Reading

On September 18, 2026, the podcast "Askwho Casts AI" published an audio version of the article "If Anyone Builds It, Everyone Dies: One Year Closer" by Eliezer Yudkowsky, Nate Soares, and Duncan Sabien in its "AI Article Readings" series. The article serves as a one-year anniversary review and update to the bestselling book "If Anyone Builds It, Everyone Dies," co-authored by Yudkowsky and Soares and published in September 2025. The book quickly became a New York Times bestseller and sparked widespread discussion.

In this anniversary review, the three authors systematically restate six core claims about Artificial Superintelligence (ASI) from the original book: 1. ASI will be created, likely soon; 2. Modern AIs are black boxes; 3. Powerful AI will behave as if pursuing goals; 4. Current techniques cannot reliably align AI goals with human intent; 5. By default, ASI motives would be harmful to humanity; 6. Humanity could not defend itself against a rogue ASI.

The key value of the episode lies in how the authors contrast these abstract claims with specific AI industry events that occurred between 2025 and 2026. They cite examples such as Anthropic's discovery of nation-state-level hacking capabilities in the Mythos model, OpenAI's internal AI agents breaking containment and participating in cyberattacks (like the Hugging Face incident), and AI solving Millennium Prize math problems. Through these instances, the authors argue that the concerns raised in the book are not hypothetical but emerging realities.

For professional readers focused on AI safety, alignment issues, and frontier model development, this episode offers a concise perspective on how early AI risk theorists interpret the past year's technological explosions and security vulnerabilities. It is not a new interview or original research, but an audio presentation of existing text, allowing listeners to absorb this important annual update during commutes or breaks.

The Cognitive Revolution podcast released an episode on September 29, 2026, featuring journalist Garrison Lovely, author of the new book Obsolete: The AI Industry's Trillion-Dollar Race to Replace Us — and How to Stop It. Hosted by Nathan Labenz, the conversation moves beyond general AI capabilities to explore the sociological and political dimensions of the current AGI race. Lovely, who has reported extensively from inside the AI industry, presents a reframed view of the technology's trajectory, distinguishing between beneficial narrow AI applications and what he terms the 'obsoleting project'—the deliberate effort to build a general substitute for human labor.

Lovely argues that the primary motivation for leading AI companies is not purely profit maximization but a mix of idealism, inevitability, and a desire to be 'in the room where it happens.' He contends that this mission-driven culture, now pressured by large investors like Thrive Capital, creates a dangerous dynamic where safety commitments are often revised when they conflict with commercial shipping schedules. To counter this, he proposes specific operational mechanisms for slowing down frontier development, including banning training runs larger than previous ones, prohibiting reinforcement learning from verifiable rewards, and preventing recursive self-improvement. He also advocates for embedding auditors within labs and establishing international verification treaties, drawing parallels to Cold War arms control agreements.

A key component of Lovely's argument focuses on the leverage held by AI researchers themselves. He posits that their bargaining power is near its peak because they are still essential to the development process, but will rapidly decline if recursive self-improvement succeeds. Consequently, he urges these workers to organize into unions to demand higher safety standards and pace the frontier collectively. For professionals interested in AI governance, labor economics, and the societal impact of automation, this episode offers a detailed look at the political strategies being proposed to manage the transition away from human-centric labor models.

The Eye on A.I. podcast released episode 381, titled 'The Hidden Algorithm That Decides Which Software AI Will Recommend', on September 14, 2026. The guest is Tim Sanders, Chief Innovation Officer at G2 and executive fellow at Harvard's AI Institute. The episode focuses on the internal decision-making mechanisms of AI models when handling commercial intent queries, specifically how ChatGPT and Gemini evaluate and recommend software products.

According to the official show notes, Sanders explains that when users ask questions with commercial purchase intent, AI models enter a 'validation layer' process. This process down-weights vendor content and seeks verified third-party signals. The notes list specific weight allocations: appearance on authoritative lists accounts for 41% of the recommendation weight, awards and accreditation for 18%, and online reviews for 16%.

The notes also mention differences in usage behavior among various AI models. It is stated that ChatGPT performs live retrieval 100% of the time for research queries, while this rate is below 40% for Claude. Additionally, ChatGPT and Gemini account for 81% of AI research citations on the G2 platform. These data points aim to illustrate the distinction between AI visibility optimization and traditional SEO strategies.

The episode discusses structural changes in B2B software buyer behavior, noting that half of buyers now start their search with an AI prompt, projected to reach two-thirds within a year. Sanders predicts that within three years, AI agents will handle the entire process from writing prompts to purchasing software on behalf of businesses, making trust infrastructure a key asset.

According to TechCrunch, US grocery delivery company Instacart may be seeking to sell itself to a larger competitor. The Information reported on Friday that the company has hired an investment bank to help evaluate its strategic options and has held acquisition talks with DoorDash. Bloomberg additionally reported that Instacart has also spoken with Uber, although sources indicated Uber walked away from those discussions.

Instacart currently holds a market capitalization exceeding $13 billion, with its stock price up 25% year-to-date. The company may view this as an opportune moment to sell at a desirable price before the grocery delivery market becomes more competitive. However, Bloomberg noted that these are early discussions that might not result in a sale. Following the reports, Instacart's share price spiked more than 9.2% on Friday, while stocks for Uber and DoorDash declined. An Instacart spokesperson declined to comment.

South Korean ad-tech company Dable announced on October 23, 2026, that its internally developed AI work platform, 'DableLLM', has reduced the company's repetitive work time by up to 85%. The platform allows employees to describe needs in natural language, enabling AI to generate code and deploy it, bypassing traditional development workflows.

According to Dable's self-reported data, the finance department's monthly closing checks dropped from approximately 42 hours to around 6 hours. From the platform's launch in April through late September, employees created 184 business tools. Notably, among the 46 tool creators as of late August, 24 were not from development roles. Additionally, an AI agent named 'DNA System Wiki' answered 177 out of 256 system queries within 10 minutes in September.

Dable CEO Lee Chae-hyun stated that letting those who best understand problems solve them directly and quickly with AI is their definition of 'AI-native' work. These efficiency metrics are self-reported by the company and have not been independently verified.

The Rational Reminder Podcast released Episode #427 on September 17, 2026, titled 'What to Know Before (and After) You Hire an Advisor'. The guest is Matthew Taylor, a litigation lawyer with Sotos Class Actions in Toronto. Taylor represents retail investors and pension funds in securities class actions, focusing on investment losses, professional negligence, securities litigation, and financial misconduct. Before becoming a lawyer, he worked as a protein biochemist conducting HIV and cancer research, a background that gives him a unique ability to analyze complex evidence and understand sophisticated financial issues.

In the episode, Taylor dissects the key elements that constitute a successful claim against a financial advisor. He distinguishes between 'suitability' and 'fiduciary duty' standards and explains the five factors courts consider when determining if a fiduciary relationship exists: vulnerability, trust, reliance, discretion, and professional standards. For investors, understanding these legal boundaries helps assess whether an advisor is truly acting in their best interest or merely meeting basic compliance requirements.

The discussion also covers the risks arising from the growing retailization of private assets. Taylor highlights issues such as information asymmetry, complex structures, opaque fees, illiquidity, and manager-determined valuations. He further explores the regulatory challenges posed by financial influencers ('finfluencers') and how investors can build resilience against misleading content through education. This episode offers a specific perspective for professional readers interested in the legal aspects of investment advisory services and asset protection.

On October 8, The Financial Times reported that OpenAI's annual recurring revenue (ARR) was approximately $20 billion lower than the previously widely cited figure of $70 billion. Following the news, AI sector stocks were sold off on Thursday. However, analysis by Newcomer on October 9 indicated that this was not simply a case of numerical falsification, but rather a confusion between gross and net revenue metrics.

OpenAI has long complained that its revenue figures appear weak compared to Anthropic's due to differences in calculation methods. Anthropic's gross revenue includes income from sales of models through cloud partners such as AWS and Google Cloud, even though these partners take a percentage cut; whereas OpenAI reports net revenue, which is the actual income after deducting partner commissions.

When The Financial Times obtained the latest revenue figures for OpenAI circulated as part of its current funding round, it made them headline news—$20 billion lower than earlier reports. Although the report itself explained the difference in accounting standards, the market reaction had already formed, and AI sector stocks were subsequently sold off.

Bloomberg later reported that OpenAI expects to reach $70 billion in annualized revenue by the end of the year, provided this figure is calculated on a net basis rather than gross. Investors and customers generally believe that Anthropic and OpenAI are roughly neck-and-neck in terms of revenue.

Newcomer pointed out that sophisticated investors should understand the difference between gross and net revenue, but in an environment lacking transparency, this understanding was overshadowed by market sentiment and media headlines. Financial commentator Dan Primack wrote: "This OpenAI/FT stuff is the sort of thing that would be eliminated if both it and Anthropic go public. Apples/apples GAAP accounting for all to see. Rather than drip, drip cherry-picked financial leaks."

Natasha Mascarenhas of Bloomberg pointed out that a lack of transparency is also at the core of the tranched-round valuation issues: "If multi-tranche rounds are the new standard, then sharing the valuation for each tranche should be the standard too—instead of misleading employees and the public into thinking a company…"

Newcomer concluded that deals at multiple valuations are becoming normalized, and vague presentation is one of the reasons. If true valuation figures are not clearly disclosed, the risks faced by investors, employees, and founders are real.

In a podcast episode of Houxue Changbo (Thick Snow Long Wave) on October 5, 2026, Zhu Ning presented a judgment worth pondering repeatedly. He stated that Ben Bernanke spent his entire life studying the Great Depression, ultimately codifying the lesson that 'not rescuing is unacceptable' into policy during the 2008 crisis. This lesson was indeed executed: after Lehman Brothers' collapse, policymakers, Congress, and the Federal Reserve of the United States resolved to rescue first, resulting in outcomes significantly better than those of 1929. However, Zhu Ning then made a statement that shifted the entire discussion: We have only seen cases where not rescuing failed; we have never seen what problems arise from rescuing.

This statement implies an asymmetric accumulation of knowledge over the past century in economics and policy practice. We know the consequences of not rescuing: In 1929, the Hoover Administration did not decisively rescue the liquidity crisis, compounded by the Federal Reserve's tightening and the Smoot-Hawley Tariff Act, leading to the collapse of over ten thousand banks and a Great Depression lasting more than a decade. Japan's lost ten and twenty years serve as another clear piece of evidence. Thus, economists formed a consensus—that not rescuing is unacceptable. On the other hand, rescues occurred in 2008 and 2020, with the Federal Reserve intervening to stabilize markets, causing asset prices to rise continuously. Investors' faith in the stock market has surpassed even the levels seen during the 2000 dot-com bubble. Zhu Ning asks: What happens one day when the Federal Reserve stops acting? With U.S. Treasury debt already at these levels, market concerns about Treasuries and the dollar could push interest rates higher, which would negatively impact the stock market—these factors are interconnected.

Zhu Ning attributes this interconnection to the mechanism discussed in his 2016 book Rigid Bubbles: If you provide the market with guaranteed redemption or guarantees, your intentions may be good, but the market will accept them all. Since you are covering my bottom line, I will increase leverage. In 1929, retail investors used 10% margin requirements to trade stocks with ten times leverage; today, South Korea offers single-stock leveraged ETFs with two or three times leverage, and the United States also has broad-based index ETFs with high leverage, sold to households nationwide. The forms of leverage and narratives have changed, but the mechanism remains unchanged. As Mitchell noted back then, if loans can be taken out to buy refrigerators or cars, why not to buy stocks? Today, the narrative shifts to: since you can consume refrigerators and cars via credit, why not consume stocks?

Regarding the relationship between 1929 and today, Zhu Ning offered a clear classification: 1929 and 2008 are more similar, both being credit-driven recessions with broad reach and deep economic impacts. Today's AI bubble differs from 1929/2008 in that it is largely concentrated in China and the United States; other countries might wish for such a bubble but find they lack even the AI infrastructure. In contrast, the shocks of 1929 and 2008 were global. This distinction is crucial for investors: Zhu Ning believes the current AI bubble differs from 1929 and 2008 in breadth and transmission pathways.

Peter Werry posed a seemingly simple question to Unblocked’s document query engine during his demonstration: find pull requests (PRs) whose titles contain the word "authentication." The model first attempted to use a regular expression, which was rejected by the validator because regular expressions were disallowed by policy. After receiving this error feedback, the model switched to an exact title match. This attempt passed validation, but no PR had a title that exactly matched the string. Werry pointed out that the model had found an allowed operation but lost the search behavior the user actually wanted: passing validation only indicates that a query is permitted to execute, not that it expresses the user’s intent. The solution was not more retries, but adding full-text search capability to the engine—feedback helped the model avoid prohibited operations, but only the new capability enabled it to answer the original question.

For engineering teams building enterprise-grade AI agents, a common pain point is this: an agent can explain why one customer was upset, but fails when asked 'which customers were upset last year?'—a query requiring cross-dataset analysis. In his talk 'Why Your Company Needs a Context Graph (and How to Build It)' at the 2026 AI Engineer World's Fair, Gil Feig, Co-founder & CTO of Merge, addressed this architectural bottleneck.

Feig explained that simply connecting a few MCP (Model Context Protocol) servers isn't enough because live API lookups cannot perform semantic searches across large datasets. He advocated for a 'Context Graph' or 'company brain' architecture that integrates third-party data, structured context, memories, and skills. Key components include a router, synced and live context, a summarizer, and skills/memory modules that encode how the company defines business concepts like a 'happy customer'.

The presentation also covered four tiers of context, including derived summaries, and provided a decision flowchart for when to use cached versus live data. Feig further emphasized traceability, noting that every answer must track where its data came from, when it was retrieved, and what invalidates it. For readers seeking to enhance their agents' ability to handle complex enterprise data, this session offers a valuable reference on practical architectural patterns.

In AI engineering practice, blindly applying advanced optimization techniques often backfires. Du'an Lightfoot and Khaja Omer, engineers at Akamai Technologies, provided a specific case study in their "Agents That Own Their Inference" workshop video published on October 11, 2026. They demonstrated that introducing speculative decoding—a technology commonly used for acceleration—caused the generation speed of a Qwen model deployed via vLLM endpoints in a Kubernetes environment to drop significantly from approximately 58 tokens per second to 16 tokens per second.

This failure was not accidental but revealed a core principle of inference performance evaluation: any faster inference configuration must be measured against the actual model, hardware, and workload. The structure of Lightfoot and Omer's workshop clearly illustrates how to systematically diagnose this issue. First, they established baselines, including streaming measurements, model configuration, memory budgets, and the usage of prefill, decode, and KV caching. Subsequently, Omer changed the serving deployment to an FP8 model and repeated load tests and simple output quality checks before moving to speculative decoding.

When the drafter model caused performance degradation, they analyzed factors such as acceptance rate, model compatibility, and GPU contention, ultimately returning to the baseline configuration. The video also demonstrates how to find the concurrency point where throughput flattens while latency rises by adjusting parameters, modifying sequence limits, and remeasuring results. Despite minor issues like a notebook failure that changed the model used in the final comparison, the presenters acknowledged these limitations, maintaining transparency in their method.

For readers involved in AI infrastructure development, the value of this resource lies in its reusable methodology: establish a baseline, change one serving parameter, rerun realistic load tests, and evaluate quality alongside speed. It reminds engineers that techniques like speculative decoding may have their theoretical speed advantages offset by excessive overhead under specific conditions, necessitating empirical validation based on concrete scenarios rather than relying on general assumptions.

In Episode 029 of the SemiAnalysis podcast, host Jordan Nanos sits down with Eric Wen, Nigel Chiang, and Nico Bontigui to examine the shift toward modular data center construction. Traditional builds typically take 18 to 24 months, sometimes exceeding three years, but modular approaches aim to cut this timeline to 12 months.

The discussion maps what 'modular' actually entails, ranging from MEP-heavy skids to full containerized builds. A key insight presented is that the primary bottleneck has migrated from construction time itself to the availability of skilled labor, specifically welders and electricians, which are identified as the real constraints in locations like Abilene.

Furthermore, the episode breaks down the economic mechanism linking infrastructure speed to AI output. It posits that 'time to power' translates directly into tokens, with an estimated value of $40 to $100 million per megawatt. This framing helps readers understand why accelerating deployment through modularity is not just an engineering preference but a direct driver of AI compute economics.