Subscribe now

Speaking on Lenny’s Podcast, Sottiaux shares his vision for the future of AI, where autonomous agents will dominate internet interactions and manual configurations will become obsolete.

He also details OpenAI’s significant investment in AI safety and alignment, emphasizing the importance of guardrails and security for these advanced models. These Big Ideas are drawn from Tibo Sottiaux’s interview on Lenny’s Podcast.

• AI agents will dominate internet interactions. Tibo Sottiaux, OpenAI’s Head of ChatGPT, predicts that AI agents will perform most internet actions. This means product developers must design for agent interaction and ensure their systems can handle the increased scale. ▶ 0:16

• AI will evolve from manual configurations to autonomous learning systems. Sottiaux states that the current trend of manually configuring AI ‘loops’ is a temporary phase. The future lies in systems like ‘dots,’ which are incredibly smart agents that operate continuously, learn from feedback, and understand user goals and preferences without explicit manual setup. ▶ 0:46

• AI safety and alignment are paramount, requiring significant investment. OpenAI prioritizes safety and alignment, investing heavily in security, guardrails, and monitoring for its models. Tibo Sottiaux notes that the majority of their investment in the API stack goes into the safety stack, rather than just raw capability development. ▶ 34:29

• User-centric skills are increasingly vital in the AI era. Sottiaux identifies ‘great taste,’ user empathy, and connecting with the target audience as increasingly important skills. He suggests that product managers, whose role often involves these qualities, are well-positioned to thrive in this environment. ▶ 23:50

• High autonomy and trust foster rapid innovation. OpenAI fosters a culture where employees, many of whom are ex-founders, are given significant autonomy and trusted to make sound decisions. This ‘bottoms-up energy’ allows for rapid development and deployment of new features, such as the decisions API.

• Ai’s promise is to enhance human focus and creativity. Tibo Sottiaux believes AI’s primary benefit is to reduce distractions and allow individuals to direct their attention more effectively. He suggests this could lead to fewer meetings, less busywork, and ultimately, more creative and productive outcomes, similar to the clarity gained from disconnecting during a holiday. ▶ 19:18

The conversation between Stripe Cofounder John, Mark Zuck and Alexander Wang (of scale.ai, now chief AI officer at Meta) explores how Meta’s AI development prioritizes the model as the product, multimodal capabilities for personal agents, and a focus on ‘talent density’ over team size.

This approach reflects Mark Zuckerberg’s leadership, characterized by boldness and a Bayesian method of updating his worldview, which has been crucial for Meta’s long-term AI investments and its ambition to become a full-stack technology company.

1. The model is the product

In AI product development, the underlying model itself has become the primary product. This means that much of the complex product work, traditionally focused on user interface and experience, now revolves around the model’s capabilities, its ‘personality,’ and its security and privacy features.

Subscribe now

For Muse, significant effort went into designing the privacy and security framework for handling user data and connected services.

2. Outsourcing product thinking to external benchmarks

External benchmarks for AI models quickly become saturated, making them less useful for guiding advanced development. A significant problem with over-reliance on these benchmarks is that it effectively outsources critical product thinking to their creators, who are often PhD students with different objectives than a product company.

This leads to a strategic shift towards internal, use-case-specific benchmarks for more effective model development, ensuring alignment with specific product goals.

3. Multimodal AI for personal agents

Meta’s strategy for AI development centers on training models end-to-end to function as highly effective personal agents. This approach emphasizes multimodal capabilities, recognizing that human interaction with the world is inherently multimodal.

This contrasts with an earlier industry consensus that prioritized coding as the paramount skill for AI, demonstrating a shift towards consumer-centric, natural interaction.

4. Talent density over team size

Meta’s AI development strategy, particularly for projects like Muse Wonderer, prioritizes ‘talent density’ over sheer team size. This means focusing on recruiting the best possible individuals and providing them with substantial compute resources.

Share

For the Muse Wonderer project, this involved a deliberate ‘reboot’ of the effort, pulling top talent from within Meta and aggressively recruiting external experts to create a high-talent-density environment.

But going back to the beginning, I think one of the things that was very interesting and challenging in standing the thing up was like, you really just wanted this like very high talent density effort.

5. Long-term AI investment: entrepreneurship on hard mode

Developing frontier AI models is akin to ‘entrepreneurship on hard mode,’ demanding extensive long-term planning and massive upfront capital investments in compute infrastructure. This pre-training phase is high-stakes and scientifically challenging, often leading to periods of uncertainty and ‘troughs of sadness’ for researchers due to the immense investment and unknown outcomes.

6. Meta’s full-stack technology company strategy

Meta aims to be a full-stack technology company, building its own data centers, chips, and low-level infrastructure to control the entire technology stack. This strategy is driven by the belief that controlling the core infrastructure and models is essential for delivering the best user experiences and avoiding reliance on external platforms.

Meta’s existing profitable core business provides the financial stability and credit to underwrite these massive, long-term AI investments. Mark Zuckerberg’s understanding of ‘power laws’ further justifies these high-risk investments, as even if a specific AI initiative fails, the underlying compute infrastructure retains significant value and can be utilized or sold.

7. The rise of agentic commerce

AI agents are poised to become a standard business component, much like websites and social media accounts are today. This will lead to a significant shift in commerce, termed ‘agentic commerce.’

Muse exemplifies this by enabling users to fully delegate purchases, reducing friction and automatically applying user criteria. This marks a ‘step change’ in shopping, offering significantly greater personalization through a holistic understanding of the user’s life, unlike siloed historical recommendation systems.

Furthermore, AI agents can expand the total addressable market for commerce by creating new demand and converting non-users into active customers, as demonstrated by new Instacart subscribers driven by Muse.

So anyone who wants to do that, we eventually want to get there. and then I think the fullness of time is that, you know, just like today every business has, a website and social media accounts and email addresses and all these things, every business is going to have agents too.

8. Mark Zuckerberg’s bayesian leadership and boldness

Mark Zuckerberg’s leadership is marked by a unique blend of boldness and a Bayesian approach to updating his worldview. Once he internalizes new information, he rapidly and decisively pivots resources and actions to align with his updated beliefs, even if traditional metrics don’t yet validate the change.

This was crucial for the rapid establishment and success of MSL, as he fully committed to the AI initiative, cleared his calendar, and empowered key talent like Alex, demonstrating a ‘very high learning rate’ and a willingness to make unconventional bets.

This reminds me of Siebel story - the largest CRM business of its time, totally unbeatable and lost to a startup (called Salesforce).

What happened? Instead of solidifying their position, Siebel created chaos. They launched a new product to compete with Salesforce’ cloud CRM.

As companies like ServiceNow look to roll out their own AI-native solutions (Flow), this history (might) repeat itself. Two teams fighting over the same client just leads to chaos: a lesson worth heeding.

Instead of launching internal competitors, incumbents might consider a hard pivot or even acquisition. Just look at what Adobe did and what Salesforce is doing now.

Tl;dr: Don’t let internal rivalry cloud your strategy. Embrace clarity instead of chaos.

What's your take on ServiceNow's launch of Flow?

The Pers…

Read more

One headline, why believe it?
Everybody wants to rule the world
All for freedom and for pleasure
Nothing ever lasts forever
Everybody wants to rule the world [Tears for fears]

Something quite strange happened last week. Anthropic’s Dario Amodei argued that AI companies need to “pace the frontier”, essentially slow down how quickly they make their most powerful models more capable. Sam Altman agreed with him.

Think about that for a second.

OpenAI and Anthropic are spending billions of dollars trying to beat each other. They are fighting for the same users, developers, talent and, ultimately, the chance to build AGI first.

Subscribe now

And now they agree that perhaps everyone should slow down.

This makes a lot more sense when you look at the AI race as a classic prisoner’s dilemma.

Assume OpenAI believes pushing AI too fast has real risks. Anthropic believes the same. Both would probably prefer a world where everyone spends more time testing models instead of shipping the next capability jump as fast as possible. This gives both enough time to prep up for enterprise adoption, which has been abysmally slow.

But neither can afford to slow down alone.

If OpenAI slows down and Anthropic keeps going, Anthropic gets ahead. If Anthropic slows down and OpenAI keeps going, OpenAI gets ahead.

Share

So both keep racing.

The result is quite sad. Both companies can end up doing something neither actually wants to do.

Which is why Dario’s proposal is interesting.

Except there is a third prisoner.

Open source.

And this prisoner isn’t sitting at the table.

Over the last two years, Chinese labs have become a serious force in open AI. DeepSeek changed the discussion around the cost of building capable models. Alibaba’s Qwen family has become one of the most widely used open model families. Moonshot, Zhipu and others continue to push models that developers can download, change and run themselves.

I would not claim that these models have simply “beaten” OpenAI or Anthropic. That is not true across the board. But they have definitely changed the economics of the race.

OpenAI and Anthropic are no longer competing only with each other. They are competing with models that can get close enough on many tasks, cost far less to use, and in some cases can run without sending anything to OpenAI or Anthropic.

Share Ashish Sinha / NextBigWhat

And that creates a much harder prisoner’s dilemma.

Sam and Dario can agree to slow down together.

They cannot make DeepSeek slow down.

They cannot make Qwen slow down.

This is where the AI safety argument starts getting very interesting.

Suppose the industry agrees that powerful AI models need outside tests before release. They need ongoing checks. Their makers need to know who can access them.

And if something goes badly wrong, the maker should have some way to restrict access.

An open-weight model cannot offer the same level of control after someone downloads it (who is going to take the blame in an org? the CTO? In that case, why would he/she even push open source models?)

Nobody has to say “ban open source.”

You can simply create safety rules that open source finds much harder to follow. And that creates a strange alignment of interests.

OpenAI and Anthropic can genuinely believe that advanced AI needs stronger safety rules. At the same time, those rules can strengthen the business model of OpenAI and Anthropic.

Both things can be true.

This does not mean Sam Altman and Dario Amodei are secretly trying to kill open source. There is no evidence for that. In fact, OpenAI itself has released open-weight models, and Dario has said he does not support banning them.

The more useful question is not what they intend. It is what happens if their view of responsible AI becomes the rule.

Because the definition of “safe AI” could slowly become a definition that favours models which remain under the control of their makers.

And then we have an even bigger problem.

Imagine OpenAI and Anthropic really do slow down. They spend more time testing frontier models. They delay releases when the risks look too high. They follow every rule they helped create.

Meanwhile, Chinese open models keep improving.

At some point, one of the American labs looks at the capability gap and asks the obvious question:

Why are we slowing down when they aren’t? And the prisoner’s dilemma starts all over again.

Except now it is not OpenAI versus Anthropic. It is closed frontier AI versus anyone who refuses to join the agreement. Which may be the real problem with trying to pace AI.

It works only if enough of the people who can push the frontier agree on what “pace” means. Sam and Dario can make peace with each other. They cannot make peace on behalf of everyone else.

And in a world where everybody wants to rule the world, that may be the part that matters most.

What’s your take?

That has turned drugs such as Ozempic, Wegovy, and Zepbound into a potential demand driver for lingerie retailers.

Subscribe now

For apparel retailers, GLP-1 drugs are not only a story about consumers eating less. They are also creating a “wardrobe reset,” as customers whose bodies and sizes have changed buy clothes that fit again. Bernstein estimates that this could add $3 billion to $13 billion in annual apparel spending.

Why lingerie is different

A loose sweatshirt can survive a few pounds of change. A bra usually cannot.

Fit is central to lingerie, and a change in band or cup size makes an existing drawer of bras less useful. Industry data already show that larger band sizes and D-cup bras are losing share, while sales are shifting toward smaller and mid-range sizes, including band size 40 and B and C cups.

Share

Victoria’s Secret CEO Hillary Super said the company had seen roughly a 3% downward shift in the bra-band and underwear sizes being sold, a trend she linked to GLP-1 use.

The store opportunity

Weight loss does more than create demand for replacement products. It brings customers back into stores to find their new fit.

More than two-thirds of surveyed GLP-1 users said sizing changes made them more likely to shop in person, according to ReturnPro data cited by Reuters.

Victoria’s Secret has added AI tools to certain fitting rooms so shoppers can request other sizes, positioning the retailer for a customer who may be trying to work out what fits for the first time in years.

And one more thing..

Under CEO Hillary Super, the brand has:

• Leaned back into its “sexy” core identity after a period of brand confusion

• Reduced heavy discounting and promotions, improving margins.

• Focused on key categories like bras and the PINK line for younger shoppers, which are now growth engines

• Brought back high‑visibility events like the runway show, reigniting cultural buzz

AI-moderated interviews could reduce this gap. AI can conduct interviews faster, and consumers seem comfortable being interviewed by AI.

Subscribe now

A survey of roughly 300 respondents across the US and UK found that 73% of UK respondents and 76% of US respondents would regularly participate in AI-moderated research. Only around 10% said they were unlikely to do so.

For consumer insights teams, this removes a major concern. The problem is not whether people will speak to an AI interviewer. Most will.

But will they be honest?

Around half of the respondents said sharing their opinions with an AI interviewer felt as authentic and reliable as speaking with a human researcher.

This does not mean AI is as good as an experienced moderator. However, the absence of a human interviewer may not affect response quality as much as we assume.

In fact, AI could work better for certain topics. People often edit themselves when another person is listening, especially when discussing money, health, relationships, or behavior they feel embarrassed about. An AI interviewer that doesn’t judge them, lose patience, or react visibly may get answers that a human interviewer doesn’t.

The biggest finding, though, is about voice.

58% of UK respondents and 62% of US respondents said they could provide more detailed answers through audio or video than through writing.

This is important because consumers don’t think in neat, survey-friendly sentences. They pause, contradict themselves, remember something mid-answer, and use language the research team may never have considered.

A text box kills much of this detail.

Voice allows people to explain themselves naturally. The AI interviewer can then ask why, follow up on an interesting comment, or ask for a specific example. This brings some depth of a human interview into research that can be conducted with hundreds or thousands of consumers.

That is the real opportunity. Not another faster survey tool, but the ability to ask follow-up questions at scale.

There is one clear warning: don’t force people to turn on their cameras.

• Audio, video, and text should be choices, not requirements.

• Respondents need to know how their recordings will be stored and used.

For consumer insights leaders, AI-moderated interviews are interesting not because they make existing research cheaper, but because they enable a different kind of research.

Instead of commissioning a new project every time the business has a question, teams can speak to consumers regularly, follow their answers in real time, and build a deeper understanding of why they buy, switch, or leave (via OPXP.AI).

Consumers appear ready for this. The question is whether research teams will use AI merely to automate surveys or rethink how frequently and deeply they can listen to customers.

“Build it and they will come” has been the virality mantra for startups, especially PLG ones. But consumers have their own lives. Your product is not nearly as important to them as it is to you. Like it or not, you still need to find a way into their lives.

A great example is the recent surprise hit Hanuman Ansh, based on the story of Neem Karoli Baba, who was also a major influence on Steve Jobs.

Subscribe now

The movie turned out to be a super flop in the first few days and then, the makers started embedding themselves in the consumer’s life..and the rest is history blockbuster!

The Hanuman Ansh movie, created with a budget of around ₹1.5 - 2 crore ($170K-$225K), made only ~₹10 lakh ($11K) in its first few days.

It was nearly written off. Then something changed, and it started to grow. Here is what the makers did: apart from the inherent virality, they took specific actions and there is immense GTM lesson for companies (btw, the movie has made ~134 crores, $15.2 million, as of today).

1) they turned community rituals into their GTM engine

Instead of buying ads, the team organized eight bhandaras (free community meals) at temples and made these the core of their campaign.

• Each bhandara served as both a spiritual act and a conversation hub about Neem Karoli Baba and the film.

• For a devotional audience, this wasn’t marketing; it was seva + satsang, the way they already gather, talk, and form opinions.

Product analogy:
For human-centric products, your “bhandara” is the ritual where your users already meet: user groups, meetups, webinars, WhatsApp communities, internal champion programs.

The question is: What value can you create in that ritual that naturally pulls people in and makes them talk?

Share

2) they aligned with the right trust nodes, not generic influencers

The makers sought guidance from Premanand Maharaj and engaged with the existing spiritual ecosystem around Neem Karoli Baba devotees.

• In this category, maharaj/satsang leaders > film stars as trust anchors.

• Their endorsements (explicit or implicit) gave the film legitimacy that no trailer or poster could.

Product analogy:
In B2B or community products, your “maharaj” might be respected practitioners, category experts, internal champions, or power users or analysts (in VC firms).

Answer: Who does your audience already trust when deciding what’s worth their time? Build for them first; the rest follows.

Sounds cheesy? Ignore and live with the hardship of sales :)

3) they designed the experience to feel like darshan, not just entertainment

Audiences began treating theaters like temples: removing shoes, folding hands, crying, and calling it a “satsang.”

• The team didn’t fight this; they kept the tone sacred, avoided gimmicky promotion, and let the experience reinforce the spiritual “why.”

• This made people feel proud to bring family and community, not just to watch a movie.

Product analogy:
If your product touches identity, faith, health, career, or money, the experience must match the emotional weight of the job. A “fun” onboarding for something serious can break trust.

Ask: Does our UX match the emotional gravity of the user’s problem?

Share

4) they were willing to be “slow” and read weak signals

The film opened on ~250 screens, dropped to ~25, and was considered dead by conventional metrics.

• But the team listened to community signals (full houses in certain towns, temple buzz, social posts) and let those dictate where to push.

• Exhibitors responded by adding shows and screens once they saw repeat attendance and strong hold, not just Day-1 numbers.

Product analogy:
Most startups kill products that don’t spike in Week 1. Human-centric products often need community discovery time.

The skill is: Can you spot early advocacy loops (repeat usage, unsolicited referrals, community chatter) and double down there, even if top-line looks weak?

Share Ashish Sinha / NextBigWhat

5) they made the audience the PR department

Director Vishal Chaturvedi said: “We hardly did any promotion and PR. The audience played the role of PR with word-of-mouth and posting on social media.”

• No big press tours, no paid influencer blitz.

• Instead: real stories, real emotions, real rituals that people wanted to share as part of their identity.

Product analogy:
For PLG and community products, your best PR is:

• Users sharing wins (“This saved my quarter”)

• Teams showcasing workflows built on your tool

• Communities treating your product as part of their story

The founders’ job is to engineer moments worth sharing and then remove friction from that sharing.

Leave a comment

The pattern for human-centric, virality products

If you translate Hanuman Ansh into a playbook for startups:

1. Start from the ritual, not the channel
   Where does your audience already gather and talk? Build value inside that ritual (bhandara = community event/value moment).

2. Win the trust nodes, not the masses
   Identify the 5, 10 people/roles whose endorsement changes minds in your category. Design specifically for them.

3. Match experience to emotional job
   If the job is sacred/serious/identity-laden, your UX, tone, and rituals must reflect that. Don’t dilute it for broader appeal.

4. Optimize for loops, not launches
   Measure: repeat usage, referrals, community chatter, champion activity. Be ready to go slow until loops prove themselves.

5. Make users your PR
   Give them stories, artifacts, and status they can share. Then get out of the way.

“Build it and they will come” works only when what you build fits a real human ritual, is trusted by the right nodes, and creates experiences people are proud to recommend.

Hanuman Ansh didn’t beat the system; it understood the humans behind it.

What about you? What part of your customer do you understand so damn well which your competition can’t even touch?

Subscribe now

The startup’s major push in 2025 was that..they are replacing ‘expensive’ humans with AI.

The 2026 pivot? No AI, 100% real

For $1,000, customers get six finished user-generated content ads made by real creators - no AI faces or voices - with script reviews, unlimited revisions, and full refunds if unhappy. The service uses AI only behind the scenes while emphasizing human authenticity, which observers note often drives better trust and conversions than synthetic clips - as per the founder’s announcement.

That is, the product has changed from ‘unlimited’ AI ads with no-expensive-humans to..6 human-filmed ads for $999 (aren’t there better agencies than this?)

Share

Why the pivot?

The reason for this pivot?

Humans are still good at selling to other humans.

AI UGC: $1 per video. 1 in 87 converts. $87 per winning asset.
Human UGC: $15 per video. 1 in 4 converts. $60 per winning asset.

The cheap option is the expensive one (via)

Is this the beginning of a new beginning? Where AI takes the backseat, i.e. becomes a co-pilot, rather than taking the driver seat?

What’s your take? What are you seeing in your AI-led ad / marketing / content campaigns?

In May 2026, a video featuring Indian Prime Minister Narendra Modi, Italian Prime Minister Giorgia Meloni, and Melody toffees gained attention on social media.

Retail investors then bought shares of Parle Industries, mistakenly associating the listed company with Melody. However, Melody is made by Parle Products, a separate and privately held company. Parle Industries had no ties to the chocolate.

The sequence was irrational but recognizable as human behavior:

• PM Modi gifts Melody to Italian PM.

• Parle makes Melody.

• A listed company has “Parle” in its name.

• Buy the stock.

Subscribe now

Synthetic consumer research is becoming one of the most funded ideas in AI. Companies like Simile (raised $200mn) are creating digital twins of people that aim to predict how consumers will respond to products, prices, features, and marketing messages.

The appeal is clear: Instead of taking weeks to recruit participants, conduct interviews, and analyze responses, a company can generate hundreds or thousands of synthetic reactions in a matter of hours.

This makes synthetic research faster, cheaper, and significantly easier to scale without needing to talk to a single customer.

But speed does not equal evidence.

The main issue with synthetic consumers is not that their answers are always wrong. It is that businesses often cannot tell when those answers are incorrect. This distinction is crucial when the research is used to make costly, irreversible decisions.

Synthetic answers lack real human evidence.

Most synthetic research systems combine large language models with demographic information, behavioral datasets, social media content, transaction data, or prior research. The model then produces responses that mimic what someone from a specific segment might say.

However, the answer still originates from a model.

There is no real customer behind the statement. There is no respondent whose situation can be examined. No interview recording exists to revisit, no behavior to observe, and no person for the researcher to question further.

A synthetic respondent might say:

“I would pay ₹2,499 for a premium protein supplement because I value clean ingredients.”

That sounds helpful. But what does it prove?

It does not show whether a real buyer will pay ₹2,499 when a competing product is available for ₹1,799. It does not reveal if the buyer will abandon the cart after seeing the delivery fee. It does not take into account advice from a gym trainer, distrust of an unfamiliar brand, concerns about taste, or a spouse questioning the monthly expense.

The model has created a plausible explanation, but it has not observed a purchase.

This makes synthetic data hard to use as evidence for high-stakes decisions. A company cannot confidently tell its board, product team, or investors that customers demanded a feature when no actual customers contributed to that finding.

Plausibility!= prediction.

Large language models are good at generating answers that a reasonable person might give.

Share

Unfortunately, humans are not always reasonable.

People contradict themselves. They forget why they bought something. They claim to care about sustainability but then choose the product that arrives tomorrow.

They say they want fewer notifications but continue opening apps designed for notifications.

They demand privacy yet trade personal information for a small discount.

Synthetic respondents are known to show less variation than real consumers and can exaggerate the relationship between demographics and attitudes. In other words, they can make consumer segments seem more consistent internally and more different from each other than real people actually are.

This is a structural problem, not just an accuracy problem.

A synthetic consumer is built from patterns. A real consumer is shaped by constraints, relationships, habits, contradictions, and moments of irrationality. These elements often drive the purchase.

The melody incident serves as a cautionary tale.

Similar errors have occurred elsewhere. After Elon Musk tweeted “Use Signal,” investors inflated shares of Signal Advance, an unrelated medical device company. Investors also repeatedly confused Zoom Video Communications with the unrelated Zoom Technologies :)

These are extreme cases, but they highlight the core problem.

Human decisions are influenced by availability bias, mistaken connections, social proof, the fear of missing out, and whatever stands out at that moment.

A model trained to create coherent behavior may consistently overlook incoherent behavior.

Consider a protein brand deciding on its next flavor.

Suppose a protein company must decide whether to launch mango, chocolate hazelnut, or unflavored whey.

Share Ashish Sinha / ProductGeeks

A synthetic panel can analyze category trends, reviews, demographic preferences, and social media discussions. It might conclude that mango will attract young Indian consumers because it is familiar, culturally relevant, and different from existing chocolate products.

That is a reasonable guess.

But the actual purchase may rely on factors the model cannot grasp:

• Does mango whey taste artificial when mixed with water?

• Does its smell become unpleasant after resting in a shaker for an hour?

• Do consumers think of mango as a refreshing drink instead of a heavy protein product?

• Will gym trainers recommend it?

• Does the bright packaging make it appear less serious than competing products?

• Will customers enjoy the first serving but tire of the flavor after ten days?

These are not just data points. They are experiences.

A synthetic consumer can describe what consuming mango protein might feel like. It cannot actually taste it repeatedly, grow bored with it, regret buying a one-kilogram pack, or leave the half-used container at the back of a kitchen shelf.

This distinction matters because the company is not deciding which idea sounds best. It is deciding which product to manufacture, stock, distribute, and promote.

Where else can synthetic research lead to mistakes?

Consider pricing.

A synthetic respondent may weigh price, ingredients, and brand reputation rationally. A real buyer may choose the most expensive product because a fitness influencer recommends it or the cheapest one because payday is still a week away.

Consider packaging.

An AI persona can evaluate the visual design presented on a screen. It cannot discover that the lid is hard to open, the scoop gets buried in the powder, or the container doesn’t fit in a kitchen cabinet.

Consider customer churn.

A model may suggest that customers cancel due to high prices. Interviews might reveal that customers actually felt embarrassed asking the support team the same question repeatedly or that a spouse objected to another subscription showing up on the credit card statement.

Consider a new feature.

Synthetic users may consistently prefer more control and customization. Real users may never figure out the settings, may feel overwhelmed, or may stick with the default because changing their habits takes effort.

Consider advertising.

A synthetic audience may understand the intended message. Real consumers may misinterpret one line, turn a screenshot into a meme, or link the campaign with a controversy that did not exist when the research data was gathered.

Synthetic research is weakest where businesses most need it: where context, behavior, and consequences matter.

The accuracy debate is a distraction.

Synthetic research companies often discuss whether their predictions are 80%, 90%, or 95% accurate.

But an average accuracy number gives decision-makers little insight.

Accurate at predicting what? Under what conditions? For which group of people? Compared to which human sample?

Is the data reliable? Trust-able?

Does the system predict average survey responses, individual choices, market share, repeat purchases, or real behavior under financial pressure?

A system might mirror broad consumer sentiment with 90% accuracy yet still fail at predicting the minority behaviors crucial for a particular product’s success. It could correctly identify chocolate as the safest flavor but miss the small but valuable group willing to pay significantly more for an unflavored, clean-label item.

Even a highly accurate model can pose risks when users do not understand where the remaining errors lie.

The issue is not whether synthetic data can resemble human responses. It clearly can. The issue is whether that resemblance stays reliable when the market shifts, the product is new, or the decision relies on an unusual human reaction.

These are often the very situations where companies seek research.

Synthetic data is useful - but only to a point.

Synthetic research should not be completely dismissed.

It can assist teams in forming hypotheses, exploring potential segments, stress-testing questionnaires, identifying obvious objections, and narrowing down a large array of concepts before engaging with customers. It can also guide researchers in deciding which questions need deeper exploration.

In these cases, synthetic data serves as a brainstorming tool rather than as proof.

The problem arises when generated responses are presented as customer evidence.

There is a big difference between saying:

“The simulation suggests that price may be a concern.”

and saying:

“Our customers told us that price is the primary barrier.”

Only the second statement requires actual customers.

A reasonable research process can use synthetic consumers early on to explore possibilities. However, before making decisions about product launches, pricing, positioning, inventory, or major investments, those possibilities must be tested with real people and, whenever feasible, actual behavior.

The question is not whether synthetic consumers are impressive. The question is whether a company should produce ten thousand units, change its pricing, or enter a new market.

For low-risk exploration, synthetic data may be sufficient (why can’t you do that with ChatGPT itself and waste money on a new tool?).

For decisions where being wrong is expensive, plausible answers are not enough. Businesses need evidence that can be traced to real people, real circumstances and real decisions.

Synthetic consumers can tell you what might happen.

(Human) Customer research tells you what people are actually experiencing.

And behavioural data tells you what they ultimately did.

What’s your take?

(Reproduced from NextBigWhat)

AI model providers have a structural incentive to migrate as much intelligence from enterprises into their model weights—the encoded intelligence of these models—as possible. Doing so allows them to lease your insights to competitors, maximize token usage, rent-seek on your high-margin workflows, or in the most extreme case: completely replace you.

As the financial pressure to grow revenue mounts, model providers have already begun directly competing with their enterprise partners.

If your incentives were aligned with model providers, why do they charge per token and not as a proportion of the value they help you create?

Every few months, the AI industry finds a new obsession. Last year it was prompt engineering. Then came agents. Then reasoning models. Today, much of the conversation still revolves around who has the smartest frontier model: OpenAI, Anthropic, Google, Meta, or whoever releases the next impressive benchmark.

Subscribe now

Palantir’s report, Institutional Sovereignty in the Age of AI*, deliberately moves away from that debate. Instead of asking which model is better, it asks a more strategic question: how can an organization ensure that AI compounds its own competitive advantage instead of somebody else’s?

The central argument is simple.

The enterprise conversation is focused on the wrong layer

Most enterprise AI discussions begin with model selection. Should we use GPT? Should we standardize on Claude? Should we deploy Gemini internally? Should we self-host Llama or Mistral? These are valid engineering questions, but they are not the most important strategic questions.

The Palantir report* argues that organizations should focus on what remains uniquely theirs: their decisions, workflows, operating knowledge, permissions, context and institutional memory. Those assets compound over years. Model leadership may change every few quarters.

Here are my key (15) takeaways from the report.

• Stop asking “Which AI model should we use?” Start asking “What part of our business should we own?”

• Institutional Sovereignty is the report’s central idea: AI should compound your organization’s advantage, not your vendor’s.

• The future belongs to organizations that own their context, not necessarily the smartest model.

• Models will become commodities. The architecture around them will become the competitive advantage.

  Share

• Model liquidity should be a board-level design principle. Never get locked into a single AI provider.

Read more

Ironically, I stopped using it last week.

[The show, Rumble with Sinha is co-hosted in partnership with Rumble, a platform that enables one to do a podcast with their AI co-host - after all, solo talks are boring!]

Not because I found a better product, but because I ended up building my own Mac client. It does almost everything I need, and that got me thinking about something much bigger than my own setup.

I don’t think Wispr Flow has a product problem. I think it has a business model problem.

The underlying technology has become a commodity. Whether you use Whisper.cpp, Parakeet, Faster-Whisper or any of the newer open-source models, the gap between a polished commercial product and a free alternative has narrowed dramatically.

That’s not a technology problem. That’s a pricing problem.

I suspect there are still plenty of people happily paying for Wispr Flow simply because they don’t know these alternatives exist. But information arbitrage doesn’t last forever. Developers are already discovering them, building their own clients, and sharing them. Over the next year or two, I expect this to become mainstream.

If that happens, I don’t think Wispr Flow should fight open source. It should lean into it.

In fact, I’d seriously consider making the consumer product free.

That sounds counterintuitive because AI products cost money to run. But I don’t think consumers are the business anyway. I think enterprises are.

The Slack analogy comes to mind. Slack didn’t start by selling to CIOs. Individuals and small teams adopted it first because it was useful. Once enough employees were using it, companies had to buy it. Wispr Flow could follow exactly the same playbook. Let individuals get addicted to voice-first computing and let them bring that habit into their workplace. That’s where security, compliance, centralized administration and enterprise workflows become valuable enough for companies to pay.

There’s another asset that Wispr Flow has which people don’t talk about enough: voice data.

Before everyone gets upset, I’m not suggesting the company secretly sells user conversations. That would destroy trust overnight.

But what if users had a choice?

Pay for complete privacy, or use the product for free and explicitly opt in to contributing anonymized speech data for model training.

That’s a much more honest trade-off.

The reason I think this is particularly interesting is because of India. Wispr Flow has publicly said that India is one of its largest markets. At the same time, every major speech AI company wants more high-quality multilingual voice data. Today, companies are literally paying people to record speech in studios because good datasets are hard to find.

Wispr Flow already sits on top of exactly that kind of usage. Every day, thousands of people dictate naturally instead of reading scripted sentences into a microphone. With explicit consent, that becomes a valuable asset.

Maybe data licensing never becomes the biggest revenue stream. Maybe it simply offsets the cost of serving free users. Either way, I think it’s a more interesting direction than trying to convince consumers to keep paying a monthly subscription for dictation.

The enterprise biz doesn’t change. In fact, it probably becomes stronger.

No company wants its sales calls, customer emails or internal strategy documents entering public training datasets. They’ll continue paying for privacy, compliance and contractual guarantees. That’s where the money is.

Consumers and enterprises don’t need the same pricing model.

What's you're take?

Everyone is talking about Kunal Shah becoming CEO of WhatsApp.

I think they’re looking at the wrong story.

The bigger story is that Meta suddenly needs WhatsApp to work.

Subscribe now

For years, WhatsApp has been one of the most successful products in the world. But it has also been one of Meta’s biggest unfinished businesses.

Meta paid $19 billion for WhatsApp more than a decade ago. Yet WhatsApp still contributes a tiny fraction of Meta’s revenue (less than 2%).

Now look at what’s happening elsewhere.

Meta is spending aggressively on AI - all of its AI launches have failed so far (remember Llama?).

Employee morale is reportedly under pressure..

As Boz, CTO of Meta recently said in an internal memo:

morale is “maybe not the worst it’s ever been in 20 years here, but it’s probably up there” — comparing it to the Cambridge Analytica scandal.

The company is reorganizing teams, moving talent around, has been recording employees clicks and keystrokes data to train its AI model and throwing enormous resources ($145Bn so far) at finding its next AI growth engine.

Which makes WhatsApp more important than ever.

And nowhere is that challenge more visible than India.

• India is WhatsApp’s biggest market by users (26% audience)

• India is one of the biggest WhatsApp Business markets (15 million active accounts).

But the monetization numbers tell a very different story.

Share

India accounts for a massive share of WhatsApp usage while contributing a disproportionately small share of revenue.

That is the paradox Meta has failed to solve.

For years, people predicted that WhatsApp Payments would dominate Indian fintech.

It didn’t happen. Not even close.

Not because the market wasn’t large enough. Not because Indians weren’t willing to adopt digital payments. Other fintech companies proved both assumptions wrong.

The real issue was execution and focus. I hope Kunal can bring it - rather, transform WA from a messaging platform to Fintech (lending becomes a feature?).

If Meta wants WhatsApp to become a meaningful business, it has to crack India fintech.

And if it can crack India, it may finally discover the monetization playbook for the rest of the world.

The question is no longer whether WhatsApp can scale.

It already has.

The question is whether Meta can finally figure out what business WhatsApp is actually in.

What’s your take?

But if you try to pin down what taste actually is, it gets slippery. It’s not just a fancy word for “expensive,” and it’s definitely not just about choosing a pretty color palette.

Subscribe now

So, what is the anatomy of taste? And why is it so important in the world of AI, where taste is considered as the (only) moat a founder can have, given that building is commoditised.

Taste is an opinionated filter

At its core, taste is a deeply honed, human-centric opinion on what quality looks, feels, and sounds like. Think of the world as a chaotic waterfall of raw data, features, and possibilities.

Taste is the filter. It’s the ability to look at a hundred different options and say: “This one matters; the other ninety-nine are just noise.” While data and algorithms tell you what is popular (the average of what people already do), taste predicts what will delight people—even if they haven’t realized it yet. It balances empathy for the human experience with the bravery to reject unnecessary complexity.

And what is it a function of? I believe, a deeper understanding of customers, their lives, the different personas, their JTBDs and eventually a love for observing things and world around you.

Taste in the Wild: 3 Unique Examples

To understand taste, look at the moments where someone chose human intuition over raw logic or industry standards.

1. iPod

When Steve Jobs walked onto a stage in 2001 and introduced the original iPod, he didn’t just launch a piece of hardware. He gave the world a masterclass in what product taste actually means.

To understand why, you have to look at how every other tech company at the time was pitching their MP3 players.

Technical Spec vs. Human Story

In 2001, the standard industry practice was to market gadgets by their technical data sheet. If a competitor had launched that exact same device, the billboard would have read:

“5GB Storage Capacity, USB 1.1 Connectivity, and 185-gram Form Factor.”

That is an engineering fact. It is accurate, it is optimized, and it is completely devoid of taste. Why? Because it forces the human brain to do the heavy lifting. The average person doesn’t know what a gigabyte feels like. They don’t know how many audio files fit into five thousand megabytes.

Apple’s taste manifested in a single, legendary phrase:

“1,000 songs in your pocket.” (Note: This evolved to 10,000 songs with later high-capacity iPods).

Why this is pure taste

Taste is the art of translation. It’s taking raw, cold capability and reshaping it into an emotional benefit that a human being can instantly visualize.

• The Competitor’s Approach (No Taste): Focuses on the machine. “Look at what our engineers built.”

• Apple’s Approach (Taste): Focuses on the lifestyle. “Look at how your relationship with music is about to change forever.”

By framing the device around the volume of your music library rather than the size of the hard drive, Apple bypassed the analytical brain and went straight for the gut. They understood that you didn’t want to buy a storage unit; you wanted to carry your entire identity around with you.

The Modern Lesson for AI founders.

Whether you are writing a landing page for a software product, pitching a startup, or creating content, the “10,000 songs” rule is your ultimate guardrail.

Share

Stop telling people how the engine works (e.g. AI agents for <X>). Taste means having the restraint and empathy to tell them exactly where the car is going to take them.

2. In Gaming: Nintendo’s The Legend of Zelda: Breath of the Wild

When this game was being developed, the trend in the gaming industry was to fill open-world maps with thousands of checklist icons, mini-maps, and blinking arrows telling the player exactly where to go. It was highly optimized, but it felt like a chore list.

Nintendo exercised incredible artistic taste by hiding the UI. They designed the physical landscape (mountains, smoke plumes, ruins) to naturally catch the player’s eye and spark curiosity. They trusted human nature over digital hand-holding, transforming a digital space into a genuine adventure.

3. “Working Backwards” (Amazon’s PRFAQ framework)

Taste isn’t just for hardware; it applies to how you build software and businesses. Take the famous Amazon framework of writing the product press release before writing a single line of code.

It takes taste to halt engineering and say, “If we can’t explain why a human being would care about this in a simple page, we shouldn’t build it.”

Taste is the Ultimate Guardrail

Whether you are writing an essay, designing an app, or cooking a meal, remember this: Data can optimize, but only taste can compose. The next time you are building something, don’t just ask, “Is this technically correct?” or “Does the data support this?”

Ask the harder, tastier question:

“Does this have a soul?”

This becomes all the more critical in today’s world where one can build anything using AI - but all that matters is the ‘taste’ to know when to stop.

They want less work.

They want momentum.

They want to get something done.

That is the mistake a lot of founders are still making.

They are mistaking a breakthrough in capability for a breakthrough in demand. Just because a model can generate cinematic video from text does not mean millions of consumers have a recurring desire for “cinematic video generation.” That is not a product insight. It is a technical milestone.

People do not want AI nearly as much as they want progress.

This sounds obvious in hindsight, but it is a lesson the market has to keep relearning every time a new technology wave arrives.

Subscribe now

A model can now generate cinematic video, realistic scenes, stylized content, characters, or entire visual worlds from text. That is technically extraordinary. But the existence of a powerful capability does not automatically imply a large, durable user need around that capability. More often than not, it just creates a new class of demos.

And this is exactly why Jobs To Be Done (JTBD) still matters so much. It remains one of the few frameworks that consistently forces builders to ask the right question:

What job is the user actually hiring this product to do?

And if you miss that distinction, you can end up building something magical that nobody actually needs often enough.

Leave a comment

The market rewards momentum, not novelty

Most consumers are not looking for tools.
They are looking for movement.

They do not wake up wanting “AI video generation.” They wake up wanting to make an ad faster, create a visual without hiring someone, send something funny to a friend, publish content without overthinking it, or avoid work that feels tedious, confusing, or creatively intimidating.

Those are jobs.

Share

“Generate a realistic 12-second clip from a prompt” is not a job. It is a technical feature that may or may not attach itself to one.

That difference is where a huge number of consumer products go wrong, especially in moments like this when the underlying technology is moving so quickly that it is easy to mistake possibility for inevitability.

Sora created video

Founders naturally over-index on what is newly possible. Users do not. Users care about whether a product helps them make progress in a way that is meaningfully better than whatever they are doing today. Faster. Easier. Cheaper. Less cognitively expensive. More confidence-inducing. More socially useful.

That is what creates retention.

Without that, you might still get excitement. You might get screenshots, social sharing, a burst of curiosity, maybe even a strong launch. But you do not get habit. And in consumer, habit is what matters.

This is why so many AI products feel exciting in week one and irrelevant by week six. They are built around what the model can do, not around what the user is repeatedly trying to get done.

That is not a model problem.
It is a product problem.

Share ProductGeeks/ Ashish Sinha

JTBD remains the cleanest way to see what is real

A lot of product frameworks become less useful the more often they are repeated. JTBD has survived because it still explains actual user behavior better than most alternatives.

Users are not “using your product.”
They are hiring it.

And they are hiring it to make progress in one or more dimensions:

• functional

• emotional

• social

That has always been true.

Share

People did not hire Instagram to “share photos.” They hired it to project identity, taste, status, and belonging. People did not hire TikTok because they wanted “short-form video.” They hired it because it is one of the most efficient boredom-killing and attention-capturing products ever created. And people are not hiring ChatGPT because they want “AI.” They are hiring it because it removes friction from writing, thinking, planning, researching, summarizing, deciding, and getting started.

That is the actual wedge in consumer AI: not intelligence, but effort compression.

This is the point many AI-first products still miss. They are designed around what the model can output, rather than around what the user is trying to avoid.

But the best consumer products rarely ask users to admire the technology (TikTok is the OG when it comes to using AI to personalise your feed..but that tech part was only limited to people interested in tech, not the normal consumers)

The best AI products make the technology disappear into progress.

Share

That is a much harder thing to build. It is also the thing that actually lasts.

People rarely know what they want. But they are constantly telling you what they hate.

This is where the Henry Ford quote, cliché as it is, still points to something important:

“If I had asked people what they wanted, they would have said faster horses.”

Consumers are usually poor at describing the solution.
They are much better at revealing the friction.

They will not tell you they want “an AI-native multimodal creative operating system.”
They will tell you (as is obvious from Reddit discussions):

• “I need this ad by tonight.”

• “I don’t know how to make this look good.”

• “I hate starting from scratch.”

• “Can this just do it for me?”

That is where product insight lives.

The best founders do not listen literally. They listen structurally.

They understand that users may not be able to imagine the right product, but they are constantly exposing the jobs they are trying to complete and the friction that is blocking them. That is usually where the real opportunities come from.

And it is also why so many “AI for creativity” products are at risk of overestimating the size of their market.

AI lowers the cost of creation. That does not mean everyone becomes a creator.

This is one of the most important questions in consumer AI right now, and I think the market is still underestimating how much it matters.

A huge number of AI-native consumer products are being built on an implicit assumption: if creation becomes dramatically easier, then a much larger percentage of people will become creators.

That sounds plausible.
It may also be wrong.

Historically, creation has always been much smaller than consumption (aka 1% rule, i.e. 1% create, 10% remix and 89% consume), even when tools become significantly easier to use. Most people do not want to create consistently. They want to consume, react, lightly remix, and occasionally produce something “good enough” with very little effort.

That was true before AI.
There is a very good chance it remains true after AI.

AI absolutely lowers the cost of making things.
But lowering the cost of making does not automatically increase the desire to make.

That distinction is subtle, but incredibly important.

Because if your product thesis depends on millions of users suddenly wanting to create on a regular basis, you may be building on top of a behavior that never actually changes.

Why products like Wabi may struggle

This is also why I am skeptical of products like Wabi (raised $20mn seed round, is a Silicon Valley darling) and the broader class of AI-native creative consumer apps.

That does not mean they cannot get traction. They can. In fact, many of them probably will. These products are highly legible to early adopters, creators, and tech-native users. They are expressive, socially shareable, and often fun in exactly the way that drives initial engagement.

But that is not the same thing as building a durable consumer business.

YouTube has “billions of monthly logged-in users”, over 20 million videos uploaded daily, and Shorts alone now averages 200B+ daily views

Because the underlying creator-versus-consumer dynamic may not actually shift very much with AI. If anything, AI may simply increase the supply of content dramatically while leaving the demand side structurally similar.

That is a dangerous asymmetry.

More supply does not automatically create more meaningful demand.
It often just creates more noise.

And what that likely leads to is not a creator revolution, but a more automated content economy:

• a small number of highly leveraged creators

• a wider layer of casual remixers

• and a very large base of users who still mostly consume

If that is directionally right, then many AI-native creative apps will eventually run into the same wall:

Creation alone is not enough.

To become durable, they probably need to become one of two things:

1. A utility

Something that helps users complete a real, recurring job they already care about

2. A distribution layer

A place where people actually come to consume, not just create

The second is much harder than it sounds.

Because AI lowers creation cost.
It does not lower distribution cost.

And in consumer, distribution remains the bottleneck.

It still decides what wins.

ProductGeeks/ Ashish Sinha

How to find opportunities in consumer AI?

So far, most of AI success story has been on SAAS or prosumer space, but I believe the next big opportunity will lie in consumer AI…

Read more

a year ago · 2 likes · Ashish Sinha

The best AI consumer products will not feel like creativity tools. They will feel like relief.

This, to me, is the bigger lesson.

A lot of AI consumer products are still being built as if the opportunity is to give users more possibility. More outputs. More optionality. More expressive range. More things to generate.

But possibility is rarely the bottleneck. Completion is.

What most users are missing is not another blank canvas. It is fewer blocked moments.

They do not need more things they could make. They need fewer reasons not to finish.

That is why the strongest AI products will likely be the ones that remove:

• hesitation

• setup friction

• execution anxiety

• taste uncertainty

• complexity

• first-step paralysis

In other words, the winners will not be products that simply help users create more.

They will be products that help users move forward.

That is a much better product category.
And usually, a much better business.

tl;dr

Sora’s shutdown is not really about Sora.

It is a reminder of a much older truth that every technology cycle eventually rediscovers:

They do not want AI for its own sake.
They want momentum.

They do not want infinite creative power.
They want to get something done with less friction.

And they definitely do not want another product that turns an impressive technical capability into one more thing they now have to learn, configure, and actively use.

That is why JTBD still matters.
Not because it is fashionable, but because it keeps pointing builders back to the only question that has ever really mattered in consumer:

What painful, recurring job can this product make easier, faster, or disappear entirely?

That is where the real companies get built.

Everything else is mostly a demo.

ProductGeeks/ Ashish Sinha

90% of consumer AI ideas will die (use this 2 X 2 to validate)

90% of consumer AI ideas look sexy but crash hard — most end up in the “tar pit” with zero real traction or revenue uplift…

Read more

9 months ago · 6 likes · Ashish Sinha

What consumer products are you building or consuming? Do share in the comments section

Subscribe now

Big qn: Is everyone running an (AI) treadmill? Going nowhere, but exhausting tokens? Read on 👇

The RevenueCat State of Subscription Apps 2026 report* highlights several unique challenges and opportunities for AI-powered subscription apps beyond their higher initial revenue.

These insights reveal why AI apps often fail to sustain growth despite strong early traction.

Revenue Premium vs. Churn Penalty
AI apps generate 41% more revenue per paying user than non-AI counterparts, largely because tools like ChatGPT have normalized $20/month pricing.

However, they churn 30% faster on average, with annual retention at just 21.1% compared to 30.7% for non-AI apps.

This structural issue stems from hype-driven acquisition where users cancel after short-term needs are met.

Share

High Marginal Costs Drive Monetization Shifts
Unlike traditional apps with near-zero serving costs, AI features powered by LLMs create significant variable expenses per user.

Leave a comment

As a result, AI developers are shortening free trials, adopting hard paywalls, pushing annual plans, and adding premium tiers specifically for AI capabilities to maintain profitability. This has led to less generous freemium models across the category.

Flood of Vibe-Coded AI Apps

AI-assisted tools have unleashed a wave of rapid launches - 14,700 new subscription apps in January 2026 alone, many AI-focused.

These “vibe-coded” apps accelerate market saturation, inflating customer acquisition costs (CAC) and worsening overall churn as undifferentiated products compete.

Success now demands workflow moats beyond basic generation features.

---

Retention Warning Signs
Watch Month 3 - 6 metrics closely, as the AI revenue premium evaporates without sticky value. Many

AI apps lose 79% of annual subscribers before Month 12 due to unmet long-term expectations.

AI Apps Command High Churn
Artificial intelligence applications are demonstrating a unique monetization profile in the current market.

While AI features allow developers to command higher upfront revenue and justify hard paywalls, these apps suffer from significantly worse retention rates compared to traditional apps. Users often churn quickly after fulfilling a specific, short-term generation need.

The Legacy App Monopoly
Despite an influx of new market entrants, older apps continue to dominate consumer spending.

What’s your take?

Is everyone running an (AI) treadmill? Moving from one cool idea to another - without going deep in one?

*Get the report

how do we deliver value without dragging customers through the operational mess behind it?

ASP tried one answer. SaaS tried another.

Now we’re in the third attempt — and this one genuinely changes the equation.

Act I: The ASP Dream That Died (1996–2003)

Application Service Providers had a simple pitch: don’t buy software, just rent it over the internet. No CDs, no on-prem servers, no giant maintenance contracts. Pay monthly, we’ll run it.

Subscribe now

I actually started my career at Aztec, one of the better-known ASP players in the late ’90s. If you were around back then, you know the optimism we had :)

But the model collapsed under its own weight.

Every customer got their own server. Broadband wasn’t reliable. Most ASPs were reselling software they didn’t build, which meant they controlled everything except the product.

Then the dot-com crash hit. The idea wasn’t wrong — the timing was.

Act II: SaaS Gets It Right (2004–2020)

The next generation rebuilt from zero, and the key insight was multi-tenancy.

One codebase. Many customers. Clean data separation. Suddenly the economics worked: high margins, simple updates, global distribution.

Share

Salesforce, Slack, HubSpot — the entire SaaS wave grew on the same promise:

But SaaS had one limitation baked into the model. It gave you capability, not outcomes.

You still needed:

• a RevOps team to run your CRM

• analysts to interpret data

• support teams to run the support tools

SaaS removed the infrastructure headache.

It never removed the human labor that operated the software - frankly, it didn’t even care for it (the renewal is all the teams cared about).

Share NextBigWhat

Act III: Service as a Software — Agents Take Over (2025–)

This decade flips the model.

• ASP: “We’ll host the software.”

• SaaS: “We’ll build and maintain the software.”

• Service as a Software: “We’ll deliver the outcome. Start to finish.”

A lot of the business logic will move to a new multi-agent tier - Satya Nadella

AI agents don’t wait for human instructions. They plan, decide, and execute workflows independently. The software layer becomes background noise.

The result becomes the product.

And Jensen Huang didn’t bother sugarcoating it:

“The IT department of every company will be the HR department of AI agents.”

---

What New Businesses Look Like

Y Combinator captured the shift in one line:

“Vertical AI agents could be 10X bigger than SaaS — because companies spend far more on employees than on software.”

That’s the real unlock.

SaaS disrupted software budgets.

AI agents will disrupt labor budgets.

This creates new categories:

• Outcome-as-a-Product

  Don’t sell contract-review software.

  Sell “reviewed contracts in 2 hours.”

  Charge for results, not seats.

• Vertical AI Agents

  Purpose-built agents that run underwriting, AR, performance marketing, etc.

  Moat = domain depth + proprietary data.

• Agent Orchestration Platforms

  The OS for managing agent fleets — handoffs, conflicts, escalations.

• Micro-Workforce-as-a-Service

  A “10-person team” that is actually 9 agents + 1 human supervisor.

  Looks like staffing. Operates like SaaS.

Arthur Mensch from Mistral put a crude number on it:

“More than 50% of current enterprise software could be replaced by AI.”

And honestly, that feels conservative.

The Loop Closes

ASP had the right intention but wrong infrastructure.

SaaS fixed the infrastructure but abandoned the service ambition.

Service as a Software brings both together — with autonomy doing the work humans used to do.

Tl;dr:

• ASP moved the software.

• SaaS rebuilt the software.

• Service-as-a-Software replaces the human labor around the software.

In style: software ate the world and AI agents will eat the software !

2030: This is how the world will look like

Humans won’t operate tools anymore; they supervise agents.

Software moves from tool → colleague.

Business models shift from license → outcome.

Legal, finance, HR, sales, support — anything repeatable is now a blank canvas for AI-native companies.

The dream ASPs had in 1999 is finally real.

It just took 25+ years, cloud maturity, and autonomous intelligence to close the loop.

We started with “software as a service.”

Then “software as a subscription.”

Now we’re inching toward “just do it for me.”

What’s your take?

This evolution is reshaping industries by making AI a central player in decision-making and operations.

Notes from YC’s video on how agent economy is shaping up.

Rise of the Agent Economy

We’re witnessing the birth of an agent economy where AI agents autonomously choose tools and services. This shift means that agents, not humans, are becoming key decision-makers in selecting development tools. This could lead to a parallel economy where agents transact and interact independently, fundamentally altering how products are marketed and consumed.

Subscribe now

Documentation’s New Role

In the agent-driven world, documentation is crucial. Companies like Resend optimize their documentation to be agent-friendly, ensuring AI can easily parse and utilize their tools. This shift means that clear, structured documentation can significantly impact a tool’s adoption, as agents increasingly rely on it to make decisions. It’s a strategic advantage for businesses aiming to thrive in this new landscape.

“”AGI is literally actually here.””

Swarm Intelligence Emergence

Swarm intelligence, where multiple AI agents collaborate to solve problems, is becoming a reality. Unlike a single ‘God intelligence,’ this model mirrors biological systems, with agents working together like humans do. This approach could redefine innovation, as agents collectively tackle tasks, leading to more efficient and creative solutions.

Agents as Economic Actors

AI agents are poised to become significant economic actors, making decisions and transactions on behalf of humans. This shift suggests a future where agents manage tasks like booking reservations or choosing services, potentially leading to an economy where agents interact with each other, possibly even developing their own currency systems.

Adapting to Agent Preferences

To succeed in an agent-driven world, developers must create tools that cater to AI preferences. This means prioritizing open-source solutions and APIs, as agents prefer these over traditional websites. Understanding and aligning with what agents ‘want’ can position developers to better serve this emerging market, ensuring their tools are chosen by AI.

“”No human involvement is the big piece.””

Frequently Asked Questions

What is the significance of Moltbook in the context of AI agents?

Moltbook represents the first AI agent-only online community where AI agents interact with minimal human involvement. This platform showcases the potential for agents to collaborate and make decisions independently, hinting at a future where AI could significantly influence various sectors of the economy.

How can developers optimize their tools for AI agents?

Developers should focus on creating tools that are intuitive for AI agents, emphasizing open APIs and clear documentation. By understanding the limitations and capabilities of agents, developers can design their tools to enhance the agents’ user experience, making them more effective in their tasks.

What are the implications of AI agents having their own economy?

As AI agents become more autonomous, they may develop their own economy, transacting and making decisions independently of human oversight. This shift could redefine how businesses operate, necessitating new frameworks for legal liability and economic interaction between humans and AI.

Behind this growth is a serious ruthless sales engine that isn’t being talked about (it’s easier for founders to buy into their PLG story and completely overlook the enterprise GTM / sales motion).

My notes from a recent interview of their GTM lead, Carles Reina (video embedded*).

Setting Ambitious Sales Quotas

At ElevenLabs, sales reps are expected to bring in 20 times their base salary as their quota. This ambitious target ensures that only high-performing individuals thrive, fostering a culture of excellence.

While this might seem daunting, it pushes the team to exceed expectations and keeps the company competitive. It’s a bold approach that might not fit every business (not everyone has the same level of PMF as Elevenlabs!), but it drives growth and accountability.

Subscribe now

Plus, ElevenLabs compensates both account executives and customer success managers for upsells, even if it means paying double.

This strategy ensures both roles are motivated to maximize customer value, leading to increased revenue. By aligning incentives, the company encourages collaboration and effort from both parties, ultimately benefiting the business. It’s a strategic investment in people that pays off through enhanced customer relationships and sales growth.

I believe that much of this strategy, i.e. of paying double is being ‘bailed out’ by their VCs :)

2. Understanding Sales Fit

Not every sales rep fits the mold, and at ElevenLabs, those who don’t meet quotas are let go.

However, it’s crucial to differentiate between those who lack product understanding and those building long-term pipelines. Some reps may need more time to develop relationships, especially in challenging industries. Recognizing potential and providing support can turn underperformers into top achievers, as seen with reps who eventually exceed their targets.

”If you don’t achieve your quota, then you’re going to be out, right?”

3. Remote Sales Culture

Building a remote sales culture requires being proactive and ensuring salespeople are on the road, engaging with clients. At ElevenLabs, remote work is embraced, but sales reps are encouraged to travel and meet clients in person.

This approach fosters stronger relationships and better sales outcomes. It’s about balancing remote flexibility with the necessity of face-to-face interactions, ensuring the team remains dynamic and connected to the market.

• Monthly pipeline review sessions at ElevenLabs are conducted openly, fostering transparency and accountability.

Read more