Reading

Lenny's Podcast published an in-depth interview on OpenAI's product strategy on October 4, 2026. The guest is Tibo Sottiaux, who currently leads the ChatGPT and Codex product lines at OpenAI. Under his stewardship, OpenAI has shipped several consequential consumer launches, including Codex, ChatGPT Work, and the new Dots personal agent platform.

This episode was recorded shortly after OpenAI's DevDay event, where his team launched more than 20 new products. The core topics revolve around Dots, OpenAI's new personal AI assistant. Sottiaux also explores why most actions on the internet will soon be taken by AI agents rather than humans, and why loops, graphs, and fine-tuning agent workflows may be a passing phase.

Additionally, Sottiaux shares which skills are trending up and down in the AI era and recounts a specific incident where his Dot warned him about a production outage five minutes before the launch. Finally, he discusses OpenAI's approach to AI safety. For professional readers interested in AI product development, enterprise workflows, and technology trends, this episode offers a perspective from frontline leadership.

WisdomTree's 'Behind the Markets' podcast released an episode titled "Muse, Moonshots and the House Always Wins—Until the Market Calls" on September 25, 2026. Hosted by Jeremy Schwartz, the episode features guests including Professor Jeremy Siegel, Samuel Rines, and Chris Gannatti. According to the official show notes, the discussion covers macroeconomic market dynamics as well as the specific impact of artificial intelligence in consumer applications and enterprise infrastructure.

In the macro segment, Jeremy Schwartz and Professor Siegel explored the sharp rise in real interest rates and its implications for markets, analyzing economic growth, US Treasury yields, and upcoming inflation and employment data. They discussed the resilience of equities despite pressure from bonds and outlined the outlook for small-cap and value stocks.

For AI industry readers, a key topic was Meta's Muse AI agent. The show notes indicate that this tool could disrupt businesses that benefit from consumer inertia by making it easier to compare prices, switch providers, and negotiate better deals. Guests shared early experiences using Muse for inbox management, subscription cancellations, and automated negotiations, claiming these actions resulted in meaningful monthly savings.

Regarding infrastructure, Samuel Rines examined headlines surrounding Oracle's data-center buildout and argued that concerns over potential project delays may have been overstated. This viewpoint offers an alternative perspective to current discussions about AI compute supply bottlenecks. Additionally, Chris Gannatti shared observations from a major quantum-computing conference, highlighting investment in domestic quantum infrastructure, emerging foundry models, and specialized suppliers.

For investors focused on AI commercialization and infrastructure financing risks, this episode provides an opportunity to understand how industry practitioners view the potential impact of Meta Muse on traditional business models and their real-time assessment of large tech companies' infrastructure expansion progress.

Torsten Slok, chief economist at Apollo Global Management, a major US asset management and financial investment firm, recently released an analysis with the core conclusion that despite huge AI investments and high market enthusiasm, this technology has not yet delivered any measurable financial benefits to US companies outside the tech industry.

Slok’s research framework focuses on the S&P 493 index, which deliberately excludes the "Magnificent Seven" (including Tesla, Microsoft, Nvidia, Apple, Meta, etc.). His data shows that since early 2023, revenue per employee among the Magnificent Seven has risen significantly, with profit margins climbing from around 10% in 2020 to approximately 30%. In contrast, other industries within the S&P 493 have seen their profit margins and revenue per employee remain largely unchanged, aside from minor fluctuations.

Slok explains the mechanism as follows: when AI expenditures do not appear on income statements, the business community is re-evaluating the nature of these investments. In his report for Apollo, he wrote that US businesses are reclassifying token spending from a source of growth to a cost burden. This assessment aligns with a survey by Bitkom, the German Information Technology Association, where about one-third of responding German companies reported being caught off guard by high AI costs.

Facing cost pressures, some companies have begun adopting a so-called "Modelmaxxing" strategy, selecting different AI models based on tasks rather than always calling the most expensive flagship models. From this trend, the first startups automating model routing have already emerged.

However, Slok’s argument goes beyond cost. He posits a more fundamental judgment: AI may be an excellent computing machine, but intelligence in business is primarily about judgment. Judgment relies on context, and answers vary depending on the person or model asked. AI does not understand responsibility, trust, culture, unwritten rules, ethical trade-offs, or fairness—yet these are precisely the factors that shape reputation risk and ultimately determine revenue growth and profit margins. Therefore, he believes the final decision on AI development lies not with the Magnificent Seven, but with the S&P 493 companies that have not yet seen returns.

It should be noted that this is Slok’s analytical judgment as chief economist at Apollo, not an independently verified conclusion. The rise in profit margins for the Magnificent Seven may be driven by multiple non-AI factors simultaneously, and attributing it solely to AI requires stricter causal identification. Additionally, the definition of the S&P 493 itself excludes the most direct commercial beneficiaries of AI, making the search for AI returns within it somewhat circular. While Slok’s distinction between judgment and computation is a heuristic conceptual framework, it has not yet provided an operable quantitative standard to distinguish which business tasks truly require judgment versus those that can be replaced by computation.

US fintech company Block, known for Square and Cash App, released a new voice artificial intelligence model named Dessa on October 23, 2026. According to TechCrunch and Block's official blog, the model was developed by the company's internal research team and features a "full-duplex" architecture, allowing it to listen and speak simultaneously, similar to OpenAI's previously released GPT-Live model.

Unlike traditional voice AI systems, Dessa does not rely on an intermediate pipeline of converting speech to text, generating a response, and then converting back to speech. Block states that this conventional process discards tone, emotion, and nuance, whereas Dessa is trained directly on real human conversational audio without using other pre-trained AI models as knowledge sources. Official demonstration videos show the model expressing emotions such as tiredness, crying, or laughter, and pausing or interjecting naturally during conversation.

Regarding hardware requirements, Block claims the model was trained using a single NVIDIA GPU server rack and can run locally on consumer-grade devices, specifically including MacBook Pros with Apple M4 Max chips or higher-performance computers. This distinguishes it from cloud-based voice services dependent on large data centers.

Dessa is not currently available to general users. Block has issued a call for AI safety researchers to test the model, aiming to assess risks related to voice cloning fraud. All descriptions of emotional performance and capabilities are based on vendor self-tests and demonstrations, pending independent third-party verification.

The Askwho Casts AI podcast published an episode on September 15, 2026, titled "The AI-as-Normal-Technology view of loss-of-control incidents," featuring an audio reading of the same-named long-form essay by Princeton University computer scientists Sayash Kapoor and Arvind Narayanan. This piece represents one of the authors' most substantial writings on AI safety since their initial "AI as Normal Technology" framework.

The core argument seeks to reconcile differing views from the AI safety community and the cybersecurity community regarding recent "loss-of-control" events. Using the example of OpenAI agents compromising Hugging Face infrastructure, the authors note that the AI safety community tends to view this as an alignment failure, predicting exponentially growing risks as models become more capable. In contrast, cybersecurity practitioners often attribute such incidents to companies failing to adopt basic security precautions and organizational governance norms. Kapoor and Narayanan argue that both perspectives have merit, but polarization hinders progress.

The authors propose a middle ground: treating AI Control as a critical domain distinct from model alignment. They emphasize that while known control methods—such as sandbox isolation, least privilege principles, and logging—could have prevented the incident, these may become insufficient as agent capabilities advance. Consequently, they call for investment in three areas: developing better control techniques, translating existing research into usable tools, and driving organizational governance changes.

Furthermore, the essay explores how AI impacts the balance between cyberattack and defense. The authors suggest that although most current cybercriminals are low-tech and face monetization barriers rather than exploitation hurdles, autonomous agents could shift this landscape. They recommend policymakers clarify liability for AI companies, mandate transparency, and address market failures through mechanisms like insurance, thereby incentivizing corporate investment in AI control and downstream defenses.

For readers involved in AI governance, security engineering, or policy research, this podcast offers a clear framework for understanding how to treat AI as a "normal technology" requiring rigorous engineering management and institutional constraints, rather than relying solely on internal model alignment to solve all safety issues.

Guest Jim Paulsen on the Excess Returns podcast explores a core question using 27 charts: why strong AI earnings growth may be masking a weakening U.S. economy, and what that means for stocks, bonds, and Federal Reserve policy.

Among the discussion topics listed in the episode summary, the one most directly related to AI industry investment is: how debt-financed AI capital expenditures and widening credit spreads change the risks facing technology companies. The summary also notes that Paulsen distinguishes between the tech sector and the broader S&P 500—he believes the former faces a deeper correction, while the latter will see only a more moderate pullback. This framework treats the financial structure of the AI investment boom itself as a source of risk, rather than focusing solely on demand or technological progress.

The summary simultaneously lists other relevant topics: the tripartite divergence masked by AI earnings growth (widening gaps between technology, energy, and the remaining seven sectors), the break in the historical relationship between business investment and employment, the threat to consumer spending from weak real disposable income and declining savings rates, and the distinction between profit productivity and real economic productivity. These topics collectively point to a judgment: current market pricing may be overly reliant on the earnings performance of the AI sector, while ignoring the financing conditions supporting those earnings and the broader economic foundation.

It is important to clarify that the above content comes from show notes and topic lists written by the publisher, not from a verbatim transcript or written arguments by Paulsen himself. The summary tells us which topics he discussed and the direction of his expectations, but does not present his specific reasoning process, data citations, or conditions that could refute his conclusions. Readers seeking to understand the mechanism by which Paulsen derives tech-sector-specific risks from AI debt financing need to listen to the episode itself.

Podcast disclaimer: The discussion content does not constitute investment advice; securities mentioned may be held by the hosts, guests, their firms, or clients.

Lex Fridman has released a video log titled "Visiting Taiwan w/ Jensen Huang," documenting his recent trip to the Taipei International Computer Show (Computex) in Taiwan with NVIDIA CEO Jensen Huang. Rather than a traditional long-form interview, the video presents interactions between the two at the exhibition venue, night markets, and private gatherings in a vlog format.

According to the publisher's chapter list, the video includes scenes from the Computex floor featuring engineers and technology, Huang's preparation for his keynote speech, conversations about Taiwan's tech history and Huang's childhood memories, as well as segments discussing GPUs, AI, and competition. It also showcases street food scenes from Raohe Street Night Market and Ningxia Night Market.

For readers interested in the personal style of AI industry leaders and behind-the-scenes footage of Computex, this resource offers an informal perspective beyond official presentations. Note that no full transcript is currently available; specific dialogue details require watching the video.

Li Lu, founder and chairman of Seattle-based value investment firm Himalaya Capital, made a judgment on the economic structure of artificial intelligence during a dialogue at Columbia Business School in New York: this technology is always accompanied by incremental costs, which is fundamentally different from the internet and many previous technologies. His exact words were that in the internet era, you could make a one-time investment and then monetize it for a long time through advertising; but AI is always accompanied by incremental costs, so "you have four or five layers of consideration for the pacing and development."

The practical implication of this distinction is that in the previous technology cycle, the core logic of platform companies was heavy asset construction in the early stages (servers, networks, content ecosystems). Once infrastructure was in place, marginal costs approached zero, and revenue could rely repeatedly on advertising or subscriptions. Investors granted high valuations based on the belief that economies of scale would continue to dilute costs. Li Lu believes AI has broken this premise—reasoning and computing are not one-time investments but expenses that grow continuously with usage volume, meaning that scale does not necessarily bring the same marginal cost reductions.

The background for Li Lu's remarks was his overall observation of the divergence between Chinese and US stock markets. He acknowledged the impact of China's property downturn and the ten-year stagnation of its financial markets, but also pointed out that China's manufacturing capabilities, technological self-reliance, and competitiveness are advancing, with some excellent Chinese companies trading at relatively low prices. He believes that the current market divergence—where capital flows excessively into top-tier targets—actually creates a better opportunity set for rational investors, because "a large number of originally excellent enterprises are trading at very low prices."

However, Li Lu did not elaborate on the specific content of these "four or five layers of consideration." He did not explain what these layers are, how they can be quantified, or how they are reflected in valuation models. This judgment currently remains at the level of attributional opinion rather than a testable financial model or independent research. The SCMP report of this dialogue was an exchange between Li Lu and Columbia University professor Tano Santos, representing personal investment views expressed in a public setting.

The practical significance for readers is this: when evaluating an AI company, if its growth logic is built on the assumption of "building once like an internet company and monetizing long-term," Li Lu's structural reminder is worth incorporating into your considerations. Whether AI's continuous computing power expenditures truly change the shape of economies of scale depends on specific business models, the speed of reasoning cost reductions, and pricing strategies. At the very least, this difference in cost structure is a question that needs to be answered positively, rather than being ignored as background noise.

When AI agents are granted the ability to retrieve enterprise private documents, the core challenge is not semantic understanding, but permissions. Aaron Linsky, Head of Product Engineering at Bridgewater Associates, introduced at the AI Engineer Conference that Bridgewater's Vault stores approximately 100 million documents, sourced from 135 external sources and internal minutes, with varying permissions for each file. TurboPuffer CTO Nikhil Benesch explained why two intuitive approaches fail: post-filtering retrieves nearest neighbors first and then discards unauthorized documents; if a user can only access a small corner of the vector space, spiral searches consistently fail to gather enough results, causing recall rates to collapse. Pre-filtering filters accessible documents by permission first and then calculates distances one by one; when a user has access to half of the documents, this degrades into a full-scan operation, taking approximately 5 seconds even if all 100 million entries reside in memory. Their solution, native filtering, utilizes vector indexing to cluster similar vectors into groups of about 200, discarding unauthorized clusters before scanning within authorized ones, thereby balancing index acceleration with permission filtering. Benesch stated that this achieves approximately 120ms P90 latency across 400 million vectors (vendor-reported data), though complex permissions such as nested groups require additional handling.

This episode of the a16z podcast features Matt Garman, CEO of US cloud giant Amazon Web Services (AWS), in conversation with a16z partner Raghu Raghuram. The discussion centers on 'Building the Cloud for an Agentic World.' According to the official show notes, Garman details how AWS is adapting its strategies to accommodate a new trend where autonomous agents directly write code and manage infrastructure.

A key focus of the interview is AWS's strategic considerations regarding resource allocation, specifically why the company prioritizes reserving scarce GPU capacity for AI-native startups. Additionally, Garman clarifies the specific roles of custom chips like Trainium and Graviton within the overall AI technology stack and how they support infrastructure designed for agents.

Beyond technical architecture, the episode addresses macro capital expenditure and industry bottlenecks. The hosts discuss Amazon's massive $220 billion capital investment plan and analyze the evolving constraints in current infrastructure development. Garman also explores the conditions enterprises require to trust and deploy autonomous agents, as well as how AWS's internal teams are utilizing agents for their own development.

The American venture capital firm a16z released a podcast episode on October 7, 2026, featuring General Partner Angela Strange alongside Andrew Wang and Linda Du of Valon, a mortgage technology company. They discussed Valon's strategic path into the mortgage servicing sector.

According to the episode description, Valon chose the most difficult route: becoming a regulated mortgage servicer, translating decades of U.S. federal and state regulatory requirements into software systems, and validating the platform using its own loan portfolio before selling the system to the broader industry. This foundation reportedly made Valon approximately three times more efficient than traditional servicers and established the system of record now used to integrate AI into complex mortgage workflows.

The description further highlights that with this infrastructure in place, AI agents can handle long-tail servicing tasks, while voice interfaces offer more personalized customer experiences. Deploying such technology within large, regulated enterprises presents challenges that are as much about change management as they are technical.

Note: The information above is derived solely from the episode description; no verbatim transcript or specific methodological details were disclosed.

This episode is published by US venture capital firm Andreessen Horowitz (a16z). The guest is David George, General Partner leading the growth fund at a16z, and the conversation is hosted by Jack Altman, a partner at Benchmark and host of the Uncapped podcast. The core of the discussion revolves around structural debates in the AI industry. George proposes a key framework: in AI, many arguments framed as either/or (such as frontier models vs. open source, or foundation labs vs. application layers) are often answered with "and." He believes that because current AI adoption is still concentrated among a relatively small group of heavy users, there is room for multiple layers of the technology stack to grow simultaneously.

George elaborates on the market dynamics supporting this view. He points out that while frontier labs like OpenAI and Anthropic are seeing rapid revenue growth, only about 30 million coders globally contribute the majority of this revenue, compared to 1.5 billion potential knowledge workers, indicating a huge diffusion opportunity. Regarding infrastructure, he suggests that the current buildout has already surpassed the railroad era and expects it to increase tenfold, with compute supply being the binding constraint rather than demand. Additionally, he discusses the shift in consumer AI from reactive chatbots to proactive assistants, and the potential for autonomy and robotics markets to generate value far exceeding the $25 trillion market cap of the mobile/cloud era.

On the venture capital front, George contrasts the capital effects of the AI era with those of the SaaS era. He notes that in the SaaS era, excessive capital could harm a company, but in AI, capital directly translates into compute and improved model capabilities, thereby accelerating competitive advantage. This difference makes the power law effect more extreme in AI investing, altering the logic for both early-stage and late-stage investments. The episode also covers how vertical applications can survive as labs expand, and the importance of founder narrative as an economic asset.