20VC Newsletter - 1st Sept 2024
Here are the transcripts and top takeaways from 20VC episodes released this last week!
Monday’s episode with Imran Khan, Founder & CIO @ Proem:
Download the full transcript:
My 6 key takeaways:
The IPO Market is NOT Closed
The issue is companies don’t want to go public because their expectations are too high.
This is the same problem with the M&A market.
When a public company is trading at 20x-30x multiple, they cannot afford to pay 50x multiple to buy a company.
Founders Should Not Obsess On Valuation
A founder’s job is to create a business with sustainable long term value.
Worrying about the business’s value today is the investor’s job.
If the founder is obsessing over valuation, they are not doing their job.
Why Bill Gurley is Wrong About Price Pops
Whenever you bring in new investors, you want them to make money.
You’re building goodwill in the long run by giving them a bit of upside.
When a company goes public they sell a very small percentage.
In the grand scheme of things, the percentage of dilution is pretty low.
It is not a DPI Issue is it a performance issue
People did not write down the bad investments to zero.
Over time people will realize it is not only a DPI issue but also a permanent loss of capital.
The actual number is greater than most people think.
If you put founders first that is BS: LPs have to come first
Your fiduciary duty is to your investors.
Any investor who says they are founder first is either clueless or lying.
I would never give money to those investors.
Why AOL is the Deal that Made Google
They gave AOL a 95% revenue share when their search was powered by Google & a 5% company warrant.
Yahoo walked away from it thinking Google will never make money.
It was the most transformative deal because it put Google on the map.
Wednesday’s episode with Arvind Narayanan, Director @ Princeton:
Download the full transcript:
My 6 key takeaways:
We Are Reaching a Stage of Diminishing Returns
Data is becoming a bottleneck because we are using up all the data currently available.
Compute allows smaller models to be built but doesn’t help as much as it used to.
I am highly skeptical that GPT-5 will be as big a leap as GPT-4 was over GPT-3.5.
Is Synthetic Data the Answer?
There are 2 ways to use synthetic data:
Improve the current quality of data we have.
Increase the quantity of data, use 1TN tokens to output 100TN tokens.
#2 will likely not happen, the quality of data is more important than the quantity.
The Paradox That Means the Cost of Compute Will Go Down
Jevons paradox: Cheaper models will get used more, and the total inference spending will actually increase.
Tasks like constantly monitoring emails & rewriting code will get used more.
The cost will be significant in the medium term even with Moore’s law.
Why LLM Evaluation is a Minefield
We’re seeing lots of LLMs with high benchmark scores but they still make extremely common mistakes.
There is so much pressure on benchmark performance that developers are optimizing the models for them.
These benchmarks don’t capture how it’ll be used in the real world.
AI is not a Weapon & It is too late for closed source
AI is getting cheaper & it is getting more accessible to everyone.
If our AI safety approach is to prevent “bad guys” from having it, we’ve already lost.
We need to think about how to use AI defensively in a world where it’s widely available.
Misinformation Is a Social Media Problem Not a Technology Problem.
The hard part of misinformation is not generating it, it's distributing it to people & persuading them.
Social media is the main medium for distributing misinformation today.
Treating this as an AI problem is distracting from all of those more important interventions.
Friday’s episode with Scott Gorlick, Former Head of Strategic Projects @ Uber:
Download the full transcript:
My 5 key takeaways:
Why Uber Beat Lyft
We had ground teams in every market while Lyft operates entirely from SF.
We were working closely with the drivers every day.
The speed & proximity gave us the edge.
Uber Left Lyft Get F****d by Regulation First
We let Lyft launch in new cities first when we were launching UberX.
We had a policy to wait 30 days to see if law enforcement comes after them.
We tried it for 4-5 cities then started launching at the same time.
Uber’s City by City Expansion Playbook
We treated every city expansion like its own startup
We had 4 key hires:
Launcher: MBA, PE, banking profile. Would move city to city
General manager: Similar profile to launcher, acts as “CEO” of the city.
Operations manager: Responsible for overseeing & expanding drivers.
Marketing manager. Responsible for partnerships & expanding riders.
This was incredibly freeing for young 20s-30s to go into a market & create it from scratch.
How the FBI Tried to Screw Uber
We couldn’t operate in SXSW Austin because of 2 rules:
Black cabs had a minimum fare of $55
There was a minimum waiting time.
We introduced pre-booking, made all the Ubers free & compensated the drivers.
The regulators would plant $20 bills in the cars to screw us over.
What Makes Travis Kalanick World Class
He moves incredibly quickly when he sees something that works.
He has a deep level of understanding across product & operations.
His ability to context switch & empower his team is unmatched.
Let us know what your big takeaways from this week’s shows were in the comments below!
Thank you for reading, and don’t miss the great guests we have next week:
Zico Kolter, Professor @ CMU
David Schneider, GP @ Coatue
Thank you for reading 20VC.
This post is public so feel free to share it.
Thanks for reading 20VC! Subscribe for free to receive new posts and support my work.