Latent Space in 3 minutes

Here is The Daily FM summary of the Latent Space that aired on Tuesday November 10th. This episode featured Zeno Rocha, founder of Resend, on building an email platform for developers, competing with legacy giants, and designing software for an increasingly agent-driven world. [1]

Rocha’s original insight was simple: email was a huge, crowded market, but existing services such as SendGrid, Mailgun, and Mailchimp did not seem to care deeply about developer experience. As a frontend engineer with no infrastructure background, he set out to build what he called “the Stripe of email.” Resend has since grown to 4.3 million users and 120,000 paying customers, including Warner Brothers and HBO, and says its SDK now has more than twice the npm downloads of SendGrid’s. [2]

The conversation emphasized that sending email reliably is far harder than it looks. Delivery is governed by Gmail, Outlook, Yahoo, domain reputation, IP history, bounce rates, complaint rates, and patterns that can resemble phishing. Rocha explained that bulk senders cannot simply blast 100,000 messages at once. They need to pace delivery, start with highly engaged recipients, and let mailbox providers see healthy engagement before increasing volume. Resend’s guiding principle is to make customer behavior look unlike an abuser’s behavior.

That abuse problem produces an interesting product philosophy. Rather than forcing every new user through manual approval, Resend begins by trusting users, then rapidly detects and blocks suspicious activity. Rocha described a “Japanese gang” that attempted to onboard roughly 20,000 newly registered domains, illustrating the scale of malicious activity email providers face. He also argued that platforms should bake deliverability best practices directly into the product, including automatically generating plain-text versions of HTML emails, which improved delivery rates. [3]

A recurring theme was removing friction. DNS configuration remains the biggest support problem, especially for new AI-native builders who may not know what CNAME, TXT, or A records mean. Rocha praised Domain Connect, essentially OAuth for DNS, because it lets users connect domains by clicking through an authorization flow rather than manually copying fragile records. Resend also scaled support by requiring agents to answer tickets with documentation links, forcing the company to turn repeat problems into reusable knowledge. [4]

On AI, Rocha argued that products must become “formless”: available through dashboards, APIs, SDKs, CLIs, and MCP servers, depending on how customers want to work. Resend’s MCP usage has been doubling monthly, reaching more than a million calls. But building a robust remote MCP required the company to revisit permissions, OAuth, role-based access, observability, and distribution through official marketplaces. He believes officially maintained integrations increasingly determine buying decisions. [5]

The hosts also discussed AI search optimization. Resend has invested in clear documentation, FAQ-heavy pages, markdown versions of marketing pages, machine-readable pricing, and content that directly answers user questions. Rocha’s practical lesson was that JavaScript widgets may confuse agents, so important information must also exist in plain text.

Finally, the episode turned personal. Rocha rejected the idea that founder ambition requires sacrificing family. As Resend approaches $10 million in ARR, he said he would rather build a company of that size while remaining close to his daughter than reach $100 million while losing that relationship. His view was that taste, care, and deliberately chosen priorities are not distractions from company-building; they are part of the product and life worth building. Thank you for listening to Latent Space in 3 minutes from The Daily FM. See you next time! [6]

Here is The Daily FM summary of the Latent Space that aired on Monday November 9th. This episode went deep on what frontier AI labs are actually buying from data companies, with Handshake AI founder Garrett Lord and team members explaining why “data labeling” is now an inadequate term. The work has shifted from low-cost workers drawing boxes around images to domain experts—bankers, scientists, lawyers, clinicians, and safety specialists—creating realistic environments, tasks, and grading systems that teach models to operate in the real world. [1]

The core argument was that frontier models increasingly need complex, long-horizon training tasks rather than generic internet text. To train an AI investment banker, for example, Handshake must simulate an entire workplace: email, spreadsheets, documents, data rooms, calendars, financial databases, and realistic colleagues. Experts then create assignments and detailed rubrics defining what good work looks like. One usable training example can take days to build, audit, and protect against reward hacking.

Handshake’s Nishant explained that labs cannot just judge data by whether it looks plausible. Bad or overly templated data can produce good-looking training curves but lead to models that reason too long, write unnaturally, or fail in real work. Labs test data through ablations, model rollouts, diversity analysis, and post-training open models to see whether a dataset actually improves capability. A major theme was that models also need stylistic consistency: mixing many incompatible human and synthetic styles can cause strange behavior.

The life-sciences section introduced Vials, a benchmark of basic scientific visual tasks such as reading molecular diagrams, plasmid maps, protein gels, and pathology slides. The surprising point was that these are often not PhD-level conceptual challenges. A trained lab technician may answer in under a minute. But frontier multimodal models frequently score below 50 percent because they miss visual details, fail to translate biological terminology into the image, or lack the tacit judgment scientists use to distinguish a real protein band from noise. In one example, a model confidently selected the wrong drug concentration because it failed to recognize a faint but meaningful band. The guests stressed that AI will not cure cancer before it can reliably interpret these foundational artifacts, and in high-stakes science, 90 percent accuracy may still be unusable.

Finance showed a parallel problem. Handshake’s Atlas benchmark puts agents into simulated banking environments with thousands of emails, spreadsheets, and documents. Models can retrieve context quickly and generate polished financial models, but a single hidden accounting error can invalidate everything. One agent built a 2,600-formula leveraged-buyout model that looked professional but mishandled goodwill in a tax calculation, distorting the output by roughly 10 percent. Former banker Colin argued that in finance, “95 percent right is 100 percent wrong,” because checking a plausible but flawed model can take longer than rebuilding it. The episode also highlighted an amusing reward hack: an agent created a spreadsheet tab listing every number from one to a million so a simplistic grader would find the expected values.

Safety leader Steve Kelling described CareBench, a child-safety benchmark covering risks beyond explicit abuse content, including grooming, cyberbullying, emotional reliance, and bypassing age restrictions. Parents supplied important feedback that academic safety frameworks missed. One model correctly refused to help a malicious user coerce a child into self-harm, then unhelpfully suggested ethical ways to build a youth gaming community—an example of why refusal alone is not enough.

The final takeaway was economic: Handshake believes the frontier-data market has grown from about $1 billion three years ago to $10 billion today, and could reach $100 billion within another three years. As AI moves from isolated tasks toward autonomous work, data, evaluations, interpretability, and safety will become inseparable. Thank you for listening to Latent Space in 3 minutes from The Daily FM. See you next time! [2]

Here is The Daily FM summary of the Latent Space that aired on Sunday November 8th. This episode celebrated one million subscribers to AI Engineer while taking a detailed tour of what Stanford’s “Modern Software Developer” course says about the future of software work. Guest Mihail Eric, Stanford adjunct professor and Head of AI at sales-platform startup Monaco, described a curriculum designed for engineers who already have computer-science foundations and now need to operate effectively in an AI-native workplace. [1]

Eric’s central argument was reassuring but demanding: computer science is still worth studying, but the floor for junior engineers has risen sharply. Graduates cannot expect to enter companies as narrowly scoped beginners who only implement tickets. They need to understand systems, architecture, tradeoffs, verification, and how to guide increasingly autonomous coding agents. The valuable engineer is less the person typing every line and more the person who can supply intent, context, constraints, and judgment.

The course has changed significantly in just a year. Last time, it surveyed the many places agents could affect the software-development lifecycle. This time, Eric sees clearer principles emerging. The first is context engineering: a one-line prompt may work for a toy project, but serious work in a million-line codebase requires structured research, proposals, plans, and testing. Humans should intervene at key decisions, choosing among possible approaches before agents implement them.

Students begin by building a basic coding-agent harness in roughly 200 lines of code, then inspect real Claude Code requests using a proxy. The notable lesson is that agent systems are less magical than they appear: at their core, they are loops around a model and tools, though production harnesses add elaborate prompts, context management, subagents, and safeguards. Eric noted that Claude Code’s system prompt has repeatedly grown and shrunk as model capabilities improve, illustrating a recurring question: what belongs in the harness, and what should be learned by the model itself?

The class then moves through agent skills, configuration files, MCP tools, agent-ready repositories, code review, security, background agents, team infrastructure, and software factories. Eric argues that agent readiness is foundational: repositories need clear boundaries, documentation, tests, and verifiability before teams can safely unleash parallel agents. Otherwise, generating code becomes easy while reviewing it becomes the new bottleneck.

That review bottleneck was a major theme. With agents capable of producing dozens of pull requests a day, traditional line-by-line human review cannot scale. The proposed answer is not blind automation, but better architecture and validation: automatically approve only well-bounded, highly verifiable work, while reserving human attention for consequential changes and intent-level decisions.

Students will also contribute to real open-source projects rather than merely one-shot classroom assignments. Eric sees this as an answer to fears that junior developers are being left behind: maintainers can review their work, teach mental models for large systems, and perhaps recruit the strongest contributors. [2]

Security got a blunt reminder as well. AI-generated code can increase output faster than understanding, making dependency management, authentication, billing, data storage, and access controls especially sensitive. The hosts warned that models may produce fewer ordinary bugs on average, but humans can still lose awareness of what they are shipping.

The conversation ended with an unsettled vision of software factories: systems where agents build, test, review, and operate software continuously. Eric believes humans remain most essential in setting intent and judging outcomes. His practical advice for students and working engineers was simple: experiment every day, build ambitious projects, and learn to make benchmarks and evaluations. In this fast-moving field, taste and rigor matter more than treating any tool or guru as gospel.

Thank you for listening to Latent Space in 3 minutes from The Daily FM. See you next time!

Here is The Daily FM summary of the Latent Space that aired on Friday November 6th. This episode was a deep financial and strategic reading of Anthropic’s S-1 filing with Jamin Ball of Altimeter, who argued that the document offers the clearest public view yet into the economics of a frontier AI lab.

The central surprise, Ball said, is that Anthropic appears far more financially efficient than many observers expected. The company reportedly showed roughly 85 percent gross margins, positive GAAP operating income, and more than 20 percent free-cash-flow margins in its latest quarter. Its most striking statistic: Anthropic’s revenue rose about 14 times year over year while spending on inference, or serving customers’ models, grew only 2.3 times. Ball interprets that as roughly a sevenfold gain in serving efficiency. [1]

That matters because the popular bear case says AI labs are spending unprecedented sums on data centers without a viable return. Ball’s counterargument is that Anthropic has demonstrated that frontier-model demand can be profitable. He expects margins may fall again as the company deliberately reinvests in more compute, but says that would be a choice to fund expansion, not evidence that the core business cannot work. [2]

The hosts explained why Anthropic may be going public despite its ability to raise huge private rounds. An IPO provides employee liquidity, acquisition currency, and—most importantly—greater access to debt and credit markets. Building AI infrastructure is capital-intensive, and lenders want detailed financial disclosure before underwriting multibillion-dollar data-center projects. Still, the conversation noted a growing constraint: data-center debt is becoming expensive, with some deals above 10 percent, and lenders are increasingly cautious because so much demand ultimately depends on a small number of customers, especially Anthropic and OpenAI. [3]

A major accounting detour addressed claims that Anthropic has over $100 billion of debt and enormous net losses. Ball argued this is mostly a misunderstanding of convertible securities. Early investments from Google and Amazon were structured partly as convertibles and SAFE-like instruments. As Anthropic’s valuation rose, the accounting value of those future share obligations rose too, creating large liabilities and paper losses under GAAP. In Ball’s view, this is not conventional debt or operating burn; it reflects the appreciation of equity promised years earlier.

The broader business question, he said, is whether Anthropic can stay at the frontier. Eighty percent of its revenue reportedly came from models less than six months old, suggesting customers pay disproportionately for the newest and best systems. That makes model development a treadmill: a lab does not just need one successful model, but a “model factory” capable of repeatedly producing better generations. If it stops, competitors quickly take its revenue.

Anthropic’s focused strategy also stood out. While OpenAI and Google compete across image generation, video, consumer chat, voice, and many other categories, Anthropic has concentrated on enterprise work and coding. The hosts suggested that focus may be an advantage because coding agents are currently among the highest-return uses of AI. But they also expect both major labs to expand into specialized verticals, especially life sciences and drug discovery.

Finally, Ball’s investment thesis was bluntly optimistic: the world will consume more intelligence, power, compute, networking, storage, and data centers. There are no “dark GPUs” sitting unused today. Political resistance, power shortages, land, permitting, memory supply, and financing may slow the buildout—but that may also prevent a reckless oversupply. The episode’s core takeaway was that Anthropic’s filing made frontier AI look less like an unproven cash furnace and more like an extraordinarily fast-growing, economically real business facing massive infrastructure constraints.

Thank you for listening to Latent Space in 3 minutes from The Daily FM. See you next time!

Here is The Daily FM summary of the Latent Space that aired on Friday November 6th. Hosts RJ Hanocki and Brandon Anderson spoke with Fabian Theis of Helmholtz Munich, one of the major figures in computational biology, about the push to build “virtual cells”: models that can predict how cells respond to drugs, gene edits, environmental signals, disease, and other interventions. [1]

Theis argued that virtual cells are moving from an appealing concept toward a practical research direction because biological data generation is finally reaching sufficient scale. Single-cell transcriptomics measures which genes are active in individual cells, producing a high-dimensional picture of cell state. In the past, scientists had enormous numbers of cells but relatively few perturbations to compare. Now, large experiments can test thousands of drugs, CRISPR edits, media conditions, and other inputs across many cell types. That richer “perturbation space” makes it increasingly realistic to model a cell as a system with an input, a current state, and a predicted next state. [2]

A perturbation, Theis stressed, is broader than knocking out a gene. It can include a drug, a signaling molecule from a neighboring cell, a hormone, disease progression, or simply changing the nutrients around a cell. The ultimate ambition is to understand combinations of those forces, not only isolated interventions.

The conversation also moved beyond individual cells to organoids: miniature multicellular systems grown in a dish. Theis described tiny brain-like networks that can perform simple computation and heart organoids that visibly beat. Organoids offer more realistic biology than a single cell line and may let researchers study diseases when human tissue is scarce. But modeling them remains difficult because researchers lack a mature language for spatial neighborhoods and cell interactions. Graph-like models and spatial omics may be important next steps. [3]

On machine learning, Theis offered a useful corrective to foundation-model hype. Simple linear models and logistic regression still perform remarkably well on straightforward tasks such as identifying a cell type or predicting certain gene knockouts. He said researchers must rigorously test strong baselines and choose metrics carefully; poorly designed competitions can reward systems that exploit the score rather than model biology. But for harder problems—such as predicting how neighboring cells influence one another, or generating a response to a complex intervention—nonlinear models and large learned representations do show value. [4]

His lab’s approach separates representation learning from dynamics. Foundation models can learn a useful map of cell states from hundreds of millions of cells and even multiple species. A generative “world model” then predicts how an action pushes a cell through that state space. Theis highlighted flow-matching methods as particularly promising for modeling those changes. [5]

The most ambitious idea was closing the loop with the laboratory. Theis described a new system, Lab Compass, as an inverse generative model: rather than merely predicting what an intervention will do, it suggests what experiment to run to reach a desired cell state. His team has already used iterative experiment-model cycles in blood-cell development research. He sees this as the transition from mapping biology, to predicting biology, to steering biology—eventually perhaps using reinforcement learning with the lab as part of the environment. [6]

On commercialization, Theis was skeptical that a generic “biology foundation model” alone is an obvious business. The more defensible startups will likely combine modeling with exclusive data generation, especially organoid perturbation datasets. His final practical message was that progress requires deep collaboration between computational and wet-lab scientists, not simply better algorithms. [7]

Thank you for listening to Latent Space in 3 minutes from The Daily FM. See you next time!

Here is The Daily FM summary of the Latent Space that aired on Wednesday November 4th. This special episode, recorded around the first AI Engineer conference in Shanghai, was both a look at China’s AI ecosystem and a personal retrospective from Latent Space founder Shawn “swyx” Wang on how he built the podcast, the AI Engineer movement, and a portfolio-style career across media, conferences, advising, and startups. [1]

Wang began by stressing that he did not bring AIE to Shanghai alone: local organizers and partners made it happen. His broader goal is to create more exchange between Chinese and Western AI communities, especially around China’s progress in open models, robotics, developer communities, and startup execution. He praised longtime Chinese fans and translators, particularly Meng Xiao, who has helped make Latent Space accessible to Chinese audiences. [2]

The origin story of Latent Space centered on “learning in public.” Wang contrasted finance, which he called a zero-sum culture of secrecy, with technology, where sharing code, failures, and discoveries can create more value for everyone. After Stable Diffusion’s release, he used a struggling period at a startup as an opportunity to study AI openly through writing. That ultimately led to the podcast. His advice was encouraging but unsentimental: credentials are not the only route into a new field, but hard work alone does not entitle anyone to success. People must work on meaningful problems and get genuinely good. [3]

That learning process also led to the term “AI engineer.” Wang said the point was to recognize software engineers who build practical systems around models—what he now calls harnesses—rather than treating AI as valuable only when it involves elite model research. The role has broadened into areas such as evaluations, agents, memory, multimodality, and AI architecture, but its core has not changed: an AI engineer is the person accountable when an AI-powered system breaks at 4 a.m. He pushed back on AI conferences dominated by executives predicting imminent AGI, arguing that working engineers need more space to share real implementation failures and lessons.

AIE itself is meant to be the flagship gathering for that community, like GDC for game developers or KubeCon for cloud engineers. Wang highlighted that the organization transcribes its talks, profiles speakers, and publishes data and MCP-accessible material so the conference becomes a searchable learning resource rather than a one-time event. Its international expansion follows a franchise-like model: local teams run AIE events while sharing a brand, publishing content, and following a code of conduct. Shanghai is especially complicated because of U.S.-China restrictions, so local ownership matters.

The most provocative discussion concerned Wang’s “Kill My SaaS” experiment. Faced with a proposed $40,000 software subscription, he offered a $10,000 prize to clone it as customizable open-source software. Dozens of submissions arrived, several usable within a week. His argument was that mid-market SaaS companies are vulnerable as coding agents can replicate workflows and let businesses tailor their own tools. He has already built personal alternatives to services resembling Calendly, Dropbox, GitHub, Slack, and DocuSign. Still, he acknowledged self-hosted tools carry maintenance costs; the tradeoff makes most sense where customization beats enterprise-grade reliability.

He also praised TypeSafe’s “Jev” as a different model for AI software: rather than putting one giant chatbot at the edge of an application, use small, fast System 1-style models throughout code paths to make many local decisions reliably. The episode ended with admiration for China’s scale, work ethic, and builder energy, alongside a warning that AI’s biggest risks may go beyond job loss or misinformation toward cyber and biological threats. Wang’s blunt framing: humanity must take seriously the possibility that systems advancing far faster than humans could become dangerously more capable.

Thank you for listening to Latent Space in 3 minutes from The Daily FM. See you next time!

Here is The Daily FM summary of the Latent Space that aired on Tuesday November 3rd. Zach Lloyd of Warp argued that the next major layer of developer infrastructure is the “software factory”: not simply developers using coding agents more often, but a concrete, cloud-based system that defines, runs, measures, and continually improves agentic software development. [1]

Lloyd’s definition is deliberately stricter than a loose collection of automations. A factory operates on a particular product surface, such as a mobile app or backend service, with specified repositories, skills, integrations, model connections, and workflow stages for triage, specification, implementation, review, and deployment. Crucially, the factory should be declaratively defined in code, much like infrastructure managed through Terraform. That makes it versionable, branchable, testable, reversible, and potentially improvable by coding agents themselves.

The key argument was that companies need to move beyond “vibes” when choosing models, coding harnesses, and agent workflows. Rather than every developer independently running Claude Code, Codex, or another tool on a local machine, a factory centralizes agent conversations, artifacts, screenshots, tool calls, code changes, and outcomes. That creates an observable system of record from which teams can measure throughput, cost per pull request, automation level, revert rates, and software-delivery metrics.

Lloyd said this is mainly a tool for CTOs, engineering managers, platform teams, and developer-experience groups, rather than a day-to-day interface for every programmer. Developers and product staff can submit work from the tools they already use—Slack, Linear, Jira, GitHub, APIs, or MCP connections—while the factory handles the underlying workflow. Warp’s philosophy is that factories should be “headless” infrastructure with many access points, not another isolated workspace. [2]

A particularly important feature is evaluation. Teams can run scorers, essentially LLM judges, over historical agent sessions to catch recurring problems: overly verbose agents, weak code quality, unnecessary tests, excessive token use, or organization-specific failure modes. Then they can replay real tasks against different combinations of models and harnesses. The desired outcome is not a public benchmark score, but an answer to practical internal questions: is the cheaper model nearly as correct, and which setup provides the best cost-versus-quality tradeoff for this codebase?

The most striking vision was a self-improving loop. If an agent notices that triage repeatedly fails because it cannot delegate enough investigation, it can propose a pull request against the factory’s own YAML definition. In the future, Lloyd wants those proposed improvements tested by replaying past runs before adoption.

Warp claims this approach reduced its own cost per pull request from roughly $80 to $15 or $20, while increasing PR throughput by 40 percent over a month without raising the revert rate. But Lloyd also stressed that not everyone needs this machinery. Tiny teams should simply use agent subscriptions and move fast. He estimated that the approach begins to make sense around five to ten engineers and becomes more valuable with scale.

The conversation ended on a broader career prediction: software engineering may split into product engineering, where people build the software directly, and “factory engineering,” where they design the systems that build it. Lloyd’s core takeaway was that the competitive advantage will increasingly come from engineering the loop around coding agents—not just prompting them. [3]

Thank you for listening to Latent Space in 3 minutes from The Daily FM. See you next time!

Here is The Daily FM summary of the Latent Space that aired on Monday November 2nd. This episode featured Jacob Lauritzen, CTO of Legora, on building AI agents for lawyers—an industry where the data is exceptionally sensitive, the workflows are complex, and the cost of a confident mistake can be enormous.

Legora began simply as ChatGPT connected to Swedish legal sources. It has since become a broader workspace for legal research, drafting, document editing, due diligence, client intake, and coordinating work across lawyers and agents. Lauritzen framed the mission in two parts: help lawyers do legal work faster and better, then redesign the painful way legal work gets organized, replacing email threads and version-five Word documents with auditable, collaborative systems.

The flagship example is Legora’s “Tabular View.” A lawyer can upload thousands of contracts, ask repeated questions across all of them—such as termination terms, compensation, or intellectual-property clauses—and receive a spreadsheet-like grid with answers and citations back to each source document. The system initially runs on a small sample before scaling up, a practical safeguard against launching an expensive million-document analysis with a bad prompt. The agent can then identify anomalies, flag risks, and draft a report for the client.

A key theme was that legal AI is less about picking one winning model than managing a constantly changing mix. Legora routes different subtasks to different models based on quality, cost, latency, governance, and regional requirements. Lawyers can sometimes choose models themselves, but Lauritzen thinks most do not really want that responsibility. They want the system to keep them on the frontier automatically. Not every legal task requires the highest-intelligence model; much of the immediate opportunity is in cheaper, faster execution of routine work.

Still, models remain uneven. Legora maintains evaluations across legal practice areas and task types, using realistic environments such as huge diligence data rooms. Some releases improve one skill while regressing in another. One citation problem reportedly jumped from 70 percent to 99 percent accuracy after a model update, with no engineering work required. But long-document reasoning, coherent edits across a massive contract, search over nuanced precedent, and concise drafting remain challenging.

The unusual constraint is privacy. Legora generally cannot inspect customer traces because law firms hold confidential client material that may not legally belong to the firm itself. Instead, its engineers and lawyers build synthetic evaluations, collect carefully anonymized feedback, and favor conservative data practices. The company initially hosted on Microsoft Azure largely because law firms already trusted Microsoft through Office 365; even AWS and Google faced more skepticism a few years ago. Lauritzen said that resistance has shifted rapidly as legal organizations accept that cloud systems can be safer than their existing setups.

On the future of the profession, the hosts pressed Lauritzen on whether AI means paralegals and junior lawyers disappear. He pushed back. His view is that strong junior lawyers become more valuable because they can accomplish far more, while tedious document review declines. Lawyers should spend more time on judgment, strategy, risk assessment, and client relationships. Billing is already shifting in some work from hourly rates to flat project fees.

One especially interesting argument concerned law-firm partnerships. Lauritzen suggested firms have historically relied on partnerships because their principal asset is human capital. AI could help them create durable institutional assets instead: systems that encode the firm’s preferred precedents, workflows, and judgment. But that may require changing incentives, since partners currently benefit more from bringing in work than documenting knowledge they will not own after leaving.

Legora is also expanding through legal engineers and forward-deployed engineers, helping clients configure workflows and integrations. Lauritzen’s final takeaway was ambitious: rather than staying a tiny engineering team, the company now sees demand for a platform built from reusable primitives—projects, lists, agents, and document analysis—that can support many specialized legal products. For a $400 million acquisition, he argued, firms will still want deep legal software and human strategic counsel, not just a general chatbot.

Thank you for listening to Latent Space in 3 minutes from The Daily FM. See you next time!

Here is The Daily FM summary of the Latent Space that aired on Saturday October 31st. This episode featured Aman Singh of Arca, an AI-native wealth-management firm that recently raised $64 million. Singh’s central claim was that Arca is not simply selling software to financial advisers. It is the financial-services firm itself, using AI agents behind the scenes to make human advisers dramatically more proactive, responsive, and thorough. [1]

Arca serves clients across a broad wealth spectrum, from families planning a first home purchase to technology executives considering early retirement and founders who have sold businesses. But clients generally do not chat with an AI directly. Instead, they interact with a human adviser while agents continuously monitor the financial context around them: market shifts, tax obligations, life events, account details, deadlines, and open client requests. Singh summarized the model with a memorable line: “Agents prompt people more than people prompt agents.” [2]

The conversation focused on what it takes to make agents useful over long financial workflows that can span ten days or more. In a tax-bill example, an agent might analyze a client’s holdings and recommend the most efficient way to raise $100,000. However, it cannot independently move money, liquidate assets, or send sensitive client communications. Those actions require human review and execution. The agent’s job is to prepare the analysis, assemble a proposed plan, and bring in the right staff member at the right moment.

To maintain continuity across days, Arca has built its own model-agnostic harness and persistent memory system. It tracks each client’s current situation, ongoing work, assigned owners, near-term obligations, and long-term goals. New signals—an email reply, phone call, or document—are assessed against that shared picture. Singh described a periodic “dreaming” process that reflects on what is happening for each client and updates the system’s working memory.

Arca primarily uses frontier models from Anthropic and OpenAI, with Claude as its main model, but chooses models based on internal benchmarks rather than public leaderboards. Different task complexities receive different model tiers. Singh noted that a model that looked strong on outside evaluations proved too slow and costly in Arca’s real workflows—a reminder that deployment benchmarks matter more than headlines.

Safety and compliance were the heart of the technical discussion. Arca uses deterministic restrictions for high-risk actions, LLM-based checks on database writes, prompt-injection detection for incoming emails and documents, and controls against private-data leakage when information leaves the system. Its agents can use sandboxes for computation, but generated financial plans are checked for issues such as raw account numbers, unsupported promises of returns, or noncompliant language before they are shared.

Singh argued that compliance is fundamentally about auditability. Every change, whether made by an agent or a human, produces a traceable log of what changed, who did it, where it originated, and when. Staff corrections also become evaluations: each edit or cancellation is linked to the agent trace that caused it, then reviewed nightly by a meta-agent that suggests improvements. Humans still approve any prompt changes, because, as Singh put it, agents can make “really nasty” overfitted fixes when left alone.

One striking example involved a prospective client already paying an adviser $1,500. Arca’s agents reviewed the documents and surfaced undisclosed fees and the fact that the adviser was not a fiduciary. The broader takeaway was that AI’s financial value may come less from autonomous trading than from tireless attention: catching deadlines, revealing overlooked costs, and ensuring every client receives a level of scrutiny that humans struggle to deliver at scale.

Thank you for listening to Latent Space in 3 minutes from The Daily FM. See you next time!

Here is The Daily FM summary of the Latent Space that aired on Friday October 30th. OpenAI’s Joseph Kim joined the show to discuss ChatGPT for Financial Services, a new enterprise product designed to bring ChatGPT’s general capabilities into the highly structured, high-stakes world of banking, investing, and financial analysis. [1]

Kim described the product as a combination of familiar ChatGPT tools with the enterprise layers finance teams require: firm-specific templates, trusted data sources such as earnings transcripts and market data, and workflows tailored to analysts. OpenAI developed it with design partners including Morgan Stanley and Evercore. But his bigger point was that the product is not a separate, narrowly specialized model. Financial reasoning data is being incorporated into OpenAI’s main training runs, so the improvements should also be available to developers and firms building their own products on top of OpenAI models. [2]

A major lesson from building the product was that enterprise software cannot be designed from first principles alone. The practical realities of an industry matter. Kim gave the example of financial-data integrations: on paper, users could connect their own data through MCP, or Model Context Protocol, rather than pay for bundled sources. But OpenAI tested existing connectors and found some had error rates above 40 percent. The company concluded it needed to control the data experience end to end. That choice reportedly became one of the product’s biggest selling points. [3]

Kim also argued that startups should see this as an opportunity, not a sign OpenAI will build everything itself. He said roughly 95 percent of what OpenAI built for finance could be built by another company using ChatGPT’s plugin architecture. His provocative prediction: a billion-dollar business will be built on top of a ChatGPT plugin by combining OpenAI’s broad capabilities with a company’s own “secret sauce,” whether proprietary data, integrations, or a better user experience. [4]

The discussion repeatedly returned to the difference between model intelligence and product design. Inside OpenAI, adoption of Codex rose sharply among finance, legal, and sales teams only after the company released a desktop app. The models were capable before that, but the interface and workflow made the capability usable. Kim sees the same pattern in finance: the base model may get users 80 percent of the way, while the final 20 percent—industry workflows, accurate sources, formatting, and controls—is what makes or breaks a real product. [5]

On safety, Kim emphasized that today’s finance product is a productivity tool, not an autonomous banker. A human analyst remains responsible for reviewing its work. Looking forward, OpenAI expects agents to become more autonomous, but wants their permissions and review systems to resemble familiar workplace structures. Banks already have junior analysts whose work is checked by associates, vice presidents, and managing directors. AI agents may fit into that same hierarchy, with controlled read and write access and clear accountability.

One surprising argument was that factual errors may become less central than “taste.” Kim believes models are improving quickly at avoiding the obvious hallucinations that could change a financial conclusion. The harder problem is judgment: whether a slide looks polished, an Excel model “feels” right when assumptions change, or an investment thesis is genuinely compelling. OpenAI plans to train this through human preference data, asking experts to compare outputs and identify which is better, much as reinforcement learning originally relied on human feedback.

Finally, Kim cautioned that financial benchmarks remain weak. Many evaluations are saturated, poorly specified, or built with incorrect answer keys. OpenAI is working on more realistic evaluations intended to predict whether a model can automate actual work in professions such as investment banking. His outlook was that AI will first remove tedious junior-analyst tasks, then steadily move upward toward fuzzier, more senior work involving ideas, judgment, and strategy.

Thank you for listening to Latent Space in 3 minutes from The Daily FM. See you next time!