Hello!
How is the new year treating you?
Delhi is COLD. Temps last night approached 4 C. Which is fine for a man with mountain genes like me. My wife’s south-Indian genes face a different challenge. I enjoy winters. It is not unusual to see me running wearing just a thin cotton shirt. My wife calls it cold when the mercury hits 20 C. In the interest of marital harmony, I have decided to maintain one room at a constant 24 C.
Winter also means I have been consuming copious amounts of ice cream - in fact, I may be one of the few weirdos on the planet who eat more ice-cream in the winters than in the summers. Anyway, I digress.
I can think of few things that have created as much excitement as ChatGPT3. It seems to be the missing link that takes one from information (Google search result) to knowledge (here is how you write a python code to build a web server).
Let’s take a look at what it means for Tech Giants:
1. Innovations - disruptive and sustaining
One of my favourite books is called The Innovator’s Dilemma by Clayton Christensen. When addressing types of innovation, Clayton writes:
“Most new technologies foster improved product performance. I call these sustaining technologies. Some sustaining technologies can be discontinuous or radical in character, while others are of an incremental nature. What all sustaining technologies have in common is that they improve the performance of established products, along the dimensions of performance that mainstream customers in major markets have historically valued.
Disruptive technologies bring to a market a very different value proposition than had been available previously. Generally, disruptive technologies underperform established products in mainstream markets. But they have other features that a few fringe (and generally new) customers value. Products based on disruptive technologies are typically cheaper, simpler, smaller, and, frequently, more convenient to use.”
The trouble is - this is often a hindsight-is-20-20 kind of categorization. a couple of examples:
PCs were a massive disruption. No one thought PCs would amount to much. IBM outsourced the CPU and the OS to Intel and Microsoft respectively. PCs became faster to the point of killing IBM’s original mainframe business. Intel and Microsoft dominated a market which dwarfed pretty much everything that came before it. Microsoft then fumbled big time because it thought of a mobile as a smaller PC. Apple on the other hand succeeded because it did not see the mobile as a smaller PC.
On the other hand, Cloud and SaaS are sustaining innovations. Small SaaS startups identified areas where the incumbents weren’t offering high quality solutions and ran with it. Sure, the makers of servers were hit but the explosion in the need for data center hardware more than made up for their losses, and in most cases, allowed for far fatter margins (with the removal of B2C-ish costs)
But one of the more surprising things that comes from the history of innovations is this - disruptive innovations are nearly always offerings of new entrants in a market. That said, those new entrants aren’t necessarily small, nimble startups. Some of the biggest winners in previous tech “epochs” have been large companies leveraging their strengths to move into, or create new spaces. Office 365 anyone?
The Innovator’s Dilemma was published in 1997. This was the year Kodak was at its peak. It’s share price was at its highest ever. The company firmly owned the present (the film business) and it had invented the digital camera. Nothing could possibly shake this behemoth.
Except - business model. Kodak made fat margins and ungodly revenues off films. Digital Cameras on the other hand did not need film. Kodak’s management was thus very incentivized to convince themselves that digital cameras would never be a big business. In some ways, they were right. Between digital cameras being invented in 1975 and their sales surpassing film camera sales in 2000, 25 years passed. What finally killed Kodak in 2012 was the phone camera. However, in the 15 year period from 1997 to 2012, Kodak made a ton of money and paid out billions in dividends.
The larger point is that even a highly innovative company can end up in the trap of defending its business model and display an ostrich like denial of the writing on the wall.
2. Google vs AI
Once ChatGPT3 and DALL-E were opened to the public, one of the first things the smarter pundits locked on was search (and the future of Google). If I can ask a question and get a complete, relevant paragraph on the fly - what does that mean? Are the “ten blue links” irrelevant? What does it mean for the paid search ad model? and Even if Google could build the same thing (it already has, but it seems to not be as polished as ChatGPT3), would it be too little and too late?
The answer seems to be a lot less binary than most of the internet would have us believe. There are a lot of different contexts one might mean when one enters a search query. and while I will not go into the details of how search engines are made, knowledge graphs and ontologies, what Google has been doing is to improve its knowledge graphs. It has been building answers instead of 10 links for a lot of the categories. However, this work of building is done category by category. As an example - meanings of words from dictionaries.
But, here is the wrench in the works - answers can often be different things.
Sometimes you know exactly what you want and just want the link. Sometimes you are looking for a range of options (reviews for hotels in Madrid), sometimes you are looking for structured data. Google has been getting better at these answers. The way Google (or any other search engine) does this is to assign engineers and ask them to build an answers product for a category. “Flights”, “Sports scores”. and so on. Almost always, it is done by hand, category by category. The biggest impact of LLMs under the hood may be the fact that they may allow for generation of these knowledge graphs for:
a far wider set of questions,
with much better (higher) levels of abstraction
at scale
without as much manual interaction.
Each of these points indicates a significant improvement in performance/cost.
Look, Google has a ton of things going on for it. Cloud, YouTube, Gmail - all are only getting better.
However, Search has always had a human as the ultimate arbiter of quality. Google provides links. A human decides whether that link is relevant or not. This same interaction extended to ads. You can pay for clicks, but the user decides if your ad was good enough by clicking on it.
Over the last few years, Google’s strategy has been to cram even more ads into search, YouTube and so on, especially on mobile interfaces. What that has been doing is increasing the bloat on Google’s offering.
In the specific context of search, this observer believes that Generative AI represents an incredibly disruptive innovation, albeit one which will be easily dismissed by Google’s managers as “not good enough”. And therein lies a huge threat - ChatGPT3 is not good enough - yet. It will be, eventually. On the other hand, Google Search seems to be on the road to become even more bloated and hard to use. So yeah, Google has Cloud and YouTube as dominant models which only have ever rising trajectories.
Search seems to be at its peak and can only go down from here.
3. Apple vs AI
Apple has a history of investing in open source tech mostly driven by the imperative to make sure its walled garden plays well with everyone else. A great example of that is the WebKit browser engine.
Apple’s efforts in AI have largely been proprietary - e.g. its engines for face recognition in photos and voice recognition. Then, Stable Diffusion came around and Apple pounced. Stable Diffusion is a remarkable achievement not because of its being open source, but because the model is remarkably small. Small enough to run on an iPhone. Here is an announcement from Apple last month:
“Today, we are excited to release optimizations to Core ML for Stable Diffusion in macOS 13.1 and iOS 16.2, along with code to get started with deploying to Apple Silicon devices…
One of the key questions for Stable Diffusion in any app is where the model is running. There are a number of reasons why on-device deployment of Stable Diffusion in an app is preferable to a server-based approach. First, the privacy of the end user is protected because any data the user provided as input to the model stays on the user’s device. Second, after initial download, users don’t require an internet connection to use the model. Finally, locally deploying this model enables developers to reduce or eliminate their server-related costs…
Optimizing Core ML for Stable Diffusion and simplifying model conversion makes it easier for developers to incorporate this technology in their apps in a privacy-preserving and economically feasible way, while getting the best performance on Apple Silicon. This release comprises a Python package for converting Stable Diffusion models from PyTorch to Core ML using diffusers and coremltools, as well as a Swift package to deploy the models.”
At this point, it seems safe to assume that this is just the jump-off. Apple has been shipping a neural engine in its chips for years. It is very likely that future Apple chips will be tuned to Stable Diffusion as well, with the model itself built into future OS releases. So, now, you have generative AI built into Apple devices. It does not have to be as good as Dall-E or MidJourney - it has to be local. In my estimation that means the walls of Apple’s garden got just a little higher. You don’t necessarily need the kind of backend chops that apps like Lensa need.
Sure, there is a big world outside of Apple devices - but the local AI capabilities will certainly have an impact on the larger market for centralized compute/services.
4. Amazon vs AI
I think Amazon’s Cloud business will benefit far more than its marketplace from AI. AWS sells GPUs in the cloud. GPUs are needed to run inferences.
Each time we ask ChatGPT or Stable Diffusion to generate something for us, the model is applied to run a ton of computing on a GPU in the cloud. This is called inference. This has a marginal cost. It’s not large, but it is not zero either, and it adds up. That marginal cost for ChatGPT is AWS revenue.
More inference → More Compute → More revenue for Amazon.
Now, Amazon’s prospects here depend on demand forecasting. Not having enough GPUs will lead to wasted opportunity, having too much is predictably having an idle asset. I think Amazon would rather err on the side of oversupply than undersupply. End usage will depend on marginal cost - and Amazon is one of the few companies that understands the impact of scale and low pricing.
Marginal costs in AI are an underappreciated challenge in the opinion of this observer. Think about it - the reason why ChatGPT is the biggest breakout of the year is because it free to the end user. That is because Microsoft gave OpenAI a sweetheart deal. If this wasn’t the case - ChatGPT would have been (and was) limited to researchers - like LAMDA is.
5. Meta vs AI
Meta has had a couple of crappy years of late. Apple’s App transparency Tracking (ATT) was a crippling blow. A great analysis of Meta’s business prospects can be found here. Meta has been making huge capex in its data centers. Historically, these data centers were about compute. Meta’s ad model is rather deterministic, as are its models used to serve content. All that needs huge compute capability. However, increasingly, it seems Meta is going after probabilistic models that are going after both targeting but also to figure out conversions.
The key determinant of success here will be if Meta is able to tune1 its models to individual users on the fly. However, this is expensive and Meta needs to figure out how to do it cost-effectively.
What I am more interested in seeing though is how things like image and text generation work for (or against) Meta in the long term. If the endgame is AI generated content, it makes for a very interesting development for Meta’s advertising business.
AI has become good enough to the point of generating copy / images and A/B testing it. There are few companies better than Meta at making this type of capability available at scale. More so because Meta’s advertising is almost always about the top of the funnel or discovery - telling folks about a product or service or app they did not know previously existed.
This means misses and low conversions, but also a ton of opportunity for experimentation and iterative learning. Yes, AI driven generation may have large marginal costs, but those marginal costs are drastically lower than a human.
6. Microsoft vs AI
Microsoft is the best placed of the lot. While it has invested almost $10bn into OpenAI and has some very, very interesting deal terms, the bigger impact will be in terms of Azure and Bing.
Microsoft will be integrating GPT into Office apps (much like the github copilot). It remains to be seen if it will be useful or a gimmick. Whatever it is, I hope it will not be like Clippy. I just hope there isn’t an additional fee for it. As an example of what it could look like, take a look at Rows. Excel with AI - yes please.
Azure stands to gain expected benefits from its GPU cloud, much like Amazon above. However, I think that the biggest gainer will be Bing.
See, Bing has mostly been an also ran. Sure, it contributes a fair bit of revenues but it is a small fraction of the dominant player’s market share. Incorporating ChatGPT into Bing will likely create an extremely credible competitor to Google search. However, Google’s moat is its distribution - something Microsoft has historically been very bad at in the context of search. That said, I think that the historical lack of a large user base set in its values/behaviour for Microsoft will likely free it up to reimagine search2.
What I would love to see is the posterboy disruption victim create a product to disrupt Google and laugh all the way to the bank.
7. Interesting reads
Schools are blocking OpenAI website.
Panos Panay (Windows “Boss”) on ML being integrated into all aspects of Windows and Office. Notice the parallels to Apple strategy.
Apple is doing AI read audiobooks.
Talking to dead legends using AI. I liked talking to Carl Sagan a LOT.
Housekeeping
As always, I look forward to hearing from you. If you liked this post, pls feel free to share this or subscribe to this newsletter using the links below. I try to write a 1000-2000 word essay once every two/three weeks.
for the want of a better word
Google will not likely touch its cash cow - search. Microsoft will make only limited changes to its cash cows - windows and office. Interesting dynamic? Do we see them gunning for each other’s dominant businesses?