Speaking

Beyond Productivity

What Happens When AI Expands Human Potential

Most AI conversations begin with time saved. That is useful, but it leaves a larger question open: what can a person now understand, create, or accomplish that previously felt out of reach? This keynote examines that question through research on work, creativity, learning, and human interaction with AI. It separates three outcomes that are often confused: stronger performance with assistance, capability people retain, and useful changes in the world. The evidence is uneven, and that is part of the story. A tool can improve an answer while weakening the learning process; a persuasive demonstration can leave the hard operational problem untouched. Eric introduces a practical framework for deciding what to test, what to measure, and where people must remain accountable. The session closes with a concrete question for the audience: which previously inaccessible possibility is worth pursuing, and what evidence would convince us that it has genuinely expanded human potential?

Audience: Executives, leadership teams, educators, innovation groups, professional associations, and cross-disciplinary audiences.

Keynote: 35–45 minutes plus discussion. Open with an apparently successful AI-assisted task, then reveal what it does not measure. Compare findings from workplace assistance, creative work, and education. Introduce assisted performance, retained capability, and realized outcomes with agency as three separate evaluation lenses. End with an audience-specific opportunity and a test that could disprove it.

Workshop: 90–120 minutes. Participants bring one activity whose constraint may be changing. Map the original constraint; specify the new capability; separate output, learning, and realized value; then design a bounded test with a baseline, measures, human decision owner, and stopping rule. Deliverable: a one-page capability experiment.

Expected outcomes

• Distinguish assisted performance from durable capability and realized value.

• Recognize a meaningful opportunity beyond throughput.

• Choose a measurable first experiment with explicit human accountability.

Research lens

• Generative AI at Work: observed workplace productivity effects in a specific support setting; not a universal return on AI.

• Human–AI combination meta-analysis: joint performance does not automatically exceed the stronger participant.

• Bastani and colleagues: a mathematics field experiment separates assisted practice from unaided exam performance. Kestin and colleagues: a structured physics tutor improved immediate learning in a short, expert-designed intervention. Different designs help explain why learning outcomes cannot be inferred from assistance alone.

Supporting research

Generative AI at Work (2025)

When Human–AI Combinations Are Useful: Systematic Review and Meta-Analysis (2024)

Generative AI Without Guardrails Can Harm Learning (2025)

AI Tutoring and In-Class Active Learning: Randomized Trial (2025)

Potential demonstrations

• A three-part scorecard for assisted output, retained capability, and realized outcome.

• A synthetic before-and-after task with identical outputs but different learning and verification paths.

 

The Distance Between Imagination and Reality

What Changes When More Ideas Become Possible to Test

Many ideas never reach the point where someone can learn whether they are useful. The cost of research, design, technical translation, or an initial prototype can end the exploration early. AI may lower some of those costs, changing which ideas become worth testing. This session follows a concept from a rough question to a tangible experiment, making the human choices visible along the way. It pairs that possibility with research showing a complication: assistance can improve individual creative output while making a collection of outputs more alike. More drafts do not necessarily create more distinct ideas. Eric examines how to preserve unusual starting points, seek alternatives, and involve the people who would use the result. Audiences leave with a practical exploration sequence: define the uncertainty, build only what can teach you something, gather a real response, and decide whether the idea deserves another round of effort before scaling it further.

Audience: Creators, founders, product teams, designers, innovation leaders, educators, and interdisciplinary builders.

Keynote: 35–45 minutes plus discussion. Trace an illustrative idea through question, alternatives, prototype, user response, and decision. Show where time or skill barriers may fall and where evidence, taste, and user access still matter. Contrast fluent variation with genuine exploration.

Workshop: 90–120 minutes. Teams select one uncertain idea, generate genuinely different approaches, sketch a minimal test, and choose the cheapest observation that could change their mind. Deliverable: a testable concept and explicit learning question, not a finished product promise.

Expected outcomes

• Identify the constraint preventing an idea from being tested.

• Use AI for exploration without accepting the first plausible direction.

• Define a prototype by what it teaches rather than how complete it looks.

Research lens

• Doshi and Hauser: short-story creativity improved for some writers with AI ideas while collective output diversity declined.

• The Cybernetic Teammate: individual and team results in a bounded product-innovation exercise; useful for concept development, not proof of production success.

Supporting research

Generative AI, Individual Creativity and Collective Diversity (2024)

The Cybernetic Teammate: Generative AI and Teamwork (2026)

Potential demonstrations

• An exploration funnel showing ideas attempted, alternatives retained, and uncertainties resolved.

• A prepared low-stakes concept transformed into three distinct prototypes, with model assistance and human selections visible.

 

The Organization of One

What Happens When Individuals Gain Capabilities Once Requiring Departments

A person with a difficult idea once needed to assemble a team before much of the exploration could begin. AI makes some research, analysis, design, and technical work more accessible, suggesting a provocative possibility: an individual with the reach of a much larger organization. This talk treats that possibility as a hypothesis to examine. Research from product innovation offers evidence that AI can help individuals bridge functional perspectives in a bounded task. It does not show that one person can replace the relationships, accountability, expertise, or execution capacity of an enterprise. Eric explores how to compose useful support around a real objective while keeping responsibility legible. Where should a person decide? What needs expert review? Which handoff requires permission? What happens when work fails? Audiences leave with a workflow map that distinguishes accessible capabilities from organizational obligations and identifies the next bottleneck that additional AI alone will not necessarily resolve.

Audience: Founders, independent professionals, small-business leaders, corporate innovation teams, and functional leaders.

Keynote: 35–45 minutes plus discussion. Begin with the functions hidden inside one ordinary project. Examine which can be assisted, which require specialist involvement, and which remain human obligations. Close with a deliberately bounded organization-of-one example and its failure plan.

Workshop: 90–120 minutes. Map a real outcome across research, making, review, authorization, delivery, and support. Assign every consequential decision to a person. Identify where a specialist, partner, customer, or team is still needed. Deliverable: a capability-and-responsibility map with a first pilot.

Expected outcomes

• Separate functional assistance from replacement of an organization.

• Design clear handoffs, review gates, and ownership.

• Identify the remaining constraint on a real project.

Research lens

• The Cybernetic Teammate, final 2026 publication: a bounded product-innovation study with P&G professionals; individuals with AI matched teams without AI on assessed proposals. The study did not observe downstream development or commercialization.

• METR task-horizon methodology: informative about evaluated agent tasks, not a measure of whole-job autonomy.

Supporting research

The Cybernetic Teammate: Generative AI and Teamwork (2026)

METR: AI Task-Completion Time Horizons and Limitations

Potential demonstrations

• A project map showing capabilities around one person and the people who retain decisions.

• A synthetic project handoff in which an AI-generated recommendation reaches a deliberate approval gate.

 

From Chatbots to Agents

When AI Moves from Answering Questions to Performing Work

The move from a chatbot to an agent changes the central question. It becomes necessary to ask what the system is permitted to do, how it knows the relevant context, and how anyone can establish whether the work actually succeeded. This session follows a simple task across planning, tool use, approval, execution, and verification. Along the way, Eric separates a plausible plan from a reliable workflow and a tool acknowledgement from a confirmed result. Research on agent performance provides useful signals, but benchmark success should not be mistaken for dependable operation in an unfamiliar organization. The talk makes the practical design choices visible: bounded permissions, clear ownership, recoverable steps, accessible audit records, and meaningful stopping conditions. Audiences leave able to identify one useful agent workflow, the evidence it would need, and the decisions a person should retain before that workflow is allowed to affect customers, money, records, or physical systems.

Audience: Business and technology leaders, product and operations teams, AI adoption groups, and governance stakeholders.

Keynote: 35–45 minutes plus discussion. Follow a synthetic work request through the full execution path. Introduce context, retrieval, memory, and orchestration only where each explains a failure or design choice. Include stale information, an unauthorized action request, and an ambiguous result.

Workshop: 90–120 minutes. Participants choose a reversible, low-risk workflow and define its task boundary, allowed tools, approval gates, success evidence, error recovery, and escalation owner. Deliverable: an agent workflow specification and an evaluation checklist before implementation.

Expected outcomes

• Explain the difference between an answer, a plan, an action, and a verified outcome.

• Choose the right level of autonomy for a specific workflow.

• Specify permissions, failure handling, and evidence before scaling.

Research lens

• METR: task time horizons measure success rates against human task duration in evaluated tasks; they are not autonomous run length or job-replacement forecasts.

• Anthropic context-engineering guidance: a first-party account of design practices, distinct from independent comparative evidence.

Supporting research

METR: AI Task-Completion Time Horizons and Limitations

Anthropic: Effective Context Engineering for AI Agents (2025)

Potential demonstrations

• An animated task path with plan, authorization, execution, and verification states.

• A sandbox demonstration that deliberately refuses an out-of-scope action and exposes a failed tool result.

 

The Future of Automotive Isn't Another App

From Fragmented Vehicle Systems to Useful Orchestration

A question such as which vehicle is ready for tomorrow's trip can cross several manufacturer apps, charging systems, accounts, and assumptions. Vehicle Conductor begins with that fragmented experience and asks whether an intelligent interface could coordinate it more usefully. This session uses the project as an honest research case: public documentation has been reviewed, but no vehicle has been connected or tested. Eric examines what a read-only readiness demonstration would need to establish before anyone should consider real commands. The discussion covers model-specific capabilities, data freshness, driver permissions, uncertain outcomes, and the difference between accepting a command and confirming a physical result. It also asks when an interface should remain quiet. Audiences leave with a clearer picture of the opportunity and its constraints, a set of useful pilot questions, and a practical understanding of why integration access, user trust, safety, and recurring value have to be demonstrated together in practice.

Audience: Automotive executives, OEM and dealer teams, mobility builders, connected-vehicle providers, and customer-experience leaders.

Keynote: 35–45 minutes plus discussion. Open with one trip-readiness question and reveal the systems beneath it. Walk through the five-brand documentation screen without implying tested compatibility. Demonstrate a synthetic readiness answer, then show how stale data or missing permissions changes it.

Workshop: 90–120 minutes. Map a readiness or ownership workflow across driver intent, data owners, signals, permissions, and outcomes. Define a read-only pilot before any actuation. Deliverable: a capability-evidence matrix, consent boundary, and proceed-or-stop criteria.

Expected outcomes

• Distinguish a shared interface from reliable cross-system orchestration.

• Identify what exact-vehicle access and live validation must prove.

• Design an initial use case with clear privacy, safety, and evidence boundaries.

Research lens

• Vehicle Conductor Phase 0 plan, reviewed September 30–October 1, 2026: five-brand documentation research and proposed tests only.

• Official Smartcar and Tesla documentation: provider-specific capabilities and access requirements require current verification.

• AAA voice-interaction research offers historical evidence about cognitive workload; it does not establish the safety of a modern AI vehicle assistant.

• Google's 2026 in-car Gemini announcement is a vendor capability description, not independent safety evidence.

Supporting research

Smartcar: Compatibility API Overview

Tesla Fleet API: Vehicle Commands

AAA: Cognitive Distraction in Ten In-Vehicle Systems (2015)

Google: Gemini for Cars with Google Built-In (2026 Announcement)

Potential demonstrations

• A five-brand ecosystem diagram with documented, unverified, and unavailable states clearly distinguished.

• Synthetic vehicle cards showing different signal ages and an explicit refusal to recommend when evidence is inadequate.

 

AI as a Thinking Partner

Expanding Thought Without Surrendering Judgment

What would it mean to use AI in a way that improves the thinking process, rather than merely supplying the next answer? This session explores practical approaches to questioning assumptions, comparing explanations, finding counterexamples, and making uncertainty visible. Eric pairs those techniques with research that complicates the promise. Human and AI performance do not automatically combine into a better result, and people's reports of reduced cognitive effort are not the same as evidence of lasting cognitive decline. The aim is to distinguish useful assistance from an easy feeling of confidence. A worked example shows how the same model can reinforce a weak premise or help interrogate it, depending on the evidence, task design, and human review. Participants leave with a repeatable thinking sequence that includes independent framing, deliberate challenge, source checking, and a final judgment they can explain without relying on the model's confidence or fluency as proof of correctness.

Audience: Leaders, analysts, educators, researchers, knowledge workers, and intellectually curious general audiences.

Keynote: 35–45 minutes plus discussion. Present two plausible answers to one decision, then expose the assumptions beneath them. Demonstrate independent framing, counterargument, evidence retrieval, and human synthesis. End with an explicit account of what remains uncertain.

Workshop: 90–120 minutes. Participants state a low-stakes decision and initial reasoning before using AI. Use the model to identify missing evidence and alternative explanations, then independently evaluate the result. Deliverable: a decision note with sources, counterarguments, and ownership.

Expected outcomes

• Recognize fluency and agreement as unreliable substitutes for evidence.

• Use structured challenge to broaden a decision without outsourcing responsibility.

• Check whether the process improved understanding, not only the final wording.

Research lens

• Human–AI combination meta-analysis: aggregate evidence against assuming automatic synergy.

• Microsoft critical-thinking survey: 319 knowledge workers self-reported confidence and effort; the design does not establish cognitive decline or causal skill loss.

• Bastani and colleagues: assisted mathematics practice and unaided exams produced different outcomes in one school. This supports measuring unassisted understanding, while remaining limited to that intervention.

Supporting research

When Human–AI Combinations Are Useful: Systematic Review and Meta-Analysis (2024)

Generative AI and Critical Thinking: Knowledge-Worker Survey (2025)

Generative AI Without Guardrails Can Harm Learning (2025)

Potential demonstrations

• A reasoning map that grows as assumptions, sources, and alternative explanations become visible.

• Two rehearsed responses to the same premise, followed by independent source checking and a human decision.

 

When Intelligence Becomes Abundant

What Becomes Valuable When More People Can Access Cognitive Capabilities

The idea of abundant intelligence is provocative precisely because its terms need examination. Which capabilities are becoming easier to access? For whom? At what cost? This session starts there, then asks how value might shift when producing a plausible answer becomes less difficult. Eric explores candidates for enduring scarcity: a worthwhile question, reliable context, earned trust, distinctive taste, access to the real world, and responsibility for action. These are propositions to test, not a settled forecast. Evidence from workplace and innovation studies shows how specific capabilities can become more accessible while leaving important differences in expertise, task performance, and outcomes. Scenario exercises then examine how an organization might respond under different assumptions about cost, reliability, access, and regulation. Audiences leave with a clearer account of what their advantage depends on today, which assumptions AI may challenge, and which human or institutional capabilities deserve investment as the landscape continues to change.

Audience: Executives, strategists, entrepreneurs, investors, educators, and public-interest or policy audiences.

Keynote: 35–45 minutes plus discussion. Define abundance conditionally, show where access remains unequal, and examine several forms of scarcity. Use two contrasting scenarios rather than a single future prediction. Close with a concrete test of the audience's current advantage.

Workshop: 90–120 minutes. Map where value comes from today, then stress-test it under lower-cost cognitive assistance and uneven access. Identify what becomes easier, what remains constrained, and what can be tested now. Deliverable: an assumption map and a small portfolio of reversible actions.

Expected outcomes

• Define the capability and access assumptions hidden in claims of abundance.

• Identify potentially durable sources of value without treating them as guaranteed.

• Choose reversible tests instead of committing to a single technology forecast.

Research lens

• Generative AI at Work: task- and experience-specific effects help complicate the claim that capability becomes uniform.

• The Cybernetic Teammate: bounded evidence about functional expertise and AI-supported innovation.

• Additional cost, access, and governance facts must be freshly sourced for the audience and event; they should not be inferred from model benchmarks.

Supporting research

Generative AI at Work (2025)

The Cybernetic Teammate: Generative AI and Teamwork (2026)

Potential demonstrations

• A scarcity map that separates answer production from context, trust, permissions, resources, and accountable action.

• Two explicit future scenarios with changed assumptions and different strategic implications.