Starting this week, I will aim to share interesting reads on tech that I've come across in the previous week (or weeks). I will group them by topics, and share short commentaries on potential implications. If you find it useful, appreciate if you can give a shoutout. Meanwhile, any feedback on how to improve is highly appreciated as well. Enjoy!
Autonomous vehicle: Robotaxi may come sooner than later
- With FSD12 that uses end-to-end neural network training using real drivers' videos, Tesla is showing remarkable progress in covering edge cases. The system also better mimics the way humans drive as it is trained on Tesla's top drivers from its ~6m cars.
- This was not possible with the previous rule-based algorithm approach, which was not scalable.
- I believe most automakers (eg. Xpeng) will move towards this direction. Those with significant user base (and therefore video footage) and end-to-end tech stack such as Tesla, will have an advantage. The next group that are in this category will be the Chinese automakers such as Nio, Xiaomi, Huawei and Li Auto.
- Also, notable that Xpeng is potentially going to ditch Lidar for a pure camera-based approach, similar to Tesla.
Gen AI: Demand for AI products/services have yet to catch up with the amount of capex investments
- This calls into question the durability of such capex spend. For context, the four tech giants (Alphabet, Amazon, Meta and Microsoft) have pledged to spend close to a total of $200bn this year, mostly on data centers and a big chunk of it went to Nvidia. Meanwhile, the largest pure-play AI company (excluding Tiktok or Meta) to date, OpenAI, is reportedly to have only generated around $3b in revenues.
- At the same time, these 4 tech giants are working on their internal AI chips with the aim of reducing reliance on Nvidia and improving margins. (eg. Amazon has Tranium, while Google has Tensor).
- With Nvidia reported having 60% of their revenues concentrated in top 10 clients, this may explain why CEO Jensen and his management team have cashed out some shares.
- Meanwhile at foundation model level, it is highly competitive and most models don’t have pricing power unless you are at the frontier level like ChatGPT-4o or Claude 3.5. This problem is further exacerbated by Meta’s open source Llama 3, which costs only $0.70 per million token output (for comparison, ChatGPT-4o costs $15 per million token output).
- Will this be a race to the bottom? We are starting to see some of this In China, where there are over 100 foundation models and the industry is going through a price war. However, training foundation models is expensive with frontier models potentially costing up to US$1b, according to Anthropic CEO Dario Amodei. Therefore only those with cash, or can raise cash, can survive this game.
- The price reduction is great for application developers. If you are a startup, there’s no better time to experiment with LLM. And it is important to start doing so as GenAI can be a great competitive weapon. ? ?
Gen AI: AI-generated digital production, from software development to video creation, is getting real. However, we are still far from being able to fully replace workers in real world setting
- After speaking to a seasoned CTO with hundreds of engineers working under him, I’ve learned that it is possible to develop a simple mobile app using GenAI today. He didn’t know Swift, but could develop the app in 2 days using Swift.
- However, automation of large scale software development is still not possible. This is true even with state-of-art multi-agent models funded by top tier VCs in the US.
- This is because a large part of an engineers’ work is communication - to discuss and align with product managers and designers. An interesting stat quoted by 1 of the big 4 tech companies stated that software developers only spend 1.5 out of 8 working hours coding.
- While developers have unanimously agreed that co-pilot have helped, it has been challenging to quantify the uplift in productivity. The CTO had no way to quantity and had to resort to doing an employee survey to figure. As you may expect, the feedback varies across different individuals.
- So the learning point for me was, while software development seems to be one of the industries on the verge of disruption from an outsider standpoint (especially when it is played up by the media), it may not be so from an insider’s point of view. Z
- Same may be true for other industries. Best to speak to end users to learn about the potential and limitations.
- That said, while we are not there yet, the trajectory towards better multi-agent task-based LLMs is definitely happening and we will have to monitor closely on what are the new breakthroughs.
Others: LLMs still have limitations due to tokenisation, and India’s stock market is bubbly with 1 in 10 stocks valued at over 100x PE due to strong retail interest.
Chief Learning Officer @ Momentum Leadership | MBA
7 个月Interesting topics. Can't wait to dive into those reads. Thanks for sharing the insights