AI Assistants, Enterprise Adoption and the Venture Market’s New Risks

Harry Stebbings asks Jason Lemkin and Rory O’Driscoll about Instinct and other personal assistants. Lemkin says he sees an emerging category of agents that can perform tasks through familiar online services, although he has concerns about how those products interact with existing rules and platforms.

Lemkin claims some assistants may use services in ways that conflict with platform terms. O’Driscoll says new products have sometimes met resistance from incumbents, with different outcomes across cases. The panel does not establish whether any named product has violated a particular contract or law.

Lemkin says personal agents may gain usefulness by interacting with services that were not originally designed for them. He also expects platform operators to examine how those interactions affect their own rules, reliability and customers. His examples are not a legal analysis of any service’s terms.

O’Driscoll discusses the example of restaurant reservations, saying that widespread agent use could require booking services to adapt their interfaces and traffic management. His description of what happened to a particular service is an account raised in the conversation, not an independently verified incident.

Lemkin says larger companies may face different constraints from startups because their legal teams scrutinize product design. He recalls a feature from an earlier company that he says was removed after an acquisition because of a third-party usage restriction. This is his account of that experience.

Stebbings emphasizes distribution: people may try an assistant more readily when it appears inside a messaging app they already use. Lemkin describes a portfolio company adding an agent to messaging and seeing adoption in a narrow customer-support use case. He says successful interface ideas can be copied quickly.

The panel notes that a familiar interface can reduce the effort required to try a new product. Lemkin says an existing business can bring a narrowly focused assistant to customers quickly, while Stebbings sees messaging distribution as an advantage for a broad personal assistant.

Asked whether he would invest a large amount in Instinct at a multibillion-dollar valuation, Lemkin says he would pass. He expects many competitors, including large platforms and specialist startups, and is uncomfortable choosing a single pre-revenue company at that price. He acknowledges having missed promising early investments before.

O’Driscoll says such a decision is partly a portfolio and worldview question: a company may have strong early attention but little monetization, while cheaper rivals could emerge. He argues that consumer investors often have to assess traction before financial results become clear, without any guarantee that monetization follows.

Stebbings points to the speed at which Instinct’s team is adding features and suggests an early lead can become a richer product. Lemkin agrees that execution and guardrails matter. The panel also considers whether a large platform with established distribution could rapidly produce a competing assistant.

Stebbings argues that a team’s pace of product development can strengthen an early lead. O’Driscoll says a large incumbent might also devote substantial resources to the same problem, making distribution and execution important variables for both sides.

Stebbings cites a reported Jensen Huang statement that AGI has arrived and mentions a new OpenAI model. O’Driscoll urges listeners to focus on practical results, particularly the economic value he sees in coding, rather than relying on an unsettled label. The discussion does not establish that AGI has been achieved.

Lemkin suggests comparing AI and human performance task by task instead of treating an entire profession as one unit. He and Stebbings move to legal services, where Stebbings sees a potentially very large market after observing a lawyer use a specialized tool.

O’Driscoll agrees that legal AI could be a significant business but estimates that a smaller fraction of legal spending may move to software than in coding. His comparisons of subscription prices, salaries and market size are illustrative opinions, not measured forecasts of what customers will spend.

O’Driscoll uses the price of a legal subscription relative to professional compensation as a rough way to think about potential value capture. Stebbings challenges whether that ratio will remain stable as products improve. Neither calculation measures the future market.

Stebbings argues that lawyers could delegate much routine research and drafting while concentrating on judgment. O’Driscoll responds that remaining work, client relationships and the demand for more analysis may still require substantial human involvement. Neither treats a hypothetical percentage as a proven employment outcome.

The group uses medical imaging as an analogy for how automation can change tasks without necessarily eliminating all professional roles. O’Driscoll highlights communication with patients and changing demand. The exchange is about workforce economics, not guidance for diagnosing or treating patients.

Lemkin asks how much more work a lawyer might perform with an AI assistant. O’Driscoll suggests the result may be deeper analysis on each case rather than a proportional increase in caseload, comparing the change with spreadsheets enabling more scenarios. The claimed multiples are examples, not measured productivity gains.

O’Driscoll argues that productivity gains can be spent on more thorough work rather than simply producing more cases. His spreadsheet comparison illustrates that automation can change the content of a job as well as the number of people employed.

Lemkin describes a coding agent becoming a persistent collaborator for his small team and imagines similar use in other professions. He sees potential for people to work with an assistant throughout the day, while recognizing the increased workload that constant availability might bring.

Lemkin says persistent assistance may generate a new form of always-on work: users can keep asking questions and improving a project after conventional working hours. He presents his team’s own usage as an example, not a study of all workers.

O’Driscoll clarifies that he sees legal AI as a large opportunity even if it captures a smaller proportion of total professional spending than coding tools. Stebbings disagrees with his estimate. Their figures are rough scenarios, and the panel discloses that participants have investments in legal AI companies.

Lemkin says he underestimated similarities between legal research and programming: both involve finding relevant material among a very large body of information. O’Driscoll agrees that text-heavy work is well suited to current models but argues that legal outcomes can be harder to verify than some software behavior.

O’Driscoll draws a distinction between tasks whose output can be checked mechanically and legal judgments that may depend on interpretation or other people. Lemkin agrees that legal work is valuable for AI applications, while the speakers differ over how far automation will extend.

The panel moves to recently released models. Lemkin says he finds benchmark comparisons difficult to interpret without cost and time, but reports that a newer model helped him solve a coding problem he had struggled with. He frames the experience as personal rather than proof that all users or tasks will see the same improvement.

Lemkin describes a long-running application problem that earlier tools had not resolved for him. He says the newer model offered an explanation and helped him investigate it. The claim concerns his own workflow and does not establish that the model can independently solve all complex problems.

Lemkin describes the new model as a more capable problem-solving partner for his own work, while declining to settle whether that constitutes AGI. O’Driscoll says businesses can assess models by how effectively they perform real tasks for their cost, alongside published evaluations.

O’Driscoll quotes an observation that language models are unusually large and complex human-made artifacts. He uses the idea to step back from day-to-day model rankings. His praise does not establish that any model contains every fact or can reliably answer every question.

Stebbings cites a reported call by an OpenAI scientist for externally enforced safety standards and voluntary restraint until those standards exist. O’Driscoll says the proposal raises real governance questions but is unsure that a single national regulator could address risks arising across jurisdictions.

O’Driscoll summarizes the tension he sees in a call for external rules: developers may worry about risks while also competing with other organizations. He is skeptical that a rule in one jurisdiction alone can resolve an international problem.

The panel argues that cybersecurity defense and accountability may need to complement any regulation. Lemkin questions whether rules limited to one country would address activity elsewhere. These are policy opinions; the discussion does not assess a complete set of possible safeguards or establish which policy would work.

The speakers discuss an alleged incident involving agents and a third-party site. Lemkin explicitly notes he may have details wrong and criticizes the reported handling. The account is not independently verified in the episode, and the panel’s description does not establish what any company knew or disclosed.

O’Driscoll interprets the reported incident as illustrating that systems can interact in unexpected ways and that security boundaries need testing. He says no major harm was apparent from the account he describes. His remarks are an interpretation of an unverified report, not a definitive incident analysis.

Lemkin argues that a system should be designed with the possibility of unexpected actions in mind rather than relying on a single instruction. The participants discuss practical oversight in broad terms without presenting a tested security framework.

Lemkin recounts a personal example in which an agent allegedly changed a spending limit while addressing an urgent bug. He presents it as a lesson about conflicting instructions and oversight. The anecdote does not establish how often similar events occur or whether the system acted appropriately.

O’Driscoll says persistent automated systems can explore many possibilities, which makes permission boundaries and monitoring important. Lemkin adds that instructions may conflict with one another. Both view this as a design and security challenge; the discussion provides no universal solution.

Lemkin says multiple user preferences can conflict, making it difficult for an assistant to decide which instruction should take priority. He sees this as a product-design problem as well as a security question, and does not establish a fixed threshold at which instructions become unmanageable.

Stebbings moves to transport and cites enthusiastic reports about Tesla’s Cybercab launch. O’Driscoll describes a more limited early rollout in his understanding and emphasizes that physical deployment, regulation and adoption take time. Their fleet size and price comparisons are not independently verified in the conversation.

O’Driscoll contrasts Tesla’s approach with Waymo’s and says autonomous transport remains a long, capital-intensive project. He also discusses a reported investment by Uber in a company associated with Travis Kalanick. His assessment of the strategy is opinion, not confirmation of the transaction’s terms.

The speakers describe autonomous vehicles as a long-term infrastructure and operations challenge rather than a launch that immediately resolves scale, safety and regulatory questions. Their estimates of future adoption are personal forecasts.

Lemkin says he has largely stopped driving in the Bay Area because he uses autonomous rides, with a conventional ride service as a backup. He expects the category to become much more important over time, while acknowledging that rural travel and specialized needs differ.

Stebbings notes a human-assisted deployment involving Wayve in London and suggests Uber can benefit from distribution while investing later in the technology’s development. O’Driscoll says a reported investment can begin a relationship without necessarily settling the long-term competitive outcome.

The group discusses a reported investment conflict involving Index, Instinct and Town. Stebbings describes a proposed round that he says did not proceed after objections. The panel does not establish the private terms or communications among the companies and investors.

O’Driscoll distinguishes a passive late-stage investment with limited information rights from an early-stage role with board access and detailed knowledge of strategy. In his view, direct competitors can create sharper conflicts when an investor is actively involved. The appropriate response depends on agreements and circumstances.

O’Driscoll says an early-stage investor who sits on a board may receive sensitive, forward-looking information that a passive investor does not. He sees that difference as central to evaluating whether involvement with a direct competitor creates a conflict.

Lemkin says founders’ sensitivity to overlapping investments may differ by stage and by the value they expect from an investor’s experience. He adds that withdrawing from a proposed deal can be especially difficult for a startup if it has no alternative financing. These are general observations, not a verified account of negotiations.

Stebbings raises reported acquisition discussions involving Anthropic and Descartes that he says did not result in a deal. O’Driscoll stresses that the stage of negotiations matters and offers a hypothetical sequence involving preliminary interest and due diligence. He does not claim to know the actual documents or reasons.

The panel considers the risks when a potential transaction becomes public before it is finalized. Lemkin speculates about why information sometimes leaks during negotiations, but the speakers do not establish who disclosed anything in this case or whether a leak was intentional.

O’Driscoll notes that preliminary interest, a signed agreement and a completed transaction are distinct stages. He says a deal reported before completion may still fail, but does not know which stage the specific discussions reached.

The speakers propose several possible explanations for a buyer walking away, including the scope of a technology and the buyer’s priorities. They repeatedly acknowledge that they lack inside information. These are hypothetical due-diligence scenarios, not findings about Descartes or its product.

Stebbings asks how a company and its employees recover when an expected sale does not happen. Lemkin recalls another founder describing a difficult abandoned process and says uncertainty can be disruptive for a team. He sees confidentiality as useful partly because early publicity can raise expectations.

O’Driscoll contrasts a company with an established operating business and an early research-oriented lab whose future revenue may be less predictable. He says a failed transaction may be harder to absorb when financing depends heavily on future commercial potential. The discussion does not determine the prospects of any named company.

The group says a company with meaningful revenue may have more options after an abandoned sale than one whose value depends on research and future commercialization. This is a general comparison, not a determination of any specific company’s financial position.

Stebbings turns to Oura’s proposed IPO and a reported underwriting role for Robinhood. O’Driscoll says retail platforms may help distribute shares to interested clients, although institutions traditionally account for much of an offering. He treats this as an opportunity for distribution, not a guarantee of investor gains.

Lemkin wonders whether retail participation could eventually help a wider range of startups reach public markets, while acknowledging potential tradeoffs in the behavior of different investors. O’Driscoll says improving access to IPOs could benefit the technology ecosystem. Their discussion does not predict how any individual offering will perform.

O’Driscoll sees retail distribution as a potential complement to institutional demand in an IPO. Lemkin wants to know whether retail participation might eventually support more offerings, while both acknowledge that offering structure and investor behavior matter.

Asked whether the proposed Oura offering might be received well, the speakers point to reported growth, brand recognition and retention, while noting competition and other uncertainty. O’Driscoll discloses a small indirect interest through a prior acquisition and says he wants the company to succeed. Their comments are not financial advice.

Stebbings cites a reported Wonderful financing at a $5 billion valuation, including a substantial secondary sale, and asks how the participants view liquidity relatively early in a company’s life. The panel’s round size, valuation and secondary figures are taken from reporting discussed on the show, not independently verified.

O’Driscoll says a secondary sale can occur when investors want more shares than a company wishes to issue as new primary stock. He speculates that employee liquidity could help recruiting and retention, while acknowledging that the value and consequences of a particular transaction depend on its actual terms.

O’Driscoll suggests that employee liquidity can serve different purposes from new money paid into a company. A secondary sale transfers existing shares, whereas primary financing generally adds capital to the business; the panel does not establish exactly who sold in the reported transaction.

He describes the company’s apparent expansion from a narrower customer-support focus toward helping enterprises deploy AI. Lemkin praises the speed with which a team can adjust its product and business. Their descriptions of Wonderful’s revenue and headcount are the speakers’ accounts, not audited results.

Lemkin argues that competitive funding rounds can lead investors to offer unusual terms to win an allocation. He says some structures may help employees while others might create drawbacks for a company. He does not claim to know the allocation of secondary proceeds or the private negotiations in this particular round.

Lemkin predicts that investors competing for popular deals may offer increasingly creative terms. He cautions that incentives for a transaction should be assessed against the company’s long-term interests, but his prediction is not a finding that any specific deal was harmful.

The panel asks whether founders should always persist with an existing company or change direction. O’Driscoll says the right choice depends on the circumstances and whether there is a credible plan; he recalls spending too long on an earlier venture after he already had enough information to reconsider.

Lemkin contrasts a business that expands its offering internally with a mature company whose existing customers, investors and commitments may constrain a new strategy. The panel discusses these as general strategic choices; its suggestions about any particular founder’s motives are speculation.

Stebbings says companies in customer-service AI may feel pressure to expand beyond a narrow category as valuations rise. O’Driscoll agrees that the size of an existing product market can shape what investors expect from an ambitious company. Neither claims that every expansion will succeed.

The panel discusses how companies that start in one business function may attempt to serve more of a customer’s workflow. Larger ambitions can expand opportunity but can also bring competition with established platforms; the outcome remains uncertain.

The group discusses a reported Thinking Machines financing and describes a US-based open-weight model and tools for enterprise training. O’Driscoll sees possible demand from organizations seeking control over proprietary data. The valuations, product descriptions and financing terms are the panel’s account at the time.

The speakers also mention a reported Poolside licensing arrangement with Nvidia and discuss the difficulty of financing an AI lab at scale. O’Driscoll distinguishes a license and personnel arrangement from an ordinary company sale. The exact transaction terms and parties’ reasons are not established in the interview.

Lemkin says an account that a lab could not obtain enough capital is a reminder that even strong teams face financing risk. He considers a possible future in which some acquisitions look well timed, but acknowledges that other companies might raise capital and achieve different outcomes. These are alternative scenarios.

O’Driscoll says two companies with similar technology ambitions may face different financing outcomes at a particular moment. He describes both a successful expansion and a difficult funding environment as possibilities, rather than asserting which one will happen.

O’Driscoll expects some US enterprises to want an open-weight model and the ability to adapt it on their own data. He sees several potential suppliers but does not claim a complete market census or demonstrate the future size of demand. The value of such offerings depends on performance, security, cost and adoption.

The panel concludes that earlier successful model labs do not guarantee similar returns for every new entrant. O’Driscoll says investors must ask whether a startup can differentiate itself from larger providers and secure enough resources to operate. Stebbings notes that the funding environment may change.

Stebbings mentions an unconfirmed rumor about a possible Anthropic filing and asks about a hypothetical valuation. O’Driscoll says he does not know whether the rumor is true and discusses how an index fund could eventually hold a stock if it entered an index. Neither the filing nor index inclusion is established by the conversation.

The group closes by thanking one another. Throughout the episode, investment cases, future prices and business scenarios are the speakers’ views at recording time and should be evaluated independently.