Bret Taylor on AI Applications, Model Costs and Building Sierra

In this October 2024 conversation, Sierra co-founder and CEO Bret Taylor discusses AI applications, model economics and leadership with Harry Stebbings. Taylor also serves on OpenAI’s board; his views reflect those roles and Sierra’s commercial interest in AI agents.

Taylor did not always expect a technology career. As a teenager he earned $4.25 an hour at a gas station, then made $400 building a mechanic’s website and began making sites for local businesses. At Stanford, an introductory computer-science course deepened his enthusiasm for software.

He believes most skills can be learned with focus, while acknowledging innate differences. Entrepreneurship also demands a tolerance for intense and uncertain conditions; he rejects the idea that there is a single founder personality.

Leadership is another learnable craft, he argues. Different teams require different ways of motivating people, and companies should invest in training rather than simply label someone a natural leader or not.

Taylor says AI is in a bubble, but compares it to the dot-com period: speculative excess can coexist with a technology that eventually changes the economy. He predicts enduring AI-native companies will emerge, while accepting that individual valuations and investor returns are separate questions.

Stebbings asks whether much higher startup valuations make the comparison misleading. Taylor acknowledges the concern and says he is approaching the question as a builder rather than deciding which investments will produce a particular multiple.

Taylor does not expect foundation models to absorb every software application. He compares AI with the cloud market, where infrastructure, developer tools and software applications all became substantial categories.

Companies generally want solutions rather than the continuing burden of building and maintaining every tool themselves, he argues. He describes Sierra as an application company that adapts others’ models to build customer-facing agents; that assessment reflects its business model.

Stebbings asks whether AI services firms will be major winners. Taylor expects some early implementation spending to decline as packaged products improve, but sees continuing value in consulting and change management when businesses reorganize around AI.

He distinguishes more commoditized foundation models from frontier models. He says lower-cost open models may suffice for some tasks, while frontier providers can temporarily distinguish themselves through meaningful improvements; the pace and durability of those leads remain uncertain.

Taylor is skeptical of most application startups pretraining their own large models because the upfront cost can distract from finding product-market fit. He explicitly distinguishes AGI research labs and large providers from companies primarily building solutions; this is his investment and product judgment, not a claim that pretraining has no value.

Asked whether model improvements will continue in dramatic steps, he says that is not guaranteed and a plateau is possible. He is optimistic because progress can come from three areas: data, compute and algorithms, including training and model design.

Taylor advocates responsible iterative deployment as a way to learn about safety, misuse and real-world effects. This is his preferred approach as an OpenAI board member; the interview does not establish that iteration alone resolves all safety risks.

Stebbings asks whether pursuing AGI conflicts with building consumer and enterprise products. Taylor argues that products and developer access can distribute the benefits of research, while acknowledging that balancing the mission and commercial choices is complex.

For application companies, Taylor prefers relatively modest training expense and inference costs tied to usage. He expects inference to become cheaper as hardware improves and techniques such as distillation produce smaller models; this is a forecast rather than a guaranteed cost curve.

He sees large cloud providers making substantial capital investments because frontier models could matter to their businesses. He expects scale to favor a small set of infrastructure providers, while acknowledging that the economics are complex.

Taylor says open models became useful more quickly than he had expected, citing Meta’s Llama 3.1 in the 2024 context. He expects a market with both proprietary frontier offerings and open models, rather than a single model for every use case.

Nobody knows the compute, data and methods needed to reach something resembling AGI or the returns required to justify the spending, he says. He believes large providers can rationally make long-term bets, while smaller startups face different financing risks.

He expects consolidation among companies pretraining models and a different, potentially more sustainable cost structure for tools and application businesses. This is his market prediction, not a known outcome.

Sierra helps brands create customer-facing AI agents, Taylor says, citing companies including Sonos and SiriusXM. He thinks conversational software could become a major interface for customer interactions; Sierra sells products into that opportunity.

He compares the shift to the point when smartphone touchscreens became useful enough to change habits. He believes conversational AI crossed an important quality threshold with GPT-4, though future adoption and the reliability of particular systems remain open questions.

His forecast is that customers will increasingly converse with a brand’s AI over the next several years. He does not expect agents to eliminate websites and apps; he sees them as an additional channel that could take over some tasks.

Text, voice and multimodal experiences can reduce friction in contexts such as messaging, driving or the kitchen. Taylor argues convenience may make conversation important even when it is not the best interface for every task.

Replacing the phone entirely is possible, he says, but the device already does many things well and is widely available. In the nearer term he expects phones to remain the main computer, potentially paired with glasses or earbuds so people look at screens less.

He considers WhatsApp well positioned for conversational commerce and expects apps may eventually include agents. Sierra’s aim, he says, is to let a brand bring its agent to whichever interfaces customers use.

The difficult technical problem is that generative AI is creative and non-deterministic, while businesses have real policies. Taylor describes a shift from enumerating every software rule to expressing goals and guardrails for an agent’s behavior.

A retail website can enumerate categories and paths; a free-form agent can encounter far more varied requests. If every answer is fixed in advance it may feel robotic, but unrestricted creativity can produce incorrect or brand-damaging responses.

He uses a subscription cancellation as an example: the agent might ask why, offer an authorized discount and process the cancellation if the customer still wants it. Goals and guardrails should define allowed offers and tone while leaving room for a useful conversation.

He expects adoption to begin with lower-risk tasks and expand as systems improve and companies gain confidence. He says Sierra customers already use agents for sales and subscription retention; these are company descriptions rather than independently audited results.

More agency can make an interaction more flexible, but also reduces control. Taylor says businesses must choose how much freedom to allow for each workflow, with tighter constraints for higher-risk situations.

He anticipates new roles such as agent engineer and AI experience architect, responsible for building agents and shaping conversations. The job categories and their future scale are his expectations.

Generative systems challenge assumptions built into traditional software engineering: identical inputs need not produce identical outputs. Taylor suggests preparing operational responses for errors as well as attempting to prevent them, while recognizing that failures cannot currently be eliminated.

Stebbings raises the risk that synthetic content will erode trust in all information. Taylor hopes AI-assisted tools and security practices will help people assess authenticity and advocates learning from deployment; he does not claim such tools can reliably verify every item.

Taylor says he sought outside funding for Sierra partly for board accountability and strategic advice. He contacted Peter Fenton, with whom he had worked previously, and values an investor relationship in which the founder can call for help on difficult decisions.

In a quick-fire round he says the speed of falling AI costs surprised him and argues attention should extend beyond models and hardware to applications. Both are his assessments at the time.

On board work, Taylor praises Fidji Simo and describes the challenge of staying sufficiently informed without taking over management. Advice must be tailored to the company and CEO, with a reliable cadence for getting the right information.

He says Mark Zuckerberg appointing him Facebook CTO changed his view of what he could do as a leader. Taylor recalls the confidence another leader placed in him, even when he was unsure he was fully qualified.

From leaders including Zuckerberg and Marc Benioff, he admires a long-term outlook and the willingness to make difficult short-term bets for a distant goal. He also emphasizes communicating that vision so employees can understand why the work matters.