Summary

The episode’s strongest investor call is that AI value is shifting above the model layer, where applications can charge for utility instead of cost-plus infrastructure. swyx says the industry went from mocking GPT wrappers to treating wrappers as “the only thing that’s interesting,” with coding, support, and deep research as leading proven forms; conversation summarization is a strong use case but not necessarily an agent. Product experience, distribution, network effects, and execution speed now matter more than owning a bespoke model.
Reasoning models reopened scaling just as pre-training seemed tapped out, but DeepSeek showed how quickly proprietary model advantages can compress. swyx argued that R1’s full reasoning traces were net-unique in open source, yet the panel separates DeepSeek’s achievement from a durable “team open source”: enterprise open-model usage was estimated at roughly 5% and going down, while one DeepSeek-driven narrative coincided with about a 15% one-day decline in NVIDIA. “There’s just different companies, and they choose to open source or not.”
Agent frameworks look overbuilt for workloads still in flux; protocols and state are the more durable bets. The panel calls today the “jQuery era” and argues MCP may matter more as an XMLHttpRequest-like protocol than another all-encompassing framework. Memory is the underappreciated layer: without persistent knowledge beyond context windows, agents are less able to become smarter or learn on the job.
Current product-market fit is concentrated in coding, customer support, and deep research, with the next wave aimed at revenue creation. swyx estimates OpenAI’s $20-to-$200 deep-research upgrade could represent billions in ARR; when Jacob asked whether access had quickly dropped to $20, swyx clarified that access had been expanded. Voice products can create value under imperfect reliability: if home-service businesses miss 50% of calls, an AI effective 75% of the time can still create substantial revenue.
Coding is becoming the first full-stack battle among model vendors, IDEs, and autonomous agents. Cursor’s roughly $9 billion-$10 billion valuation makes acquisition harder, while Anthropic, Cognition, and others may converge across “inner-loop” IDE work and “outer-loop” cloud agents. The remaining moat is operational schlep—codebase integrations, security approvals, infrastructure, and enterprise trust—not merely model quality.
App defensibility is conventional but compounding: network effects, brand, integrations, and relentless velocity beat “train your own model” stories. Chai’s marketplace between users and model providers is swyx’s network-effect example; elsewhere, category leadership has reportedly doubled ACVs even where investors expected pricing compression. Jacob’s framing is “the thousand small things” that make a product delightful and let it exploit every new model within three to six months.
The largest technical fork is whether reinforcement learning can move beyond verifiable domains. If RL works for coding and math but not law, sales, or marketing, the result could be fully autonomous technical agents alongside human-dependent creative copilots: “to write the most basic sales email,” a person may still be the tastemaker. Bob McGrew’s reported reliability rule also leaves an infrastructure question: each move from 90% to 99% to 99.9% was described as requiring another order of magnitude of compute.
The explicit public-market pair was long Google and short Apple, although Jordan sees Apple’s Private Cloud Compute as an important exception. Google’s models are gaining practical usage—Flash reportedly wins most days in swyx’s daily frontier-model bake-off—but fragmented developer surfaces still impede adoption. Apple Intelligence disappointed badly, while PCC may help bring on-device-style privacy to cloud workloads; Jacob argues that AI will need single-tenant guarantees in multitenant environments.

Deep dive

1. Reasoning rescued scaling, but DeepSeek compressed the moat

The opening surprise was the abrupt transition from Ilya Sutskever’s NeurIPS-era “scaling is dead” mood to reasoning models: “It’s so over” became “we’re so back” almost immediately, with inference-time compute appearing to replace pre-training as the scaling frontier.
swyx found the timing “suspiciously neat.” Strawberry had reportedly been underway for roughly two years, and Noam Brown’s arrival at OpenAI was presented as evidence of a major strategic bet: the work was ready just as pre-training appeared tapped out. “If there were an Illuminati, this would be what they planned.”
Open models present a paradox. Ankur from Braintrust estimated enterprise usage at about 5% and declining. Jacob’s explanation was that enterprises remain in use-case-discovery mode and keep choosing the strongest available model; swyx said DeepSeek showed him how short a model company’s exclusive product-building window could be.
The disagreement on DeepSeek is worth preserving: Alessio called replication “an order of magnitude cheaper” than invention and therefore overhyped, while swyx argued R1’s full reasoning traces were net-unique in open source. His harder correction: “There’s no team open source,” especially if DeepSeek stops releasing models and everyone else merely distills from it.

2. Legacy product surfaces caused incumbents to miss the AI-native turn

swyx’s surprise was that Zapier, Airtable, Retool, and Notion failed to capture AI software creation while Bolt and Lovable did. The incumbents had distribution and low-code DNA, but used natural language to improve existing actions, documents, and tables instead of rebuilding software creation “from whole cloth.”
Timing may have mattered as much as imagination: Bolt and Lovable reportedly went from $0 to $20 million in three months as models crossed a usability threshold. Their numbers transformed an architectural curiosity into a market investors could no longer ignore.
Jordan’s negative incumbent case was Apple Intelligence. Apple appeared ideally positioned to build a personal assistant, yet produced viral-but-bad message summaries and a BBC-linked false summary saying a man had shot himself when he had not: the consumer product fell far below the panel’s expectations.
His Apple exception was Private Cloud Compute, which carries on-device-style privacy into larger cloud workloads. Jacob’s mechanism: GPUs make dedicated VPC deployments impractical, so AI needs “single-tenant guarantees in multi-tenant environments”—an architectural opening as consumer AI handles more sensitive data.

3. Protocols and memory look more durable than agent frameworks

The framework critique is not that LangChain lacked velocity—it shipped relentlessly as abstractions changed—but that workloads remain too unstable for developers to choose a framework they can trust for years. The panel’s analogy: agents are still in a “jQuery era,” while the market keeps building more jQuery before discovering its React.
swyx’s modification was that even React may be premature: “The thing is the protocol and not the framework.” MCP resembles XMLHttpRequest, the enabling primitive behind AJAX, because a protocol can survive shifting implementations without freezing today’s agent assumptions into a monolith.
Memory is the missing agent primitive. OpenAI’s Agents SDK and Cloudflare’s agent work omitted persistent “memory memory”—facts, preferences, and knowledge extending beyond a context window—even though MCP’s initial servers included an implementation swyx recommends as a starting point.
Alessio pushed back that memory may be difficult rather than ignored, citing active work around LangMem, Letta, and Zep. swyx acknowledged the difficulty and cited those efforts, but kept the investment call: stateful abstractions should become standard, and “anything stateful should be interesting to VCs” because the business model resembles databases.

4. General models are squeezing new labs toward proprietary data

Jacob remains surprised by the number of new companies training broad models; swyx had thought that category died at the end of the prior year. The basic objection is strategic: every entrant wants AGI and benchmark leadership, but “they can’t all do it,” and no missing general capability clearly demands another lab.
Alessio’s reading of Noam Shazeer’s answer was that test-time compute may eventually let models perform AI engineering and discover the next algorithmic breakthrough themselves. That leaves little room for undifferentiated challengers unless they possess a genuinely different method.
The plausible exception is unique data generation: robotics, biology, and materials companies can gather specialized data through physical systems, wet labs, and experiments. Even there, the unresolved “bigger is better” question persists—smaller vertical models may be cheaper, while the next general model may simply outperform them.
BloombergGPT supplied the cautionary example. Bloomberg concluded closed models were better and stopped the model effort, but retained the valuable assets around it: its data pipeline, assembled team, fine-tuning capability, and roughly 12 or 13 teams applying generative AI across the company. “Everything but the model survived.”

5. Coding is becoming the first full-stack model-versus-app battle

Model companies are moving into products partly because expensive research creates near-term monetization pressure, but every move creates a new enemy. Search threatens Perplexity; deep research crowds application vendors; coding products raise the question of whether Anthropic remains Cursor’s supplier or becomes its competitor.
Cursor’s roughly $9 billion-$10 billion valuation has made it “hard to acquire,” in swyx’s view; at $5 billion-$6 billion, joining OpenAI and contributing coding data might have been plausible. Now it likely must remain independent while diversifying away from any single model provider.
Alessio argued Devin may be more exposed than Cursor because Anthropic could pursue autonomous agents without building an IDE. swyx’s pushback was Claude Code, while Jacob expected both sides to converge: IDEs will add agents, and agents will expose interactive coding environments.
The moat differs by loop. Inner-loop tools operate inside the IDE and a Git commit; outer-loop agents work between commits and must integrate codebases, satisfy security requirements, and endure enterprise “schlep.” That gives Cognition a possible moat—and perhaps favors Sourcegraph, whose decade of contracts and codebase trust is difficult for two-year-old startups to reproduce.

6. Product-market fit is real, narrow, and expanding

swyx originally set a $100 million bar to rebut claims that generative AI products were unreliable toys. His examples were Copilot, Jasper’s writing category—with the caveat “Jasper—no longer”—and Cursor as a coding agent.
Deep research is the newest validated form: long-running agents produce reasoned reports across OpenAI, Gemini, Grok, Perplexity, and vertical products such as Brightwave for finance. swyx estimates OpenAI’s $20-to-$200 upgrade could mean billions in ARR; when Jacob asked whether access had been dropped to the $20 tier, swyx clarified that access had been expanded.
Customer support joins coding and deep research as a “killer agent” category. Alessio reads Brett Taylor’s choice to build Sierra—despite his enthusiasm for developer tools and ability to fund almost anything—as strong evidence that support contains durable budgets, alongside Decagon and similar entrants.
Google is becoming the usage counterweight. Gemini handles swyx’s YouTube summaries, native image generation, and thinking workloads; Flash “wins most days” in his manually run daily model bake-off. Google Cloud, Vertex, and AI Studio fragmentation still hurts developers, but both swyx and Jacob chose the same trade: long Google, short Apple.

7. The next application wave must move from savings to revenue

The first wave sold cost reduction into work companies already outsourced to BPOs, especially support. Alessio sees the narrative shifting with Jensen Huang’s GTC wording: last year, “the more you buy, the more you save”; this year, “the more you buy, the more you make.”
Jacob’s concern is that outsourced cost centers may also face the fiercest price competition because buyers already accept lower performance for savings. Revenue-producing tools could prove more defensible: a go-to-market product that reliably creates incremental sales can charge more because its ROI is visible.
Voice AI illustrates why imperfect systems can already expand revenue. Home-service businesses such as electricians reportedly miss 50% of incoming calls; an agent effective only 75% of the time still captures substantial demand. The next bottleneck then appears: can the electrician hire, train, and deploy enough people to fulfill twice as much work?
The proposed next forms include screen-sharing assistants, outbound sales, hiring, finance, personal AI, and focused education products. swyx’s education lesson from trying to make Python a Singaporean national language was institutional resistance—teachers were unprepared—so he favors narrow products such as Speak over attempts to rebuild education all at once.

8. Durable app moats are networks, brands, and compounding execution

swyx’s clearest answer on app defensibility was “network effects.” Chai, a Character.AI competitor, lacks a proprietary-model advantage but connects users with outside model providers; that marketplace may hurt near-term differentiation while making the company robust to future model changes.
Brand is already producing unexpected economics in enterprise categories. Where investment committees predicted commoditization, pricing compression, and shrinking ACVs, Jordan says some ACVs have doubled because customers pay a premium for the category-defining company already invited into every buying process.
Jacob called unique datasets and self-trained models an early “head fake.” The real defense sounds less magical: “the thousand small things” behind delightful UX, broad product coverage, and organizational velocity. Every model release creates an existential three-to-six-month race to exploit new capability before a competitor does.

9. Agent infrastructure matters, but applications capture the richer margin

The attractive “LLM OS” layer sits around models: E2B-style code execution, memory, search, security, and other tools agents require. It is distinct from bare-metal serving; the application layer has nevertheless been “way more interesting” because vendors charge for customer utility while infrastructure often collapses toward cost plus.
Alessio’s security thesis is symmetrical: wherever AI improves offense, defense must respond in email, identity, red teaming, and binary inspection. Models add semantic understanding to historically syntactic rules—reasoning about what code is trying to accomplish, not merely matching isolated instructions.
Research labs will absorb some infrastructure. swyx’s heuristic is that every checkbox in ChatGPT’s custom-GPT builder represents a startup, and those checkboxes are becoming APIs; OpenAI Search was compared with Axon at roughly five times the price for the same amount of research, leaving room for multi-provider strategies.
swyx struggles with standalone fine-tuning, AI DevOps/SRE, and real-time voice infrastructure as venture-scale categories. Jordan disputed the SRE skepticism: full autonomy is distant, but even a 10% reduction in mean time to recovery can support meaningful revenue. swyx acknowledged the incremental case while maintaining that basic anomaly detection predates LLMs.

10. Reward, reliability, silicon, and identity remain unresolved

swyx’s largest question is whether RL can succeed outside verifiable domains. Coding and math have checkable rewards; law, contracts, marketing, and sales require taste. Failure would divide the market between autonomous agents in verifiable work and copilots elsewhere, producing the odd world of automated software but human-written sales emails.
Bob McGrew’s reported “rule of nines” makes reliability a compute problem: moving from 90% to 99%, then 99% to 99.9%, costs an additional order of magnitude each time, with the progression described as taking roughly two to three years. That raises both hardware availability and NVIDIA’s continued dominance as ecosystem-level questions.
swyx sees a theoretical opening for transformer-specific silicon because GPUs remain general enough for gaming, crypto, and AI. The bet requires stable workloads—and therefore confidence that transformers persist—but post-2019 or 2020 entrants can specialize more aggressively than chip companies designed before transformers had clearly won.
The emergent protocol problem is agent authentication: when Operator or another agent visits a site, it must indicate that it is an agent acting for a user rather than the user themself. swyx calls this a necessary “new SSO effectively for agents,” then allows the uncomfortable possibility that Sam Altman’s answer was early: “Maybe you have to scan your eyeballs.”

Shortwave Rides the Tidal Wave: Inbox Agents, Hyper-Growth & Hiring AI Managers, with CEO Andrew Lee The AI Cold War, Signalgate, CoreWeave IPO, Tariff Endgames, El Salvador Deportations