Headlines: Sustainability of AI investments & how it compare with dot com era, Robotaxi's in San Francisco, New deep learning architecture called TTT to reduce compute

AI Businesses

• Sequoia capital story about the gaping hole in AI revenues coupled with fast depreciating value of the investments already made into AI infrastructure, somewhere someone is going to lose money - very likely the early investors (Sequoia Capital)

• Goldman Sachs report on the AI boom in the US market and potential risks of over-investment (Goldman Sachs)

• Financial time report on Robotaxi services in San Francisco and economics surrounding the success (Financial Times)

• Techcruch article on a new class of deep architecture know as Test Time Training (TTT) which will vastly improve the training efficiencies and reduce the enormous power and compute requirements of today's Transformer architectures (Techcrunch)

• Anything that will solve supply side constraint on AI chips / computing platforms will be welcome. Worthwile news from OpenAI & Broadcom on building a new AI chip. (Bloomberg)

Resources from this week

• How does fundrasing work with VCs and what are the incentives for Limited Partners and General Partners (Explorepodcast)

• What to expect when 'AI scaling' plateu? AI scaling - ie the efficacy of LLMs are driven by the scaling lows of underlying inputs - data, model& hardware. How will infrastructure look like after AI scaling (Nandu Anilal)

• How do we define 'Agents' what are the range of possibilities in Agentic systems. Andrew Ng speaks about in this article. (Andrew Ng)

Sobering thoughts

• Labor exploitation in digital industries is one of the taboo topics - with $0.3 an hour as on online worker to to $800/month outsourced employee they all tend to be below minimum wage but still competed out by a global workforce running into millions. The range of tasks of photo labelling to social media moderation of in appropriate content (Realstories)