23 July Weekly Reading Summary: Where will be the opportunities for AI startups? What’s next after when Moore’s law reaches its limits?
As per my previous post (https://www.linkedin.com/feed/update/urn:li:activity:7218091211893485568), I will aim to share interesting reads that I've come across in the previous week (or weeks) with links embedded in square brackets [ ]. The purpose is to keep us up to date on the latest developments globally and try to translate as to what it means for us in SEA. I will group them by topics, and share short commentaries on potential implications.
LLM: How should we think about AI and where will be the opportunities for AI startups?
According to Sequoia US [1], AI would be analogous to the cloud in the internet era. That’s because like the cloud, AI enables new distribution and business models, but it is not a consumer front end application on its own.
The interesting part about Sequoia’s comment was, they think the opportunity for startups would be to pick a services industry and AI-enable it (instead of building an AI version of existing software, like how startups back in 2010s took on-prem software and built cloud-equivalents).
If I were to guess the reasons behind this thinking, it is probably:
(a) services are large categories and for the first time, we can see significant productivity uplift akin to what steam engine did to physical world;
(b) AI still cannot complete 100% of the work, in fact it’s far from it. On the other hand, customers are only concerned about the quality, speed, and cost of the deliverables. Therefore, services enabled by AI can potentially provide meaningful competitive advantage on those 3 fronts;
(c) software incumbents are already building GenAI products (eg. Box, Salesforce) into their platforms and therefore there’s less surface areas for startups to attack. That said, most of these incumbents have yet to proof to users are willing to pay for the new LLM features that they have introduced (eg Salesforce missed its first revenue target since 2006 inspire having introduced Einstein)
At Vertex, we are also believers in the thesis of AI-enabled services as outlined in the article by our Managing Partner Ben Matthias [2].
However, personally I do think that we are just in the early inklings and more digital creation industries (eg. Video production, 3D modeling, CAD design) could be disrupted as AI can reduce the marginal cost of digital content creation to nearly zero (like how the internet reduced content distribution marginal cost to nearly zero). Any process of creation, that can be represented as digital files, could potentially be touched by AI. Therefore, GenAI should be viewed as a capability amplifier for the digital creation tools, just like how engines/motors have done for the physical tools.
With motors, carriages turn into cars, bicycles turn into motorbikes, sewing machines go from manual to powered, sickles become combine harvesters, and the list goes on. Notice that the problem statements remain unchanged, but the introduction of motors unlock productivity by a quantum leap. The same may happen in software and we are already seeing it with Canva, Figma or HeyGen.
The next frontier is probably creation of 3D objects (think AutoCAD, Dassault, Unity, Cadence and others). Interestingly, funding has started to into foundation models that are working on “spatial intelligence” such as World Lab by Fei-Fei Li [3]. New research such as Gaussian Splat and Shape of Motion [4] are trying to turn 2D images into 3D/4D objects. Notably, Nvidia is also pouring significant resources into Omniverse [5]. The confluence of all these, plus the emergence of Apple Vision Pro, just made spatial computing so much more real and exciting.
Robotics is also seeing a resurgence with interesting developments leading to companies such as Figure and Skild receiving significant funding [6, 7, 8].
However while everyone is getting too bullish on GenAI, it is important to balance that with a healthy dose of skepticism. I found that this Twitter thread by Bill Gurley has asked really good questions [9]. Will LLM continue to scale linearly, exponentially, or sub-linear? What are LLMs’ limitations eg. Can it do deterministic work? Does it handle numbers well? Is it cost reduction or revenue generation? Small vs large model? 100% worth a read.
Also, will data training run out, or is synthetic data going to solve it? [10]
How about SEA? For any companies to succeed, I believe they need to have an advantage in certain dimensions. For SEA companies, my current hypothesis is (a) its products/services need to an element of localisation that the large foundation model companies may find it difficult to expand due to the difficulty in getting the data, or not an immediate priority eg. local dialects, for instance Singlish or Manglish; or (b) its products/services will need to have real-world physical integration (i.e. an online-offline element). Drawing an analogy with the Vietnam War, SEA is unlikely to develop the best stealth bombers or drones to win the war. Instead, SEA companies can leverage their in-depth local knowledge to develop innovative solutions, much like the Vietnamese built tunnels.
If you are startup building these products, feel free to reach out.
https://youtu.be/8mqNaOuRdkA?si=WqJXysQ7vaHrtVjt
https://shape-of-motion.github.io
https://www.youtube.com/watch?v=-eiL9HzvJh0&t=152s
https://twitter.com/bgurley/status/1811864315177455695
Semiconductor: What’s next after Moore's law reaches its limits?
So, it was announced that Apple has secured all of TSMC’s 2nm capacity for the coming year. That’s not surprising as every year Apple is always at the frontier of the most advanced nodes.
But what was more interesting is the mention that Apple plans to adopt SoIC (System on Integrated Chips) packaging for its M5 chip in 2025, which will involve stacking multiple chips with different functions to form a compact three-dimensional structure, called advanced packaging.
According to McKinsey, advanced packaging which was introduced around year 2000, is now gaining significant momentum as the next breakthrough in semiconductor technology [2]
Despite Moore’s law, which in 1965 posited that the number of transistors on a microchip would double every couple of years, node advancement is now reaching its limits. As a result, technical advances on the front end of chip manufacturing are slowing, and the economically viable maximum size of a die, and thus its performance, are becoming more limited. New approaches in back-end technology that combine multiple chips offer a promising solution. Advanced-packaging techniques that have arisen over the past two decades—including 2.5-D, 3-D, fan-out, and system-on-a-chip (SoC) packaging—promise to fill the void by supplementing the wire-bonding and flip-chip technologies of the previous half century [2].
In the past, 2.5D and 3D were used more in high performance computing due to the cost involved. Now we are starting to see this going into consumer technology. TSMC has also announced their Foundry 2.0 strategy, which is intended to better leverage its advanced packaging capabilities for market expansion by encompassing a wider range of activities compared to the traditional wafer manufacturing industry [3]
In other semiconductor news, continuing from the discussion around Nvidia’s growth and profit durability, OpenAI is reported to be developing their own chip as well with Broadcom [4] (as per last week’s post, other tech giants have already started work on their own chip). And following the footsteps of Tesla, important Chinese EV players Nio and Xpeng are also moving away from Nvidia for its self-driving chips [6].
Others news: Robotaxis in China, Google's acquisition of Wiz and others
Robotaxi is already rolling out in China. This is just the start and from pure-play robotaxi companies (ie. companies that don’t manufacture and sell EVs). Soon we will see more EV manufacturers going into this game. https://www.scmp.com/tech/policy/article/3270693/shanghai-put-driverless-robotaxis-roads-despite-pushback-taxi-drivers-wuhan?utm_source=feedly_feed
Google’s potentially most expensive acquisition to date: https://techcrunch.com/2024/07/15/googles-kurian-approached-wiz-23b-deal-could-take-a-week-to-close-source-says/
But the management team declined the offer. Gungho https://techcrunch.com/2024/07/22/wiz-walks-away-from-googles-23b-acquisition-offer-read-the-ceos-note-to-employees/
Economist, Lecturer, Writer (SCMP, BT, CNA, Asia Times), Entrepreneur | Views are my own, reposts are not endorsements
5moGreat initiative! Looking forward to reading more
Building Companies, At the Intersection of Insights, Marketing & Technology, Growth mindset, Humanity
5mogreat content . thank you