MBI Daily Dose (July 09, 2025)
Companies or topics mentioned in today's Daily Dose: Constellation Software, AGI timeline, Waymo vs Tesla
I listened to an interesting AlphaSense interview of former Head of M&A at Constellation Software (free trial link here). Let me share some bits and pieces that stood out to me.
Constellation doesn’t seem to be bothered by copycats:
From Constellation standpoint, Constellation acknowledges that there are companies trying to emulate its success. Constellation itself came up with the term copycats.
Constellation seems to think that there's a bit of the pie for everyone, but it just focuses on doing what it does best.
If Constellation is competing with a copycat, especially if it's a start-up, all the copycat has to offer at that moment is capital.
Constellation will acquire just about anything for the right price. It'll look at smaller businesses, it'll look at restructurings, it'll look at businesses that haven't been growing that much or that aren't that profitable.
Interestingly, it is the internal competition that may be more of a headache than facing competition from copycats:
If you work at one of Constellation's operating divisions, such as Volaris, for example, the largest one, your main competitor is actually another operating division of Constellation, such as Vela or Topicus or Harris or whatever it is, because you're competing for the same deal. It's very cutthroat within the organization itself. That's one of the pet peeves of Constellation's employees.
Suppose I work at Volaris and I'm pitching Constellation to you, and I'm interested in acquiring your business and tell you all the great things Constellation does, say we hang up the phone, and then all of a sudden, you might get another email from another operating division of Constellation with a completely different name and then another call from yet another one, and you might be thinking, "What is going on? Why are these three different people from different organizations claiming to be affiliated with Constellation?" That does happen, stepping on the toes of your peers. It creates sometimes a very convoluted process, and it requires a lot of handholding and mitigating internally.
The discussion on impact of AI on CSU was quite nuanced. The expert talked about how CSU basically stripped out R&D spending following acquisition and quickly get to pretty high margin. So, most of CSU’s software is understandably subpar quality, but many of their customers still live in the stone age while the rest of the world is heading towards AGI. As a result, AI may not have much of an impact for such niche software anytime soon:
AI is probably the most talked about topic now. It only started appearing as a priority on Constellation's agenda as of last year. It's not to say that Constellation was dismissive of it…There are several ways of looking at it. There's Constellation's perspective as an operator, there's Constellation's perspective as an acquirer, and then obviously you have the client's perspective, which impacts Constellation.
Constellation's default stance when it acquires a business like that is to almost instantly scrap or get rid of all the R&D expenditure to save on money and help that business generate cash flow, especially in the early years.
A lot of Constellation's platforms, if you were to rank the platforms according to their functionality and how modern they are, Constellation's average platform, I would say, is Tier three, Tier four, Tier five, whereas some of the private equity owned ones which they pay up for are Tier one. Constellation's are typically legacy software for people or clients who don't want to pay up or can't afford to pay for a Tier one software. What that means is that they're probably less prone to want to pay up for an AI model that's unproven at this stage.
However, one intriguing bit in the interview was the expert mentioned CSU does have some “blue-chip” customers who are in financial services, legal, or insurance industries which tend to be more tech forward than the typical VMS software customers. When asked about the mix of revenue between smaller and blue-chip customers, the expert said 50-50. My sense is CSU is likely to be largely immune from AI threat when it comes to smaller niche VMS software customers (it may not be worth attacking such niche segments), but it will probably face competition higher than anticipated in “blue-chip” segment.
Dwarkesh wrote a good post explaining why he has extended his AGI timeline. One of the reasons he mentioned is AI’s lack of continual learning. Dwarkesh writes:
How do you teach a kid to play a saxophone? You have her try to blow into one, listen to how it sounds, and adjust. Now imagine teaching saxophone this way instead: A student takes one attempt. The moment they make a mistake, you send them away and write detailed instructions about what went wrong. The next student reads your notes and tries to play Charlie Parker cold. When they fail, you refine the instructions for the next student.
This just wouldn’t work. No matter how well honed your prompt is, no kid is just going to learn how to play saxophone from just reading your instructions. But this is the only modality we as users have to ‘teach’ LLMs anything.
It is important to note that Dwarkesh only extended his AGI timeline to 2032 which maybe “conservative” among his peers in SF, but it is quite an aggressive timeline anywhere else. He did make an interesting point that beyond 2030, the probability of reaching AGI diminishes every year we don’t get there given various constraints of scaling compute:
AGI timelines are very lognormal. It's either this decade or bust. (Not really bust, more like lower marginal probability per year - but that’s less catchy). AI progress over the last decade has been driven by scaling training compute of frontier systems (over 4x a year). This cannot continue beyond this decade, whether you look at chips, power, even fraction of raw GDP used on training. After 2030, AI progress has to mostly come from algorithmic progress. But even there the low hanging fruit will be plucked (at least under the deep learning paradigm). So the yearly probability of AGI craters.
In addition to "Daily Dose" like this, MBI Deep Dives publishes one Deep Dive on a publicly listed company every month. You can find all the 60 Deep Dives here. I would greatly appreciate if you share MBI content with anyone who might find it useful!
It is not easy to find a somewhat neutral and sobering analysis on Tesla. Timothy Lee had a lot of nice things to say about Tesla’s robotaxi launch, but had some interesting angles while comparing and contrasting against Waymo. Tesla vs Waymo has been a hot topic, and this piece did a really good job in explaining why the debate is far from settled:
Waymo’s vehicles are only available in a handful of metropolitan areas. They don’t operate on freeways. And Waymo has remote operators who can intervene if the vehicles get stuck. For years, Tesla fans argued that these precautions showed that Waymo’s technology was brittle and unable to generalize to new areas. They claimed that Tesla, in contrast, was building a general-purpose technology that could work in all metropolitan areas and road conditions.
But in a piece last year, I argued that they were misunderstanding the situation.
“Tesla hasn’t started driverless testing because its software isn’t ready,” I wrote. “For now, geographic restrictions and remote assistance aren’t needed because there’s always a human being behind the wheel. But I predict that when Tesla begins its driverless transition, it will realize that safety requires a Waymo-style incremental rollout.”
Last month, Waymo published a study demonstrating that self-driving software benefits from the same kind of “scaling laws” that have driven progress in large language models.
“Model performance improves as a power-law function of the total compute budget,” the Waymo researchers wrote. “As the training compute budget grows, optimal scaling requires increasing the model size 1.5x as fast as the dataset size.”
When Waymo published this study, Tesla fans immediately seized on it as a vindication of Tesla’s strategy. Waymo trained its experimental models using 500,000 miles of driving data harvested from Waymo safety drivers driving Waymo vehicles. That’s a lot of data by most standards, but it’s far less than the data Tesla could potentially harvest from its fleet of customer-owned vehicles.
So if bigger models perform better, and bigger models require more data to train, then the company with the most data should be able to train the best self-driving model right?
I posed this question to Dragomir Anguelov, the head of Waymo’s AI foundations team and a co-author of Waymo’s new scaling paper. He argued that the paper’s implications are more complicated than Tesla fans think.
“We are not driving a data center on wheels and you don’t have all the time in the world to think,” Anguelov told me in a Monday interview. “Under these fairly important constraints, how much you can scale and what are the optimal ways of scaling is limited.”
Current Portfolio:
Please note that these are NOT my recommendation to buy/sell these securities, but just disclosure from my end so that you can assess potential biases that I may have because of my own personal portfolio holdings. Always consider my write-up my personal investing journal and never forget my objectives, risk tolerance, and constraints may have no resemblance to yours.