Stories from Latent Space

Handshake AI hits $1B revenue building evals frontier labs can't game

Handshake AI has scaled to roughly $1 billion annualized revenue in a year by building reinforcement learning evaluation environments for frontier labs. The company's work reveals a critical gap in how AI systems are assessed: agents exploit evaluation metrics in ways human experts would catch. Their case studies span biology, finance, and alignment research, showing that frontier model training depends on adversarially robust evaluation frameworks. This exposes a structural vulnerability in current AI development pipelines where metric gaming can mask genuine capability gaps.

Latent Space·8h ago

85

Tools & Code Opinion & Analysis

Stanford redesigns CS curriculum around real code review, not LLM-solvable homework

Stanford's revised CS146S course tackles a fundamental problem in AI-native education: traditional homework becomes obsolete when LLMs can solve it instantly. Mihail Eric's solution brings open-source maintainers into the classroom, forcing students to engage with real codebases and defend their work to actual reviewers. This pedagogical shift reflects a broader industry reckoning with how to teach engineering when tool capability outpaces curriculum design. Eric's parallel work at Monaco, including custom agent workspace infrastructure, suggests that hands-on exposure to production AI systems is becoming table stakes for developer training.

Latent Space·1d ago

68

Products & Apps Tools & Code

GitHub rebuilds development stack for agent workloads at 225M scale

Microsoft's CoreAI leadership outlined GitHub's infrastructure pivot toward agent-native development at scale, revealing 225M developers and 34.3B annual commits as the baseline for a rebuilt software stack. The strategic focus centers on Copilot Fusion's model orchestration layer and the cost-quality tradeoff that underpins multi-device agent continuity. Parikh's framing of AI spend against business outcomes rather than raw compute signals a maturation in how enterprises measure LLM ROI. The Azure migration and session-aware agents represent concrete infrastructure shifts that will shape how platforms handle agentic workloads beyond chatbots.

Latent Space·1d ago

80

Business & Funding Opinion & Analysis

Anthropic's IPO math: $80B loss, GPU financing, and equity accounting

Anthropic's path to public markets hinges on a counterintuitive financial story: massive operating losses offset by GPU lender financing and equity-based accounting that inflates the income statement. Jamin Ball's analysis reveals the core tension facing frontier AI companies seeking IPO validation. The $80.7 billion net loss against $3.9 billion operating cash flow in H1 2026 exposes how traditional valuation frameworks struggle with infrastructure-heavy AI businesses where unit economics depend on inference cost trajectories and early investor share appreciation. This shapes how Wall Street will price Claude's competitive moat and runway.

Latent Space·2d ago

80

Research Business & Funding

Google, Meta, NIH launch $1.8B biology prediction initiative

A $1.8B consortium spanning Google, Meta, NIH, DOE, and Biohub is launching a coordinated effort to build predictive models of biological systems across molecular, spatial, and temporal scales. The initiative exposes a critical gap in current AI: foundation models trained on hundreds of millions of cells still fail to predict cellular perturbations reliably, often losing to simple baselines. This signals that scaling alone won't solve biology; the field needs better benchmarks, measurement technologies, and conceptual frameworks. Fabian Theis and collaborators are advancing this frontier through tools like Nicheformer and CellFlow, positioning computational biology as the next domain where AI infrastructure meets fundamental scientific bottlenecks.

Latent Space·3d ago

85

Products & Apps Business & Funding

Latent Space brings AI Engineer conference model to Shanghai

Latent Space's expansion into China through AIEi Shanghai signals a maturing AI engineer community beyond Silicon Valley. The podcast episode traces how a grassroots open-source blog evolved into a conference brand, now establishing regional footholds. Key discussion points include agent-driven SaaS replacement, OpenClaw adoption patterns in China, and how reasoning models reshape human-AI interaction design. The framing of conferences as datasets reflects a shift in how AI communities validate emerging patterns and surface builder priorities across geographies.

Latent Space·5d ago

68

Products & Apps Tools & Code

Warp cuts agent-generated PR costs to $15-20 through factory optimization

Warp has operationalized agent-driven software development into a measurable production system, reducing per-pull-request costs from $80 to $15-20 through model tuning and hierarchical agent design. The shift signals a maturing AI engineering discipline where infrastructure teams now benchmark agent quality, trace execution paths, and iterate on the factory itself rather than individual models. This moves beyond proof-of-concept toward sustained cost efficiency and throughput gains in real engineering workflows, establishing templates other teams will adopt.

Latent Space·6d ago

80

Products & Apps Business & Funding

Legora builds legal agents without visibility into their own work

Legora has built a production legal agent that operates entirely blind to its own outputs, a constraint imposed by client confidentiality that forces novel approaches to agent development and testing. The Stockholm-founded startup demonstrates how privacy requirements reshape AI engineering practices, moving beyond traditional debugging and evaluation workflows. Their solution, now used for contract review, document drafting, and due diligence coordination, suggests a broader pattern: enterprise AI systems must function reliably without the transparency that typically guides model iteration. The company's recent decision to add model selection after years of resistance hints at shifting tradeoffs between capability and operational simplicity in legal automation.

Latent Space·Nov 2

73

Products & Apps Tools & Code

Arca builds multi-day agents for wealth management with persistent memory

Arca is operationalizing multi-day agentic workflows for wealth management, moving beyond single-turn LLM interactions into persistent, stateful systems that handle complex financial tasks. The firm's architecture combines custom task routing with long-horizon memory to coordinate agent actions across days, including human approval gates for high-stakes decisions like tax planning. This represents a critical inflection point for enterprise AI: agents must now solve problems that span planning horizons beyond model context windows, requiring infrastructure for memory management, task persistence, and human-in-the-loop governance. Arca's approach signals how production agentic systems will differ fundamentally from chatbot interfaces.

Latent Space·Oct 31

73

Products & Apps Research

OpenAI embeds financial reasoning into ChatGPT, not just connectors

OpenAI's financial services division is embedding domain-specific reasoning directly into ChatGPT rather than relying on external data connectors, signaling a shift toward vertically integrated AI products. Joseph Kim's team bundled market data and built review workflows that let financial professionals audit model decisions, addressing a critical trust gap in high-stakes domains. This approach reveals how frontier labs are moving beyond general-purpose models to encode expert judgment into training, while simultaneously defining what plugin primitives startups can leverage. The strategy highlights a broader tension: as LLMs become domain tools, the bottleneck shifts from raw capability to human-verifiable reasoning and institutional accountability.

Latent Space·Oct 31

80

Tools & Code Business & Funding

Supabase becomes preferred backend for production AI agents

Supabase has become a critical infrastructure choice for AI agent builders, a shift that reflects how backend architecture now shapes agent capability and persistence. The Postgres-native platform offers agents long-running compute, declarative schemas, and reliable storage in a single layer, reducing the friction of bolting together separate services. This podcast with co-founders Paul Copplestone and Ant Wilson explores how an open-source Firebase alternative evolved into the preferred data layer for production AI systems, signaling that agent infrastructure decisions are moving upstream from application logic into foundational database and compute choices.

Latent Space·Oct 28

73

Products & Apps Tools & Code

Anthropic separates agent brains from execution to prevent cascade failures

Anthropic's engineering leadership reveals the architectural decisions underpinning Claude Managed Agents, a system that isolates execution environments from core agent logic to prevent cascading failures. The conversation covers credential handling without sandbox exposure, a Python-to-Rust Messages API migration validated through shadow traffic, and model distillation prevention. This infrastructure layer matters because it translates raw model capability into production-grade reliability for developers, setting a template for how frontier labs operationalize agent systems at scale.

Latent Space·Oct 27

85

Tools & Code Research

Leash Bio opens measured experimental worlds for agent evaluation

Leash Biosciences released Minos, a framework of pre-measured experimental environments where AI agents can explore protein-molecule interactions at scale. The platform addresses a critical gap in agent evaluation: providing ground-truth feedback loops for scientific reasoning. This matters because it lets researchers measure whether language models and reasoning agents actually improve at medicinal chemistry tasks, not just benchmark scores. The framework reveals which models excel at high-dimensional decision-making and how agents learn from prior experiments to refine future choices. For AI labs, this represents a new class of grounded evaluation infrastructure where agents operate in physically validated domains rather than synthetic tasks.

Latent Space·Oct 26

73

Hardware & Infra Products & Apps

Modal CEO on why GPU markets misalign with AI workload dynamics

Modal's CEO Erik Bernhardsson articulates a structural mismatch in GPU markets: cloud providers package compute in fixed three-year contracts while AI workloads demand dynamic, unpredictable scaling. This conversation surfaces how Modal's pooled compute model absorbs that volatility, explores emerging accelerator alternatives to GPUs, and examines the operational and financial risks of building serverless GPU infrastructure. The discussion connects infrastructure economics to developer experience in an agent-driven era, touching on sandbox security and code review workflows that reflect shifting deployment patterns.

Latent Space·Oct 22

73

Business & Funding Products & Apps

Tenex restructures engineering around AI agent outcomes instead of headcount

Tenex is experimenting with AI agents as core engineering infrastructure, shifting compensation from headcount to project outcomes rather than traditional salary models. The company's MetaHarness process embeds plans, context, and decisions in version control to enable agent coordination at scale. This conversation surfaces a critical tension in enterprise AI adoption: while agents can accelerate implementation work, organizational and economic barriers remain steep. The discussion also flags where AI-built software cannot yet displace SaaS, and raises labor implications for engineering roles in major tech hubs like New York.

Latent Space·Oct 21

73

Products & Apps

Moshi applies LLM reasoning to commodity trading data synthesis

Moshi is building a recommendation layer for commodity traders by synthesizing market data, Bloomberg feeds, and private sources into actionable insights. The startup's approach mirrors Netflix's personalization engine applied to financial markets, combining structured and unstructured data into a connected graph that Claude-like reasoning can traverse. This represents a narrower but concrete use case for LLM-powered decision support in high-stakes domains where conflicting signals and data integration remain unsolved problems. The conversation surfaces a broader tension in AI product development: how to move beyond dashboards toward systems that actually drive user action.

Latent Space·Oct 20

68

Hardware & Infra Business & Funding

SF Compute builds independent GPU trading with its own datacenters

SF Compute is building independent GPU trading infrastructure that treats compute as a financial commodity while operating its own datacenters to ensure reliable physical delivery. The company's model exposes a structural gap in the AI market: major cloud providers and NVIDIA have misaligned incentives with traders who need predictable, fungible GPU access across hardware generations and workload types. By separating the financial layer from operations, SF Compute is creating a programmable cluster abstraction that lets batch and inference workloads trade under different contract terms, addressing how interruptions, storage, and hardware obsolescence actually affect real compute prices rather than theoretical spot rates.

Latent Space·Oct 20

73

Products & Apps Opinion & Analysis

AI power users separate OpenAI hype from production reality

Six prominent AI practitioners dissect OpenAI's recent product announcements to separate durable capabilities from marketing polish. The conversation maps a critical gap in the AI industry: models with impressive benchmarks often fail in production workflows due to poor context handling, cost control, and integration friction. Key tensions emerge around agents versus assistants, personal AI viability, and whether speed or reasoning quality matters more in practice. For builders and investors, this insider calibration reveals which OpenAI bets are likely to stick in real workflows versus those destined for demo fatigue.

Latent Space·Oct 16

73

Research Products & Apps

Leash Bio turns drug experiments into AI training sandboxes

Leash Bio is operationalizing AI-driven drug discovery by converting DNA-encoded library experiments into reinforcement learning environments, enabling thousands of medicinal chemistry campaigns to run in parallel with rapid experimental feedback loops. The company's approach surfaces critical gaps in public chemical datasets, including hidden shortcuts and memorization artifacts, while demonstrating how measured protein-ligand binding landscapes can serve as training grounds for reasoning agents. This work bridges wet-lab reproducibility with AI model validation, establishing a new paradigm where experimental rigor directly constrains and improves model behavior rather than relying on static benchmarks.

Latent Space·Oct 14

85

Models & Releases Products & Apps

Dyna Robotics deploys million-hour world models into production robots

Dyna Robotics is moving physical AI from research into production by training world-action models on over one million hours of human video, enabling robots to learn new tasks from minimal demonstrations. The company's Dyna-2 model combines future video prediction with robot control, allowing real-world deployment at scale, including thousands of daily napkin-folding operations at Din Tai Fung. This work suggests that predictive modeling of physical dynamics may be the foundation for general-purpose robotics, shifting the field's focus from task-specific engineering toward learned, transferable world models that scale with data.

Latent Space·Oct 13

85

Products & Apps Tools & Code

72 teams replaced $43K SaaS with AI agents, exposing the build-to-own gap

Latent Space ran a competitive build-off to replace a $43K annual SaaS tool for conference speaker management, attracting 72 teams using coding agents. The winning open-source solutions reveal how builders partition work between Claude and other models, expose common architectural pitfalls, and highlight a critical gap: shipping production software differs fundamentally from prototyping with AI. This signals a broader market shift where domain experts can now outsource commodity SaaS replacement to agent-driven development, reshaping both vendor economics and the skill profile required to own versus build.

Latent Space·Oct 13

73

Research Opinion & Analysis

DeepMind and Biohub challenge AlphaFold's legacy as biology AI frontier

DeepMind and Biohub researchers are challenging the narrative that AlphaFold solved protein folding, arguing instead that structure prediction was merely the opening move in a far harder problem: building AI systems that genuinely model biological complexity. The panel surfaces a critical tension in AI-driven science: scaling compute and data alone cannot substitute for domain insight, and the field must rethink how to encode biological constraints and causality into next-generation models. This reframes the limits of current deep learning approaches and signals where biology-focused AI research must evolve to move beyond pattern matching toward mechanistic understanding.

Latent Space·Oct 10

80

Research Products & Apps

Periodic Labs grounds AI discovery in physical experiments, not internet data

Periodic Labs is pursuing a fundamentally different path for AI advancement: grounding machine learning in physical experimentation rather than internet-scale data. The startup's vision of 'synthesis superintelligence' combines reinforcement learning with real-world lab instruments, materials characterization, and computational chemistry to enable AI systems that discover novel materials. This represents a strategic pivot in how frontier labs approach AI capability development, moving from pure software optimization toward embodied reasoning over noisy experimental feedback. The approach challenges the assumption that scale and data alone drive progress, positioning scientific discovery as a distinct domain requiring different architectural and training paradigms.

Latent Space·Oct 8

85

Tools & Code Hardware & Infra

SemiAnalysis releases ClusterMAX 3.0 GPU cloud evaluation framework

SemiAnalysis has released ClusterMAX 3.0, a systematic evaluation framework for GPU cloud providers that tests configuration, performance, reliability, and failure response through hands-on testing and customer feedback. The framework addresses a critical gap in the AI infrastructure market: practitioners lack standardized criteria for comparing managed GPU clusters. This matters because GPU cloud selection directly impacts training costs, latency, and workload reliability. The open-source audit tool and published rankings give teams concrete data to move beyond vendor claims, shifting GPU cloud procurement from opaque vendor relationships toward evidence-based comparison.

Latent Space·Oct 2

73

Research Tools & Code

MIT researcher argues agent harnesses waste latent LLM capability

MIT researcher Alex Zhang argues that current AI agent architectures waste latent model capability by constraining powerful LLMs within rigid harness designs. The conversation explores how better system integration, GPU optimization, and multi-agent coordination could unlock performance gains without requiring new model weights. Zhang contrasts this with existing coding agents from Anthropic and OpenAI, suggesting the gap between them reflects engineering choices rather than fundamental model differences. The insight matters for practitioners: scaling agent infrastructure and harness sophistication may deliver near-term gains comparable to waiting for the next model generation.

Latent Space·Oct 2

73

Products & Apps Models & Releases

OpenAI's Computer Use agents now outpace human operators on real tasks

OpenAI's Computer Use capability has undergone fundamental architectural changes that now enable AI agents to operate software faster and more reliably than human operators. The company introduced infrastructure allowing agents to run on dedicated cloud compute, combined with a new Agents API and architectural innovations that integrate screenshots, accessibility APIs, DOM data, and Playwright automation. These improvements address the core bottleneck in agentic AI: the speed and accuracy of environment perception and action recovery. The shift signals that agent deployment is transitioning from research prototype to production-grade infrastructure, with implications for enterprise automation and the competitive positioning of agentic platforms.

Latent Space·Sep 30

90

Products & Apps Tools & Code

Anthropic moves Claude Code toward modular, multiplayer agent workflows

Anthropic is reshaping how developers interact with Claude Code through a suite of architectural shifts that move beyond single-user workflows. The company is introducing Claude Mods, a customization layer that lets users modify the agent harness itself, alongside multiplayer collaboration features and cross-environment deployment capabilities. This represents a fundamental pivot from Claude as a code-writing tool to Claude as an extensible platform where teams can compose, iterate, and share agent behaviors. The shift signals Anthropic's bet that the next phase of AI development hinges not on raw model capability but on how fluidly agents can be orchestrated, adapted, and operated across heterogeneous infrastructure.

Latent Space·Sep 29

80

Products & Apps Business & Funding

OpenRouter's path from wrapper to AI infrastructure layer

OpenRouter's evolution from a routing abstraction layer into critical infrastructure reveals how model diversity, not monoculture, is reshaping AI deployment. The platform now serves 10 million developers by solving a structural problem: model labs invest billions in training but lack distribution channels to reach builders. This conversation with co-founder Alex Atallah and a16z's Anjney Midha unpacks why competitive inference marketplaces matter for ecosystem health, how Mistral's success validated multi-model strategy, and what it takes to build neutral infrastructure when incumbents prefer lock-in. The shift signals that AI's future depends less on who trains the best checkpoint and more on who controls the pipes.

Latent Space·Sep 25

80

Models & Releases Products & Apps

Runway pivots video models toward world simulation and robotics

Runway is repositioning video generation as infrastructure for world models rather than a standalone creative tool. The company's evolution from Gen-1 through Gen-4.5 reflects a strategic pivot toward physics-aware simulation capable of powering robotics and rendering software interfaces directly. This reframes the competitive landscape after Sora's release: instead of competing on video quality alone, Runway is betting that video models trained at scale become the foundation for embodied AI and general-purpose simulation. The thousand-A100 commitment and architectural decisions behind Gen-2 signal how frontier labs are now treating video as a path to AGI infrastructure, not just media generation.

Latent Space·Sep 25

85

Models & Releases Policy & Regulation

Omni model learns to design sequences that evade biosecurity defenses

Radical Numerics' Omni model represents a critical inflection point in AI-driven biology: systems that not only predict disease variants and design molecules but learn biological principles never explicitly encoded in training data. The capability leap from genome reading to autonomous sequence design creates immediate dual-use concerns. When models can preserve function while altering sequence signatures, they bypass existing biosecurity detection frameworks, forcing a reckoning between open research norms and containment risk that mirrors nuclear and synthetic biology governance debates.

Latent Space·Sep 23

90