読書

ElevenLabs has released its latest text-to-speech models, Eleven v4 and Eleven v4 Turbo. According to the official video description, v4 is built on an entirely new architecture that interprets tone, pacing, emotion, character, and context to generate speech that feels performed. It understands who is speaking and what just happened, simulating the overlap, timing, and reactive energy of real conversations in multi-speaker scenarios, even without explicit audio tags.

The release includes two models based on the same research: Eleven v4 is designed for produced content such as narration, character work, voiceover, and dubbing. Eleven v4 Turbo is the low-latency version tailored for voice agents, live assistants, and interactive experiences, running at around 100 ms median latency while maintaining the same expressive range and bidirectional input streaming capabilities.

Technically, v4 follows audio tag sequences far more reliably than v3, including stacked tags in a single line. Voice cloning stability has improved, with speaker identity holding across long-form work, regenerations, and dialogue turns. Enhanced context stitching ensures regenerated lines fit seamlessly into surrounding chapters, which is critical for audiobooks. Eleven v4 covers over 90 languages, with the strongest gains over v3 reported in Japanese, Brazilian Portuguese, Mandarin, and Cantonese, plus better International Phonetic Alphabet (IPA) support for tricky pronunciations.

Crusoe published a video case study on September 24, 2026, featuring Yutori. In the video, Yutori Co-founder and Chief Scientist Dhruv Batra explains how his 14-person team works with Crusoe from training through inference.

According to the video description, Yutori's Navigator n2 model was trained in part on Crusoe Cloud. The model reads screens like a person and clicks, scrolls, and types accordingly, allowing it to work on websites and apps that lack APIs.

Batra notes that Yutori uses Crusoe Managed Inference with a Tailored Deployment and a cluster-wide KV cache for its cache-heavy, multistep workload. The video claims this architecture enabled a roughly 10x scale-up on the same growing footprint, serving every Navigator generation from n1 to n2.

Additionally, the video cites OSWorld 2.0 benchmark data, stating a cost of approximately $1.46 per task for Yutori, compared to about $13 to $40+ per task for frontier models. These figures and performance claims originate from joint marketing materials by the vendor and customer and have not been independently verified.

Insilico Medicine released a short documentary titled "Cell Paper 2026: Short Documentary" on YouTube on September 21, 2026. The video accompanies the company's landmark study, which was selected as the cover story for the September 17, 2026 issue of Cell magazine. It presents a complete open-source AI toolkit designed to evaluate and accelerate longevity research.

According to the publisher's description, the documentary explores how domain-specific language models and multi-agent platforms move beyond conventional chatbots. Specifically, it demonstrates how these technologies are used to autonomously plan research workflows, evaluate multi-omics data, and identify candidate targets for anti-aging interventions.

For readers interested in the application of AI in biological sciences, this resource provides visual context on how Insilico Medicine and its international research partners integrate AI into addressing the complex challenges of aging. While no full transcript or detailed argumentation is available, the official description clarifies the core topics and technical applications, making it suitable as a quick introductory material for understanding the design philosophy of the open-source toolkit.

The official Replit channel published a video update titled "New Agent Models, Muse Connector & VR Apps | This Week in Replit" on September 26, 2026. The video outlines three main changes: first, the addition of GPT-6 Sol, GPT-6 Luna Fast, and Claude Opus 5.5 to Replit Agent, with availability depending on user plans; second, Replit becoming an approved connector within Meta's AI assistant Muse, allowing users to push app-building tasks from Muse chats to Replit; and third, support for building VR applications for Meta devices, announced at Meta Connect.

As a resource selection, this video is suitable for developers and technical decision-makers seeking quick insights into Replit's latest product movements. It provides official confirmation regarding model availability, ecosystem integration (specifically with Meta), and support for new hardware platforms. Note that the video description does not include specific technical implementation details, performance benchmarks, or pricing changes, serving primarily as a feature launch announcement.

Vessel AI, a South Korean GPU cloud infrastructure provider, announced on October 16 that it has signed a business cooperation agreement with Korea Investment & Securities (KIS) to facilitate GPU financing.

Under the agreement, KIS will assist in expanding investment scale and fundraising, while Vessel AI will support business execution based on its experience in operating GPU cloud platforms. Both parties stated that specific investment structures and amounts will be determined through consultation for each future project.

Vessel AI noted that building and expanding GPU infrastructure requires large-scale capital investment, making a stable financing base a core element of its business expansion. The partnership aims to combine financial institutions' fundraising capabilities with Vessel AI's 'neocloud' business model.

This episode features a panel from the Frontier Data Summit 2026 with Alex Shaw, Founding Member of Technical Staff at Laude Institute, creator of Harbor, and co-creator of Terminal-Bench, and Sanmi Koyejo, Director of the Stanford Trustworthy AI Research lab. The moderator was Francesca Vera, AI Community Manager at Snorkel AI.

According to the publisher's chapter list and description, the experts discussed why agent performance is hard to predict, why aggregate benchmark scores mislead, how benchmarks get saturated or exploited (such as through reward hacking), and what is required to maintain evals over time. Topics also included broken RL environments, applying measurement science from other fields, and how open versus private benchmarks should coexist.

For readers tracking AI capability progress, this resource provides context on the limitations of current evaluation systems. It does not present new experimental data or conclusions but offers insights from core builders in the field on why existing leaderboard mechanisms may fail to reflect true model capabilities and what problems infrastructure solutions like Harbor aim to address.

ElevenLabs has published a video introduction on its official channel regarding the Eleven v4 Turbo voice model, focusing on how it empowers real-time conversational experiences within the ElevenAgents platform. As the company's claimed fastest and most emotive voice model, v4 Turbo's core selling point lies in reducing median inference latency to approximately 100 milliseconds while maintaining a broader range of emotional performance and stability during long calls.

The video indicates that this low-latency capability enables AI agents to handle sensitive scenarios requiring high empathy and immediate response, ranging from urgent requests in medical institutions to transaction disputes in banking. Additionally, Eleven v4 Turbo supports over 90 languages, allowing global brands to reach users in different languages with a unified voice. Architecturally, the model is co-optimized with ElevenLabs' existing transcription and turn-taking models, aiming for faster responses and more natural interruption handling.

For professional readers evaluating AI voice interaction infrastructure, this video provides key performance metrics and application scenario definitions from the vendor's perspective. It is important to note that claims such as 'native-level accuracy' and 'most emotive' are self-declarations by ElevenLabs; the video itself does not include independent benchmark comparison data or detailed technical whitepapers, primarily focusing on intuitive product capability demonstrations and market positioning.

Episode 2 of the Unsupervised podcast features Boris Sofman, co-founder of Bedrock Robotics, and Dave Ferguson, founder of Nuro, to discuss the critical evolution of "Physical AI" in the context of heavy autonomous machinery. Both guests bring significant expertise from the autonomous vehicle sector, focusing on how large-scale autonomous machines can operate safely alongside humans.

According to the episode description, the conversation spans technical perspectives from both Nuro and Bedrock Robotics, addressing three key areas: the high-stakes safety challenges of human-machine interaction, the engineering rigor required for real-world reliability, and the impact of automation on labor dynamics. This resource is recommended for readers interested in the practical deployment constraints and safety standards of physical AI.

This resource is a talk video from Frontier Data Summit 2026, presented by Steven Dillmann, a Stanford PhD student and Project Lead for Terminal-Bench-Science.

Terminal-Bench-Science aims to evaluate AI agents' performance in real scientific research workflows, covering five major scientific domains: life, physical, Earth, mathematical, and engineering sciences.

The video covers the scientist-led task review process, examples and results from the 0.1 release, leaderboard results across models and domains, and a roadmap for how researchers can contribute tasks.

The project received support through Snorkel AI’s Open Benchmarks Grants. For researchers and developers interested in the actual capability boundaries of AI agents in scientific discovery, this video offers direct reference material on benchmark construction methods and preliminary evaluation results.

In the 2009 shareholder letter, Warren Buffett provides a detailed explanation of how Berkshire Hathaway measures its performance and justifies the use of per-share book value as the central metric. Although the ideal standard for measuring our yearly progress would be the change in Berkshire’s per-share intrinsic value, Buffett notes: "The ideal standard for measuring our yearly progress would be the change in Berkshire’s per-share intrinsic value. Alas, that value cannot be calculated with anything close to precision, so we instead use a crude proxy for it: per-share book value." He acknowledges this measure has flaws, as book value often underestimates intrinsic value for most companies, but at Berkshire—especially within its insurance operations—the discrepancy is substantial. Nevertheless, he and Charlie Munger still consider it the most useful tool for tracking changes in intrinsic value.

The letter also offers a profound insight into long-term performance versus short-term volatility: "Defense over offense." By comparing data from 1965 to 2009, Buffett finds that Berkshire significantly outperformed the S&P 500 in years when markets declined, while not necessarily achieving large gains during market rallies. He concludes: "In other words, our defense has been better than our offense, and that’s likely to continue." The core of this strategy lies in avoiding permanent capital loss, rather than chasing the highest return each year. Additionally, he notes that as the company grows larger, future advantages will shrink: "But huge sums forge their own anchor and our future advantage, if any, will be a small fraction of our historical edge."

For contemporary investors, this letter provides a clear framework for evaluating their own investment systems: Do they have a rational and unshakable standard for measuring success? Buffett emphasizes: "That keeps us from the temptation of seeing where the arrow of performance lands and then painting the bull’s eye around it." If an investor assumes they possess stock-picking skill simply because they beat the market in one year, while ignoring drawdowns during bear markets, they may fall into the trap of hindsight bias. True long-term compounding comes from survival capability in adverse conditions and steady value growth—not speculative gains during bull markets.

"Berkshire Hathaway Shareholder Letter 2009" was published by Warren Buffett in February 2010 (covering fiscal year 2009). It is recommended for readers seeking to establish a rational performance evaluation standard and understand the primacy of risk control in value investing. It is advised to begin with the "How We Measure Ourselves" section, focusing especially on the candid discussion of book value’s limitations and the data-backed support for the defensive strategy.