Business: Chinese startup WeRide IPO in the US, AI pricing strategies in SaaS offerings, Top 10 SaaS KPIs in charts, Deep Mind's robots beating humans in table tennis, Elon's X.ai releasing Grok-2 and its rise to the LLM leaderboard
Technology: 'AI Scientist' - an LLM writing research papers for $15, Falcon Mamba LLMs matching Transformer LLMs in evaluations, Anthropic becomes the first LLM provider enabling prompt caching saving 90% in opex for subscribers
Resources: Software 2.0 - Andrej Karpathy's 2017 article, Best Deep Learning book - Dive into Deep Learning
AI in Businesses
China’s autonomous vehicle startup WeRide seeks US IPO at $5B valuation. WeRide has commercial operations in Beijing, Singapore and the UAE already. Will they pause serious competition with the US players, particularly Tesla's FSD planned to launch in October 2024 (Rebecca Bellan, Techcrunch)
How AI is priced into SaaS offerings. It's a difficult balance to be made between outcome-based and subscription-based pricing strategies. In addition, pricing for AI features could be applied to the core product OR as a premium tier OR as an add-on. a16z article discusses the key trends in this space. The trend is still subscription-based until the companies can find a way to measure outcome-based (Sarah Wang et al, Andreessen Horowitz)
Ten best SaaS growth charts worth tracking - Revenue / Employee to CAC payback period. (Kyle Poyar, Growthunhinged.com)
Google Deep Mind robots are beating humans in table tennis. Robots are winning against 'beginner' level players and have ~50% success rate against medium-level players! “This is the first robot agent capable of playing a sport with humans at the human level and represents a milestone in robot learning and control,” the paper claims (Brian Heater, Techcrunch)
X.ai released its Grok-2 LLM model with 2 versions - a full-scale and a mini version. The product doesn't come with any free tier pricing. Those who want to use can buy it as a premium feature to an X.com subscription. What is amazing is that the overall ELO score is already very close to the GPT-4o & Gemini models within 1 year of its existence. Grok-3 is expected by the end of the year, more to come! (X.ai blog)
Technology updates from AI
AI scientist who can write research papers from your prompts and develop novel ideas and assertions. The joke of the past is becoming a reality. Each idea is implemented and developed into a full paper at approximately $15 per paper. It's not just writing the paper - AI scientist does idea generation, experimental iteration, paper write-up and even automated paper reviewing. (Skana.ai blog)
Transformers, based on the attention mechanism, are the dominant architecture used in all the strongest large language models today. Yet, the attention mechanism is fundamentally limited in processing large sequences due to the increase in compute and memory costs with sequence length. Various alternative architectures, in particular State Space Language Models (SSLMs), tried to address the sequence scaling limitation but fell back in performance compared to SoTA transformers. Mamba is one such architecture introduced in this research paper. But now, Falcon Mamba, an implementation of Mamba architecture achieved comparable results with top transformer models. (Huggingface blog)
Anthropic introduced prompt caching on its API, which remembers the context between API calls and allows developers to avoid repeating prompts. No other LLM provider has built this capability within their models, so far it was down to the developer to implement caching in their LLM application workflow. The prompt caching feature is available in public beta on Claude 3.5 Sonnet and Claude 3 Haiku, but support for the largest Claude model, Opus, is still coming soon. This is a game changer for LLM app developers and cost savings to the tune of 90% (VentureBeat, Anthropic blog)
There are some interesting financial benchmarks in this article towards the end which are super useful. For example, the below chart indicates the EV/NTM ratio for a set of SaaS startups (Enterprise Value / Next Twelve-month revenue). Other key metrics are - NTM revenue growth, Gros Margin, CAC payback, NRR & Rule of 40 (Zachary Dewitt, Notorious)
Resources
Andrej Karpathy's November 2017 article explains the paradigm shift that has come about with neural network architectures in the world of software. Andrej presciently called it Software 2.0. The article is as relevant today as it was back then (Andrej Karpathy, Medium.com)
Dive into Deep Learning is by far the best book I have across in Deep Learning. Not just that, it is free and available online as a website in itself. The book builds on code examples which you can run right into a runtime. Written by Deep Learning experts and researchers from Amazon, this book is a feat in itself in terms of its theoretical coverage and practical application (Amazon AI)
Comments