This resource guide captures key insights from a workshop on navigating AI adoption in business. Whether you're just starting your AI journey or looking to accelerate your organization's progress, you'll find practical frameworks, real-world case studies, and actionable strategies.
What You'll Find Here:
Mental models for understanding AI's capabilities and limitations
A proven framework for identifying high-impact AI opportunities
Real case study: How AI transformed 35 hours of work into minutes
Practical guidance on tools, security, and implementation
Resources for continued learning and next steps
This isn't theory—it's battle-tested advice from building AI solutions in regulated industries.
Navigating the AI Revolution
A practical guide for to unlock exponential gains through artificial intelligence
Semi-Pro Roller Coaster Rider
1997
Innovative Tech
Medical & General IT Services
1999
Mp3Car
Telematics, Community, E-commerce
2013
Whitebox
Raised $55m, E-commerce, Grew to 500 Employees
AI Venture Studio
Today, I'm building multiple AI products that solve real-world problems. These aren't experiments—they're working solutions serving actual customers in regulated industries.
AI Diagnostics
Smart phone-based diagnostic tools at Curiedx.com bringing medical analysis to your pocket
Food Verification
TrayVerify system ensuring accuracy in food service operations
Synthetic Board of Directors
AI board advisor providing strategic guidance and diverse perspectives (Sopheva)
What You'll Learn
Psychology Shift
You need to fundamentally relearn how to learn in the AI age
Why It Feels Impossible
Understanding the overwhelming volume and pace of change
Making AI Work
Operationalizing AI: security frameworks, daily workflows, practical examples
Interactive Engagement
Let's keep this conversational—questions and discussions throughout
Why You Can't Keep Up
The disorientation you're feeling isn't a personal failing. It's a completely rational response to an unprecedented volume of change. Let's quantify exactly what you're up against:
Newly Funded AI Companies
Per week - Serious commercial products entering the market
Public AI Apps Published
Per week - Many are experiments, not viable businesses
No-Code AI Agents Created
Per week - Most are not maintained long-term products
Half-Life of Learning was a decade +
The Old World
- Learn a skill or technology early in your career
- That knowledge remains valid and valuable for a decade or more
- Continuing education happens a few times per year
- Deep expertise compounds over time
The New Paradigm
- Knowledge is valid for 3-6 months before major shifts occur
- The specific tool you mastered yesterday might be obsolete next quarter
- Learning must be continuous, not episodic
- Expertise is now about adaptation speed, not accumulated knowledge
The Old Learning & Growth Model: The Rocket Equation
Bigger Team
Requires a proportionately larger HR department and support staff.
More People
Demands enhanced leadership skills and management infrastructure.
Growth
Necessitates acquiring significantly larger office spaces and facilities.
Scale
Requires developing more comprehensive systems for team training and onboarding.
Expansion
Implies implementing more complex organizational structures and processes.
Learning AI: The Fuel Generates More Fuel
AI Learning: The Fuel Generates More Fuel
Learn New Tool
Initial time investment
Work Faster
Complete tasks in less time
Free Up Time
Hours available for learning
Learn More
Accelerated skill acquisition
Repeat Faster
Each cycle spins quicker
The J-Curve Reality
You need to be honest about the dip. When you start, you WILL be slower. That's the cost of admission to exponential gains. You'll struggle. You'll make mistakes. You'll wonder if it's worth it.
The best time to start was 2023. The second best time is today.
Never Felt This Far Behind?
You must rethink your learning strategy.
"I've never felt this much behind as a programmer... I have a sense that I could be 10x more powerful if I just properly string together what has become available over the past year. A failure to claim the boost feels decidedly like a skill issue."
—Andrej Karpathy, Former Director of AI at Tesla, Co-founder of OpenAI
This isn't about having more hours in the day. It's about fundamentally restructuring how you approach learning, experimentation, and skill acquisition. The opportunity isn't just to keep up—it's to compound your capabilities at an accelerating rate.
A few of Ethan Mollick's Ideas
- Wharton Professor specializing in entrepreneurship and innovation
- Co-director of Wharton's Generative AI Labs
- Author: Co-Intelligence: Living and Working with AI (2024 NYT bestseller)
- TIME's 100 Most Influential People in AI (2024)
The Jagged Frontier
Why AI is superhuman at some tasks but fails at seemingly easy ones
Reverse Salient
How single bottlenecks hold back entire systems—and unlock many use cases when fixed
Secret Cyborgs
Why over half your employees are hiding their AI usage—and what to do about it
These three frameworks will help you understand why AI feels so inconsistent, where breakthroughs come from, and what's actually happening in your organization right now.
The Jagged Frontier: Why AI Feels Inconsistent
Reverse Salient: Watch the Bottlenecks
Ethan Mollick borrowed this term from historian Thomas Hughes, who studied how technological systems evolve. A reverse salient is a single bottleneck holding back an entire system. In military terms, it's the part of your line that hasn't advanced as far as the rest—your weak point.
Historical Example: Electricity
Edison invented the light bulb, but it was useless without power transmission. The bottleneck wasn't the bulb—it was the transmission infrastructure. Once transmission lines were built, dozens of electrical applications became viable overnight.
Ability Gaps
AI simply can't do it yet. Examples: reducing hallucinations, specific reasoning tasks, nuanced judgment calls. These tend to get fixed as models improve.
Institutional Constraints
Your processes limit what's possible, regardless of AI capability. Examples: approval workflows, regulatory requirements, org structure. These require redesign.
Edge Cases
AI handles 99% of cases perfectly, but the 1% requires human intervention. Examples: unusual customer scenarios, ethical judgment calls. These need hybrid solutions.
Secret Cyborgs: Unleash your team!
Here's an uncomfortable truth about your organization right now: Your employees are already using AI extensively. Over half of them just aren't telling you about it. Mollick calls them "secret cyborgs"—people who've integrated AI into their workflow but keep it hidden.
Hidden AI Users
Employees who don't disclose AI usage to employers
Performance Improvement
On 1 in 5 tasks where workers use AI
Why They're Hiding
Policy Fears
Company policy explicitly or implicitly bans AI usage, creating legal/compliance anxiety
Devaluation Concerns
Fear that "AI-assisted" work will be seen as less valuable or less legitimate
Job Security
"If I automate 90% of my job and tell anyone, will I be fired or have my compensation cut?"
These secret cyborgs are your early adopters, your internal AI experts, your competitive advantage—if you can get them to come forward. Here's how to enable and embolden them:
Public Acknowledgment
Explicitly state that anyone can contribute to the AI agenda. Make it clear this isn't just for technical roles—it's for everyone.
Psychological Safety
Create explicit protection for experimentation. Make it abundantly clear people won't get in trouble for trying AI tools and sharing what they learn.
Active Rewards
Publicly recognize and reward people who bring forward AI solutions that benefit the business. Celebrate both successes and intelligent failures.
Inaction Is a Decision: You Have Five Options
1. Go Out of Business
Ignore AI entirely. Watch competitors gain 2-10x productivity advantages. Eventually lose customers who demand modern capabilities. Not recommended, but it's technically an option.
2. Retire or Quit
Decide this transition isn't for you. Exit now while your business still has value. Hand it off to the next generation or sell while you can still get a good price.
3. Sell the Company
Strategic exit before the value gap widens. Find a buyer who's ready to make the AI investment. Get out while the getting is good. Perfectly legitimate choice.
4. Execute Disaster Plan
Recommended: Replace yourself in day-to-day operations and become Chief AI Officer. Pretend you have a serious health condition—bring in your second-in-command NOW. Dedicate 50-75% of your time to intensive re-education.
5. Hire Someone (Last Resort)
If you absolutely must delegate this, hire an AI explorer. But understand: they won't know your business, your customers, your risk tolerance. You'll still need to invest heavily in your own education to manage them effectively.
Where to Start: Finding Your AI Opportunity
Why Passion Matters
AI adoption involves a steep learning curve, countless failures, and persistent iteration. If you don't genuinely care about the problem you're solving, you'll quit at the first major setback. Passion provides the resilience to push through the J-curve dip. Pick something that energizes you. Something you think about in the shower. Something that frustrates you enough that you're willing to invest 20+ hours learning how to fix it. Why Underperformance Matters Don't pick something that's already working great—the improvement won't be visible or meaningful enough to justify the effort. You need a clear "before and after" where the gains are undeniable. Look for processes that are: tedious (people avoid doing them), expensive (eating resources), slow (bottlenecking operations), or inconsistent (quality varies wildly). The intersection of these two circles—passion AND underperformance—is your AI opportunity. That's where you have both the motivation to persist and the potential for transformative impact. Use the AI Idea Scoring Tool (link in resources) to systematically evaluate multiple opportunities and choose the one with the highest value for your specific situation.
What is Selfware? Examples: Tools I've Built
One of the most powerful shifts in AI is the emergence of what I call "selfware"—custom tools you build for yourself to solve your specific problems. We're entering an age where saying "I need a tool that does X" increasingly results in a working tool appearing within hours.
Blood Lab Data Extraction
Automated extraction and analysis of lab results. Turns complex medical PDFs into structured data I can track over time. Built in an afternoon.
Custom Teleprompter
Video recording tool that scrolls my script at exactly the pace I speak. Needed it for a video shoot, had Claude Code build it while I wrote the script.
CRM Enrichment Tool
Multi-AI system that enriches sparse contact data with websites, LinkedIn profiles, and email addresses. 95% time savings on lead research.
Celeste
AI with guardrails for my 7 year old daughter. (link)
Initial Data: From PDFs to Actionable Lists
The first hurdle in CRM enrichment is often scattered and unstructured data. Our process begins by standardizing this information, making it ready for the next steps.
Our initial dataset was locked in various PDF documents. We automated the conversion of these unstructured PDFs into a clean, actionable Excel file, forming the essential foundation for our enrichment process.
Before: Raw PDF
After: Structured Excel
Let's Make the AI work….
Number of Employees
Gain insights into company size, helping you tailor your outreach strategy and understand potential resource allocation.
Revenue Data
Assess financial health and market position, enabling more targeted and relevant sales proposals.
LinkedIn URL
Facilitate direct access to key contacts, company profiles, and networking opportunities for personalized engagement.
Email Address
Ensure direct and efficient communication with decision-makers, reducing bounce rates and improving outreach effectiveness.
Company Description
Quickly understand a company's core business, mission, and values to personalize your messaging and build rapport.
Recent News Articles
Stay informed about current events, achievements, or challenges related to the company, providing timely talking points for sales and follow-ups.
Have AI be your SDR: Score leads
Example:
- How many miles from X
- Last time this person had a regulatory citation
- Public financial filings
- Customer Reviews
Impact: More Than Just Numbers, 35 hours in minutes
This Is What AI Adoption Looks Like
The CRM enrichment case study you just saw—35 hours compressed into minutes—isn't unique. It's what becomes possible when you systematically identify and execute on AI opportunities.
But here's the problem most leaders face: You know AI matters. You just don't know where to start, or which opportunities are worth your time.
That's why I built Sopheva.
Sopheva Virtual Chief AI Officer delivers what a full-time AI executive would: daily AI opportunities tailored to your specific business, weekly research on what's actually working, and a scoring system to prioritize where to focus.
No generic advice. No "AI strategy decks." Just actionable opportunities that move your business forward—for $500/month instead of a $300K salary.
Ready to stop feeling behind? Visit sopheva.com/caio
How It Works: Multi-AI Approach
Resources: Try It Yourself
I've put together a comprehensive Google Drive folder with everything you need to explore this CRM enrichment approach yourself. You don't need to be a programmer to understand the concepts—the value is in seeing how the problem was broken down and solved systematically.
How-To Videos
Step-by-step walkthrough of the setup process, architecture decisions, and implementation details
Code Examples
Working code you can adapt for your specific use case. Heavily commented to explain what each piece does.
Sample Data
Input data examples and output results so you can see exactly what the transformation looks like
This connects directly back to the opportunity framework we discussed earlier:
- Passion: I care about sales effectiveness and lead generation. Having accurate contact data directly impacts revenue.
- Underperformance: Manual data entry wasn't getting done. It was sitting on to-do lists indefinitely because it's tedious and time-consuming.
- Result: A tool I actually use regularly, that saves real time, and that makes me more effective at a core business function.
All resources, including the Google Drive link, are in the "Stuff to Read" section at the end of this presentation.
This List of Tools Is Wrong.
Google Gemini
Best for: Image processing, calendar/file integration, massive document analysis
Why I use it: Handles huge context windows. Take a photo of a flyer, say "add this to my calendar"—it extracts everything. Excels at processing 100+ page documents.
OpenAI / ChatGPT
Best for: Video generation (Sora/BEO3), conversational AI, code generation
Why I use it: Leading-edge capabilities in video creation. Strong general-purpose performance across many tasks. Excellent API for integration.
Anthropic / Claude
Best for: Data analysis, graphics generation, browser control, software development
Why I use it: Claude Code is my primary tool for building software. Exceptional at reasoning through complex problems. Strong safety guardrails.
Supporting Cast (Specialized Tools)
- Mac Whisper: Voice-to-text transcription. I use this constantly for dictating prompts and capturing thoughts. PC has similar alternatives.
- Granola: Automatically records calls and generates comprehensive notes. Saves hours per week.
- HeyGen: AI avatar video generation for scalable video content without recording.
- Superhuman: AI-powered email search. Good tool but reached its half-life—better options emerging.
- Gamma: Made this slide deck gamma.app
What Is a Context Window?
If you only learn one technical concept from this session, make it this one: understanding context windows is essential to getting good output from AI. It fundamentally shapes how you organize your thinking and structure your requests.
For Long Documents
Some AIs can process 100+ page documents. Others max out at a few pages. Know your tool's limits and structure accordingly.
For Conversations
In a long chat thread, older messages eventually "fall out" of context. The AI forgets them. Start fresh when context gets stale.
For Prompts
Include relevant examples, constraints, and background. But don't dump your entire company history—just what's needed for this specific task.
Navigate Uncharted Waters
Without Halucinating
Cite Your Sources
Always ask the AI to show its sources, especially when dealing with factual information. This allows you to verify accuracy, understand the AI's data foundation, and deepen your own understanding of the topic.
"Extract Verbatim Quotes"
This ensurces you can are getting the exact material from the source. Then you can actually find (Ctrl-f) to find the source, and surrounding words
Screw Driver required demo
How I built an investor slides (Link)
How I used AI this week
Building These Slides
Used voice transcription to capture my talking points. Fed transcripts to AI for organization and structure. Iterated on layout and flow. You're looking at the output right now.
Investor Deck Creation
Same process as these slides—started with voice notes about our value proposition and market opportunity. AI helped structure the narrative arc and create supporting visuals.
Financial Modeling
Built scenario analysis for three different growth paths. AI helped structure the model logic and identify assumptions that needed validation.
Meeting Transcripts
Automatically captured and processed for action items, decisions, and follow-up questions. Distributed within an hour of meeting end.
Folder Reorganization
Had AI analyze my chaotic file structure and propose a new organization system based on how I actually work, not how I think I should work.
Deep Market Research
Hospital industry analysis with zero prior knowledge. Fed AI three market research reports and got investor-ready insights with citations.
Product Comparison
"I need to buy a new camera. Build me a comparison chart of the top 5 options across these criteria."
How Do I Cook Bacon?
Even mundane questions. "What's the best method for crispy bacon without making a mess?"
Did you say cooking?
Learn From Your Errors. You're going to make 10x more.
Most people treat errors as pure waste—time lost, work thrown away, frustration without value. This is exactly backward. Every failure with AI contains valuable data about how to improve your approach, refine your prompts, or adjust your expectations.
The winners aren't the people who make fewer mistakes. They're the people who've built systems to extract maximum value from their inevitable failures.
The Error Mindset
- Identify what you'd do differently
- Update your approach based on learnings
- Share insights with your team
- Treat failures as experiments, successes, not setbacks
Why Errors Are Valuable
Each mistake teaches you something about the jagged frontier—where AI capabilities end and human judgment begins. Failed prompts show you what context was missing or what constraints weren't clear.
Building Error Recovery Systems
Instead of just fixing errors manually, use AI itself to help you recover and learn. Turn failures into documentation. Extract patterns from what went wrong.
Celebrate your failures!
Then….Use AI to Recover From Your Errors
My Mistake this week: The AI-Powered Recovery & The Silent Demo Video
Harvest Session Data
Used Claude Code to extract every prompt I'd given it during that session from the conversation history
Generate Documentation
Had it organize all the prompts chronologically with the outputs and reasoning at each step
Better Than Original
Ended up with a comprehensive 36-page written tutorial—better documentation than the video would have been
This is the pattern: when something goes wrong, immediately think "How can I use AI to recover value from this failure?" Often you'll end up with something better than if the mistake had never happened. And learning something?
Employees, Security & Confidentiality
This comes up constantly in every conversation about AI adoption: "What about our confidential data? Won't AI companies use our information to train their models? Isn't this a massive security risk?"
Let's address this directly with facts, not fear.
The Door Lock Analogy
Your data is like your house. AI tools are like your front door. You can lock the door. Every major AI tool has settings that prevent your data from being used for model training. Just like a locked door keeps people out, these settings keep your data private.
The real question isn't "Is AI secure?"—it's "Have you locked the door?"
ChatGPT / OpenAI
Location: Settings → Data Controls → "Improve the model for everyone"
Action: Turn OFF to prevent training on your conversations
Claude / Anthropic
Location: Settings → Privacy → "Allow training on conversations"
Action: Turn OFF to opt out of model training
Google / Gemini
Location: Activity Controls → Web & App Activity
Action: Turn OFF to prevent data collection for training
How to Hire Someone
(Last Resort Option)
Why This Is Last Resort
- You know your business - where risks are acceptable and where they're not
- You know your customers - what they value and what would alienate them
- You know where failure is tolerable - which experiments are safe vs. dangerous
- You know your culture - how to introduce change without breaking what works
What an Outside Hire Won't Have
- Deep understanding of your industry dynamics
- Relationships with key customers and partners
- Intuition about what will work in your specific context
- Political capital to drive controversial changes
- Authority to make strategic bets with real resources
If You Must Hire Your AI Explorer
If hiring is truly your only option, here are the traits and mindsets that predict success. Note: these are dramatically different from traditional job requirements. You're not hiring a project manager or a developer—you're hiring an explorer.
The Critical Traits (In Priority Order)
Resourcefulness
Can work effectively without a manual or clear instructions. Comfortable in ambiguity. Finds paths forward when there's no obvious route.
Understanding Why (Not Just How)
Doesn't just follow recipes—understands underlying principles. Can adapt approaches to new situations because they grasp the reasoning.
Persistence
Doesn't quit at the first wall. Views obstacles as puzzles to solve, not stop signs. Comfortable with multiple failed attempts before success.
Curiosity
Genuinely excited to explore and learn. Doesn't need external motivation to dive deep into understanding how things work. Intrinsically motivated.
Problem-Solving Orientation
Default response to challenges is "How can I solve this?" not "This is too hard." Finds paths, not excuses. Sees constraints as creative challenges.
Creativity
Sees unconventional solutions others miss. Willing to try approaches that seem weird. Doesn't get trapped by "the way it's always been done."
Communication Skills
Can translate technical findings into business language. Makes complex topics accessible. Comfortable presenting to leadership.
Technical Comfort
Doesn't need to be a programmer, but needs genuine comfort with tools and technology. Not intimidated by unfamiliar interfaces.
Attention to Detail
Catches AI hallucinations and errors before they become problems. Verifies outputs systematically. Doesn't blindly trust results.
The Critical Mindset Shift
What to measure: Experiments run, lessons learned, knowledge shared with the team, quality of documentation. What NOT to measure: Success rate, speed of wins, immediate ROI.
Your Homework: What You Must Do Next
For Advanced Users
- Skim the CRM enrichment video - Use my code as an example and starting point
- Build your own "selfware" tool - Pick a problem from your passion/underperformance intersection
- Focus on architecture - Study the multi-AI pattern with confidence scoring
- Document your process - What worked, what failed, what you'd do differently
For Newer Users
- Skim Ethan Mollick's work - Start with his Substack "One Useful Thing"
- Push AI until it breaks - Start with personal hobbies to reduce pressure
- Find your jagged frontier - Learn where YOUR specific edges are
- Keep a learning log - Document what works and what doesn't
Free Tool: AI Project Scoring System
Comprehensive Scoring
Scores up to 5 AI projects simultaneously (1-10 scale across 18 criteria)
Customizable Weights
Produces a single weighted score for easy comparison and flexible evaluation
Automated Insights
Auto-generates top 3 strengths and bottom 3 weaknesses per project
Financial Tracking
Tracks costs (labor hours, dollars, hard costs) and projected ROI
Consistent Evaluations
Built-in scoring rubrics for consistent, unbiased assessments
Why Use It
Rigor & Consistency
Brings rigor and consistency to AI prioritization across the organization
Common Language
Creates a common language for leadership teams to discuss AI investments
Risk Mitigation
Surfaces hidden risks before committing valuable resources to projects
Objective Comparison
Enables objective "apples to apples" comparison of different AI initiatives
Zero Cost
Completely free to use — no software purchase or subscription required. Link to the free scoring tool: AI Project Scoring System (Excel)
You are an AI project evaluation assistant helping business leaders systematically assess potential AI implementation opportunities. Your role is to guide users through a structured scoring framework that evaluates projects across multiple critical dimensions.
For each AI project idea submitted, evaluate and score (1-5 scale) across these categories:
Impact Potential (Weight: 25%)
- Revenue generation or cost savings magnitude
- Number of people/processes affected
- Strategic alignment with business goalsFeasibility (Weight: 25%)
- Technical complexity and current AI capability
- Data availability and quality
- Integration with existing systemsTime to Value (Weight: 20%)
- Speed to initial prototype/proof of concept
- Path to production deployment
- Learning curve for teamResource Requirements (Weight: 15%)
- Budget needed (tools, talent, infrastructure)
- Team time commitment
- Ongoing maintenance burdenRisk Level (Weight: 15%)
- Regulatory/compliance concerns
- Data privacy and security issues
- Reputational risk if it fails publiclyProvide a weighted total score (out of 5.0) and a clear recommendation: Pursue Now, Explore Further, or Deprioritize. Include 2-3 specific next steps for high-scoring projects.
Stuff to Read / Resources
Essential resources for continued learning beyond today's session. I've curated these carefully—every item here has proven valuable in my own learning journey.
TLDR AI Newsletter
What it is: Daily AI news summary delivered to your inbox. Filters the firehose down to what actually matters.
Why it's essential: Staying current without drowning in noise. 5-minute daily read that keeps you informed.
Link: https://tldr.tech/ai
Ethan Mollick's "One Useful Thing"
What it is: Substack newsletter from Wharton professor covering practical AI applications for business.
Why it's essential: Best single source for understanding AI's business implications. No hype, just practical insights.
Link: Search "One Useful Thing Substack"
Book: Co-Intelligence by Ethan Mollick
What it is: Comprehensive guide to living and working with AI, grounded in research and real-world examples.
Why it's essential: The single best book for business leaders trying to understand AI strategically.
Two Ways Sopheva Can Help You
Sopheva Advisory Board — For Your Loneliest Decisions
You're about to make a decision that keeps you up at night. Your team can't help—they're too close. Your board meets quarterly. Your coach is great, but doesn't know your industry.
Sopheva Advisory Board simulates a room of world-class advisors (Bezos, Buffett, Dalio, and others) who stress-test your thinking in real-time. Get perspectives you'd never consider. Spot blindspots before they cost you.
$300/month self-serve | $750/session facilitated
Try it free at sopheva.com
Virtual Chief AI Officer — For Structured AI Adoption
You know you need to "do something with AI," but hiring a full-time AI executive isn't realistic yet. And you don't have time to chase every shiny new tool.
Sopheva CAIO delivers daily AI opportunities tailored to your business, weekly research on what's working, and a scoring system to prioritize where to focus. It's like having a Chief AI Officer—without the $300K salary.
$500/month
Learn more at sopheva.com/caio
Not sure which fits? Book a 15-minute call with Alex to walk through your situation.
Schedule at calendly.com/sopheva/alex
Deep Thinks
01:38:20
YouTube
AI: What Could Go Wrong? with Geoffrey Hinton | The Weekly Show with Jon Stewart
As artificial intelligence advances at unprecedented speed, Jon is joined by Geoffrey Hinton, Professor Emeritus at the University of Toronto and the “Godfather of AI,” to understand what we’ve actually created. Together, they explore how neural networks and AI systems function, assess the current capabilities of the technology, and examine Hinton’s concerns about where AI is headed. 0:00 - Intro 1:36 - Geoffrey Hinton Joins 5:13 - Machine Learning & Neural Networks 10:20 - How Neural Concepts
05:15:01
YouTube
Dario Amodei: Anthropic CEO on Claude, AGI & the Future of AI & Humanity | Lex Fridman Podcast #452
Dario Amodei is the CEO of Anthropic, the company that created Claude. Amanda Askell is an AI researcher working on Claude's character and personality. Chris Olah is an AI researcher working on mechanistic interpretability. Thank you for listening ❤ Check out our sponsors: https://lexfridman.com/sponsors/ep452-sb See below for timestamps, transcript, and to give feedback, submit questions, contact Lex, etc. *Transcript:* https://lexfridman.com/dario-amodei-transcript *CONTACT LEX:* *Feedback*
35:47
YouTube
Ryan McClelland NASA "From Text to Spaceship: Advancing AI in Aerospace" at CDFAM NYC 2024
This presentation at CDFAM Computational Design Symposium in NYC, 2024, introduces the innovative ‘Text-to-Spaceship’ concept by Ryan McClelland at NASA Goddard, focusing on the pivotal role of AI in transforming text-based science objectives into mission designs. We discuss how leveraging current and near-term AI technologies can accelerate the entire mission development process, from initial concept through to hardware realization. Specific attention is given to AI-driven computational desig
Tools Reference
Detailed breakdowns of each major AI tool and when to use it
Appendix: Tools - Google Gemini
Google Gemini excels in specific areas where its deep integration with Google's ecosystem provides unique advantages. Best for: image processing, personal assistant tasks, and analyzing extremely large documents.
Standout Capabilities
Image Processing
Exceptional at extracting information from images, diagrams, and screenshots. Can read handwriting, process charts, analyze photos.
Calendar/Files Integration
Deep hooks into Google ecosystem. Take a photo of a flyer, say "add this to my calendar"—it extracts everything automatically.
Large Context Window
Can process massive documents (100+ pages) in a single pass. Excellent for analyzing comprehensive reports or contracts.
Best Use Cases
- Processing PDFs with tables, charts, and images
- Extracting data from invoices or receipts
- Summarizing lengthy research reports
- Managing personal scheduling and tasks
- Analyzing visual content at scale
Limitations
- Less strong at pure code generation
- Reasoning capabilities lag behind Claude/GPT-4
- Requires Google account and ecosystem
Appendix: Tools - OpenAI / ChatGPT
OpenAI's ChatGPT (especially GPT-4 and newer models) excels at conversational AI, code generation, and creative tasks. Recently added video generation through Sora/BEO3 models.
Standout Capabilities
Video Creation (BEO3)
Leading-edge video generation from text descriptions. Can create professional video content from scripts and concepts.
Conversational AI
Natural, contextual conversations that maintain coherence over long interactions. Excellent at understanding nuanced requests.
Code Generation
Strong programming capabilities across many languages. Good at explaining code and debugging.
Best Use Cases
- Creating video content from scripts
- Writing and editing business content
- Brainstorming and ideation sessions
- Building prototypes and MVPs
- Customer service automation
Limitations
- Can be verbose and repetitive
- Sometimes lacks depth in specialized domains
- Data cutoff means no real-time information without plugins
Remember: by the time you read this, specific model versions may have changed. Focus on understanding the relative strengths rather than memorizing version numbers.
Appendix: Tools - Anthropic / Claude
Claude (especially Claude 3.5 Sonnet and Claude Code) is my most-used tool for serious work. Excels at data analysis, graphics generation, reasoning tasks, and software development. Strong safety guardrails reduce hallucinations.
Standout Capabilities
Excel & Data Analysis
Exceptional at processing structured data, creating formulas, analyzing patterns. Can work with complex datasets effectively.
Graphics Generation
Creates charts, diagrams, and visual assets. Better than competitors at structured visual output.
Browser Control
Can interact with web interfaces, fill forms, extract data. Useful for automation tasks.
Claude Code (Power Tool)
Advanced software development capabilities. This is what I use to build actual working applications. Requires technical comfort but incredibly powerful.
Best Use Cases
- Building custom software tools
- Complex data analysis and modeling
- Research synthesis across multiple sources
- Creating detailed documentation
- Tasks requiring strong reasoning
Limitations
- Slower than some competitors
- More expensive per token
- Sometimes overly cautious/conservative
If I had to choose only one AI tool to use for the next year, it would be Claude. The combination of reasoning, coding, and data analysis capabilities covers the majority of my high-value use cases.
Appendix: Tools - Supporting Cast
Beyond the major AI platforms, these specialized tools handle specific workflows efficiently. They're the supporting actors that make the overall system work smoothly.
Mac Whisper: Voice to Text
What it does: Converts speech to text with remarkable accuracy. I use it constantly for dictating prompts, capturing thoughts, and creating initial drafts.
Why essential: Speaking is 3-4x faster than typing. Removes the friction between having an idea and getting it into text form.
PC alternatives: Windows has similar voice transcription tools built in, or try Otter.ai
Superhuman: AI Email Search
What it does: AI-powered email search and management. Natural language queries to find messages.
Why I mention it: Good example of a tool that's reached its "half-life"—it was cutting edge 6 months ago, but better alternatives are emerging. Still useful but no longer differentiated.
The lesson: Tools you love today may be obsolete next quarter. Stay flexible.
Granola: Call Recording & Notes
What it does: Automatically records calls and generates comprehensive notes with action items, decisions, and key discussion points.
Why essential: Saves hours per week on meeting documentation. Let you be fully present in conversations instead of frantically taking notes.
ROI: If you have 10+ hours of calls per week, this pays for itself in the first week
HeyGen: AI Avatar Videos
What it does: Creates video content using AI-generated avatars. Input script, choose avatar, get professional-looking video without filming.
Why useful: Scalable video content creation. Record once, generate variations. Good for training content or repetitive explainer videos.
Limitation: Still has "uncanny valley" feel. Best for internal use or where polish isn't critical.
Backup Interactive Exercises
Additional exercises available if time permits or group energy suggests pivoting
Appendix: Backup Exercise - Stump Rob Challenge
Best for: Skeptical audiences who need to see both AI's capabilities and its genuine limitations. Time: 15-20 minutes. Energy level: High.
How It Works
"Give me your best shot. What's something you're convinced AI can't help with? I'll either show you how it can work, or honestly explain why you're right that it's beyond current capabilities."
What Makes This Work
- Embrace failures openly: "You got me, that's a genuine limitation"
- When AI succeeds: Explain WHY it worked and what made it tractable
- When AI fails: Explain the underlying reason, not just "it's bad at this"
- No fake victories: Don't pretend AI can do things it can't
Common "Stumpers" That Actually Work
- Highly specific industry jargon (often works with sufficient context)
- Tasks requiring physical presence (can assist with planning)
- Real-time data without tool integration
- Ethical judgment calls
Genuine Limitations to Acknowledge
Real-Time Information
Without tool integration, AI doesn't know what happened after its training cutoff or what's happening right now. Stock prices, news, weather—requires plugins.
Human Judgment on Ethics/Values
AI can analyze ethical frameworks but shouldn't make final calls on morally ambiguous decisions. That's inherently human territory.
Tasks Requiring Accountability
When someone needs to be held responsible for an outcome, that must be a human. AI can advise, but humans decide and own consequences.
Appendix: Backup Exercise - Live Prompt Surgery
Best for: People who are already using AI but not getting good results. Time: 15-20 minutes. Energy level: Medium.
The Setup
"Who's been using AI and feels like they're not getting great results? Tell me what you're doing—no judgment, we're all learning here. Let's rebuild your approach together."
The Surgery Checklist
What's the Actual Goal?
Often different from what they're asking for. Uncover the real desired outcome, not just the surface request.
What Context Is Missing?
AI can't read your mind. What background information would a human need to do this well? Probably AI needs it too.
What Constraints Aren't Specified?
Tone, length, format, audience, style—these aren't optional details, they're essential parameters.
Should This Be Multiple Steps?
Complex requests often work better as a sequence of simpler prompts rather than one giant mega-prompt.
Key Teaching Moments
- Show how vague prompts get vague results
- Demonstrate adding specificity without overcomplicating
- Explain when to use follow-ups vs. one big prompt
- Highlight the importance of examples
Before/After Template
Before: "Write me a marketing email"
After: "Write a 150-word email to existing customers announcing our new product. Friendly but professional tone. Focus on how it solves their top pain point: [specific problem]. Include a clear CTA to schedule a demo. Here's an example of our typical voice: [sample]"
Appendix: Backup Exercise - What Would Rob Do?
Best for: Teaching strategic thinking through concrete scenarios. Time: 20-25 minutes. Energy level: Medium.
How It Works
Present a business scenario drawn from pre-event survey responses. Collect 2-3 audience approaches first, then share my approach with explicit trade-off analysis.
Sample Scenarios (Customize From Survey Data)
Customer Complaint Response
"You need to respond to 50 customer complaints about a product issue. How do you use AI? What's automated, what needs human review, where are the risks?"
Board Meeting Prep
"You're preparing for a board meeting and need to synthesize 6 months of financials into talking points. Time pressure. High stakes. How do you approach it?"
Competitive Analysis
"A competitor just launched a similar product. You need comprehensive competitive analysis by tomorrow morning. What's your process?"
Novel Hiring Challenge
"You're hiring for a role you've never hired for before. You don't know what good looks like. How can AI help you avoid costly mistakes?"
Discussion Structure
Collect Audience Suggestions
What would you try first? Validate all suggestions—there are many valid approaches.
Share My Approach
Here's what I would do and why. Walk through the reasoning, not just the steps.
Compare Trade-offs
"Sarah's approach is faster, mine gives more control—depends on your situation and risk tolerance."
Appendix: Backup Exercise - Real-Time Build-Along
Best for: Demonstrating the iterative nature of AI work. Time: 20-25 minutes. Energy level: High.
How It Works
The audience collectively directs what to ask AI. Collaborative prompt building—audience suggests, I type, we iterate on output together. Shared ownership of the result.
Good Starting Points
Draft An Actual Email
"Let's draft an email someone here actually needs to send. Who has something real on their to-do list?"
Analyze A Real Problem
"Someone describe a problem you're facing. Let's work through it together and see what insights emerge."
Create A Decision Framework
"What decision are you struggling with? Let's build a framework to think through it systematically."
Facilitation Tips
Take ALL Suggestions Seriously
Even odd ones. When suggestions conflict, explain trade-offs and let the group decide. Build consensus, don't dictate.
Celebrate When AI Gets It Wrong
"Great, now we learn how to fix it." Bad outputs are learning opportunities. Show the recovery process.
Keep Pace Quick
Don't over-discuss before trying. "Let's just see what happens" is a good default. Build momentum through rapid iteration.
Energy Management
- Type fast, even if there are typos—it keeps energy up
- Narrate what you're doing so people follow your thinking
- When you get stuck, say it out loud: "Hmm, this isn't working, let's try..."
- Let the group suggest pivots when we hit dead ends
Appendix: Backup Exercise - Show Your Work
Best for: Demystifying the process and normalizing struggle. Time: 15-20 minutes. Energy level: Medium.
The Format
Walk through 2-3 real problems I solved recently, showing the mess—dead ends, iterations, frustrations, and eventual breakthroughs. Not the polished result. The messy reality.
Stories to Prepare (Customize With Your Own)
When AI Gave Terrible Output
My first three attempts were garbage. Here's what was wrong, why it happened, and how I eventually recovered. This took 45 minutes, not 5—and that's okay.
Down the Rabbit Hole
I went down a complete rabbit hole for 2 hours before realizing I was solving the wrong problem. Here's what made me realize it and how I pivoted.
Simpler Than I Made It
I overcomplicated this dramatically. The solution was way simpler than my initial approach. Here's what I learned about when to step back and simplify.
What to Emphasize
The Struggle Is Normal
- "My first three attempts were garbage"
- "I almost gave up here, but then..."
- "This took me 45 minutes, not 5—and that's okay"
- "I felt really stupid at this point"
Why This Matters
- Counters the "AI should be instant" expectation
- Normalizes frustration and iteration
- Shows expertise is in the recovery, not the first attempt
- Makes failure feel less personal
"The difference between me and a beginner isn't that I don't make mistakes. I make tons of them. The difference is I've learned how to recover quickly and extract value from the failures."
Appendix: Backup Exercise - Behind the Curtain Architecture
Best for: Those interested in systems-level thinking about AI adoption. Time: 15-20 minutes. Energy level: Medium.
Topics to Cover
Model Selection Strategy
"I use Claude for data analysis, Gemini for document processing, ChatGPT for video generation. Here's why—it's about matching tool capabilities to task requirements."
Workflow Integration
"Here's how AI fits into my actual daily process—not theory, but the real flow of work from initial idea to finished output."
When NOT to Use AI
"These tasks I still do manually and here's why—sometimes human-only is actually more efficient or appropriate."
Architecture Principles
Right Tool For The Job
Don't try to use one AI for everything. Each has strengths. Match capabilities to requirements.
Build Review Checkpoints
Don't trust AI outputs blindly. Design verification into your process at critical junctures.
Humans in the Loop
Keep humans involved for judgment calls, ethical decisions, and anything where being wrong has serious consequences.
Discussion Prompts
- "How would you structure this workflow for your specific business?"
- "Where would you put the human checkpoints in your processes?"
- "What's the cost of being wrong at each step?"
- "What tasks should never be fully automated?"
Appendix: Backup Exercise - Live Debugging
Best for: Teaching failure analysis frameworks. Time: 15-20 minutes. Energy level: Medium.
The Setup
"Who has a story where AI just completely whiffed? Gave you something useless or confidently wrong? Let's figure out why and how to fix it."
The Debugging Framework
Was the Goal Clear?
Vague input produces vague output. If you said "write something good," AI has no idea what "good" means in your context.
Was Context Sufficient?
AI can't read your mind. What background information did it need that you didn't provide? What did you assume was obvious?
Was It the Wrong Tool?
Some tasks need different models. Using image-generation AI for data analysis won't work well. Matching tool to task matters.
Was It a Hallucination?
Did AI make up information confidently? This happens especially with specific facts, dates, or statistics. Requires verification layer.
Were Expectations Realistic?
Some tasks are genuinely hard for AI. If experts struggle with it, AI probably will too. Adjust expectations accordingly.
Common Failure Patterns
- Asking for too much in one prompt
- Missing context that user assumed was obvious
- Requesting precision in areas where AI isn't reliable
- Not iterating when first output was off
- Using wrong model for the task type
- No verification step for critical information
- Expecting perfect output on first try
- Giving up after one failed attempt
"Every AI failure teaches you something about how to work with it better. The question isn't 'why did this fail?'—it's 'what did this failure teach me about my approach?'"
Hiring Resources
Detailed interview questions and frameworks for hiring AI explorers
Appendix: Interviewing for AI Explorer Traits
These interview questions are designed to surface real behaviors and patterns, not rehearsed answers. The "two specific examples" framing is intentional—anyone can have one lucky story, but two examples reveal a genuine pattern of behavior.
Interview Questions By Trait
Persistent
Question: "Can you share two specific examples of when you continued pursuing a task despite multiple setbacks? What kept you going when it would have been easier to quit?"
Listen for: How they describe the emotional experience of setbacks. Do they frame obstacles as challenges or as insurmountable walls? Do they show resilience or do they need external motivation?
Curious
Question: "What's the last new skill or technology you learned, and what motivated you to explore it? How did you get over the learning curve?"
Listen for: Did they seek out learning or was it forced on them? Do they describe learning as exciting or as a chore? How do they handle not knowing something?
Problem Solver
Question: "Can you describe two specific times when you encountered a complex problem, there was no documentation, and you had to solve it with little help from others? How did you approach solving it?"
Listen for: Do they default to asking for help or figuring it out themselves? How systematic is their problem-solving approach? Do they get frustrated or energized by ambiguity?
Tech Savviness
Question: "Tell me about two specific times when you fixed a computer, phone, or other tech problem because you couldn't get support or didn't have time to wait for help."
Listen for: Are they comfortable with technology or intimidated by it? Do they experiment and tinker or immediately call for help? How do they handle unfamiliar interfaces?
Red Flags to Watch For
Warning Signs
- Vague answers without specific examples
- Blaming others for obstacles or setbacks
- No examples of self-directed learning
- Needing clear instructions for every task
- Giving up quickly when things get hard
Green Flags
- Rich, detailed stories with specific outcomes
- Taking ownership of both successes and failures
- Evidence of learning outside work requirements
- Comfort with ambiguity and uncertainty
- Multiple attempts before seeking help
Appendix: AI Spending vs. Manhattan Project
To understand the scale of AI investment, it helps to compare it to history's most ambitious technical projects. The numbers are staggering—and accelerating.
Historical Context
The Manhattan Project—America's crash program to develop atomic weapons during WWII—was the most expensive single-purpose project in American history when adjusted for inflation. Total cost in today's dollars: approximately $30 billion.
That was considered an almost incomprehensible investment at the time. It represented a national commitment of resources unprecedented in peacetime research.
AI Investment Scale
Global AI spending in 2023 was already 5x the Manhattan Project at $154 billion. The projected 2028 spending of $632 billion is 21 times the Manhattan Project investment.
But here's what's even more striking: When I first made this comparison in November 2024, the 2028 projection was $336 billion. In just over a year, that forecast nearly doubled.
If you're wondering whether AI is "real" or just hype, look at where the money is going. Six hundred billion dollars in annual spending doesn't happen for temporary trends. This is a permanent transformation of how work gets done.
Appendix: Disorienting, Blistering Adoption Pace
This chart explains why AI feels so disorienting. It's not just you—the pace of technological adoption has compressed from millennia to months.
The Acceleration Pattern
Ancient Technologies
Farming/Agriculture: ~5,000 years to spread globally
The Wheel: ~2,000 years for widespread adoption
Measured in millennia
Industrial Era
Electricity: 166 years
Telephone: 40 years
Personal Computer: 16 years
Measured in decades
Digital Era
Internet: 14 years
Smartphone: 7 years
AI: Measuring in months
ChatGPT: 100M users in 2 months
Each wave of technology adoption is faster than the last. The disorientation you feel isn't weakness—it's a rational response to an irrational pace of change. Your brain evolved for agricultural timescales, not AI timescales.
Understanding this progression helps contextualize why everything feels so overwhelming. We're experiencing technological change at a pace that has no historical precedent. The rules for how to adapt are being written in real-time—by people like you who choose to engage rather than retreat.
AI Adoption Insights: Workshop Resource Guide