Sarah Guo

The indiscriminate “SaaSpocalypse” trade mistakes a real long-run shift for an immediate extinction event. Gil argues that a fleet-management product like Samsara—with in-cab hardware, distribution, enterprise sales, and support—will not simply be replaced by a weekend vibe-coded app, though support can use vibe agents. Guo says Fortune 100 change management and security make a weekend-built CRM likely implausible. Decagon and Sierra do demonstrate a genuine move from per-seat software toward usage-priced support agents, but “this isn’t gonna be every single SaaS company.”

AI-native companies show that code can become abundant without making the rest of company-building abundant. Guo cites portfolio companies with hundreds of millions of dollars of revenue and fewer than 50 engineers, yet rapidly expanding from zero to nearly 100 salespeople: “Vibe sales is not happening.” Gil expects engineering productivity to be absorbed by enormous unmet software demand rather than simply eliminating engineering teams.

The investable engineering bottleneck may shift from writing code to allocating trustworthy human attention. If agents generate enormous volumes that nobody reads, “nobody deeply understands the code base,” leaving production systems fragile and full of “vibe coding slop.” Guo calls agent-first engineering management, testing, smart review, and formal verification “open season around this really, really big problem.”

A month of AI hype blurred partnerships, demos, and planted behavior into evidence of autonomous markets. Gil disputes claims that agents already choose vendors: partnership defaults that provision particular tools resemble Airtable silently running on AWS, not independent purchasing judgment. The “Malt book” Reddit-like forum example also seemed partly human-generated or planted for marketing. Guo’s sharper equity-research critique is that “the theory of competitive advantage didn’t just like poof, disappear”; production, distribution, change management, and product completeness still matter.

The underlying economics remain extraordinary even after stripping away the hype. Using Capital IQ data and projections, Gil’s team puts the AI labs’ journey from $1 billion to $10 billion of revenue at roughly one year, versus about 20-plus years for ADP and Adobe, while public projections imply $10 billion to $100 billion could take the labs three to five years. Meanwhile, GPT-4-equivalent pricing fell from roughly $37 to $0.25 per million tokens in 21 months—150x—and o1-equivalent pricing from $26 in December 2024 to $0.30 in November 2025—88x in 11 months.

AI could pull far more GDP into technology while concentrating much of the resulting value in a power-law head. Gil says the leading tech companies’ share rose from roughly 4% of US GDP in 2005 to about 12% today, with plausible 2035 scenarios ranging from 15%-20% to 30%; the top eight tech companies already represent about $23 trillion and well over half of the S&P’s value. Guo argues that expanding technological surface area could let the tail dominate, while Gil says there may be more $100 billion businesses but “the head and torso aggregate almost all the value.”

Faster growth does not confer SaaS-era durability: AI may compress a decade-long displacement cycle into one or two years. Gil warns that many startups get only about a 12-month window at peak value, so boards should schedule unemotional exit reviews once or twice annually; only a very small group should “never, ever, ever sell.” The strongest defense is a bundle spanning five or 10 aspects of the same vertical or application, reinforced by platforms, ecosystems, networks, or hardware—because “if every two years is 10 years,” a cloneable point product is a precarious control point.

Tony Zhao places robotics “in between the GPT moment and the ChatGPT moment”: the field appears to have a scalable recipe, but has not yet converted scale into a great consumer product. Classical sense-plan-act systems required bespoke interfaces for every task and environment; newer learning methods aim to scale data and models instead of repeatedly rebuilding task-specific systems. His bet is that robotics will follow other AI fields: more relevant data should improve usefulness.

Sunday’s central asset is an in-the-wild data engine with almost 10 million long-horizon trajectories, collected by more than 500 people. Cheng Chi’s UMI work replaced lab-bound teleoperation with a GoPro and gripper, while Diffusion Policy stabilized multimodal imitation, ALOHA made dexterous data collection more intuitive, and ACT/action chunking plus transformers helped make bimanual tasks scalable. A three-student collection produced 1,500 espresso-serving clips in roughly two weeks.

The apparent simplicity of glove-based collection hides full-stack execution difficulty. Sunday iterates the data device, robot controls, automated filtering and cleaning, calibration, training pipeline, and mechanical design together; the glove has gone from V0 to V5 with around 20 iterations per version. As Zhao recalls, the founders initially feared someone could “just take our glove,” but learned that “things are so much harder than we thought.”

Sunday rejects humanoid completeness when simplification can produce a useful robot faster and more cheaply. Its friendly, cartoon-like robot uses three fingers because people often use those fingers together for chores such as grasping handles or opening dishwashers. Perception lets the robot correct inexpensive, compliant, imprecise hardware, supporting a mechanically safe and compliant design while retaining sufficient accuracy.

The founders currently see imitation learning as more sample-efficient for manipulation, while reinforcement learning works well for locomotion. Ground contact is comparatively tractable to simulate, but reproducing a deforming hand, transparent cup, orange juice, reflections, and physical properties is extremely difficult and expensive. For manipulation, the behavior—put the hand in front of the cup and close with suitable force—can be easier to demonstrate than the world is to simulate.

Commercialization begins with a 2026 in-home beta, with shipping to the masses in 2027 or 2028 possible but explicitly dependent on reliability, capability, safety, and price. Prototypes currently cost $6,000-$20,000; at a few thousand units, Sunday expects material cost likely under $10,000 as costly low-volume cladding shifts to injection molding, implying a selling price around that level. Zhao envisions more than 1 billion home robots within a decade, while Cheng frames one possible future as the marginal cost of labor in homes approaching zero.

The founders’ rule for evaluating robotics videos is “make zero assumptions. No Priors.” Verify autonomy, then treat the exact object, person, environment, and demonstrated sequence as the proven capability—not evidence the robot can generalize. Sunday’s evidence includes long-horizon table cleanup, fragile glass handling, zero-shot trials across about six Airbnbs, espresso operation, and sock folding, though Cheng says the team shattered many glasses during development.

Elad Gil sees a subset of AI markets consolidating, with apparent winners for the next two years—even if today’s leaders may not be five-year winners. Foundation LLMs, medical scribing, coding, and customer success have recognizable frontrunners; sales productivity, financial-analyst tooling, and accounting remain open, and Sarah Guo adds pharma. Gil wonders whether the unresolved markets lack the right product approach or still need better models. After years when “the more I learn, the less I know,” he finally sees “a nice breather in terms of uncertainty.”

The emerging application playbook combines a valuable vertical workflow, proprietary data, distribution, and users who can create, derive, or extend knowledge from that data. Guo places Abridge and OpenEvidence in that category while remaining unsure how sales will be won in a historically fragmented market. Coding suggests the ambiguity is narrowing: plausible entry points now include Cursor, Codeium/Windsurf, Cognition, and Microsoft Copilot.

AI startup consolidation may become strategic defense against incumbents, not an admission of defeat. Gil advises the top two startups in a category to consider merging to stop the startup-to-startup war and shift the fight toward three or four incumbents. Guo says founders, boards, and investors often treat even considering a merger as capitulation, though she calls it “capitulating in service of winning.”

Gil argues that several enormous biotech markets are commercially obvious but structurally neglected. Examples include deriving sperm or eggs from reprogrammed adult cells, maturing a woman’s remaining oocytes, treating skin aging and hair loss, restoring near vision or hearing, and regrowing teeth through genes such as USAG-1. The startling implication of cell-derived reproduction is that “any adult” could potentially have children with another—and cells obtained from someone by swapping some off them in a handshake could potentially enable reproduction involving that person.

Biotech’s financing architecture selects for assets that pharma will buy, rather than enduring new companies serving unconventional demand. Gil says incubators may invest $40 million for roughly 40% ownership, then steer programs toward cancer, cardiovascular disease, or neuroscience pipelines sought by acquirers. Regulatory capture and scientists’ view that fixing wrinkles is “low status” further strand usable science outside the venture funnel.

Gil presents world models and reinforcement learning as a possible route beyond text prediction, not primarily as near-term gaming products. Agents need planning, tool use, feedback, alternative reasoning paths, and self-evaluation, while behavior cloning is brittle outside demonstrated paths. The challenge is building “a copy of the universe”—an environment rich enough to teach adaptable problem-solving, yet cheap and diverse enough to avoid overfitting. Gil is unsure about immediate commercial applications; Guo’s extrapolation from Go and molecular evolution is that unconstrained search could produce unusual but superior code, molecules, or strategies.

Tahoe 100 shifts single-cell perturbation data toward machine-learning scale. Vivo reports 100 million single-cell measurements across 50 cancer models from different patients, 1,200 drug treatments, and 60,000 drug–cell or drug–cell-line interactions, versus roughly 1–2 million publicly available perturbational data points before it. Dave Burke compares the hoped-for impact to an “ImageNet moment” for cellular biology.

The differentiated asset is causal, diverse, consistently generated data—not another pile of healthy-cell observations. Perturbations provide a before-and-after signal, while disease models expose states absent from normal tissue. Four people performed consistent work across 60,000 experiments with effectively no batch effects. The team argues that scale and breadth can make the data itself generate hypotheses and surprises.

Virtual cells address the systems-level question that protein models leave open. Protein language models can learn folding, binding, and structural biology, but a drug acts on a protein embedded in a cell, tumor, and broader system. The proposed virtual cell is a “notional CPU” that predicts how a genetic edit or chemical changes the transcriptome—and eventually solves the inverse problem of moving a diseased state toward health.

Current virtual-cell models remain commercially immature, with predictive performance on differentially expressed genes on the order of 10% and no accepted benchmark. Tahoe 100 represents roughly 200–300 billion training tokens under current encodings. Dave describes roughly 1 trillion tokens as a comfortable comparison point from other domains, while stressing that biology’s scaling laws remain uncertain. Hani Goodarzi places protein models beyond GPT-3 maturity but cellular models “closer to GPT-1 than 2.”

Open sourcing is an operating strategy, not philanthropy detached from Vivo’s economics. Tahoe 100 joins Arc’s roughly 230 million-cell scBaseCamp collection to form a 330 million-cell Virtual Cell Atlas; the community can train and critique models while Vivo continues building its Mosaic data-generation platform. Nima’s framing is “a new stake in the ground”: a small internal team can recruit an external research ecosystem without building a large organization.

If virtual cells work, the potential prize is simultaneous compression of target risk, chemical search, and experimental time. Roughly 90% of drugs fail in clinical trials; the guests argue failures reflect both poor drug matter and selection of the wrong target. A sufficiently accurate model could search tens of millions of compounds, generalize across patient contexts, and help teams “measure twice and cut once”—but at 10% accuracy, Patrick Hsu cautions, “you’re just simulating noise.”

The technical inflection may be now, but therapeutic proof will remain slow. Chinese biotechs’ cost, pace, antibody manufacturing, and IND-enabling packages are pushing US companies toward leaner teams and stronger vendors; Nima calls it “morning in bio” and urges building now rather than promising results in three to five years. Sarah Guo stresses that treatments commonly require 11-plus years, and even moving success from 10% to 30% leaves a “law of small numbers” that may take a decade-long window to validate.

Wildfire is not an intractable climate-only problem in Bill Clerico’s account; it is an adaptation and infrastructure problem with addressable causes. Climate change lengthens hotter, drier seasons, but forests are roughly three times denser after a century of “maximum suppression.” Meanwhile, 98% of climate capital goes to mitigation and just 2% to resilience and adaptation—a split he sees finally shifting.

The highest-leverage prevention is more prescribed fire and mechanical thinning, but regulation and local risk aversion make both painfully slow. Clerico’s own 170-acre timber treatment took three years and six biologist visits; Sarah Guo cited average environmental reviews of 4.7 years for simple projects and 7.5 for complex ones. Florida burns 2–3 million prescribed acres annually, while California targeted only 100,000: “The forests are going to burn…we can do it on our terms or they can do it on their terms.”

Housing growth has converted wildfire into an urban balance-sheet problem. Homes in the wildland-urban interface have increased 46% since 1980, yet Paradise largely rejected tougher codes while rebuilding after the 2018 Camp Fire. Once flames cross a community’s perimeter, Clerico says the event becomes an “urban conflagration,” spreading house to house and becoming vastly harder to stop.

Utility mitigation offers unusually concentrated loss reduction: utilities cause about 11% of ignitions but 50% of damage. Vegetation management, sensitive breakers, public-safety power shutoffs and undergrounding can help, though undergrounding costs $3–4 million per mile. The constraint is not utilities’ aversion to CapEx—they can mark spending up roughly 15%—but regulators and ratepayer advocates resisting already-high bills.

California’s insurance retreat follows regulated prices that stopped reflecting risk, not simply insurer unwillingness. Until recently, admitted-market rates could not include reinsurance costs—which doubled or tripled over five to ten years—or forward-looking catastrophe models, relying instead on 30 years of history. Clerico’s diagnosis: the state “choked rates so low that…we broke the system,” pushing policies toward the FAIR Plan and non-admitted carriers such as Stand, one of his companies.

Drones could become a meaningful wildfire-response category over the next five to ten years as payload, swarming and flight regulations improve. Clerico expects pre-positioned systems to attack camera- or satellite-detected ignitions before they spread, or carry suppressant over blocked roads for structure defense. The low-tech version already exists: he keeps 20 gallons of Barricade gel and argues WUI homeowners should consider their own applicator, pump and water supply.

The enduring bottleneck is political memory and incentive design, not a shortage of isolated technical fixes. Elad Gil proposes a “Smokey Bear 2.0” that advocates for healthy fire, while Clerico focuses on changing public attitudes toward prescribed burns, faster permits for safer rebuilding and insurance pricing that rewards hardening. Elad also frames wildfire as a neglected national-security exposure, asking what would happen if 30 fires threatened military bases or nuclear power plants and overwhelmed the system.