Transcript (edited by A.I.)
Note: this was transcribed by Otter.ai and cleaned up by OpenAI’s ChatGPT. It therefore may not be 100% accurate, but it’s pretty darn close and saved us hours of time 🙂
So, the title of my presentation: AI, AI, O. That’s a joke—like E-I-E-I-O. I could make an Old McDonald joke here.
I’ve got about 25 minutes and 214 slides to get through. You think I’m joking? No, really—I do have 214 slides. So I’m going to move fast and give you some different ideas, like Wayne was talking about: an unusual set of perspectives.
My background is in civil engineering. I got my Bachelor’s and Master’s in civil engineering and practiced for about a decade. Then in 1997, this thing called the internet came along—you may have heard of it. It’s going to be big one day.
Through a very odd series of circumstances, I found myself in the internet space. I’ve partnered with, sold to, and worked for a bunch of big tech companies: Google (who we heard from this morning), Oracle (who acquired one of my previous companies), Facebook, Twitter, Apple, and so on. I’ve been in and around big tech for a long time—maybe too long. Some might argue that’s not healthy.
I did have a PE license, but it was impractical to keep it up since I wasn’t actively practicing. I let it go—and then came full circle, back to civil engineering. Not because I didn’t like engineering, but because the internet was just too big to ignore. Now, 25 years later, I look up and think: “Wow, what an interesting spot to be.”
Over the past 24 months, since founding Beacon, I’ve had probably 100,000 conversations with people across DOTs, consultants, the transportation community—maybe some of you in this room. I really appreciate that perspective.
Now, I’ll tell you up front: this is not going to be nearly as technical as some of the other AI presentations, like the one next door on sensors. But it will be the only one featuring:
- The Pope in a puffer coat
- An old man hugging a unicorn
- The elephant in the room
- PalmPilot
- Terminator and Skynet
- How to fry an egg (as determined by AI)
- The world’s longest cow
- A really cool heat map
My goal is to provoke some thoughts—just like Wayne said. Let’s get a discussion going, whether in this session, in the hallway, or even during cornhole tonight. I’ll be around and I love talking about this stuff.
Three Conclusions
In preparing this talk, I’ve come to three conclusions:
- AI is bad.
- AI is good.
- AI is misunderstood.
All three of those statements can be true—and I would submit to you that they are true. If you’re trying to figure out, “Is this good or bad?” the answer is: yes. It’s all of the above.
The Internet Parallel
There’s a striking parallel between the early days of the internet and where we are today with AI.
Back in 1999:
- People had AOL accounts.
- We used 56k dial-up modems. Remember the gurgly-boing modem sound?
- Yahoo was basically a directory—and it was a multi-billion-dollar company.
- AltaVista was an amazing search engine—until Google came along and ate their lunch in 1998.
The internet evolved quickly:
Dial-up → Broadband → Search (Google) → Video (YouTube) → Mobile → Social → Cloud
Back in 1999, trying to have a conversation about “cloud” computing? Not happening. It wasn’t even a concept.
Check this out: In 2000, Palm (the PalmPilot company) was worth more than Apple, NVIDIA, and Amazon—combined. If you ever want a reminder of how fast things change…
Trillions of Dollars and Engineering Economics
We have no idea where AI is going to take us, but there will be hundreds of billions of dollars invested and trillion-dollar companies formed.
Why is NVIDIA worth $2 trillion? Because they sell valuable chips everyone wants. And while money isn’t everything—it’s how business keeps score.
In engineering school, I took Engineering Economy and… that was it. No finance, no market dynamics. But to understand what’s going on in AI, you’ve got to at least understand the economics.
So—AI is Bad
Let’s talk about the dark side of AI.
Copyright:
There’s a fantastic Atlantic article titled Generative AI is Challenging a 234-Year-Old Law. We’re going to have multiple Supreme Court cases. What does “imitate” mean when done by a machine?
Deepfakes:
The Pope in the puffer coat? Totally fake—but looked real. Now, with Pixel phones, you can alter an image and literally make someone appear to be a drug user. That’s dark. It used to take Photoshop skills. Now it just takes malice.
Fake Content:
A Wall Street Journal reporter built a propaganda machine for $105 using AI. Created fake content, images, and could have monetized it.
Absurd AI Mistakes:
- Sauce vs. dressing? AI said: “Sauces add flavor. Dressings protect wounds.” Technically… true?
- “A room with no elephant” became: a room with an elephant.
- “Vegetative Electron Microscopy” is now cited in 20+ scientific papers—because two unrelated columns of text lined up.
- World’s longest cow? AI stitched together two cows into one 5.4-meter cow.
- Unicorn horn through the head of an old man in an otherwise stunning Michelangelo-style image.
- Step-by-step images of frying an egg? Pure chaos. AI doesn’t know how to cook—it’s just tossing symbols and patterns at the wall.
AI Can’t Do Math:
One of my favorite charts: a heat map showing large language models failing basic multiplication. Red and yellow = failures. Multiplying two 20-digit numbers? Forget it.
AI Can’t Read Clocks or Calendars:
Ask AI what time’s on a clock face? Wrong.
What day of the week is Christmas? Only two models got it right—and possibly by luck.
Ethical Concerns:
- Anthropic (AI company) tells applicants not to use AI to submit resumes.
- Lawyers + AI = 10,000 lawsuits in 100 jurisdictions? Not hard to imagine when AI lowers the cost of filing to near-zero.
Thoughtful Contrarians:
- Gary Marcus
- Missy Cummings
- Philip Koopman
Especially Cummings and Koopman on autonomous vehicles. Remember: a person put their Tesla in full self-driving mode… and played a game on their phone. Don’t do that.
So—AI is Misunderstood
We fear Skynet. But remember:
- It’s fiction.
- It’s flawed. (Terminator had a 120-year battery and a heat sink for power? Really?)
- A really pissed-off single mom beat Skynet—twice. So… I’m betting on the humans.
Words > Numbers in Transportation
Back to transportation. Huge technical efforts on sensors and AI, yes—but I’ll argue that words might matter more than numbers.
Employees are drowning in information:
“I’ve got documents in email, SharePoint… I don’t even have time to think.” —DOT staffer in Alaska
We’ve been working with Jennifer Portanova and team at NCDOT. Great tools like TEPPL are available, but they’re like digital filing cabinets—you still have to read and synthesize everything.
We built a proof of concept for them to actually synthesize and summarize. That’s the exciting part. It’s not magic, and it’s not perfect, but it can save time and reduce tedium.
Real Use Cases
At Georgia DOT, we’re helping TMC operators using a chat-like interface. SOPs are hundreds of pages, with critical info on geography, data, and procedures.
Some real test queries:
- “Where’s the nearest rest area to Tift County?”
- “What’s the criteria for contacting GEMA/HS?”
Even silly ones like:
“Fire engine en route to fire encounters light traffic at a traffic light. What do the SOPs say?”
SOPs are word-based, vague, complex. But LLMs can help surface information.
More Use Cases
We’ve worked on document discovery and estimation spreadsheets too.
One engineer asked:
“Can you create a decision matrix for school speed zones based on curb & gutter, time of day?”
AI spat out a good-enough answer in 10 seconds.
“That’s 50–80% correct,” he said.
Exciting? Yes. Terrifying if unchecked? Also yes.
You can’t just assume AI = correct. But you also can’t ignore the power.
Spreadsheets? Remove the words, and they become meaningless. AI can help decipher large volumes of semi-structured data—but only with context.
It’s Harder Than It Looks
Slapping a search box on top of a database? Easy. Making it work with large PDFs, documents of varied structure and formats? Much harder.
You also need domain-specific knowledge. You don’t ask a literature professor about structural engineering—and vice versa.
AI Agents
Agentic software = compound AI tasks:
“Go to five sites, compare products, check shipping, buy the best gift.”
Google’s Project Mariner does this. Others will too.
One agent completed a 50-step workflow—months of work—in 30 minutes.
But not everyone agrees. Ezra Klein of the New York Times says AGI is almost here. Others say he’s dead wrong.
Final Thoughts
Even I can’t keep up. But try small steps:
- I changed my default search engine to ChatGPT.
- Very jarring at first. But now, 2 out of 3 times, I get better results than Google.
This quote really stuck with me:
“I want AI to do my laundry and dishes so I can do art and writing. Not the other way around.”
Another:
“AI sees the past, not the future. It sees your data trail, not your human story.”
And finally:
Integrity, trust, patience, humility—these things still matter.
Humanity still matters.
So let’s update my three conclusions:
- AI is bad.
- AI is good.
- AI is misunderstood.
But what if AI could… ?
Thank you.
