I’m scouring the world for startups in 3 areas right now. If you’re building here, DM me…
AI Data for Construction — I dream of a world where robots make our houses. Houses will be cheaper and better-built than ever.
But to get there, we need data. Lots of data.
For robots to build a house, they need enormous amounts of video of humans building houses.
So I’m looking for startups that are capturing video from construction sites. The easiest way is probably with Meta Ray-Bans, which are already used by several startups to collect video in other areas.
I recently did an investment in Cortex AI, which captures video from factories. But I don’t know of anyone doing this on construction sites.
Bulk B2B Purchases — Walmart can undercut the corner store because it buys in enormous volume. What if the corner store could do that?
I’m looking for startups that are bundling purchase orders from small businesses. These bundled orders could give small businesses the same great prices as the big boys.
AI Cybersecurity — AI does some wonderful things, but it also presents serious security risks.
Hackers can use an AI browser like Atlas for prompt injection. This lets them take over your system and do whatever they want.
I’m looking for startups building security solutions to counter the threats posed by AI.
There’s a lot of money here. If you can meaningfully improve security, enterprises will give you their grandma.
Wrap-Up
One of the most fun parts of this job is dreaming up new startups. When I come up with a new idea, it’s usually not long until I find a startup doing it.
Helping robots build homes, small businesses compete with the big boys, and companies keep their data safe are three of the best opportunities I can think of.
If you want to start a startup but you’re not sure what to do, try one of these ideas. And when you do, shoot me a message!
Chinese AI startup MiniMax is going public this week at a $7 billion valuation. But its model flopped in my testing.
I ran MiniMax through three tests with real world questions. These are harder to game than benchmarks.
Let me show you where MiniMax does well and where it struggles…
Round #1: Coping With Long Winter Nights
It’s cold and dark most of the time here in North Jersey these days. What are the best ways to prevent Seasonal Affective Disorder?
Minimax has some good ideas, like using a light box. I actually have a drawing tablet, which Andrew Huberman recommended.
But Minimax doesn’t cite any sources. How can I rely on this answer?
I’m giving this round a C.
Round #2: The Top Stocks of 2025
2025 was an incredible year for markets. But I’m curious which stocks did the best.
Can MiniMax help?
Minimax gave me a list that looked great and even provided citations.
But when I clicked the source, it was completely wrong. Minimax was looking at the best performing stocks of 2024.
This is a common problem even among good AI models. They don’t know what day it is!
I don’t understand why this happens. It seems like an easy thing to fix.
I’m giving this round an F.
Round #3: Handicapping the 2028 Election
We’re seeing a lot of dissent in the Republican Party these days.
Secretary of State Marco Rubio is raising his profile by handling Venezuela. Meanwhile, former Congresswoman Marjorie Taylor Greene has attacked Trump for intervening overseas.
With these different factions forming, who will get the nomination in 2028?
Let’s ask Minimax…
Vance is the clear frontrunner. MiniMax likes Rubio for vice president.
Minimax made an interesting point: if one of them failed to get the presidential nomination, the other would likely get vice president. It could wind up Rubio-Vance, not just Vance-Rubio.
Minimax cited quality sources, including The Hill. With original thinking and good sourcing, I’m giving this response an A!
Wrap-Up
MiniMax earned a C overall in my testing.
Its responses are inconsistent. One will be excellent, the next useless.
If you’re relying on these outputs to power your application, Minimax just isn’t reliable enough. If you want something open source, Kimi or DeepSeek are better choices.
China is about to launch its first AI model IPO, Zhipu AI. Investors may be excited, but Zhipu’s product is weak.
China’s first IPO of an LLM startup, Zhipu AI, will start trading Thursday. The IPO is expected to value Zhipu at nearly USD $7 billion.
This morning, I ran it through a series of tests. I found Zhipu to be well behind top American models.
Let me show you where Zhipu falls short…
Round #1: Learning About AI Chips
NVIDIA and AMD introduced new AI chips at CES this week. How do these chips differ?
Zhipu gave a strong answer, emphasizing NVIDIA’s better software.
But I would have liked to see Zhipu cite better sources. It mostly cited Substacks and Yahoo Finance articles rather than technical specs.
I’m going to give this round a B+.
Round #2: Zhipu the Personal Trainer
I’ve done strength training twice a week for years. Every quarter, I do a deload week to give myself some rest.
But is that enough?
Zhipu recommends taking a deload week every 6-8 weeks instead. But it doesn’t cite any sources.
How can I rely on this answer?
When I run the same query through Grok, it answers faster and cites 25 sources. That’s the bar for an AI model today.
I’m giving this round a C.
Round #3: Zhipu the Sleep Coach
I just got back from a wonderful trip to Wisconsin. I saw friends and family over the holidays, which made me really happy.
There was only one downside: my sleep. The system I have dialed in at home just isn’t the same on the road.
Let’s ask Zhipu how to get my sleep back on track…
Zhipu gives some interesting ideas, like getting light in the morning. But the sources it cited were all in Chinese.
Zhipu is supposed to be bilingual in English and Chinese, so citing sources in the wrong language is a serious problem.
I’m going to give this round a C as well.
Wrap-Up
Overall, Zhipu earns a C+ in my testing.
Its answers are decent, but sourcing is weak. It’s hard to tell if Zhipu’s answers are correct or not.
These responses would have been excellent two years ago. But today, Grok and Gemini blow Zhipu out of the water. Zhipu is also behind Chinese models like Kimi and DeepSeek.
I had fun playing with Zhipu, but I won’t be using it again any time soon. It’s just too far behind.
This is the definitive biography of Michael Jordan. I’ve read several other books on MJ, but none digs as deep.
I loved playing basketball as a kid, but I wasn’t that good. Still, I looked up to Michael Jordan.
If you want to excel in anything, Jordan will give you the blueprint.
The Rest of the List…
For the remainder of the list, I grouped books by theme.
If you have a strong interest in investing, biotech or fiction, feel free to jump to those sections!
Business and Investing
Hetty — Hetty Green was one of the most important investors at the turn of the last century. Called the “Witch of Wall Street,” she worked alone and amassed a fortune worth billions in today’s money.
Gambling Man — How Masayoshi Son came from obscurity to be the richest man in the world, lost a fortune, and made it back.
The New Tao of Warren Buffett — “Cryptocurrencies are like rat poison squared.” That quote alone was worth the price of admission!
Damn Right! — This biography of Charlie Munger gives great advice, like the importance of avoiding self-pity.
Warren Buffett Speaks — “… we like great companies with dominant positions, whose franchise is hard to duplicate, and has tremendous staying power or some permanence to it.“ This series of quotes from Buffett break down his investment philosophy.
Memos from the Chairman — Ace Greenberg was the chairman of Bear Stearns in its heyday. If they had stuck to his simple advice of serving the customer and being frugal, they’d still be around today.
Common Stocks and Uncommon Profits — This slim volume from 1958 is one of Buffett’s top book recs. Much of his approach of investing in high-quality companies at reasonable prices comes from this book.
eBoys — How Benchmark hit eBay and became one of the top firms in Silicon Valley.
Talent — Tyler Cowen and Daniel Gross explain how to spot top talent.
Grinding It Out — Ray Kroc’s account of building McDonald’s.
The Default Line — Go behind the scenes as the EU deals with the fallout from the financial crisis. I ordered this book from England years ago. Since then, I’ve read it three times, often near market peaks.
The Courage to Act — Ben Bernanke’s account of steering the Fed through the financial crisis.
Boomerang — Michael Lewis tours the crisis-struck West in the wake of the financial crisis.
Population Plunge Meets AI Revolution
One Billion Americans — Matthew Yglesias explains that for the United States to stay number one, we have to grow our population big time.
Empty Planet — Exploring the fertility crisis from Korea and Japan to the developing world.
Homegrown — The story of Oklahoma City bomber Timothy McVeigh. One surprising fact: had the GM plant in his hometown not closed, he might have become a normal factory worker like his father and grandfather.
Janesville — When Janesville, Wisconsin’s GM plant closed, the town never truly recovered. AI is going to cause a lot more Janesvilles.
Industrial Society and Its Future — The manifesto of the Unabomber. I find most of what Kaczynski says to be off base, but it does get me out of the techno-optimist echo chamber.
Breakneck — Dan Wang explains that although China is catching up to the West, its future is bleak.
Politics
Original Sin — How Joe Biden and the Democrats lost the presidency.
I was laying on the couch, eyemask on, earplugs in, trying to imagine the next great startup. “What about collecting data from factories to train robots?” Well, I found it…
Cortex AI gathers video from factories and workshops. It sells that data to robotics companies to help train robots.
To teach robots to work in factories, we’re going to need tons of video. But there isn’t much video available from real facilities.
Robotics companies are forced to use data from labs and simulations. That data isn’t realistic.
Several startups collect video from inside people’s homes to feed to AI models. Cortex’s approach is different — they are focused on industrial data.
Getting Real World Data
Cortex AI is building the marketplace for industrial robotics data.
Factories and workshops make money by providing video. In businesses with tight margins, this additional revenue stream is huge!
Robotics companies love this data because it’s from the real world. This lets their robots learn faster.
The Perfect Founder
The founder, Lucas Ngoo, has one of the most impressive backgrounds of anyone I’ve invested in.
Lucas was the co-founder and CTO of Carousell, a marketplace for used goods in Southeast Asia. Carousell was backed by Sequoia India and is valued at $1.1 billion.
Lucas’s background is perfect because he’s already built a super successful marketplace.
Carousell is in a very different market — used goods vs. AI data. But a marketplace is a marketplace, and Lucas has seen this movie before.
What About Jobs?
Robots will replace some humans in factories. This is inevitable.
Imagine if we had refused to adopt tractors because they would replace farm workers. Almost everyone would still be doing backbreaking labor on farms for very little money.
Technological progress has some cost in the short run. But in the long run, we are way better off.
Wrap-Up
I’m delighted to have a little slice of Cortex AI’s recent seed round!
It’s rare that I see a fantastic founder building in just the right market. When I do, I’m itching to place a bet.
This investment showed me the value of daydreaming.
I learned this technique from studying Cyan Banister and James Simons, two top investors. They’re both known for spending hours just thinking.
I spent hours imagining big opportunities. When I saw Cortex AI, I knew I’d found what I was looking for.
If I hadn’t prepared my mind by daydreaming, I might have missed it!
If you work in robotics, check out Cortex AI’s data and take your system to the next level. And if you run a factory, look into selling your data!
This is the last post of the year. I’ll see you guys on Monday, January 5th.
Ever feel lost on how to improve your product? Here are three ways to get the customer feedback you need…
Send an Amazon Gift Card. Customers are busy. Getting them to respond to your messages is hard.
But there’s nothing wrong with a little bribery!
If you offer folks a $50 Amazon gift card for a 15 minute feedback call, you’d be surprised how many people will do it.
Be sure that you put the offer right in the email subject line so they’ll see it.
Ask for Feedback Inside the Product. Why not meet your customers where they already are: inside your product?
Put a little link inside it that lets them offer feedback. Pop-ups are annoying — just put the link on the side or bottom of the screen.
Just let them type free text. Don’t make them answer a long survey.
Send Some Swag. If you have their address, why not send them a little company swag?
Make sure it’s something high quality that they’ll actually want to use. A well-made coffee mug, a butter soft hoodie, or a high quality hat.
Make your logo discreet. You don’t want to turn people into a billboard.
Inside the package, you can put your phone number. Tell them you’re really looking forward to hearing their thoughts about the product and you want to make it better for them.
When you give someone a gift, they want to reciprocate. It’s human nature.
Wrap-Up
Sending gift cards and swag to a few dozen customers will probably only cost you a couple thousand dollars. The feedback they give you could be worth millions.
Your existing customers have the secrets you need to get your company to the next level. You just need those customers to share them with you.
If you have a little money kicking around, using it to get customer feedback is one of the best investments you can make.
Here’s how a typical angel investor loses $250,000:
Start with $250,000 bankroll.
Gives $100k to founder.
“We’re launching soon!”
A year later, still not launched.
“We need more money.”
Gives another $100,000 to founder.
Six months later, product still not launched.
“We’ll be ready to launch any minute! We just need a little more money.”
Angel is afraid of losing his $200,000 investment.
Gives founder the last $50,000.
Company goes bust.
$250,000 vaporized.
This is what happens for most angels. They only invest in one company. The founders are not technical, and the product never launches.
Their money goes up in smoke.
Here’s how we can do better…
Focus on Launched Products
I never invest in a software product that’s pre-launch.
Especially with vibe coding, anyone can launch a product. If they haven’t done so, this tells me everything I need to know about them. They are not motivated or skilled.
At least 90% of startups will never launch a product. If I simply remove those startups from consideration, I take 90% of the potential zeros out of my portfolio.
Invest in Builder Founders
Why don’t products launch? Most of the time, it’s because the founders don’t have the necessary skills to build a product.
The rule that YC uses is to back teams of two to four founders, at least half of them technical. That’s as good a rule as any.
Every day I see early-stage startups with a dozen people on the team. Few, if any, are builders. What the heck do these people do?
Focus on people who can actually build software. It’s awfully hard to have a pizzeria if nobody knows how to make pizza.
Spread Your Bets
We’ve eliminated the non-builders that are never going to launch a product. But even for builder founders that manage to launch and get some customers, the startup game is brutally difficult.
So we need to do the opposite of the typical angel — we need to spread our bets.
Take that $250,000 bankroll. The better way to divide that is 50 bets of $5,000 each.
If I make between 30 and 50 investments, I should be able to hit at least one unicorn. But if I only make one investment, my odds of hitting a unicorn are practically zero.
In our business, returns come from a handful of companies. You have to hit a unicorn in order to get a good return.
Diversification gives us an opportunity to hit something big.
Wrap-Up
If you’re thinking about getting into angel investing, be aware that you could lose every cent.
You worked hard for this money. Don’t hand it to just anybody.
Focus on builder founders with launched products. Spread your bets widely.
If you do that, you have a chance to beat the odds and nail a killer investment.
The Allen Institute for AI recently released its most powerful model ever: Olmo 3. Olmo is cheap to train, making it perfect for anyone training their own model.
Olmo 3 is 2.5 times more efficient to train than Llama 3.1 based on GPU-hours per token. Olmo is also much more transparent than other models.
Most so-called “open source” models, like Llama or DeepSeek, are really just “open weight.” They give you the weights, but they won’t show you how they trained the model (“model flow”) or give you the training data.
Olmo is 100% open.
I popped up Olmo 3 on OpenRouter this morning. Let’s see what this thing can do!
Round #1: Researching Marketplace Startups
Yesterday, I was researching the background of a really incredible founder. He co-founded a unicorn marketplace.
This got me thinking…how many people in the world can say that?
Olmo gave a solid estimate on the number of unicorn marketplace startups and the number of co-founders. Its best guess is that there are a couple hundred people in the world who have co-founded a billion-dollar marketplace.
Olmo’s response was solid, but it didn’t cite any sources. This makes it hard to rely on the answer.
Competing models like DeepSeek have nailed sourcing. At this point, it’s table stakes.
I’m giving this round a C.
Round #2: Learning About the Fiber Buildout
We are in the midst of a massive AI buildout. Big tech companies are spending a fortune on GPU’s, data centers, and training data.
How big can this get?
For comparison, I asked Olmo to tell me about the fiber buildout in the 1990s. This time, I used its “Search” function to see if we can get some citations in the response.
Olmo estimates that Telecom spent $50 billion in the US on the fiber build-out. But it doesn’t cite any sources, so it’s hard to know if these figures are accurate.
I got vastly different numbers from Grok and Gemini, both around $500 billion.
Again, the lack of sourcing makes Olmo’s outputs a lot less useful. I’m giving this round a C as well.
Round #3: Planning My Christmas Trip
I’m beginning to feel a little sorry for Olmo. So, I decided to give it something a little easier.
I’m flying back to Wisconsin on Christmas Eve. The airport is likely to be packed to the gills.
Can Olmo give me some tips for surviving the chaos?
Olmo reminded me that my Global Entry membership lets me use the TSA Pre-Check line. I forgot all about that!
That was a very helpful tip that could save me hours. I’m giving this round an A!
Wrap-Up
Overall, Olmo 3 scores a C+ in my testing.
Olmo is definitely not the most powerful AI model out there, or even the most powerful open source model.
But the outputs are decent, and it’s very cheap to train.
Try using Olmo in your application and see if it gives you the outputs you need. If so, it may make sense to switch.
The most powerful model isn’t always the best choice for your application. Sometimes, cost matters.
I tried to sell software to an AI simulator. I wound up crushing my quota and getting promoted! My biggest lesson: you can’t sell to every buyer the same way.
Every day, I’m evaluating founders’ sales skills. The founder is the first salesman for the startup.
If they’re good at it, the company can scale fast. If not, the company will struggle.
But there’s one little problem: I’ve never actually done sales myself. So when Jason Lemkin from SaaStr launched Vaporsell.ai, I had to try it!
Here are three buyers I sold to and what I learned…
John: The Busy Buyer
“Give me your value prop in 30 seconds.”
John is on his way to the airport and doesn’t have much time. He wants me to get straight to the point.
Some might consider John rude. But I live in the NYC area, so I’m used to people being blunt.
I actually sold to John pretty successfully. I just had to keep my responses short and specific.
When a response required more time, I offered to send more detailed information to his email. In the end, I was able to close a large contract.
I actually like buyers like John. Getting to the point makes things easier for both of us.
But John wasn’t my favorite customer. No, that’s Gavin…
Gavin: The Gullible Buyer
“That’s amazing! What else can this software do?”
Gavin is like a breath of fresh air. He’s already really excited about the product.
It would be hard to screw this sale up. Everything I say, Gavin likes.
If you have PMF/market pull, you’ll meet a lot of Gavins. You’ll be the leader in your area and people will be predisposed to buy.
What worked for me here was to not overthink it.
Gavin wants the product. Just give quick and cheery answers to his questions, and soon he’ll be signing on the dotted line.
I wish they were all Gavins! But then where’s the challenge, right?
Speaking of challenge, it’s time we meet Sarah.
Sarah: My Most Hated Buyer
“I don’t buy software from amateurs.”
I was feeling pretty good about my sales skills. Then I met Sarah.
Sarah is a VP at a major startup. Her deal could have put me way over my quota. I needed this!
But Sarah is the buyer from hell. She’s prickly, rude, and super technical.
Even though she was fake, I could feel my cortisol rising.
I would try to answer her questions, but she would always say my answer wasn’t specific enough. My technical sophistication simply didn’t match hers.
Try as I might, I was never able to sell to Sarah. She actually called me an “amateur” and hung up.
I’m glad I never have to speak to this jerk again!
But Sarah taught me something useful. I don’t need every single buyer.
I never sold her a dime’s worth of software, but I still crushed my quota. Take that, Sarah!
Wrap-Up
Using Vaporsell gave me new empathy for anyone in sales. Selling is friggin’ hard.
My biggest takeaway: you can’t sell to every buyer the same way.
Some buyers, like John, want a quick rundown. Others, like Sarah, want you to go deep.
In the end, the main thing that helped me was practice.
This makes me want to prioritize founders that have experience in sales. They’ll be better at closing customers and at fundraising, too!
Try Vaporsell and improve your sales skills. It’s better to lose a fake deal than a real one!
Tremendous 