1 of 42

Getting Value from A.I.

Dr. Jon Krohn

Co-Founder & Chief Data Scientist at Nebula

New York University — November 25th, 2024

2 of 42

3 of 42

Getting Value from A.I.

Dr. Jon Krohn

Co-Founder & Chief Data Scientist at Nebula�Best-Selling A.I. Author

Ok so you know who I am now. Why I am here in front of you?
In this talk, I’ll detail how you can derive commercial value from A.I.
Why does A.I. matter to you? Each passing year, A.I. adoption becomes increasingly critical commercial but things are only just starting to get exciting now. Here’s why…

PROBABLY REMOVE ALL OF THIS: “Throughout the talk, I’ll be using visual approaches to convey the critical points of how particular A.I. techniques work and then I’ll associate each of those techniques with real-world applications that will enable Hg and its portfolio companies to overcome short-term uncertainty and drive long-term value with A.I. A.I. is critical to the success of your company and if you’re not thinking about it already, you need to now because your competitors are. “

4 of 42

1875

1997

To provide you with context around how the speed of technological change — particularly w.r.t. A.I. — presents an unprecedented opportunity, let’s take a quick journey through the life of Jeanne Calment.

This is Jeanne Calment
Born in France in 1875.
Telephone patented: 1
Lightbulb patented: 4
Steam turbine invented: 9
Wireless communication with radio waves: 20
Zeppelin and Wright Brothers fixed-wing flight: 20s
Except general anaesthesia, essentially all modern medicine was developed after her birth

X-rays: 20s
"gene" term was coined: 30s
Penicillin's antibiotic properties discovered: 50s

In terms of information tech:

First programmable computer: 60s
Transistor: 70s
Email: 90s
Basis of modern Internet: 100th birthday in 1975

Jeanne passed away in 1997, 122 years young, making her the oldest person whose age has been verified by official documents.
The change that Jeanne witnessed is incredible: telephones, lightbulbs, engines, wireless communication, the splitting of the atom, computers, digital cameras, cell phones, the Internet, tremendous automation in agriculture, industry and the home, the coining of "gene" through to the sequencing of genomes, the eradication of many fatal but once-commonplace diseases, antibiotics, antidepressants, organ transplants, brain imaging, and so much more.
This dramatic technological progress corresponds with dramatic, quantifiable improvements to the quality of human life on earth. For example:

Two centuries ago, vast majority on earth regularly hungry

Today, fewer than one in ten in extreme poverty

Two centuries ago, small portion of adults could read

Today, vast majority can

Two centuries ago, nearly half of all children died before reaching the age of five

Today, only a few percent do

I couldn't find data stretching far enough back in France, but in neighbouring Germany the average life expectancy when Jeanne was born was 38; by the time of her death, expected lifespan had more than doubled to 77.

5 of 42

1875

1997

today

2140s

?

6 of 42

Our minds quickly adapt to change so it’s easy to lose sight of how crazily rapidly our ways of working and ways of living change due to technology.
Indeed, while Jeanne Calment might not have been able to predict innovations a century out, she probably could have made decent predictions a decade out.
Today, that would be much tougher.
Due to a confluence of factors — such as:

More human brains on the planet than ever before
A higher proportion of human brains than ever before having the luxury of spending time solving cognitive problems instead of, say, tilling the fields
Those human brains being connected in real-time such that a new discovery, academic journal article, or piece of open-source software can be made instantly available to billions of other brains

…thanks to these opportunities for inventiveness and problem-solving, the rate of technological progress is much more rapid today than ever before.

7 of 42

8 of 42

Getting Value from A.I.

Deep Learning A.I.
Replaceable vs Augmentable Tasks
How to Implement A.I. Successfully

TELL WHAT TELL

45 min talk / 15 min Q&A
While A.I. developments in the coming years will be so rapid and paradigm-shifting that the exact timeline and specifics are unknowable [pause], over the next 45 minutes I’ll be detailing how you can nevertheless harness A.I. today [pause] to drive significant… commercial… value. [pause]
While you’re listening over the next 45 minutes, I’d like you to reflect on how the value-enhancing A.I. applications I describe relate to your career aspirations.
Then, you’ll have 15 minutes in which you can ask me anything, such as questions that are specific to your circumstances.
To provide you with food for reflection on specific A.I. applications that are relevant to you, I’ll cover:

What Deep Learning is and how its driving the A.I. revolution (WHAT IT IS)
Examples of work that is replaceable with more affordable A.I. versus work that can be augmented and done more efficiently with A.I. (HOW IT WORKS)
Concrete tips on how to implement A.I. projects successfully (THE IMPLICATIONS)

TELL WHY LISTEN

These topics are important to you because:

It will help you select your portfolio companies
Successful adoption of A.I. will drive efficiencies within your companies
Thanks to those increased efficiencies, you’ll be able to exit at higher valuations [SMILE]

9 of 42

Getting Value from A.I.

Deep Learning A.I.
Replaceable vs Augmentable Tasks
How to Implement A.I. Successfully

10 of 42

Artificial Intelligence

Click to edit Master title style

11 of 42

Artificial Intelligence

Machine Learning

Click to edit Master title style

12 of 42

Traditional ML vs Deep Learning

Deep Learning

Click to edit Master title style

13 of 42

Artificial Intelligence

Machine Learning

Representation Learning

Click to edit Master title style

14 of 42

Artificial Intelligence

Machine Learning

Representation Learning

ANNs / Deep Learning

Click to edit Master title style

15 of 42

Click to edit Master title style

16 of 42

playground.tensorflow.org

17 of 42

Artificial Intelligence

Machine Learning

Representation Learning

ANNs / Deep Learning

Machine

Vision

Click to edit Master title style

18 of 42

Click to edit Master title style

19 of 42

playground.tensorflow.org

20 of 42

Click to edit Master title style

21 of 42

Traditional ML vs Deep Learning

Deep Learning

Click to edit Master title style

22 of 42

Click to edit Master title style

23 of 42

Artificial Intelligence

Machine Learning

Representation Learning

ANNs / Deep Learning

Machine

Vision

Deep Reinforcement Learning

Click to edit Master title style

24 of 42

Click to edit Master title style

25 of 42

Artificial Intelligence

Machine Learning

Representation Learning

ANNs / Deep Learning

Machine

Vision

NLP

Deep Reinforcement Learning

Click to edit Master title style

26 of 42

Artificial Intelligence

Machine Learning

Representation Learning

ANNs / Deep Learning

Machine

Vision

NLP

Transformers

Deep Reinforcement Learning

Click to edit Master title style

27 of 42

Click to edit Master title style

28 of 42

28

Getting Value from A.I.

Deep Learning A.I.
Replaceable vs Augmentable Tasks
How to Implement A.I. Successfully

29 of 42

30 of 42

31 of 42

32 of 42

A reminder that this section of the talk is about replaceable vs augmentable tasks.

The replaceable tasks — i.e., repetitive ones — are now in the rear-view mirror.

Tasks that can be augmented with A.I. include those that require creativity, leadership, and decision-making:

ChatGPT

Ideate and draft copy for contracts / business plans / pitch decks / emails / marketing material / social media posts… I may have even used it for drafting elements of this very presentation! [Share]
Summarizing articles / web pages / complex research in shorter format / simpler terms
Planning / Suggesting solutions to specific problems

Elon Musk software engineer plausible work
How to increase revenue or profitability in specific circumstance
How to retain customers
What product feature to consider building next
What A.I. products you could be building given your data, industry, other unique defensible (moat) aspects of the firm

Motivation / Therapy

OpenAI Codex

[video game]
Text-to-code in dozens of programming languages, e.g., through GitHub CoPilot
ChatGPT / GPT-3 can do this too but Codex is specialized for it

Of course, we’ve all been augmenting our intelligence with Google for years and its capabilities of surfacing relevant, well-summarized expert recommendations to specific business questions improves with DL NLP Transformer examples

Google especially useful still today to confirm ChatGPT hasn’t confidently lied to you

Augmentation is where I personally see the most opportunity in the B2B SaaS space.

Here are concrete examples from my own B2B SaaS company, Nebula:

Search for talent (plus open LinkedIn profile) or sales leads without TA expertise… NLU instead of keywords… same principle applies to any kind of document search/recommendation that could be improved by using language vectors of smeared meaning as is done with modern DL/Transformer-driven NLP than with keywords (Syntheia)
Generate relevant images, videos, or language automatically in-platform (JD builder)
Use LLMs (e.g., GPT-3) in place of model development (suggest relevant job titles, employers, educational institutions)

33 of 42

33

Getting Value from A.I.

Deep Learning A.I.
Replaceable vs Augmentable Tasks
How to Implement A.I. Successfully

34 of 42

35 of 42

Approach / Philosophy

1. Start with simple, explainable models:

Heuristic
Regression
Semi-supervised
Late-stage only: DL, Gradient-boosted trees

2. R&D Culture

Acceptance of high expense and high project risk
The best projects are high-risk but can be balanced with lower-risk projects that have more finite, predictable timelines… and these are more common thanks to Foundation Models (covered in Augmentation)
Lab meetings, tons of whiteboarding, journal club, going to conferences
Executives/investors w/o hands-on A.I. experience aren’t sending every blog post they read on A.I. to DS team… and executives’ gut instincts on A.I. are often way off

My speech coach told me “nobody will ever ever understand this… do not include it”... but based on everyone I’ve been chatting to today, I think

2. How in order for an A.I. project to be profitable, it can’t be open-ended — you’ve got to structure it to have a specific revenue goal from the onset.

Binary classifiers are easiest to track success with (each quadrant of confusion matrix can be associated with a specific income or cost)

36 of 42

Data Engineers
ML Engineers
MLOps
Data Product Manager
Software Engineers
UI/UX Specialists
QA Leads

People (ensure this is as high-level as possible… 2-3 minutes max)

Data scientists (not BI/analyst) are on the team

Especially in uncertain times, versatile DSs that can be “full-stack”:

engineering own data
developing models
deploying
pitching/selling

Finding versatile, full-stack DS talent like this is hard, but that’s why talent-sourcing platforms like my Nebula company exist

Lots of critical roles that cannot typically be filled by DS in order to ensure successful A.I. product:

Data Engineers
ML Engineers
MLOps
Data Product Manager: know importance of designing data/A.I. products with real data
Back-end & front-end software engineers
UI/UX Specialists with A.I. experience
QA Leads with A.I. experience

Data science leadership is available (can be cut if running low on time):

A.I. R&D is challenging; most projects fail, so someone who can steer efficiently is critical
Lots of “gotchas”: selection bias, measurement bias, biases against particular demographics, Simpson’s paradox, statistical significance, feature drift… that a production model becomes a “high-limit credit card of platform technical debt”
Needs to have weight directly upon C-suite in order for team to get resources they need to compete projects that drive value (profit center not cost center)
Can prototype A.I. product externally if needed:

Tribe
Merantix Momentum

37 of 42

38 of 42

Code versioning (Git)
Data and model versioning (MLFlow)
Containerization (Docker, Kubernetes, Kubeflow)
Code review
Automated testing (Jenkins)
Data pipeline orchestration (Dagster, Luigi)

39 of 42

39

Getting Value from A.I.

Deep Learning A.I.
Replaceable vs Augmentable Tasks
How to Implement A.I. Successfully

jonkrohn.com/talks

All right, so those are my top tips for getting value from A.I.

The key points that I’d like you to take away:

Introduced you to the nuts and bolts of what deep learning is. Until just the past few years and, in many cases, the past few months what I’ve shown you today wouldn’t have been possible. So you don’t need to be concerned if you haven’t been injecting A.I. into your commercial processes yet, but now is the time to get going on it.
I detailed examples of B2B SaaS-relevant tasks that — thanks to deep learning — can be replaced or augmented by machines with far shorter implementation times than ever before and the tasks can then be done more quickly, more cost-effectively, and more accurately, driving efficiencies for your business. All of you are exceptionally busy people and so some of the ideas I shared today might even help you automate or augment parts of your workday, thereby driving efficiencies for you individually too.
Your appetite whetted for implementing as much DL as possible, in the final section I provided you with practical roadmap to implementing A.I. successfully so that you can waste as little time, energy, and capital as possible getting A.I. projects off the ground.

The slides for today are available on my website at jonkrohn.com/talks so that you can be sure not to miss any of these tips and you’ll also be provided access to the recording through Hg, my YouTube channel, or perhaps both.

40 of 42

1875

1997

today

2140s

?

41 of 42

41

Q&A

42 of 42

Stay in Touch

jonkrohn.com to sign up for email newsletter

linkedin.com/in/jonkrohn

youtube.com/c/JonKrohnLearns

twitter.com/JonKrohnLearns