What Makes an AI Application Worth Building?

The model improves before the product settles

An enterprise team chooses the strongest model it can access, finds a useful application, then faces another model release and starts exploring again. This recurring problem appears early in the Unsupervised Learning × Latent Space crossover: teams still discovering what works may prioritize capability over the attractions of open models.

The opening retrospective contrasts optimism about reasoning models with disappointment in enterprise open-model adoption. One participant sees reasoning models as a rapid reversal of the pessimism surrounding Ilya’s NeurIPS talk. Another questions that tidy narrative, recalling that Strawberry had been in development for roughly two years. The shift toward inference-time scaling—spending more computation when using a model—had a longer development history than the public mood suggested.

On adoption, a participant recalls Ankur from Braintrust estimating that open models accounted for roughly 5% of enterprise model usage and that their share was declining. This was a secondhand estimate without a specified sample or measurement period. Yet DeepSeek’s rapid progress suggested a separate competitive change: the window in which a model provider alone could build products around a valuable capability was shrinking.

Another participant distinguishes DeepSeek’s achievement from progress by a unified open-source movement. Individual companies decide what to release, and building from DeepSeek through distillation is a lower bar than independently training a competitive model from scratch. With Llama 4 still awaited, the extent of that independent competition remained uncertain. Even a participant who viewed DeepSeek mainly as strong execution of existing ideas singled out R1’s full reasoning traces as a distinctive contribution.

Finance contacts asking why they had not been alerted to DeepSeek expose a gap between technical attention and market attention. One participant accepts responsibility: covering a development is insufficient if listeners miss why it changes the stakes. The conversation then turns to AI engineering above the model layer, work that augments model capabilities and turns them into useful products. The startup advice is direct: establish that people want the product before making differentiation the primary concern.

Why new builders surprised the incumbents

Bolt and Lovable make the product opportunity concrete. Why did established low-code products such as Zapier, Airtable, Retool and Notion—with users and distribution already in place—not capture it? One proposed explanation is attachment to existing product boundaries. Zapier used natural language to create Zapier actions; Notion used AI to write documents and fill tables. Both improved familiar workflows, while new entrants explored what users could build from scratch.

Timing offers another explanation: perhaps the new builders began just as models became capable enough to support their products. The discussion leaves both possibilities open—freedom from an inherited product design and a favorable capability threshold.

Jordan’s disappointment is Apple Intelligence. He expected Apple’s position in personal assistants to produce stronger execution, and cites a BBC-related summary that falsely reported a man had shot himself. The failure grounds the discussion in reliability: distribution and a promising interface still need trustworthy output.

Stable foundations: protocols, privacy and memory

Jordan considers agent frameworks overhyped because the workloads they organize keep changing. He praises LangChain’s release velocity and willingness to revise abstractions, but wants foundations developers can keep for years. His analogy is a jQuery era awaiting a React-like shift; he sees Mastra, on the TypeScript side, moving in a promising direction.

Another participant favors standardizing protocols before frameworks, naming MCP as the leading example. The comparison shifts to XMLHttpRequest enabling Ajax: agreement on a common mechanism could support experimentation while broader application patterns remain unsettled. The discussion makes this architectural argument without explaining MCP’s protocol details.

Jordan is more optimistic about Apple’s Private Cloud Compute, whose promise he describes as extending on-device security properties to cloud computation when larger models are needed. Another participant argues that GPU scarcity pushes even enterprises accustomed to private deployments toward shared infrastructure. That creates demand for the protections of a dedicated environment while multiple customers share the underlying resources.

Swyx identifies persistent memory as another missing foundation. He wants stored facts about a user, organized as a knowledge graph, that remain available beyond the model’s current context window. In his assessment, the OpenAI and Cloudflare agent frameworks discussed lacked this kind of memory. He recommends the memory server among MCP’s initial servers as a starting point and argues that better abstractions could help agents learn on the job.

The pushback is that memory may be difficult rather than neglected. Swyx acknowledges efforts such as LangMem, while maintaining that persistent memory should become a normal part of the agent stack. He also sees an investment case: stateful AI has business characteristics resembling databases.

The case for specialized model companies

Several participants doubt the opportunity for another general-purpose model company seeking broad benchmark leadership. A stronger case is a company clearly best at a particular application. Robotics, biology and materials science offer possibilities because they require distinctive data, including data generated through wet labs and physical experiments.

That raises a question for existing labs: why expand into deep research and Operator rather than supply specialized models to other businesses? One participant suspects high costs create pressure to monetize quickly. A finance model could enable customers; search and research applications can make those customers competitors. But the technical case remains unsettled: specialized models might offer lower cost and latency without exceeding the quality of improving general-purpose models.

BloombergGPT illustrates how useful assets can outlast a model effort. A participant reports that the team concluded closed models were better and stopped the model project, while retaining the team, collected data and data pipeline. No benchmark or precise competing model is identified. The lasting value in this account is the capacity to work with domain data and implement whichever models prove useful.

Coding agents compete on more than coding

Anthropic and Cursor bring the provider–customer tension into focus. One participant expects Anthropic to expand into coding applications and Cursor to diversify its model dependencies. These are expectations at the time of the conversation. The competitive question is whether a superior model can launch a strong application from a standing start—or whether developers will prefer a better product even with a slightly weaker model.

Participants disagree over whether Devin is more exposed than Cursor. One imagines agents handing work to an integrated development environment when deeper interaction is needed; another challenges the evidence for that division, while a third expects the capabilities to converge. They distinguish two places an agent can operate: - Inner-loop coding: inside the development environment, within the work leading to a Git commit. - Outer-loop coding: in the cloud, between commits.

Outer-loop agents require codebase integration, security agreements, infrastructure and commercial support. One proposed defense for Cognition is its willingness to do that work, which the speaker believes model providers may prefer to avoid. Sourcegraph is proposed as another contender because existing enterprise contracts and trusted access to codebases could help it compete. The advantage depends on how deeply these products become embedded—and whether competitors undertake the same work.

What works now, and what merits an early bet?

A seed investor and a market analyst approach product–market fit differently. The investor seeks clear demand with unresolved technical feasibility: customers would want the product if it worked. Already-proven categories can be too late, while pursuing an underappreciated idea can signal founder conviction. The analyst needs to explain what works today and judges that only a few application forms qualify.

Deep research is proposed as one such form: an agent takes on more of the reporting task and returns a more reasoned report. Brightwave supplies the finance-specific example, alongside Gemini and Perplexity research products. This judgment of product–market fit rests on the participant’s assessment, without adoption or revenue measurements.

Pricing produces disagreement. One participant speculates that upgrades from $20 to $200 subscriptions generated billions in annualized revenue for OpenAI, without supplying subscriber counts or revenue figures and while questioning retention. Another interprets rapid expansion to the $20 tier as possible competitive pressure or a limit to premium pricing. The alternative explanation is deliberate distribution: offer smaller versions of premium capabilities at lower tiers to increase exposure and become the default choice.

Gemini provides a personal example of changing adoption. One participant reports increasing use for YouTube summaries, native image generation and Flash Thinking. They also describe manually comparing frontier models for the daily AI News recap, with Flash winning most days. That is a recurring workflow judgment, without specified model versions or a scoring rubric.

For developers, the obstacle is integration: fragmentation across Google Cloud, Vertex and AI Studio, plus the reported inability to use Google’s thinking models in Cursor at the time. Impressive capability can still lose everyday use through access friction. Two participants nevertheless favor Google over Apple in the public-company discussion; they express a directional preference without establishing that they hold those financial positions.

Compare the agent with the actual baseline

Customer support, with Decagon and Sierra as examples, is another proposed early market for agents. One participant suggests scanning a large company’s income statement or organization chart for knowledge-work functions that agents could progressively take over. The economic opportunity may broaden as applications move from cutting costs toward generating growth.

The distinction is willingness to pay for quality. Buyers who already outsource work such as call-center operations may accept some performance tradeoffs for lower costs, encouraging fierce price competition. A tool with a clear connection to incremental revenue might command a premium for better results. This is a hypothesis about future application economics, not an established split between markets.

Voice scheduling makes the baseline concrete. A participant reports that home-services businesses they examined missed 50% of calls, then proposes an agent effective 75% of the time. The argument is that imperfect automation can improve on unanswered calls. The first percentage is an observation with no specified sample; the second is hypothetical, with no defined success measure or measured revenue uplift. A separate anecdote describes a bootstrapped team focused exclusively on veterinary appointment scheduling, usually attending veterinary conferences to meet its customers.

Another participant challenges the revenue inference: an electrician receiving twice as many calls may lack capacity to perform twice as much work. Better intake can move the bottleneck into staffing or training, creating possible follow-on opportunities. The practical lesson is to check the next capacity constraint before equating more successful interactions with more revenue.

Education tests the limits of model progress

Beyond coding, support and research, the proposed opportunities include assistants that watch a shared screen and offer help, outbound sales, recruiting, personalized teaching, finance and personal AI. Outbound sales is described as approaching product–market fit; personal AI is considered harder to monetize. Voice and conversation summarization earns a place among useful applications even though it involves less autonomous action.

Education sharpens the question of whether progress requires better models or more time to build with existing ones. One position emphasizes institutional and human barriers. Another points to private education and spending outside school as opportunities distinct from public education.

A participant recounts trying to make Singapore’s education system more AI-native, with the ambition of making Python a national language. The reported Ministry of Education response was that teachers would be unprepared. Speak’s English teaching in Korea offers a more focused product approach: choose a specific learning problem and execute it well before attempting to address education broadly.

Defensibility accumulates around the application

One proposed source of defensibility is network effects: interactions among users and other participants can strengthen a product beyond its single-user experience. Chai Research supplies the example. Contributors submit models, while the application mediates between users and model providers and handles the flow of money. The argument is that this network could remain valuable as individual models change, even if lacking proprietary model technology is a short-term disadvantage.

Enterprise network effects may be harder to build. Brand, product breadth and development speed offer other advantages: becoming associated with a category can open customer conversations competitors struggle to enter. Another participant rejects proprietary training or a unique dataset as the default application defense, emphasizing accumulated improvements in usability and functionality.

Model releases add recurring urgency to this familiar software-business argument. A participant describes releases every three to six months as potentially existential events: if one company fails to discover what a new model enables, a competitor may get there first. Adaptation speed becomes part of the moat, alongside the improvements already embedded in the product.

Infrastructure must earn its place

The infrastructure thesis centers on code execution, memory and search around models. A participant discussing execution discloses an investment in E2B. Cybersecurity offers another proposed opportunity: defensive uses of AI could follow offensive uses into email security and identity, while making security testing more useful. Binary inspection illustrates the intended change—from matching instructions or patterns toward inferring what code is trying to accomplish overall.

Platform competition is close by. The custom GPT builder’s capability checkboxes are proposed as a map of startup opportunities that can also become platform APIs. Independent search providers might retain demand from customers comparing providers and prices, while bundled search appeals to existing OpenAI customers. A participant predicts an OpenAI code-execution API, extending the question of whether a lab earns more by selling infrastructure or packaging its capabilities inside ChatGPT.

Reservations about model serving concern capital efficiency. Together AI and Fireworks are cited as businesses doing well; success and capital requirements are separate questions. The broader argument favors applications because they can charge for delivered utility, whereas much infrastructure tends toward pricing based on cost plus a margin.

Standalone fine-tuning companies raise a question of business scope. Writer AI is offered as a broader enterprise offering in which fine-tuning can be one component. Real-time voice infrastructure attracts interest but also uncertainty about eventual market size. The participants advocate holding such investment opinions loosely and learning from public mistakes as capabilities change.

AI for operational reliability produces a sharper disagreement. The skeptic argues that log monitoring and anomaly detection are established capabilities that do not inherently require language models. The counterargument asks whether AI can shorten incident resolution before fully autonomous operations become possible. A hypothetical 10% reduction is offered as potentially valuable. The evaluation target becomes incremental improvement in recovery time, with full autonomy a separate, longer-term possibility.

Verification, compute and delegated identity

The closing technical questions begin with reinforcement learning beyond verifiable tasks. A participant sees success where outcomes can be checked, but asks whether the approach can extend to contracts, legal documents, marketing and sales. An outbound-sales simulator is proposed as a way to explore conversational outcomes, not as a demonstrated solution.

If that extension proves difficult, the imagined future is uneven: autonomous agents in verifiable domains, copilots elsewhere, with humans still supplying judgment and taste. One participant imagines autonomous coding alongside basic sales emails that still need human help. The ability to evaluate success could matter more than how intellectually impressive a task appears.

Reliability also raises a compute constraint. A participant recalls Bob McGrew describing a rule of nines: moving from 90% to 99% reliability, then to 99.9%, requires an order-of-magnitude compute increase at each step. With no task, reliability measure or compute setting specified, this remains a recalled heuristic. It motivates questions about hardware supply and whether alternatives such as AWS Trainium can challenge an advantage the speaker associates with NVIDIA’s CUDA ecosystem.

The case for specialized silicon rests on workload stability. A participant argues that stable transformer operations could justify embedding them in dedicated chips. They expect state-space models to appear alongside transformers, preserving much of that opportunity. The argument identifies a condition under which specialization makes sense, without identifying a likely winning company.

The final infrastructure requirement is agent authentication. When an agent visits a website on a person’s behalf, the site needs to distinguish the person from their delegate. One participant calls for a shared sign-in mechanism for agents while questioning whether every other software layer needs reinvention. Suggestions involving cryptocurrency and iris scans remain speculative.

The people and communities behind the conversation

The quick-fire closing names John Carmack as a dream guest. Another choice is a supporter identified only as Andre, credited with helping the AI engineer movement gain attention by predicting that AI engineers would outnumber machine-learning engineers. Doomers, a play about OpenAI’s board drama, prompts a different reflection: one participant considers it largely fictionalized and cautions that oral histories can also distort complex events.

News recommendations include swyx’s Twitter and Shawn’s link sharing in the Latent Space Discord, described as having roughly 40–50 topic channels. That community was also Latent Space’s starting point. Originally called Dev/Invest, it grew from developer-tools investing conversations during COVID. The podcast idea emerged during an OpenAI visit in October 2022, and OpenAI became its first guest.

One participant values in-person conversations in San Francisco while recognizing the risk of being used to circulate false information. Independent judgment remains necessary. Private and public discussion can overlap surprisingly closely, they say, though some private information cannot be repeated until it becomes public.

Latent Space closes with its combination of newsletter, community and podcast, then emphasizing YouTube growth and better video production. Conferences add another way to draw out insights: some people open up in a long conversation, while others respond to a prestigious stage. The organizer points ahead to the June AI Engineering World’s Fair, predicting it would be the largest technical AI conference.

Listeners are invited to follow Latent Space’s newsletter and both shows on YouTube. Jacob closes by describing Unsupervised Learning, Redpoint Ventures’ AI podcast, as an examination of what works today, what may work next and what those changes mean for businesses.