読書リサーチレーダー投資フレームワーク
ログイン / 新規登録
すべて技術製品投資株式市場その他

読書

2026-11-18133 投稿

更新

Benchmark: Coding Agent Harness Choice Can Cause 5x Cost Variance for Same Model

トピック · AI Coding Agent Cost Benchmarks重大
2026-11-18 06:13 GMT+8

According to the newly released FrontierHarness Eval v1.1 benchmark by Runta and Baseten, the same AI model can cost up to five times more to run depending on the coding agent harness it is paired with. The study tested 22 configurations across five open-weight models, executing 27,280 runs and consuming approximately 83 billion tokens.

While the model itself accounted for 92.8% of cost differences, the harness choice remained significant. Rankings shifted by a median of 11 positions when the model changed, and costs varied by as much as fivefold even when the model stayed the same. For instance, with the DeepSeek model, Cline CLI cost five times as much as OpenCode v1 but delivered only a two-percentage-point higher pass rate.

A major finding concerns resource waste: nearly 65% of estimated inference spending went to runs that either failed verification or timed out. Specifically, 52.7% was spent on failed runs and 11.8% on timeouts, leaving only 35.5% for successful executions. Among the top five best-value harnesses, 68% to 82% of unsuccessful runs ended with the agent submitting a final answer that failed verification rather than timing out.

The research also identified step count as the primary driver of cost, showing a statistically significant correlation (0.51) between steps and cost per successful task. Mini-SWE-agent, despite having the smallest initial prompt, averaged 38 steps per successful task, ranking 17th in cost efficiency. In contrast, the most cost-effective harnesses relied heavily on shell commands and rarely used subagents or context summarization.

ソース:The New Stack

リサーチ

Anthropic Discloses Post-IPO Ownership: Amazon Holds 14.7%, Google 13.7%

トピック · Anthropic IPO & Ownership重大
2026-11-18 06:29 GMT+8

US-based AI company Anthropic disclosed new details on Tuesday regarding its largest investment agreements with major cloud service providers Amazon and Google in connection with its initial public offering (IPO). According to the disclosure, Amazon’s total stake will come in at just shy of 15%, specifically 14.7%.

Notably, Amazon’s potential ownership could be significantly higher. When Amazon invested billions in Anthropic in 2023 and 2024, it obtained promissory notes that could convert into equity. These notes carry conversion prices of about $2.09 or $5.05 per share, depending on the year of investment. Amazon has opted to convert a portion of them into a special class of non-voting shares as part of the IPO, giving it about 15% ownership. If it converted all held notes right now, it would own just under 24% of Anthropic, precisely 23.8%.

Meanwhile, Google will own just under 14%, specifically 13.7%, also in non-voting shares. These holdings come from the conversion of notes as well as the “private placement” sale of shares as part of the IPO. A group of major tech companies—including AMD, Broadcom, Google, Micron, Microsoft, Nvidia, and TSMC Partners—have agreed to buy $26 billion worth of this non-voting stock. Google and Nvidia each agreed to purchase $10 billion worth.

While many investors stand to make billions from Anthropic’s IPO, Amazon is one of the biggest winners. This does not even count the revenue AWS stands to earn from Anthropic’s commitment to spend more than $100 billion on its cloud services over the next decade. This ownership structure highlights the deepening capital and business ties between frontier AI labs and cloud infrastructure providers.

ソース:TechCrunch

リサーチ

OpenAI Releases GPT-6.2 Astra and Sol, Cutting Astra Costs by 20%

重大
2026-11-18 05:21 GMT+8

OpenAI released two new AI models, GPT-6.2 Astra and GPT-6.2 Sol, on November 17, 2026, designed for use in ChatGPT Work, Codex, and the OpenAI API. This launch comes just days before rival Anthropic's planned initial public offering (IPO).

According to OpenAI, GPT-6.2 Astra is tailored for complex professional tasks, such as creating presentations, spreadsheets, and documents, claiming stricter adherence to templates and industry-specific standards. Compared to the previous generation, GPT-6 Astra, the prices for input and output tokens have dropped by 20%. In the "Agent's Last Exam" benchmark, OpenAI claims GPT-6.2 Astra achieved a new top score, exceeding GPT-6 Astra by 5.2 points while costing less than a third of the previous model.

GPT-6.2 Sol is positioned as a cheaper everyday variant, maintaining the same price level as its predecessor, GPT-6.1. OpenAI states that this model is smarter and more efficient for multi-step agentic workflows and is used internally as the standard model for both technical and non-technical tasks. Although OpenAI describes Astra as the best AI model for researchers, in the Artificial Analysis Intelligence Index, GPT-6.2 Astra scored 62.1 points, slightly trailing Anthropic's Claude Fable 5.2 (64.6 points), albeit at a fraction of the cost.

ソース:t3n.de

リサーチ

OpenAI CEO Sam Altman: AI Industry Will Continue Without Jensen Huang, No Concerns About Funding

重大
2026-11-17 10:33 GMT+8

On November 16 local time, OpenAI CEO Sam Altman publicly commented for the first time on the resignation of Nvidia founder and former CEO Jensen Huang at the Semafor Silicon Valley Summit. Altman stated that while Huang has always been his hero, "heroes have flaws," and Huang's departure will not prevent the entire AI industry from continuing to move forward.

Addressing concerns that interim CEO Colette Kress might scale back investments and impact AI industry liquidity, Altman expressed optimism. He noted that Nvidia is now a massive corporation whose guarantee capabilities are still trusted by the market. Reports previously indicated that Nvidia agreed to provide up to $105 billion in financial guarantees for OpenAI's Ohio data center project, with total investment commitments in the AI ecosystem exceeding $100 billion.

Altman also responded to recent controversy over his comments on RAM. He clarified that his original words were misunderstood, stating he supports individuals having large memory in their PCs and did not advocate concentrating all scarce memory resources in data centers. Additionally, he revealed that OpenAI has begun using AI to accelerate its own model development but has not yet initiated fully autonomous Recursive Self-Improvement (RSI) to ensure safety.

Regarding capital markets, Altman said he is not worried that competitor Anthropic's IPO launch will drain funds, arguing that demand in the primary market far exceeds the supply of quality assets. He emphasized that while OpenAI will eventually go public to allow ordinary people to share in AI's economic benefits, the company is not in a rush during this period of industry turbulence.

ソース:量子位

リサーチ

Prediction Market Platform Kalshi Plans Stock Trading Launch via Acquired Broker, Challenging Robinhood

トピック · Kalshi Stock Trading Expansion重大
2026-11-17 23:35 GMT+8

According to reports from The Wall Street Journal and CNBC, US-based prediction market platform Kalshi is preparing to offer stock trading on its platform. The company reportedly plans to leverage its acquisition of broker FB Securities to launch equities and exchange-traded funds (ETFs) "within the next two months," positioning itself directly against rivals like Robinhood.

The rumored expansion into traditional securities trading comes as Kalshi faces growing legal battles over its core prediction market business across the United States. New York, Nevada, New Mexico, Washington, Connecticut, and several other states have accused Kalshi of operating illegally in their jurisdictions, while Utah recently filed criminal charges against the company and its CEO. Meanwhile, the U.S. Commodity Futures Trading Commission (CFTC) is pushing back against state efforts to regulate prediction markets, arguing it has "exclusive" authority. Nine Democratic senators wrote a letter to the CFTC this week urging it to stop interfering with state bans.

ソース:The Verge

リサーチ

Fireworks AI CEO Lin Qiao: Two Standard Modes for AI API Data Retention, Advocating Ephemeral Processing When Workflows Allow

トピック · Fireworks AI Data Retention重大
2026-11-14 10:46 GMT+8

In AI infrastructure, data retention is one of the most misunderstood terms. In a November 2026 post on X, Lin Qiao, co-founder and CEO of Fireworks AI, proposed a clear framework for thinking about this: AI inference and training APIs follow two standard design patterns, distinguished by who manages the data lifecycle and where.

The first pattern is customer-managed retention: upload, use, delete. Users create datasets or files, submit jobs, and then explicitly delete the data. The platform keeps the data stored and available until it receives a deletion instruction. Batch APIs are a typical example—OpenAI Batch, AWS Bedrock Batch, GCP Gemini Batch, and Fireworks Batch all follow this standard. The platform cannot know when the user no longer needs the data unless the user defines it; deletion is the user's responsibility.

The second pattern is ephemeral processing: send, compute, return. Users send data, the platform uses it during request processing, and returns the result without retaining the content afterward. Standard inference APIs and serverless training fall into this category. Applications own their conversation history, with no separate platform cleanup step required.

Lin Qiao used an intuitive analogy: saving a spreadsheet to a cloud drive versus passing numbers to a calculator. The former requires the service to keep the file for later use, while the latter only requires computing the answer and returning it immediately.

She advocates that ephemeral processing should be prioritized whenever workflows allow, not for performance or cost reasons, but to avoid human error. Ephemeral processing leaves no forgotten datasets and no missed deletion calls. Every retained copy of data adds something that must be protected and eventually deleted. If data never reaches disk, it cannot be taken from disk later.

The practical implication of this framework is that zero data retention (ZDR) is not a vendor label but an endpoint attribute. A single platform can provide endpoints for both modes. OpenAI’s Chat Completions defaults store to false, falling under ephemeral processing; however, its Batch and Files endpoints involve customer-managed retention, with data persisting until deletion. Amazon Bedrock’s Converse API uses content solely to generate responses, whereas its Batch inference reads and writes to user-specified S3 locations. Google’s Gemini text generation is stateless when store is false, but its Interactions API has store defaulting to true.

However, this framework has clear boundaries. Fine-tuning based on datasets requires data at rest, making customer-managed retention indispensable in these scenarios. Additionally, even when zero retention is declared, both OpenAI and AWS note that content flagged as CSAM may still be retained. Default settings, approval requirements, and exception conditions vary significantly across platforms, so developers must verify settings per endpoint.

It is worth noting that the article explaining this framework on MarkTechPost is explicitly labeled as sponsored content, with the Fireworks team providing thought leadership and resources. While Lin Qiao’s framework serves as a clear technical classification, her preference for ephemeral processing aligns with Fireworks’ own product positioning for its ephemeral APIs, indicating a potential conflict of interest.

ソース:X (formerly Twitter)

リサーチ

French AI Firm Mistral Preparing for European IPO, Not This Year, CEO Says

重大
2026-11-18 05:49 GMT+8

Arthur Mensch, CEO of French artificial intelligence startup Mistral, announced that the company plans to go public in Europe and is currently preparing for an initial public offering (IPO). Speaking at Semafor’s "Silicon Valley & The World" event, Mensch stated, "The natural place to list is Europe." He clarified the timeline, adding, "We’re not going to do this this year, but we’re definitely getting ready for it."

Mistral, which primarily develops open-weight AI models, positions itself as a provider of sovereign AI, offering an alternative to closed US models and open Chinese models. Mensch noted that most of the company's investor base and half of its customer base are European. While other European tech companies such as Arm, Spotify, and Klarna have listed in New York to access higher valuations and a larger pool of investors, Mensch emphasized that Europe remains the preferred jurisdiction for Mistral's listing.

ソース:semafor.com

リサーチ

SEC reportedly opens civil probe into Nvidia over investor disclosures

トピック · Nvidia SEC Probe重大
2026-11-18 05:48 GMT+8

According to The Wall Street Journal, the U.S. Securities and Exchange Commission (SEC) has opened a civil probe into AI chipmaker Nvidia to investigate whether the company's disclosures misled investors. This report was widely cited by tech media, including TechCrunch, on November 17, 2026.

The investigation comes as Nvidia faces multiple crises. Former CEO Jensen Huang resigned following reports of extramarital affairs, one involving an alleged Chinese spy who served as an executive at Singapore-based data company Megaspeed. Megaspeed has been under investigation by the U.S. Department of Commerce since late 2024 on suspicion of diverting Nvidia chips to China in violation of U.S. export laws.

Beyond the SEC probe, Nvidia is under pressure from Congress and the administration. Treasury Secretary Scott Bessent accused Nvidia of knowingly selling powerful chips to China to "goose their revenues and their earnings." Congress has also demanded that Huang answer questions about how his personal relationships might have influenced his stance on loosening export restrictions.

The news of the SEC probe coincides with Nvidia's Q3 earnings report. Despite revenue hitting $115.8 billion, up 103% year-over-year, the company's stock has fallen more than 10% since Huang's departure. The SEC's involvement further exacerbates investor concerns regarding corporate governance and legal compliance risks.

ソース:TechCrunch

リサーチ

The First Cost Lever for Generative AI: Determine if a Step Needs a Model Before Deciding Which One

重大
2026-11-18 05:14 GMT+8

The most overlooked cost lever in generative AI systems is not which model you choose, but whether a step actually requires calling a model at all. The Amazon Web Services (AWS) machine learning blog notes that teams often find a model call effective for one step during prototyping and then adopt it as the default for subsequent steps—including input validation, request body formatting, and routing selection. While costs go unnoticed at prototype volume, at production scale these low-value model calls become one of the largest expense items in the architecture: costs grow with request count, latency stacks up layer by layer in chained calls, and identical inputs may yield different results, making testing and auditing more difficult.

The blog proposes a three-question decision framework requiring sequential answers for each workflow step: First, can success be verified by a rule, such as format, pattern, keyword, or numeric range? Second, can the model write the rule once, such as a regular expression or JSON schema? Third, does the step select one answer from a fixed list, such as routing, classification, or a yes/no judgment? If the answer to all three questions is no, the step is open-ended and should be handed off to an appropriately sized model.

The core insight of the framework is that most routing and classification steps fall between rules and open-ended generation. Rules cannot cover every case, but these steps also do not require generating text—they only need to provide one answer from a predefined list of fixed options. The blog categorizes these steps as 'decision rather than generation,' advocating for a component that restricts output to allowed options and provides an escalation signal: confident answers proceed directly, while uncertain ones escalate to a larger model. On AWS, this can be implemented in four ways: using a small Amazon Bedrock model with structured outputs to limit answers to enumerated options plus an 'unsure' option; using embedding similarity to match the nearest labeled example, using the similarity score as an escalation signal; deploying the experimental open-source decision model Strands Decider, which returns calibrated confidence; or training a classifier on Amazon SageMaker AI to return category probabilities with low latency and overhead once labeled data is available. Regardless of the method chosen, the escalation signal must be validated against your own labeled data and thresholds set accordingly.

Before the decision framework lies an even simpler optimization: a deterministic fast path. Many messages carry unambiguous signals, such as 'refund' or 'reset password,' which keyword matching can resolve in single-digit milliseconds. If rules handle 95% of requests, only one in twenty requests reaches the model—this is before any savings from model selection are even realized. The blog recommends using Choice states in AWS Step Functions to branch before invoking any model, using Lambda to host rules requiring more logic, and using API Gateway request validation to reject malformed input before computation begins, sending only the residual cases that rules cannot resolve to constrained decision calls or models.

ソース:Amazon Web Services

リサーチ

Nvidia Packs Optical Components into Switch Chip Packages to Target Data Transfer Bottlenecks in AI Clusters

トピック · Nvidia CPO Integration重大
検証済み 2026-11-18 05:27 GMT+8

The bottleneck in large-scale AI clusters is shifting from computation to data transfer. Gilad Shainer, Senior Vice President of Networking at Nvidia, said in an interview with Scientific American that the rule is simple: use copper whenever you can. But copper is only suitable for short-distance, low-bandwidth scenarios. When tens or even hundreds of thousands of compute nodes and memory units in a data center need to communicate with each other at extremely high speeds, the problems with copper become apparent: electrons carry charge and repel each other, requiring power to drive them; copper itself acts like a filter, smearing clear electrical pulses into blurred noise.

Photons do not have this problem. They carry no charge and hardly interact with one another. A laser beam passing through glass fiber optic cables experiences negligible signal attenuation over several kilometers. However, data inside chips is in electronic form. To hand it off to photons for transmission and then back to electronics for processing, a conversion device is required—a transceiver.

The traditional approach is to make transceivers into pluggable modules inserted into the front panel of switches. Craig Thompson, Vice President of Interconnect Products at Nvidia, explained the fatal weakness of this method: physical connectors can loosen and move, having pins, solder joints, and epoxy, and every insertion or removal affects adjacent modules. Shainer provided a figure: the presence of humans in data centers is itself responsible for 5% of downtime.

The idea behind Co-Packaged Optics (CPO) is to integrate the optical engine directly into the package of the switch chip. Nvidia’s latest switches—the Quantum-X and Spectrum-X—feature small photonic chips arranged on both sides of the central large chip, communicating face-to-face with electronic chips. Fiber optic cables plug directly into the front panel of the switch, eliminating the need for pluggable modules. Thompson stated that their modeling shows this can improve Mean Time Between Interruptions (MTBI) by at least 10 times.

Improvements in power consumption are equally specific. Old-style pluggable transceivers each carried their own lasers and required a dedicated chip to clean up electrical signals, consuming about 25 watts per unit together. A single switch holds dozens, and these modules can consume 10% of the total power in an entire data center. Placing the optical engine next to the chip means signals barely need to travel over copper, allowing the same electricity bill to buy more computing power.

From a strategic perspective, Nvidia designs approximately 85% of the graphics processors used for AI globally and has already sold most of them. If it simultaneously controls the way these chips communicate, competitors will find it difficult to replace Nvidia from AI infrastructure. The article also mentions that Nvidia has invested in laser manufacturers Coherent and Lumentum to help them expand capacity, and is collaborating with chip foundries to develop manufacturing processes for photonic circuits.

ソース:Scientific American

リサーチ

Supply Chain AI Startup Pedestal AI Raises $4.7 Million Seed Round Led by Eclipse

トピック · Pedestal AIクイックテイク
2026-11-18 02:04 GMT+8

Supply chain AI startup Pedestal AI announced on November 17, 2026, that it has emerged from stealth mode with $4.7 million in seed funding. The round was led by Eclipse, with participation from RRE, Converge VC, Timescale Ventures, and VitalStage. The company plans to use the funds to expand its product and engineering teams, grow its marketing organization, and support additional client deployments.

Pedestal AI focuses on using agentic AI to remove friction from the delivery process. According to TechCrunch, the company reduced missed deliveries by 20% for a major furniture retailer by deploying agents that call customers 30 minutes before delivery to collect special instructions, falling back to text messages if calls fail. This system also gathers data on friction points to further optimize operations.

CEO Raj Aggarwal stated that modern AI systems are particularly useful for handling messy, disparate data sitting in legacy supply chain systems. As alumni from autonomous driving teams launch startups in this space (such as Atomic and Loop), Pedestal AI's entry highlights investor interest in applying reinforcement learning to high-volume, measurable outcomes like last-mile delivery.

ソース:TechCrunch

リサーチ

Meta Executive Posts Deepfake Video of OpenAI CEO to Promote Muse, Citing AI Labeling Policy

重大
2026-11-18 05:00 GMT+8

Meta Chief AI Officer Alexandr Wang recently posted an AI-generated video on Instagram to promote the company's consumer AI agent, Muse. The video features Wang, Meta CEO Mark Zuckerberg, and OpenAI CEO Sam Altman in a four-way video call. When Altman joins the call, he is depicted wearing a furry hooded suit and face paint resembling Jolly, the anthropomorphic mascot of Muse. The clip is captioned "When the homies meet about openai product strategy," referencing the ongoing AI agent race.

This action sparks discussion on the use of AI-generated content (AIGC) in commercial competition. While AI imitations of real people led to strong public rebukes and legal threats just a few years ago, AI content has now become part of the texture of social media. Meta spokesperson Tracy Clayton responded that Wang's post is labeled as AI content, in accordance with Meta's policy. Clayton also noted that Altman has made his likeness available for AI content generation, citing the flood of fake Altman videos when OpenAI launched its Sora app.

It is worth noting the difference in Meta's AI labeling requirements for static images versus dynamic videos. While videos must be labeled, static images currently have no such mandatory requirement. Previously, an unlabeled AI-generated image of actress Mia Goth spread widely on Instagram, drawing criticism from her representative regarding the protection of public figures' likenesses. In contrast, Wang's video involving a competitor's CEO is considered compliant due to its AI label.

This event reflects a trend where major tech executives are increasingly using generative technology to mock or depict competitors when promoting their own AI products. It serves as a test of platform AI content policies and marks a normalization of deepfake technology in corporate marketing.

ソース:The Verge

リサーチ
次へ →
ログイン / 新規登録
読書リサーチレーダー投資フレームワーク

Practical advice for developers: map every endpoint you call to one of the two modes; prioritize ephemeral processing endpoints when you only need results; plan deletion calls for each upload and record documented expiry times in your documentation; audit persistence flags like store; define ZDR by processing path in contracts, then verify the settings in your account.

When a model is indeed needed, choosing the cheapest and fastest model that meets the quality floor is often not the newest or largest one. The blog suggests defining four requirements before comparing models: quality floor (the minimum acceptable output quality on a golden evaluation set), cost ceiling (the maximum paid per request or per successful outcome), latency target (the p95 or p99 response time the experience can tolerate), and compliance constraints (such as data residency requirements). Amazon Bedrock offers multiple model families and sizes behind a single API, supporting intelligent prompt routing and model distillation, allowing you to select the right size for each task without changing your integration.

The blog concludes with a four-step practical path: first, tag every invocation (deterministic code, constrained decision call, or model call) and track the cost per successful outcome; second, move rule-able steps to run on Lambda or Step Functions, which usually yields the largest and safest gains; third, add constrained decision calls with escalation signals for fixed-option steps; and finally, run the golden evaluation set before switching models, keeping models behind configuration switches for easy rollback. The core principle is simple: ask whether a step needs a model first, then ask which model to use.

Competitors are not absent. Broadcom is already on its third generation of co-packaged switches, while AMD and Intel are also advancing similar technologies. However, Nvidia’s advantage lies in scale: the number of chips it sells is sufficient to support building the entire silicon photonics ecosystem it needs.

A further direction is optical interconnects within the chip itself. Processors are limited by the physical size of photolithography masks; a single chip cannot exceed approximately 850 square millimeters. To make them larger, multiple small chips must be stitched together. These small chips communicate via fine copper wires, and the denser the wires, the more severe the crosstalk. Sudip Shekhar, an electrical engineer at the University of British Columbia, believes photons may perform better than copper in these short-distance hops.

It should be noted that the data regarding the 10x MTBI improvement and power savings cited in the article are results of internal Nvidia modeling, not independently verified by third parties. The article’s author also mentioned that the Nvidia souvenir he received—a model of a 2026 server rack—still contained network interconnect components from before the era of co-packaged optics, indicating the rapid pace of this technology's iteration.