Y Combinator · Published 2025-06-18

Alexandr Wang: Building Scale AI, Transforming Work With Agents & Competing With China

privateReview: unreviewedConfidence: high
scale aialexandr wangAgentic workflowsscaling lawsai evaluationreinforcement-learningscaling-lawsbenchmark-contaminationhallucinationHuman-in-the-loopmulti-agent
Open on YouTube ↗

Summary

Overview

  • Speaker: Alexandr Wang
  • Channel: Y Combinator
  • Main topic: The evolution of Scale AI, frontier AI models, agentic workflows, scaling laws, and global AI competition.
  • Purpose: Provide founders, engineers, and researchers with insights into scaling frontier AI businesses, understanding AI safety and capability curves, and navigating enterprise and geopolitical AI landscapes. Alexandr Wang, founder and CEO of Scale AI, discusses his early days at MIT and Y Combinator, founding Scale AI in 2016 initially focused on autonomous vehicle data, and navigating the paradigm shifts towards large language models and agentic AI systems. The conversation covers scaling laws, the transition from data labeling to high-value full-parameter fine-tuning, the future of work with autonomous AI agents, enterprise AI adoption, geopolitical competition with China, and the ongoing importance of data quality and human-in-the-loop oversight.

Topic Map

Founding Scale AI and Early Autonomous Vehicle Focus

  • Explanation: Alexandr recounts his transition from Quora and MIT to Y Combinator in 2016, initially exploring chatbot use cases before pivoting to data labeling and infrastructure for autonomous vehicle companies like Cruise.
  • Key claims:
    • ML engineers made more than software engineers by 2014-2015.
    • Early chatbot ideas failed due to lack of quality training data and compute.
    • Focusing on the narrow problem of autonomous vehicles provided the initial high-growth wedge for Scale AI.
  • Examples:
    • Cruise becoming Scale AI's largest customer within months of launch through word-of-mouth.
  • Terminology:
    • data labeling
    • autonomous vehicles
    • machine learning engineers
    • scaling laws
  • Why it matters: Demonstrates how focusing on an extremely concrete, high-value vertical problem can anchor a startup before broader technological waves hit.

Scaling Laws, GPT Models, and Frontier AI Capability Curves

  • Explanation: Discussion on how scaling laws for LLMs became clear around GPT-3 in 2020, shifting the AI industry from data annotation for specific tasks to building and fine-tuning general-purpose foundation models.
  • Key claims:
    • Scaling laws are shockingly effective, with capabilities growing predictably with compute and data.
    • Pre-training gains are beginning to shift toward reinforcement learning and reasoning-based scaling curves.
    • LLM hallucinations occur when models lack correct context, correct prompting, or attempt tasks beyond their current capability step.
  • Examples:
    • The shift from GPT-1 and GPT-2 novelty demos to the breakthrough capability of GPT-3 and GPT-4.
  • Terminology:
    • scaling laws
    • pre-training
    • reinforcement learning
    • hallucination
    • fine-tuning
  • Why it matters: Explains the foundational mechanics driving rapid advancements in AI capabilities and the shifting bottlenecks in model training.

Agentic Workflows and the Future of Work

  • Explanation: Exploring how AI is transitioning from passive chatbots and assistants to autonomous agents capable of managing multi-step workflows, debugging code, and conducting deep research.
  • Key claims:
    • Prompting gets users 90% of the way there, but reinforcement learning agentic loops get you the final 10%.
    • Human managers of AI agents will retain agency by defining high-level vision, setting metrics, and handling edge cases.
    • Coding serves as the ultimate case study for agentic workflows, transitioning from assistant tools to fully autonomous multi-agent systems.
  • Examples:
    • Codex and agentic developer workflows executing complex multi-file code refactors autonomously.
  • Terminology:
    • agentic workflows
    • reinforcement learning
    • prompt engineering
    • agent deployment
    • codex
  • Why it matters: Redefines organizational productivity and labor leverage, suggesting that future economic growth will be driven by humans managing swarms of specialized AI agents.

Human's Last Exam and Advanced Benchmarking

  • Explanation: Overview of 'Human's Last Exam', a benchmark built in partnership with the Center for AI Safety to evaluate frontier models against the absolute hardest scientific and mathematical problems.
  • Key claims:
    • Existing AI benchmarks saturate quickly as models improve.
    • Human's Last Exam compiles novel, expert-level scientific and mathematical problems designed by top researchers.
    • Model performance on extreme benchmarks has risen from single digits to over 20% in months.
  • Examples:
    • Advanced PhD-level math and quantum physics problems that have never appeared in textbooks or on the public internet.
  • Terminology:
    • benchmark saturation
    • Human's Last Exam
    • Center for AI Safety
    • evaluation
  • Why it matters: Highlights the need for rigorous, non-contaminated evaluation tools as AI models approach and exceed human expert performance in specific scientific domains.

Geopolitical AI Competition and China

  • Explanation: Analysis of the competitive dynamics between US and Chinese AI labs, focusing on data access, manufacturing capacity, energy production, and open-source models.
  • Key claims:
    • Chinese labs make exceptionally fast progress due to massive government-backed data labeling initiatives and industrial access.
    • US advantages lie in raw algorithm innovation and cutting-edge hardware, though compute and energy constraints pose challenges.
    • Espionage and model weight leakage remain significant vectors for technology transfer.
  • Examples:
    • Chinese foundational models rapidly closing the performance gap with Western frontier models.
  • Terminology:
    • espionage
    • open-source models
    • compute constraints
    • energy production
    • foundation models
  • Why it matters: Frames AI development not just as a commercial race, but as a critical national security imperative.

Key Points

Hire People Who Give a Shit

  • Explanation: Alexandr emphasizes that company culture and long-term success rely heavily on hiring intensely passionate, deeply committed individuals who care about their work.
  • Evidence: Scale AI's survival and growth through multiple pivots hinged on mission-driven employee dedication.
  • Practical implication: Prioritize intrinsic motivation and dedication over credentialism when hiring early team members.

The Pivot from Data Labeling to Enterprise AI Platforms

  • Explanation: Scale AI expanded from a pure labeling service into a full-stack enterprise platform providing reinforcement learning, fine-tuning data, and evaluation infrastructure.
  • Evidence: Partnerships with major enterprises, defense agencies, and top labs like OpenAI.
  • Practical implication: Startups must evolve their value proposition as core industry bottlenecks shift from data collection to model alignment and deployment.

Agentic Infrastructure Replaces Manual Workflows

  • Explanation: Moving from human-in-the-loop chatbots to autonomous multi-agent loops transforms organizational efficiency across legal, software development, and defense.
  • Evidence: ThunderForge and defense AI applications deployed in the Indo-Pacific region.
  • Practical implication: Organizations must redesign internal processes to operate via agentic task delegation rather than traditional bureaucratic management.

Frameworks, Models & Processes

The Complexity Curve of AI Adoption

  • How it works: Organizations progress from basic prompt engineering and assistive tools to fully agentic multi-step workflows and fine-tuned proprietary base models.
  • Components:
    • Prompt engineering
    • Retrieval-Augmented Generation (RAG)
    • Parameter fine-tuning
    • Autonomous multi-agent deployment
  • When to use: When scaling enterprise AI integration from experimental proofs-of-concept to core operational infrastructure.

Examples & Case Studies

Scale AI launching 'Human's Last Exam' with the Center for AI Safety.

  • Illustrates: The necessity of creating non-contaminated, expert-level benchmarks to accurately track frontier model reasoning capabilities.
  • Lesson: Traditional benchmarks quickly become obsolete as models saturate existing test suites.

Cruise finding Scale AI via early launch materials and web presence.

  • Illustrates: How building a focused API for human labor or data can capture major market players early in a technology wave.
  • Lesson: Clear positioning on product hunt and developer channels can attract high-value enterprise clients organically.

Actionable Takeaways

  • Immediate:
    • Focus on rigorous evaluation frameworks rather than saturated public benchmarks.
    • Design workflows around agentic autonomy rather than simple chat interfaces.
    • Prioritize data quality and domain-specific alignment over raw parameter size alone.
  • Strategic:
    • AI will fundamentally transform organizational structures into agent-managed work units.
    • Geopolitical competition in AI requires balancing compute capacity, energy infrastructure, and talent retention.
    • Enterprise moats will be built on proprietary domain data and customized fine-tuned workflows.
  • Questions to investigate:
    • How will benchmark saturation impact frontier model training incentives?
    • What are the long-term macroeconomic effects of widespread agentic labor substitution?
    • How can democratic nations maintain a sustainable edge in AI hardware and energy production?

Claims Worth Verifying

  • ML engineers made more than software engineers as early as 2014-2015. (historical market fact)
  • Top models scoring on Human's Last Exam have risen from 7-8% to over 20%. (technical benchmark metric)

Notable Quotes

"The AI industry really continues to suffer from a lack of very hard evals and very hard tests that show really like the frontier of model capabilities." (at 0:31) "The biggest thing is you just have to really, really, really care." (at 0:45) "The wisdom of that was I think twofold. I think like first, and from talking to people who are like there at the sort of like the Genesis moment of this business, like one thing, probably the most important thing, was that they had conviction that that that the sort of like underlying business model data would basically be this like this like infinitely large and growing market." (at 53:02)

Compressed Summary

  • Scale AI evolved from autonomous vehicle data labeling to enterprise AI fine-tuning and evaluation.
  • Agentic workflows represent the next phase of enterprise productivity, shifting humans from operators to managers of AI swarms.
  • Human's Last Exam provides a rigorous benchmark for testing frontier model reasoning against elite expert-level scientific problems.
  • Geopolitical competition with China highlights critical dependencies in data labeling, compute, and energy infrastructure.
  • Keywords: scale ai, alexandr wang, agentic workflows, scaling laws, ai evaluation
  • Core insight: Frontier AI advancement is shifting from raw parameter scaling to sophisticated agentic reasoning loops, requiring rigorous new evaluation paradigms and deep enterprise data alignment.

Core insights

7
Mechanismhigh noveltymoderate evidence

Prompting/context engineering has a hard ceiling: it delivers roughly 90% of agent task completion, but the final 10% requires reinforcement-learning agentic loops (iterative feedback/reward-driven self-correction) rather than better prompts.

Why it matters

It implies prompt-only harnesses will plateau on long-horizon or self-verifying tasks, so reliability budgets should assume an RL/training loop—not prompt refinement—closes the last mile.

Generalization

For any agent task with a measurable success signal, expect a prompt/tooling ceiling and plan an optimization loop (RL, self-critique, or search) as a distinct architectural layer.

Prompting gets users 90% of the way there, but reinforcement learning agentic loops get you the final 10%.
Open source video
Mental Modelhigh noveltymoderate evidence

The dominant scaling curve is migrating from pre-training toward reinforcement learning and reasoning-based scaling, meaning the bottleneck and the data needed are changing (from annotation volume to reward/reasoning signal).

Why it matters

Architectures and data pipelines optimized for pre-training corpora may not capture the value if incremental capability comes from RL/reasoning scaling; tooling should be re-based on reward and reasoning-signal infrastructure.

Generalization

As a paradigm shifts from one scaling axis to another, the supporting infrastructure (data, evaluation, feedback) must be rebuilt around the new bottleneck rather than extended from the old one.

Pre-training gains are beginning to shift toward reinforcement learning and reasoning-based scaling curves.
Open source video
Failure Modemedium noveltymoderate evidence

Hallucination is framed as a boundary-condition failure with three distinct triggers: missing correct context, incorrect prompting, and attempting tasks beyond the model's current capability step—implying different mitigation paths (retrieval/context, prompt repair, capability gating) rather than one fix.

Why it matters

Treating hallucinations as a single category leads to one-size-fits-all mitigations; splitting into context, prompt, and capability failures lets an agent runtime route to retrieval, prompt-repair, or task-decomposition/escalation respectively.

Generalization

Decompose model error taxonomies into causes with distinct remedies so the runtime can dispatch a targeted intervention instead of a generic retry.

LLM hallucinations occur when models lack correct context, correct prompting, or attempt tasks beyond their current capability step.
Open source video
Practicemedium noveltystrong evidence

Existing AI benchmarks saturate quickly, so evaluation must use novel expert problems that never appeared on the public internet or in textbooks (as in Human's Last Exam), to avoid contamination.

Why it matters

Contaminated or saturated benchmarks give false capability signals; non-public, freshly authored, expert-level items are required to measure the frontier honestly.

Generalization

Any evaluation suite decays as models train on its distribution—sustained capability measurement requires a continuously renewed, non-public item pool rather than a fixed benchmark.

Existing AI benchmarks saturate quickly as models improve.
Open source video
Human's Last Exam compiles novel, expert-level scientific and mathematical problems designed by top researchers.
Open source video
Empirical Resultmedium noveltymoderate evidence

Model performance on extreme expert benchmarks has risen from single digits to over 20% in months, indicating a fast-moving capability curve at the frontier that shortens the useful lifetime of any evaluation snapshot.

Why it matters

If frontier scores move this quickly, evaluation cadence must match capability cadence; infrequent evaluation will systematically mislabel current model ability and misinform deployment decisions.

Generalization

Evaluation infrastructure should be continuous and versioned, since a benchmark's discriminating power has a short half-life at the frontier.

Model performance on extreme benchmarks has risen from single digits to over 20% in months.
Open source video
Tradeoffmedium noveltymoderate evidence

The core industry bottleneck has shifted from data collection to model alignment and deployment, requiring vendors to evolve from simple data labeling into full-stack platforms spanning reinforcement learning, fine-tuning data, and evaluation infrastructure.

Why it matters

It shows that value capture migrates to whatever the current scarce input is; systems should be designed with swappable capability layers rather than betting on one static bottleneck.

Generalization

Build platform boundaries around the current bottleneck but keep the value layer pluggable, because the scarce resource (data → alignment → deployment) shifts as the field matures.

Startups must evolve their value proposition as core industry bottlenecks shift from data collection to model alignment and deployment.
Open source video
Mental Modelmedium noveltyweak evidence

Human oversight of agents is reframed away from doing the work: managers retain agency by defining high-level vision, setting metrics, and handling edge cases, while agents execute multi-step workflows.

Why it matters

This defines the human-facing interface of an agent system as goal/metric specification plus exception handling, which is a concrete design target for control planes rather than an abstract 'human-in-the-loop' claim.

Generalization

Agent control planes should expose vision/goal definition, metric configuration, and escalation-to-human for edge cases as first-class primitives.

Human managers of AI agents will retain agency by defining high-level vision, setting metrics, and handling edge cases.
Open source video

Deep dives

5

Closing the last 10% of agent task completion with RL agentic loops

Research question

On verifiable long-horizon agent tasks, does an RL/reward-driven agentic loop raise completion above a prompt-only baseline, and what is the marginal gain per unit of optimization compute?

Why

If prompting has a hard ceiling, reliability budgets and runtime architecture should include an explicit optimization layer rather than only better prompts.

Prompting gets users 90% of the way there, but reinforcement learning agentic loops get you the final 10%.
Open source video
Pre-training gains are beginning to shift toward reinforcement learning and reasoning-based scaling curves.
Open source video
Source video

Benchmark saturation half-life and non-contaminated evaluation cadence

Research question

How quickly do frontier benchmark scores saturate, and at what cadence must newly authored, non-public expert items be introduced to preserve discriminating power?

Why

Saturated or contaminated benchmarks produce false capability signals, so evaluation infrastructure must be continuously versioned and renewed.

Existing AI benchmarks saturate quickly as models improve.
Open source video
Human's Last Exam compiles novel, expert-level scientific and mathematical problems designed by top researchers.
Open source video
Model performance on extreme benchmarks has risen from single digits to over 20% in months.
Open source video
Source video

Runtime classification and routing of hallucination triggers

Research question

Can an agent runtime reliably classify errors into missing context, incorrect prompting, or capability gaps, and does routed intervention reduce hallucination rates more than generic retry?

Why

One-size-fits-all mitigation wastes retries; distinct causes map to retrieval, prompt repair, or capability gating.

LLM hallucinations occur when models lack correct context, correct prompting, or attempt tasks beyond their current capability step.
Open source video
Source video

Pluggable value layer for shifting AI industry bottlenecks

Research question

As the scarce input moves from data collection to alignment and deployment, which platform boundaries let a system re-center without rebuilding the whole stack?

Why

Value capture migrates to the current bottleneck, so systems should keep the value layer swappable rather than betting on one static scarce input.

Startups must evolve their value proposition as core industry bottlenecks shift from data collection to model alignment and deployment.
Open source video
Source video

Agent control plane for vision, metrics, and edge-case escalation

Research question

Does exposing goal/vision definition, metric configuration, and structured escalation as first-class control-plane primitives reduce human handling time while preserving task success?

Why

It defines human oversight as a concrete interface rather than an abstract human-in-the-loop claim.

Human managers of AI agents will retain agency by defining high-level vision, setting metrics, and handling edge cases.
Open source video
Source video

Article ideas

4

The Last 10% Is Not a Prompting Problem

Agent reliability plateaus because the final 10% is an optimization problem; teams should treat RL/search/self-correction as a separate runtime layer, not a prompt-refinement task.

Angle

Architectural separation of prompt/tool harness from outcome-driven optimization.

Source video

Your Benchmark Is Already Decaying

Fixed public benchmarks are unsafe as capability signals at the frontier; sustained measurement requires continuously renewed, non-public expert item pools.

Angle

Evaluation as living, versioned infrastructure rather than a static leaderboard.

Source video

Hallucination Needs a Router, Not a Retry

Splitting hallucinations into context, prompt, and capability failures enables targeted runtime interventions that outperform generic retries.

Angle

Error taxonomy to runtime dispatch.

Source video

Design for the Next Bottleneck

AI platforms should keep the value layer pluggable because the scarce input shifts from data collection to alignment and deployment.

Angle

Platform architecture and business model alignment.

Source video

Project ideas

4

RL Agentic Loop Harness

movement-lab

Adding a reward-driven agentic loop to a prompt-only agent raises success on verifiable long-horizon tasks from roughly 90% to above 95%, with measurable cost per additional point.

Proof of concept

Implement a benchmark of verifiable multi-step tasks, run a prompt-only baseline, then add an RL/self-correction loop over outcomes; compare success, attempts, and token cost.

Measurement

Task completion rate, variance across runs, attempts to success, tokens/compute per solved task.

Source video

Benchmark Half-Life Tracker

beyond-evals

Frontier benchmark scores saturate with a measurable half-life, and a non-public expert-item pool retains discriminating power longer than public benchmarks.

Proof of concept

Track scores of successive model versions on a fixed public benchmark and a freshly authored private benchmark; estimate time to saturation and discrimination loss.

Measurement

Score trajectory, saturation half-life, discrimination gap between public and private items, contamination rate.

Source video

Hallucination Trigger Router

new

Classifying agent errors into missing context, incorrect prompting, or capability gap and routing to retrieval, prompt repair, or capability gating reduces hallucination rate more than generic retry.

Proof of concept

Instrument an agent runtime to log and label errors, build a classifier, and A/B test routed versus generic interventions on a task suite.

Measurement

Hallucination rate, resolution rate by trigger type, retry count, latency, cost.

Source video

Agent Control Plane for Edge-Case Escalation

gatehouse

A control plane exposing vision/goal definition, metric configuration, and structured edge-case escalation reduces human intervention time while maintaining or improving task success.

Proof of concept

Build a minimal control plane for a multi-step agent workflow; compare manager effort and success against a baseline where humans monitor every step.

Measurement

Human minutes per completed workflow, escalation rate, edge-case resolution time, task success rate.

Source video

Architectural implications

3

The bottleneck shifts from data collection to model alignment and deployment.

Before

Platform value concentrated in labeling/annotation throughput and data volume.

After

Value concentrates in RL, fine-tuning data, and evaluation infrastructure.

Consequence

Systems should keep the value layer pluggable so the architecture can re-center on whichever input is scarce as the field progresses.

Source video

Prompting reaches ~90% while RL agentic loops close the final 10%.

Before

Reliability handled reactively inside the prompt/harness layer.

After

Reliability handled by a distinct optimization loop (RL/search/self-correction) over measurable outcomes.

Consequence

The runtime gains an explicit optimization-and-feedback layer separable from prompting and tool invocation.

Source video

Coding is used as the leading edge where assistants transition to autonomous multi-agent systems.

Before

Developer tooling as single-assistant, human-driven editing.

After

Multi-agent systems performing multi-file refactors autonomously.

Consequence

Code becomes a high-signal proving ground whose agent patterns (multi-agent orchestration, autonomous refactor) likely propagate to other domains.

Source video

Tradeoffs and failure modes

4

Prompt-only agent design

Benefit

Fast to build and requires no training loop.

Cost or risk

Plateaus around 90% completion, leaving the hardest last 10% unsolved without an RL/optimization loop.

Prompting gets users 90% of the way there, but reinforcement learning agentic loops get you the final 10%.
Open source video
Source video

Public benchmarks

Benefit

Convenient, reproducible, widely comparable.

Cost or risk

Saturate quickly and risk contamination, giving false capability signals near the frontier.

Existing AI benchmarks saturate quickly as models improve.
Open source video
Source video

Frontier capability transfer

Benefit

Rapid dissemination of techniques and open weights.

Cost or risk

Espionage and model weight leakage serve as vectors for technology transfer, undermining advantage.

Espionage and model weight leakage remain significant vectors for technology transfer.
Open source video
Source video

Static data-labeling business model

Benefit

Clear initial wedge and high early growth via a concrete vertical.

Cost or risk

Becomes a low-value commodity once the bottleneck moves to alignment and deployment.

Startups must evolve their value proposition as core industry bottlenecks shift from data collection to model alignment and deployment.
Open source video
Source video

Open questions

4

What mechanism actually closes the last 10% of agent task completion that prompting cannot reach?

Why unresolved

The summary states the 90/10 split but gives no implementation detail on the RL agentic loop.

Research direction

Prototype reward/feedback-driven agent loops on tasks with measurable success and compare against prompt-only baselines.

Source video

How quickly does a frontier benchmark lose discriminating power, and at what cadence must new non-contaminated items be authored?

Why unresolved

The summary gives a single data point (single digits to >20% in months) without a decay model.

Research direction

Track score trajectories across benchmark generations to estimate a capability-curve half-life for evaluation assets.

Source video

Which failure cause—missing context, wrong prompt, or capability gap—dominates real agent errors, and can it be detected at runtime?

Why unresolved

The summary offers a three-way taxonomy but no distribution or detection signal.

Research direction

Instrument agent runtimes to classify hallucination triggers and measure the proportion of each type with targeted interventions.

Source video

At what task complexity should a single assistant be replaced by a multi-agent autonomous loop?

Why unresolved

The summary asserts assistants evolve into autonomous multi-agent systems but gives no threshold or criterion.

Research direction

Run controlled comparisons of single-agent vs multi-agent orchestration on multi-file refactor tasks, measuring success, cost, and failure modes.

Source video

Key claims

8
causalVerification needed

Reinforcement-learning agentic loops, not prompting, close the final 10% of task completion.

Evidence

Prompting gets users 90% of the way there, but reinforcement learning agentic loops get you the final 10%.

Question

On controlled agentic tasks, does adding an RL loop meaningfully raise success above a prompt-only baseline, and by how much?

Source video
predictionVerification needed

Pre-training gains are shifting toward reinforcement learning and reasoning-based scaling curves.

Evidence

Pre-training gains are beginning to shift toward reinforcement learning and reasoning-based scaling curves.

Question

Do marginal capability gains increasingly derive from RL/reasoning compute rather than additional pre-training data?

Source video
causalVerification needed

LLM hallucinations arise from missing context, incorrect prompting, or tasks beyond the model's capability step.

Evidence

LLM hallucinations occur when models lack correct context, correct prompting, or attempt tasks beyond their current capability step.

Question

Can hallucination errors in production be reliably attributed to one of these three triggers, and does targeted mitigation reduce them?

Source video
factualVerification needed

Existing AI benchmarks saturate quickly as models improve.

Evidence

Existing AI benchmarks saturate quickly as models improve.

Question

What is the observed saturation half-life of commonly used benchmarks across successive frontier models?

Source video
opinionVerification needed

Model performance on extreme expert benchmarks rose from single digits to over 20% in months.

Evidence

Model performance on extreme benchmarks has risen from single digits to over 20% in months.

Question

What are the exact benchmark, model versions, and dates behind the single-digit to >20% trajectory?

Source video
comparativeVerification needed

The industry bottleneck has shifted from data collection to model alignment and deployment, requiring evolved value propositions.

Evidence

Startups must evolve their value proposition as core industry bottlenecks shift from data collection to model alignment and deployment.

Question

Does spending and demand in the AI tooling market show a measurable shift from labeling toward alignment and deployment infrastructure?

Source video
causalVerification needed

Espionage and model weight leakage are significant vectors for AI technology transfer.

Evidence

Espionage and model weight leakage remain significant vectors for technology transfer.

Question

What concrete incidents or forensic evidence substantiate model weight leakage as a technology-transfer vector?

Source video
opinionVerification not requested

Human managers of agents retain agency by setting vision and metrics and handling edge cases.

Evidence

Human managers of AI agents will retain agency by defining high-level vision, setting metrics, and handling edge cases.

Source video

Connections

4