Reading

The National Highway Traffic Safety Administration (NHTSA) announced on Monday that it has opened a defect-petition review into claims that 2026 Tesla Model Y robotaxis and Cybercabs exceeded speed limits in school zones while operating in autonomous mode.

According to Business Insider, an unnamed petitioner submitted three videos as evidence. The agency stated it will evaluate the complaint before deciding whether to grant a formal investigation. No crashes or injuries have been reported, and approximately 1,000 vehicles are estimated to be affected.

Rival Waymo previously faced similar scrutiny and recalls over school zone safety issues. Tesla did not immediately respond to requests for comment.

On October 1, 2026, the Odd Lots podcast invited Luke Kawa, Head of Markets at Sherwood News, to discuss the current market state. Kawa argued that the most significant feature of markets is speed — the pace of both news events and asset price fluctuations has accelerated. He noted that at the time of recording, the S&P 500 was 0.5% off its record high, yet only 51.2% of its constituents were above their 200-day moving averages. He compared this to the day after the dot-com bubble peak in March 2000, when similar proximity to highs coincided with extremely narrow breadth. However, he cautioned that similar low breadth appeared in 1998 before the market rallied for two more years, so breadth alone is not a sufficient predictor of tops but reveals structural fragility.

Kawa attributed this divergence to the dominance of AI capital expenditure. Citing Goldman Sachs research, he noted that approximately half of this year's expected earnings growth is driven by hyperscaler capex. This implies that market pricing increasingly relies on expectations of continued AI hardware and infrastructure investment rather than broad consumer demand. In this environment, rising interest rates impact AI companies (which can issue debt at scale) far less than rate-sensitive sectors like homebuilders. Kawa also observed that despite higher nominal yields, hyperscaler credit spreads did not widen significantly, indicating the market still views them as relatively safe borrowers.

On fixed income, Kawa observed that rising global bond yields primarily reflect strong growth expectations. Notably, US Treasury breakeven inflation rates remained relatively stable despite higher nominal yields. Kawa expressed surprise at this, suggesting it might reflect investor confidence in central banks' ability to control inflation, or that bonds have become a 'residual asset' — bought only when growth concerns arise. He further argued that for bonds to be competitive with stocks, real yields need to rise, which competes with the real return expectations offered by AI investment.

Regarding retail investor behavior, Kawa mentioned Robinhood data showing net single-stock purchases peaked in late June, coinciding with Micron's earnings release. Although September is typically a slow month for retail trading, recent rebounds have been strong, recovering to approximately 89% of the June peak. He noted that retail investors are actively chasing volatility, particularly through options with four days or less to expiry. Kawa mentioned that every day when reviewing the data, there is a ticker in the top 25 that he does not recognize, reflecting the market's pursuit of extreme volatility.

On potential risk signals, Kawa argued that if the AI boom slows or financing becomes overly creative, the market faces adjustment risks. Currently, hyperscalers' free cash flow has turned negative, relying on debt to sustain capex. He referenced Greg Ip's observation at the Wall Street Journal that the AI boom is unusual in that it is not being financed by the fruits of AI itself — unlike the dot-com era, current construction funding comes primarily from tech platforms with large free cash flows from their core businesses. Kawa argued that if AI applications fail to deliver expected productivity gains and downstream consumer demand, the current valuation logic will face fundamental challenges. He summarized that the market needs to find someone to punish, and currently the punished are traditional consumer and cyclical stocks with slower growth but higher valuations.

AI weather startup WindBorne Systems has won a contract with the US Air Force worth up to $49 million for data collected by its fleet of autonomous weather balloons, marking its largest single deal to date.

WindBorne operates approximately 700 balloons globally, feeding atmospheric sensor readings into its deep-learning weather models. The contract was awarded without competitive bidding, intended to support constellation growth and hardware R&D.

Although the contract ceiling is $49 million, the Air Force pays only as needed. So far, $1.7 million in orders have been placed, with executives expecting more. CTO Andrey Sushko stated the government determined WindBorne was the only provider meeting their operational needs.

In August, CEO John Dean disclosed a $37 million Series B raise. This major military contract provides critical backing for the transition from venture funding to stable government revenue, though the gap between the contract ceiling and current orders remains a key watchpoint.

Scarlet, a certification body for AI and software medical devices, has raised $50 million in a Series B round led by Headline, with participation from Notion Capital and existing investors including Atomico.

The funding aims to bridge the gap between rapid AI model development and slower regulatory processes. According to Tech.eu, companies can train diagnostic models in weeks but often wait 12 to 18 months for certification.

Nancy Xiao Mclaughlin, Partner at Headline, will join Scarlet’s board. The new capital is intended to scale certification capacity to meet the growing demand for AI-driven medical devices.

In an article published on October 12, 2026, Adam Newgas presents a specific argument about how AI is reshaping the creative economy. He observes that while current AI is not entirely sterile—capable of remixing and random exploration—it remains weaker at generating true 'novelty' compared to its other strengths. This technical characteristic leads to a counter-intuitive result: the supply-and-demand relationship for creative content has inverted.

Newgas points out that historically, 'idea guys' were reviled as the lowest form of tyros because ideas were cheap while execution was difficult. However, as AI tools have incredibly shrunk the gap between 'idea' and 'output', execution is no longer the primary bottleneck. Instead, truly unique 'best ideas' that can capture limited attention have become scarcer than ever.

To illustrate the consequences of this trend, Newgas cites an art history precedent: the invention of photography made realistic representation easy, forcing painting to flee in the opposite direction toward highly conceptual or symbolic pieces. He suggests AI is having a similar effect on media like movies and articles. When high-quality execution can be easily achieved via script drafts plus personalized rendering, cultural consumption may shift toward consuming 'ideas' directly.

Therefore, Newgas predicts that future social credit will accumulate more to the person whose idea it was, rather than the inevitable high-quality execution that follows. Short posts may be valued more than movies, and conjectures more than proofs. This shift implies creators need to reassess their core value proposition: whether to polish perfect outputs or continuously produce original concepts with high distinctiveness.

A persistent challenge in developing AI tutoring systems is obtaining effective student feedback quickly and cheaply. Traditional methods rely on slow, expensive human studies or generic large language model role-playing, which often fails to simulate a specific student's actual capability boundaries, leading to biased tutor training. Microsoft Research introduced StudentSim on October 12, 2026, to address this bottleneck.

The core of StudentSim lies in training personalized AI student simulators from publicly available, de-identified learner records. Unlike simple prompt-based role-playing, it employs a two-stage training pipeline: first learning common response formats and error patterns across pooled student records, then specializing on individual student data. This approach aims to preserve the learner's actual ability boundary rather than allowing the model to demonstrate knowledge far beyond that student's level.

To evaluate these simulators, the research team proposed the StudentSimEval benchmark, covering chess, second-language English writing, and mathematics. The assessment focuses on two dimensions inspired by educational research: 'behavioral fidelity,' measuring whether the simulator matches a specific student's recorded responses on held-out problems, and 'guidance responsiveness,' measuring whether the simulator updates its response after reading tutor guidance. These dimensions correspond to independent performance and performance-with-support in Vygotsky's zone of proximal development.

Experimental results show that StudentSim outperforms leading baselines across all three domains. In chess, for example, StudentSim achieved a behavioral fidelity of 0.5150 and guidance responsiveness of 0.9067, compared to 0.2316 and 0.7186 for GPT-5.4. While Maia2, a specialized chess move prediction model, had higher behavioral fidelity (0.4535), its guidance responsiveness was extremely low (0.2721). This indicates StudentSim better balances mimicking specific student behavior with reacting to teaching interventions.

The study also validated whether such simulators could optimize AI tutors. In a reinforcement learning experiment, tutors trained using StudentSim as the reward source received the highest ratings in blind tests, achieving 90.5% accuracy, significantly higher than the no-RL baseline (75.7%) and the condition using GPT-5.4 as the reward source (71.6%). This demonstrates that simulators trained on real learning data can provide useful feedback signals for improving tutor strategies.

Despite showing value as an early-stage R&D tool, Microsoft Research explicitly states that StudentSim is not a product, is not deployed in classrooms, and is not intended to replace teachers or real learner studies. Human studies remain essential for measuring final learning outcomes, safety, and classroom fit. StudentSim's role is to help researchers test tutor strategies, compare models, and refine training signals before moving to more expensive studies with real students.

AI research often treats evolutionary algorithms as an inefficient fallback, used only as random search when gradients cannot be computed. Akarsh Kumar, a PhD student at MIT and researcher at Sakana AI, refutes this view in a Machine Learning Street Talk podcast episode description. He argues that selection preserves partial solutions, meaning mutations only need to produce useful improvements about 1% of the time for the entire search process to make continuous progress. This mechanism repositions evolutionary algorithms as a primary tool for complex, non-differentiable problems.

Kumar is a PhD student working with Phillip Isola at MIT and the first author of the Fractured Entangled Representation paper, collaborating with Kenneth Stanley, Jeff Clune, and Joel Lehman. In the episode, he explained to host Tim Scarfe that the path a learner takes may shape the structure of what it learns, offering a view of intelligence distinct from the statistical one.

To illustrate the mechanism concretely, the episode description introduces Kumar's lead work on ASAL (Artificial Life via Simulation and Language). Traditional approaches attempt to predict rule behavior, whereas ASAL directly runs simulations and then asks a foundation model what happened. The team applied the method to map all 262,144 Life-like rules and found that the most interesting and open-ended worlds are not uniformly distributed but concentrated on a small island.

The above content comes from the podcast episode description rather than a full transcript. The specific derivation of the 1% figure, complete ASAL methodology details, and specific metrics for the island finding are not expanded in the readable material. This is an attributed explanation; the core mechanism (selection preserving partial solutions) has broad consensus in evolutionary computation, but specific numerical claims and experimental conclusions require independent verification against the original paper (arXiv:2412.17799).

In the field of Natural Language Processing (NLP), Emily M. Bender and Batya Friedman proposed the concept of "Data Statements" in 2018. The paper points out that as technology becomes widely integrated into social applications, researchers need to critically consider how design decisions impact people. The field currently faces two core problems: insufficient generalizability of scientific results and potential real-world harm from an ethical standpoint. Many NLP systems fail or reinforce existing biases against uncovered groups because their training data comes from specific populations.

A Data Statement is defined as a description of dataset characteristics, intended to provide developers and users with context to understand the scope of experimental result generalization, the appropriateness of software deployment, and any biases that may be reflected in the system. Drawing on practices for standardizing research populations in psychology and medicine, the authors advocate for incorporating Data Statements into NLP papers, experiment reports, and system documentation. This increased transparency helps address ethical issues such as exclusion, overgeneralization, and underexposure.

The paper details key roles involved in dataset creation, including speakers, annotators, curators, and stakeholders. For example, indirect stakeholders do not directly use the system but are affected by it, such as when search engine algorithms determine whether their web content is visible. By clarifying these roles and dataset characteristics, Data Statements help the field better understand and describe results, enabling more rigorous scientific and engineering practices.

Consider a hiring screening system trained on historical resume data. If the data shows a low proportion of female candidates for computer programming positions and lower ratings for them, the system may learn and amplify this bias, ranking female candidates lower. Data Statements require disclosure of the source population of training data and potential biases, allowing developers to identify such risks and make adjustments before deployment to avoid unfair treatment of specific groups by the system.

"Data Statements for Natural Language Processing: Toward Mitigating System Bias and Enabling Better Science," written by Emily M. Bender and Batya Friedman, was published in 2018. It is suitable reading for NLP researchers, engineers, and technology policy makers. It is recommended to start with the introduction section of the paper to understand the background and necessity of proposing Data Statements, then delve into Section 5 for specific proposal details.

In his 2003 shareholder letter, Warren Buffett explicitly stated that the core metric for measuring company performance is intrinsic value per share, not book value. He wrote: “It’s per-share intrinsic value that counts, however, not book value.” This view corrects the common investor misconception of equating accounting figures with real wealth. Buffett emphasized that although book value calculations have flaws, at Berkshire, it can still serve as a slightly understated indicator of long-term intrinsic value growth.

As the company's scale expanded and its asset structure changed, simply comparing annual performance against the S&P 500 index became less relevant. Buffett noted that because equity holdings dropped from 114% of net worth in the 1980s to an average of 50% between 2000 and 2003, stock market volatility had significantly reduced its impact on net worth. Therefore, he focused more on whether the company could achieve intrinsic value growth exceeding that of index funds in the future; otherwise, management would be unnecessary.

Regarding capital allocation, Buffett revealed the constraints imposed by scale. He pointed out that when valuations are similar, they strongly prefer owning businesses over stocks. However, in recent years, it has been difficult to find significantly undervalued stocks, and the surge in available capital has exacerbated this difficulty. He warned that investment managers often profit more from accumulating assets than from managing them. This explains why Berkshire shifted toward acquiring companies with durable economic characteristics, managed by honest and capable leaders, at reasonable prices.

For readers today, consider an investor facing a company with a massive market cap but slowing growth. One should not look solely at historical stock price appreciation or book profit growth rates. Buffett's logic suggests we must examine whether the company has lost the flexibility to allocate capital due to its excessive size, and whether its core business can still generate intrinsic value above market averages. This perspective helps identify companies plagued by the "curse of scale."

The Berkshire Hathaway Shareholder Letter 2003 was written by Warren Buffett in 2003. It is suitable for readers who wish to deeply understand core value investing concepts, particularly the definition of intrinsic value and the challenges of large-scale capital allocation. It is recommended to start reading with the sections defining "Operating Earnings" and "Intrinsic Value" to establish a correct evaluation framework.