Business: OpenAI 1o - complex reasoning model, First civilian spacewalk and the cost, open source consolidation in observability, Warren Buffet's 10 stocks for his 90% wealth, Oracle's supercomputer with Blackwell GPU delivering 2.4ZettaFLOPS, Klarna shutting down all salesforce licenses, AI vision interview with Andrej Karpathy

Technology: Lilian Weng's blog on LLM Autonomous Agents, Moonshot projects in the computing hardware industry, RAG framework from Stanford

Long Read: Goldman Sachs white paper on equity investment strategies to counter concentration risks following the AI-led boom in US stocks

Resources: Lilian Weng's blog on the evolution of various technologies and architectures in deep learning and LLM building

AI in Businesses

• OpenAI has released a new model to perform complex reasoning with significant improvements in evaluation scores involving maths, coding, and science research problems. This model is by far the most expensive one from OpenAI with $15/Million input tokens and $60/Million output tokens (~4x GPT-4o) (Official Blog, OpenAI)

• SpaceX enables the first civilian space walk. Jared Isaacman, an American tech billionaire estimated to have paid $200m (£153m) to fellow billionaire Elon Musk for all four seats aboard the SpaceX craft (Nadine Yousif & Brandon Drenon, BBC News)

• OpenTelemetry is creating an open-source architecture in the observability space connecting all key players (Splunk, Snowflake, Databricks, Datadog, Dynatrace, Elastic, Cribl) - the biggest beneficiary will be enterprise customers who could control the observability costs spread across their pipelines (Chris Zeoli, Data gravity-Substack)

• Warren Buffet has 90% of his investments ($280B overall) in these 10 stocks (AAPL - Apple Inc., BAC - Bank of America Corp., AXP - American Express, KO - Coca Cola Co., CVX - Chevron Corp., OXY - Occidental Petroleum, KHC - Kraft Heinz Co., MCO - Moody’s Corp., CB - Chubb Limited, DVA - DaVita Healthcare Partners) (Andrew Lokenauth, X.com tweet)

• It's not just the hyperscalers, Meta & Tesla who are building supercomputers and data centres - the old horse Oracle Corp too in supercomputers! - at that a massive scale - 130K Blackwell GPUs - 2.4 ZettaFLOPS (Anton Shilov, Tom'shardware.com)

• Klarna shuts Salesforce and plans same for Workday - Beginning of the end of SaaS packaged software as we know it? of course, that's too dire a conclusion given its just one company reporting this - that too an agile tech company with the right resources and talent - but what if more companies take to reducing licenses with their in-house gen AI capabilities (MSN.com)

• There is no one like Karpathy when it comes to providing a commentary about what is going on in the AI industry - he shares his assessment of key topics - there is so much packed in this 45min interview - here is a summarised transcript (No Priors podcast, Youtube, Collationist.com)

Technology updates from AI

• Lillian Weng's blog on LLM Powered Autonomous Agents. This is dated by more than a year but still provides the best single document one can get on Autonomous Agents working in collaboration with LLMs (Lilian Weng, GitHub Page)

• Moonshot Approaches Harnessing New Physics is a blog from semiconductor analyst Mackenzie. He discusses physical systems from photonics -> quantum -> in-sensor and end uses from AI data centres -> ultra-low energy edge intelligence - all initiated by start-ups. With 145 start-ups tracked with ~$15B invested in them. (Mackenzie, mackenziemorehead.com)

• DSPy from Stanford has created ripples in the RAG frameworks arena with its innovative modular approach towards prompt generation (much in line with how neural network modules work) and capabilities for optimising prompts from training prompt data sets (just like neural network optimisations). Creating programmability and learning into prompt generation is a significant step forward. (Valliappa Lakshmanan, ex-Google AI Solutions Director, LinkedIn Post)

Long Read

• Goldman Sachs published a paper on equity strategies to counter the market concentration in equities following the AI-led boom in US stock markets. The paper puts things in perspective starting with 17th-century canal companies (railways, telegraph, telephone, radio, personal computers etc in between) to modern-day internet companies. There are some very useful metric comparisons in the report - such as the figure below showing the changes in the top 10 companies in the S&P500.

• Premise: "In equity markets, narratives have the power to attract and direct much-needed capital. However, they can also amplify interest to the point of monopolising investor attention at the expense of other opportunities, leading to unrealistic expectations about future profits and leaving companies vulnerable to a sharp de-rating"

• Problem: "Mostly, the infrastructure left behind in the wake of the initial investor surge and capex leads to the emergence of new products and services. These are often underestimated or poorly anticipated."

• Assessment 1 - Not a bubble yet: "Despite the significant interest that AI has generated, it still does not appear to have driven a bubble in valuations which sets it apart, so far at least, from previous narrative investment cycles like the internet in the late 1990s"

• Assessment 2 - Capital-intensive industries are low margin: "Nevertheless, the AI winners of today are no longer capital-light businesses. Just as we saw with the networking companies of the internet, AI is driving a major capex boom and threatens to stifle the high rates of returns that have characterised the sector over the past 15 years and which current valuations imply will continue"

• Assessment 3 - Risk concentration: "While these companies may be less highly valued than in other narrative-led bubble periods, the scale of market dominance is greater this time. The 10 biggest stocks have their highest share of the market for many decades at over one-third of the index, while the five biggest companies are worth 26% of the total value of the S&P 500"

• Assessment 4 - Competition will drive into profit margins: while absolute returns remain good for the dominant companies, these strong returns fade over time, and they often remain solid ‘compounders’. Importantly, however, the returns are generally negative for dominant companies if an investor buys and holds them as other faster-growing companies come along and outperform

• Recommendation: "The paper provides valuable recommendations on diversification - not just out of the tech sector and magnificent 7 stocks but across equity markets globally and across industries which are somewhat not in popular sight - pharma, healthcare, robotics, cyber security, banking and financial services, consumer products ~100 companies for diversification at page 31 "

(Goldman Sachs Strategy Paper - Aug 2024, Portfolio Strategy Research Team)

Resources

• Lilian Weng's blog is perhaps the best blog I have across describing the evolution of various technologies and architectures from the research community - on Deep Learning and Large Language Model training