of software and so there the sort of Jevons paradox will also kick in more where making it cheaper will actually make you end up using a lot more of that good and hence the demand for people creating that good or service will go up also.

[Host / Moderator]

That's where I've started to lie. We're seeing many more startups, we're Seeing people who are laid off start companies. So the question I ask myself is, do we still have enough problem surface area for humans to be productive? So is there more software incrementally for people to build that would make people's lives better, make things cheaper, better, faster, more entertaining? And the answer is yes now?

[Eric Bernhardson (Modal Labs CEO)]

Absolutely. I think the optimistic take is there was, you know, 10,000 people building a metaverse that very few people used. Maybe hopefully they can go and start startups that actually contribute, you know, to the society in more meaningful way.

[Host / Moderator]

You made a product people didn't want. Congratulations. Now you have to shut it down. I mean, it could be as simple as that. But they're also studying everybody's behavior with key capture and some desktop software they created in house to do training. And this has people shook. So Tenay, what do you think of this concept of. At the same time Zuckerberg is announcing the 8,000 layoffs, he's saying, and we're recording your computers to do reinforcement learning for our models at a pretty granular level. But don't worry, it's anonymized, but we're recording it and we know when people leak information to the press, so maybe it's not so anonymized. What did you think of those two announcements together?

[Tanay (Whisper Flow CEO)]

Again, I look at them all very differently and I have a couple of friends who work at Meta.

[Host / Moderator]

What's the back channel?

[Tanay (Whisper Flow CEO)]

The back channel is people are not happy. People are not happy. There is like some version of this that companies do, which is you have your, your MDM system that they can use to kind of lock your computer. And so many of the companies, especially during COVID started installing time tracking and like software to see what applications are using if you're actually working or not. But usually that happens in low trust environments. Like, hey, you're working with a bunch of contractors, you want to make sure the contractors are doing work. Makes sense.

[Host / Moderator]

Or a regulated industry.

[Tanay (Whisper Flow CEO)]

Or regulated industry. This is one of the times when you see it here. And so the sense generally is like these folks felt that they were in a higher trust environment and now once you start putting guardrails like this, it, these are slightly sour taste in their mouth. And I can totally understand why Meta is doing this. Right? They need to collect the data from somewhere and they've gotten a lot of backlash from collecting it from their users over the years. And so that was this. But the, the one thing I would have changed, that is at least change the messaging somehow that it'd be a bit more opt in because I think company culture matters a lot and I wouldn't place that above the cost of data.

[Host / Moderator]

The company culture seems to be broken over there or to some extent, or maybe it's just so cataclysmic right now when you're saying goodbye to people every 18 months. It's just hard for anybody to enjoy that. It's like being in this extended, never ending war. Right, Richard, you've been there.

[Richard Socher (Recursive Superintelligence CEO)]

There's a survivor guilt for people in these kinds of situations often and a lot of their friends go and so on. I think what's interesting here is that if you are an entrepreneur and you just care about the outputs of your organization, then you love AI. If you get paid by the hour and everything you do is owned by the company and you don't own any equity and significant shares of that company, then you may actually hate AI because yeah, when you do something, the AI will record it and then Meta will clearly say, well, we can now do what this person has been doing, so they're going to be fired. Right. And so I think that could be a push to many more solopreneurs entrepreneurs, because then they can love AI and AI is the future versus being feeling like the AI is just learning what they're doing and then it's going to do it for them. I think that it might push more startups to be formed.

[Host / Moderator]

Eric, this is the power of equity and everybody rowing in the same direction. If you worked at Meta and they were like, yeah, you have $10 million in equity, $5 million in equity, a million dollars in equity, and they're like, you lost your job, but the company's massively profitable and the Stock's going up 7% a year, it's going up 12% a year now it's like, okay, well I'm making. If I have $2 million in RSUs, you're telling me it's going up 12% a year, I'm making a quarter million dollars not working here, but you gave me enough equity. It's almost like what they're doing with these Trump accounts, they're trying to get every American into the stock market. So if there is job loss and companies are becoming more efficient, at least you can root for them from the sidelines and be like, yeah, I own some shares in that. It takes the edge off.

[Eric Bernhardson (Modal Labs CEO)]

Yeah, yeah, I think that's right. Yeah. And of course it's sad when people lose jobs, but you know, hopefully the economy gets better over time and people find other jobs. And I'm very Optimistic about the future.

[Host / Moderator]

Here's an interesting chart that we threw up on the screen just a minute ago. We'll throw it back up here. This is software developer jobs. These had come crashing down to since mid 25. As you see the software developer jobs are that light blue line that just does a complete V and then you have the overall job postings. And these are on indeed and Bloomberg, I guess and indeed did this together. Okay, so here is a job board posting showing a trend. Who knows if this is correct or not, but 125 million job postings down to hundred million overall. But software jobs coming plummeting down to 62 million and then back up to 74 million. So theoretically since this time last year, we've seen a lot 10 million more job postings on this one particular job board. I don't know if this is representative, but what we're seeing here is,

[Tanay (Whisper Flow CEO)]

I

[Host / Moderator]

think what we're seeing overall is people are picking the number. They like to support their narrative. If you're in the Trump administration or you're a Republican and you're want to build the and you're building software and tools, you want to believe jobs are coming back. And if you're a doomer, you represent the unions or maybe you're a Democrat who hates. Yeah, the right. You think that this is complete, utter job destruction. My P doom is pretty good right now. I think anybody coming out of college with who says they're an AI first employee and they know how to use Perplexity computer or Grok computer or Cloud cowork, anybody with, I don't know, 60 days of experience using those tools, I think gets a job instantly.

[Tanay (Whisper Flow CEO)]

I would hire somebody like that to work side by side with me 20 hours a day, every day. I'm actually looking for somebody like that right now. It's like, please come help me because I have so much issue. That position is so you're young. So like Ralph is like say like 22 to 25. You are willing to basically like be beside me or sleeping in the office. It's like, okay.

[Host / Moderator]

Crazy energy to be the Something.

[Eric Bernhardson (Modal Labs CEO)]

Yes.

[Tanay (Whisper Flow CEO)]

Gonna be chief of staff. I want them to be high emotional maturity. Right. Because they're going to be managing my, my inbox and doing honestly a lot of things on my behalf. They're going to be my second brain essentially. And there's a good amount of automation that I do. Like at this point, I don't open my inbox anymore because I built a whole inbox tool for myself and a lot of These tools I built for myself, I build inside the product. And I just want somebody high bandwidth to do that. Because for me, I had my younger brother, he's seven years younger to me, he's 21 right now. I just want another one of him to just be maniacally focused on this. And he's starting his own company, he raised money for it, so he's doing his thing. So I can't get him anymore. But that's what I'm a backup brain.

[Host / Moderator]

Yeah, like you want an agent. I look at this as like almost like guys used to have like a butler who like helped them dress like I'm, you know, and helped their make sure their cars and their, you know, shotguns were all tweaked perfectly. Like somebody who is like your, your body man standing next to you doing your driving and your security and everything. There's this need for somebody to do that, but to make software for you, like to be a software butler who just makes you bespoke software to solve all of your problems. But yeah, anybody coming out of school, Eric, with any kind of familiarity with AI is going to get a job.

[Eric Bernhardson (Modal Labs CEO)]

Yeah, I'm not so sure I actually would take the other side of it. Yeah, please. I mean like, I think what you saw like during the ZURP era with like, you know, bootcamp grad could just, you know, spend like a couple of months like learning basic coding and then get a job. I do think it's gotten harder. I do think, you know, in order to get it like an entry level job today, you do need to spend a little bit more time like learning the sort of ladders of abstraction and writing a lot more code. And I have a feeling that part of the market is a little bit harder. I'm still bullish on software engineers. Like, I do think there's going to be an enormous amount of job creation, but I do think the bar in my opinion probably has gone up a little bit.

[Host / Moderator]

I love the idea of a software valet. Here's another chart. Just try to make sense of this, folks. This is all workers and their unemployment rate, which is at a pretty close to our lifetime low. This is a little bit earlier in the year 4.2. I think it's 4.4 right now overall. And then recent college graduates age 22 to 27, they're 5.6% unemployment. And then all college graduates 3.1%. So it's hard to understand this data except that recent college graduates almost never trailed all workers. They were more hireable. So there's something about college age people recently out of the workforce that are, you know, maybe their unemployment's up 20% or something. But it doesn't seem to be cataclysmic. Like we might see if something like self driving works. Suddenly we, we have self driving work.

[Richard Socher (Recursive Superintelligence CEO)]

I think as usual, averages kind of are hard to interpret because you could have multimodal and bimodal distributions. My hunch is the young graduates that are bringing AI into organizations are highly sought after. And then there are some graduates who, they studied history and the history used to be fine. If you want to go to McKinsey and just become a general consultant or something, you could have, they're okay with that, be okay with that. But now they don't know AI either so and they're like ramping them up. It will take them years before they get to an AI and so I think it'll likely be bimodal. And the average now is slightly increasing because there are more people studying social sciences and so on than like computer science and other vintages.

[Host / Moderator]

If we look at these like wine vintages, you could have a season where like there's a group of people out there who are not AI first. And since they're the recent graduates are you could have recent graduates and you know, outpacing these non recent graduates, you know, the people who maybe graduated five years ago pre chat GPT and they weren't using chat GPT to cheat in school and Claude code to, you know, to do all their assignments, which was super weird. I don't know if you guys saw all the booing at the commencement addresses for AI Feels like there's just generations of people who, it just gives them the ick. I think as the kids say. The Pope also says AI needs to be disarmed a bit. Quote, artificial intelligence needs to be disarmed. Freed from the logic that turned it into an instrument of domination, exclusion and death. It must be at the service of all and the common good. That's Pope Leo. He warned AI threatens to normalize an anti human vision and said the concentration of digital power in the hands of a few private actors must be counted. Here's a video of that moment. Artificial intelligence needs to be disarmed. The word is strong, I know, but deliberately chosen because this moment needs words capable of attracting attention, awakening consciences and indicating paths forward for humanity. The Church has long been working for nuclear disarmament, aware that every great technical power can affect people's lives and so must be accompanied by adequate moral discernment and public Control nuclear disarmament remains a service to peace and the dignity of the human family. In a similar sense, artificial intelligence now demands to be disarmed, freed from logics that turn it into an instrument of domination, exclusion and death. ERIC clearly, you know, the services you're providing and this infrastructure is, you know, in the service of domination and death.

[Eric Bernhardson (Modal Labs CEO)]

I really think.

[Host / Moderator]

How are you planning to burn in purgatory for all time? Eric?

[Eric Bernhardson (Modal Labs CEO)]

Yeah, exactly. I mean, I do think that there's something where tech doesn't fully aware, isn't fully aware of the PR battle here. Like there is something happening where like there's. The general public is very nervous and I don't think it's, you know, I don't think, you know, big figures are doing a great service when they're going out and talking about massive job losses and probability of doom and all these things. Like I look at all our customers and the amazing value creation that's happening. Like we have customers who are working on new drugs, like drug discovery that could cure amazing diseases we haven't been able to cure before. Like we have other customers that build amazing consumer products. And so I feel like there's this very pessimistic scenario leading to this backlash to data centers and all these things. The problem with tech is they have this sort of slightly autistic truth telling

[Host / Moderator]

desire and they are very dramatic and

[Tanay (Whisper Flow CEO)]

self motivated, very dramatic.

[Eric Bernhardson (Modal Labs CEO)]

I feel like the tech industry almost needs to hire a couple of PR people and then engineers just need to chill out a little bit and try to be a little bit more optimistic or else there's going to be some massive backlash against this technology.

[Host / Moderator]

Richard, you're nodding here. How does it feel to be building tools to dominate, exclude and create death and destruction amongst the populace of humanity? According to the Pope, Death, exclusion and domination. We have to overturn this AI.

[Richard Socher (Recursive Superintelligence CEO)]

There's a lot to, there's a lot to unpack here.

[Host / Moderator]

I mean, pretty dramat words from the Pope and he says, hey, I'm being dramatic for a reason. Do you think he believes what he think he does?

[Richard Socher (Recursive Superintelligence CEO)]

I, I think so. I think the future needs better marketing. I think that much is clear. I agree with Eric. I think LA is, you know, because they have a lot of illustrators and, and people and special effects and you can now create lots of special effects in virtual worlds very quickly. There's a lot of anti AI sentiment there and they're trading most stories when you watch the studio, right, and others they're like literally say, yeah, fuck the AI. And so On a lot of storytelling of dystopian futures and not a lot of positive storytelling where you can have drama, but in an otherwise optimistic and positive future. I think that's true. I think it's also true that AI is only as good as the systems, the people and the data that influence it. It is a general purpose tool. And just like the Internet, you can focus on all the horrible torture porn that's on the Internet, I'm guessing it's there, you read about it sometimes and we should make that illegal. And that is clearly regulated correctly. So. And you can focus on the Internet just having horrible things and bullying and so on on it, or you can focus on the Internet bringing amazing communication to everyone and connecting the world and bringing knowledge and making it super cheap to learn whatever you want to learn nowadays. And both can be true at the same time. Right? This is omni use technology. And maybe a last point is I do think it's dangerous when CEOs say this is the most dangerous technology ever. Also we are the only ones who can do it really, really safely and you should trust us and give us billions of dollars to make it as good as possible. And then at the end of that, people will just remember the it's super dangerous. And by the way, I don't trusted random tech CEO to make it better. And so why are they even building it in the first place? And then it's just like a negative loop. And I am worried about some of those patterns that we're now seeing. Attacks on, you know, people's houses here in San Francisco, like Sam Altman's house and so on. I think as CEOs amp up this pdoom conversation, pretend like this is this like super hardcore technology that's bad now. At the same time, I think it makes sense to not focus on the negative applications. And some of those regulations as they like really influence people's lives, do need strong regulation. We do need regulation on how AI is used in warfare. It's not something I want to work on. I don't think super intelligence should be used in warfare because that could literally lead to the kind of Skynet like scenarios. But you do need to regulate AI when it gets actually impactful for people's lives. And then in other areas, I think it's best not to overregulate too soon before you really know all the implications. There are a lot of stories where you can say, oh, AI has been really bad for mental health for these five cases. You make those very public. But there are no stories where many people actually improve their mental health by talking through their problems with an AI, those are sort of silently happening as well. So there's a lot of gray in between the black and white.

[Host / Moderator]

Yeah. And some companies maybe aren't doing themselves any favors or maybe if they want to get more attention.

[Eric Bernhardson (Modal Labs CEO)]

I think it's worked really well for marketing for, for those companies.

[Host / Moderator]

It's worked really well for Anthropic with Mythos and like the valuation of the company and like, hey, the delusion of grandeur is like a known psychological phenomenon where people want to be important. They want to believe that they created the technology that then solves every problem in humanity. Here's Christopher Ola talking about the real possibility that human labor will go away.

[Tanay (Whisper Flow CEO)]

His Holiness call for discernment is profoundly timely.

[Richard Socher (Recursive Superintelligence CEO)]

I wish to name three questions where

[Tanay (Whisper Flow CEO)]

I think the Church's voice is especially needed. The first is our duty to the global war. There is a real possibility that AI will displace human labor at a very large scale. If that happens, supporting those displaced will

[Richard Socher (Recursive Superintelligence CEO)]

be a moral imperative of historic proportions.

[Host / Moderator]

Okay, so today you're hiring like crazy and the product is helping people communicate faster, better, more accurately. What do you think about the Pope's comments and anthropics, you know, pronouncements that like Mythos is going to too powerful for anybody to have. It's going to hack everything. It's all omni omnient and can see through all hacks and problems. And you know, this sort of, these are these delusions of grandeur or these, you know, thoughtful, you know, missives that we should consider as a society.

[Tanay (Whisper Flow CEO)]

There's a couple of people I know who think about these things deeply and they sound basically how Dario, how Dario talks about these topics. And what I realized is once you really understand the meat of what they're actually talking about, it makes a lot more sense. But there's essentially like two kinds of people. One who's very philosophical, very intellectual, like a big brain person. Like that's how I just mentally think about it. And then there's people who are just more tactical, more in the present. And a lot of times you hear that and you hear like, oh, there's this going to be this, this cognitive displacement of labor. It doesn't resonate and it sounds like a far fetched fear, but there's actually a lot of truth to it. And so what I found at least is being able to translate that into a way that is much more like, resonates a lot more with people on what it actually tactically means today. That way they can actually understand it. And that is one thing that is like my, my personal wish for, for these folks. Because if you read some of what Dario, if we, if we go two years into the future and we read all the things that Dario and Anthropic say today, it will make so much sense. But there's that communication gap where the way they, they phrase it out isn't how the average person consumes, consumes it. And so those get missed. And this has been historically what has happened with philosophers in the world. I know because my co founder is like this, right? He says these very profound things to me that make sense to me six months later. And then I have to work with him to figure out how can you tell me these things in ways that enter my brain. So I understand it now instead of it making sense six months later. And that's when I started to realize this trend emerge and got me appreciation for oh, this is what you're trying to say there. But this is definitely not landing with anybody the way you wanted it to. And it feels more like fear mongering when it's something completely different.

[Host / Moderator]

I think it's a well said and I think maybe for our last topic here, talk a little bit about the explosion in Chinese models and their token usage. There's a data source called OpenRouter and they are claiming that Chinese AI models hit 9 trillion tokens the week of May 18th, up 20% basically week over week versus US models up 16% week over week at 5 trillion tokens. And the Chinese models have led weekly usage for four consecutive weeks. Deep seq version 4 flash now the number one model globally at 3.4 trillion tokens. Top US models Claude Opus Sonnet and then Gemini 3 flash preview all at or at 1.9, 1.9 and 1.1 trillion respectively. Quinn 3.7 Max also dropped this week. Eric, are you seeing American companies starting to embrace these models yet or.

[Eric Bernhardson (Modal Labs CEO)]

Yeah, at a massive scale. I mean the Chinese models are very good and I wish, wish there were better open source models coming out of the U.S. but the reality is there isn't.

[Host / Moderator]

Why is that? Why are we seeing two different approaches, all of us, but for the audience?

[Eric Bernhardson (Modal Labs CEO)]

Some people claim, and I think there's some legitimacy to their argument that the Chinese models, they're all distilled on US proprietary models and or subsidized by the Chinese government in order to build better models. I think there is some logic to that argument and I think it's worth Discussing those things. I also think to some extent that the Chinese teams are very good. If you read the Deep Seq paper, it's just sentence after sentence of just profound research that is very, very deep. And so I don't know. I wish we had better models in the US that are open source, but today we don't.

[Host / Moderator]

You're using these models currently or working them, building off them. What can you tell us a little bit about the Chinese model you think about them?

[Tanay (Whisper Flow CEO)]

We're experimenting with some of them and they are really good. They are really good when it comes to what's their open source. And I mean, same as Eric, I think the way capitalism is set up in the US none of these companies are incentivized to build open source models. And then this is what we have to eventually lean onto. But the one issue that we saw happening with them is they have very interesting and strange biases that come into these models. And so we can't deploy them as is. So that's again, because we end up being the layer between what you say and what eventually gets written out. And the last thing you want is any kind of censoring or misinterpretation of what you said to go out. So we decided not to deploy these models directly as the core thing that's running, but using them a lot for some other internal tools.

[Eric Bernhardson (Modal Labs CEO)]

Presumably you can fine tune them though, and still get okay results. Have you looked into that?

[Tanay (Whisper Flow CEO)]

You could. There's all these little edge cases that comes up. One was really funny, like Taiwan.

[Eric Bernhardson (Modal Labs CEO)]

Is Taiwan a country?

[Host / Moderator]

Tiananmen Square? How do I get there?

[Tanay (Whisper Flow CEO)]

Yeah, see those things? Those things, we just don't get into that. But one was really funny. This didn't happen with the Chinese model, happened with another model. But it had this very little nuance where if you said, hey, ladies with a Latino accent, it would convert it and write, hey, Mahamas. And this is such a rare edge case. I do not know how that came into the model. We found one little thing that came in from there, but really with the breadth of usage, it's very hard to test all of this. And so you need to have like massive, massive eval sets to actually get it to work.

[Host / Moderator]

This could be actually a feature set. You could just be able to set whisper flow. Like, hey, I would like everything I say to go through like a Gen Z filter and just drop in, you know, Riz and whatever. I want mine to go through the urban dictionary. I want mine to go through like

[Tanay (Whisper Flow CEO)]

the 70s last last week. People are really Enjoying that one.

[Host / Moderator]

Awesome. All right, listen, another amazing episode of this week in AI. Thank you, Richard, Tanay and Eric. You guys are all hiring in the AI job apocalypse. So take from that what you will, dear audience, but maybe you could each give a little plug for who you're trying to hire and how people can find out more.

[Eric Bernhardson (Modal Labs CEO)]

We're hiring primarily systems engineers and we're based in New York. So if you're on the west coast, west coast or on the east coast, we're also hiring sf, but primarily in New York, hiring a lot of people who love building complex systems, low level performance optimization, stuff like that.

[Host / Moderator]

Right. And Eric is giving everybody who comes and joins a pair of tickets to the Knicks finals. So if you refer a sith admin, he'll buy you a pair of tickets to the finals. I'm joking. I just made that up.

[Tanay (Whisper Flow CEO)]

Richard.

[Richard Socher (Recursive Superintelligence CEO)]

So at Recursive, we're hiring also infrastructure folks, folks who can wrangle tens of thousands of GPUs. And that's mostly it. Some very, very strong researchers that are able to build very large frontier models. We have a few roles too. And [e-mail], we're looking for a lot of senior backend engineers, folks that want to wrangle very large search indices to actually make these models up to date, accurate and have citations.

[Tanay (Whisper Flow CEO)]

And today we're hiring a lot of ex founders, product leads, GMs for all the new business units that we're building. Hiring researchers, engineers across the stack, salespeople and marketing and growth folks. And the key things that we really care about is if you're somebody who is a first principles thinker, you're low ego and you're hungry to win, then you're going to fit right in.

[Host / Moderator]

And we are hiring associates in training. So we like to train up associates out of school to learn the craft of being a great investor. You can go to x.com launch and you can DM launch to get more details. If you follow our x.com launch, previously known as Twitter. We'll see you all next time. Bye bye.