독서리서치레이더투자 프레임워크
로그인 / 회원가입
전체기술제품투자상장시장기타

독서

2026-11-19129 게시물

새로고침

OpenAI Internal Report: Researcher Daily Token Costs Exceed Salaries, But Complex Engineering Success Rates Remain Low

중대
2026-11-19 21:08 GMT+8

OpenAI published its second internal research acceleration report on November 19, 2026, detailing the deep embedding of AI agents in the development of new models. The data shows a surge in the use of internal tools like Codex and Dots. At regular API list prices, the median researcher consumed $2,400 worth of tokens per day in October, up 62 percent from the previous month. Among the top 10 percent of users, this theoretical value doubled to $23,000 per day.

The report notes that these agents run the equivalent of 5.7 workdays per researcher per calendar day, and about eight in pre- and post-training teams. More than two-thirds of tracked researchers ran at least four agents in parallel at peak usage. Notably, the theoretical API equivalent of this agent runtime was twice as high as researchers' regular salary costs in September (excluding equity packages, which make up 70 to 80 percent of total compensation).

Despite high adoption, the capability boundaries of AI agents remain clear. According to an Epoch AI taxonomy, systems perform well on simple tasks: autonomous success rates for training-run monitoring stand at 79 percent, and for independently answering technical questions at 72 percent. However, for more complex engineering work, the success rate for building evaluation and reinforcement-learning environments sits at 29 percent, and for writing code for inference and serving systems at 25 percent.

Internal surveys indicate extreme reliance on these tools. In a non-representative survey of 101 employees, researchers said at the median that they would rather give up 40 percent of their team's staff or 30 percent of their compute capacity than go three months without current coding agents. A separate quick poll of 28 employees found that 27 preferred the latest internal model over a mid-career human engineer.

Qualitative interviews revealed downsides: models lack 'taste' for good research ideas and code quality, often producing 'AI slop' and unusable data artifacts that require close human oversight. Key conceptual ideas for GPT-6 Astra came exclusively from humans. OpenAI cautions that current human-AI collaboration data may not extrapolate linearly, pointing to potential abrupt advances in deep-learning research similar to recent breakthroughs in mathematics.

출처:The Decoder

리서치

YC-backed AI Startup LATO Raises $4.7M Seed to Automate Private Equity Due Diligence Interviews

간단한 요점
2026-11-19 17:50 GMT+8

LATO, a Y Combinator Summer 2026 startup, has raised $4.7 million in seed funding co-led by Air Street Capital and Axiom Partners. The round also included participation from Y Combinator, Montis VC, Inception Fund, and angel investors such as Paul Graham and co-founders of ElevenLabs, Vercel, Snowflake, and Rogo.

LATO uses AI agents to conduct and analyze customer and industry expert interviews, aiming to accelerate commercial due diligence for private equity investors. Traditionally, investment teams spend weeks finding and arranging cross-border interviews with target market customers and industry experts to assess demand, competition, and pricing assumptions. LATO's platform allows investment teams to gather this primary data in hours rather than weeks by automatically identifying respondents, interviewing them in their native languages, and analyzing the conversations alongside public information and fund data.

According to the company, its agents have conducted more than 1,000 expert interviews across over 35 countries since its first project in late July 2026. The startup also reports an annualized revenue run rate exceeding $1 million approximately 100 days after its first study. Its customers include investment funds Verdane, Blume Equity, and North Island. The new capital will be used to expand engineering, research, and sales teams and to broaden the range of studies its agents can run independently.

출처:Tech Funding News

리서치

LOIs in M&A Transactions: Why Sellers Lose Negotiation Leverage After Signing

중대
2026-11-19 20:31 GMT+8

In M&A transactions, most people focus on the signing of the definitive agreement and the public announcement, but the moment that truly determines how much money the seller receives often comes earlier—the moment the LOI (Letter of Intent, or term sheet) is signed. The author of the OnlyCFO blog, a CFO with hands-on M&A experience, made this judgment a core argument in his November 2026 analysis of the full M&A process: after the LOI is signed, terms basically will not improve because the exclusivity clause has already handed the negotiation initiative to the buyer.

The logic behind this mechanism is not complex. Exclusivity clauses in LOIs typically require sellers not to evaluate or seek competitive bids for 15 to 90 days. Buyers demand this clause because due diligence is costly; they do not want to invest significant resources only to discover there are other parties negotiating. Once the exclusivity clause is signed, the balance of power shifts fundamentally: sellers lose their ability to pressure the buyer with competing offers, while buyers gain ample time and information advantages to re-examine transaction terms.

The author emphasizes specifically that although price terms in an LOI are themselves non-binding, the practical effect is nearly equivalent to locking them in. He writes in the article: 'Terms (especially price) basically never get better after the LOI is signed.' If other matters are not clarified during the LOI stage—such as what counts as a debt-like item—buyers will usually exploit such ambiguity at the definitive agreement stage to secure more favorable conditions. 'Buyers typically win post-LOI if it wasn’t clear.'

Therefore, what really requires effort to negotiate at the LOI stage is not the price number itself, but the definitions and boundaries that will be reinterpreted in the final agreement. The author outlines several specific negotiation points:

First, the scope of debt-like items. This is the category most easily expanded by buyers in hindsight. The author suggests that if a buyer attempts to categorize certain items as debt-like to lower the equity consideration, the list must be fixed during the LOI stage, leaving no room for ambiguity.

Second, the classification of deferred revenue. Some buyers push to include deferred revenue as a debt-like item, meaning it is fully deducted from the consideration. The author believes sellers should strive to have deferred revenue classified as working capital rather than a debt-like item. If the buyer refuses to budge, a compromise is to calculate it based on cost to serve (deferred revenue multiplied by the COGS percentage). For companies with high gross margins, this compromise is far more favorable than deducting 100% of deferred revenue.

Third, double counting between working capital and debt-like items. Buyers sometimes count the same amount into both working capital adjustments and debt-like deductions. The author reminds sellers to confirm at the LOI stage that there is no double counting.

출처:onlycfo.io

리서치

AI Sovereignty Analyst Govrin: No Solution for Hyperscalers Regarding the US CLOUD Act

중대
2026-11-19 20:24 GMT+8

During the SC26 Supercomputing Conference, Amit Eyal Govrin, lead analyst for sovereign AI at theCUBE Research and chief executive officer of Agentcy Labs, released the results of an assessment covering seven hyperscale cloud providers, 40 product offerings, and five regions. The most surprising discovery for Govrin was not that any specific provider performed poorly, but that every provider faced an intractable legal risk when confronted with the same issue: the United States CLOUD Act.

Govrin’s central argument is that AI sovereignty is not a single checkbox but a stack of control layers that must be examined one by one. He breaks sovereignty down into five pillars: territorial (the physical location of data centers), operational (jurisdiction over daily operations and incident response), technological (the actual location where models and data are processed), legal (applicable judicial jurisdiction), and financial (costs and contractual constraints). The territorial layer is the easiest to satisfy; providers can simply claim compliance by building a data center in the target region. However, Govrin points out that the operational layer is far more complex than it appears.

He cited the Amazon Web Services (AWS) Frankfurt region as an example: as a German customer, your data is indeed stored within the European Union, and AWS operates a sovereign region in Brandenburg. However, if that region experiences an outage, PagerDuty alerts may not be routed to operations personnel located in Germany, but instead sent to Seattle or Hyderabad. Geographic attribution of data and the attribution of actual control are two different things.

Regarding the legal layer, Govrin believes that differences between national laws are too vast to be assessed using a binary pass/fail metric, so he evaluates it separately. This leads to the CLOUD Act issue: the act allows United States courts to issue data subpoenas to companies under their jurisdiction, regardless of where the data is physically stored. Govrin states that if a European entity uses an American cloud provider's AI services and a United States court issues a production order requiring the provider to hand over data, there is no room for defense; the data effectively becomes data belonging to the United States government.

This assessment utilized 17 yes-or-no tests, with the legal layer handled separately, drawing on over 300 sources and hundreds of buyer interviews. Google’s air-gapped solutions scored relatively well, but the CLOUD Act issue applied equally to all providers.

It should be noted that this framework comes from theCUBE Research, led by Govrin himself, whose commercial interests align with sovereign AI consulting. Furthermore, the judgment that 'no solution exists' for all providers may partially reflect design choices in the evaluation framework—cloud providers might consider CLOUD Act compliance a legal matter rather than a technical control issue, and thus exclude it from their sovereignty commitments. However, from the buyer's perspective, if no provider can answer this layer, it indeed constitutes a systematic blind spot in the evaluation.

출처:SiliconANGLE

리서치

Australian design software company Canva adds switch to turn off generative AI features

간단한 요점
2026-11-19 20:30 GMT+8

Australian graphic design software company Canva has added a new setting that allows users to turn off its generative AI features. This toggle disables tools such as Magic Expand for photo extension and image generation via the Canva chatbot. However, it does not remove all AI capabilities; features like the background remover and Magic Grab, which manipulate existing elements within a design, remain available because they do not generate new content.

The update also introduces an option to see less AI-generated content in Canva’s library. When enabled, this setting prioritizes human-created templates and graphics in search results. Canva states that only a small portion of its library consists of AI-generated elements, which will be labeled in their metadata. These controls are available to both free and paid users.

출처:The Verge

리서치

Survey: 90% of US Employers Reject AI-Looking Resumes, Yet Half Use AI to Interview

중대
2026-11-19 20:00 GMT+8

New data from hiring systems firm Greenhouse highlights a contradiction in the US job market: 90% of employers report having rejected real candidates because their applications appeared fake or AI-generated, yet half of the surveyed companies are now using AI-powered voice technology for early-stage interviews to reach more applicants.

As job seekers increasingly rely on AI to polish their materials, employers find it harder to distinguish genuine talent from automated submissions. The report notes that 86% of employers have adjusted their hiring processes in response, with 57% adding an extra human interview stage. Meanwhile, 81% of employers have observed applicants attempting to bypass AI screeners through tactics like hiding text in resumes or claiming skills they do not possess.

Despite this suspicion toward AI-styled applications, companies are turning to AI to manage volume. Greenhouse states that 40% of employers using AI interviews have identified strong candidates they would have otherwise missed. In a tough labor market, only 7% of US applicants say they would turn down an AI interview, suggesting that access to any interview may outweigh preferences for human interaction.

출처:ZDNET

리서치

Blue Tech Startups Raise Nearly $10 Billion in First Three Quarters of 2026, Surpassing Full-Year 2025 Total

중대
2026-11-19 20:00 GMT+8

According to the "Rethinking the Blue Economy" report by Mind the Bridge, produced with support from Crunchbase, blue tech startups have raised almost $10 billion through the third quarter of 2026. This figure already exceeds the $6.2 billion raised in all of 2025, marking a significant upward trajectory for the sector despite prevailing trends in the global innovation landscape.

The primary drivers of this funding surge are exceptional deals involving defense and infrastructure companies with maritime applications. Anduril Industries raised $5 billion in 2026, while Helsing secured $1.8 billion, accounting for a substantial portion of the total capital influx. This concentration suggests that investors are prioritizing technologies that address geopolitical instability, energy security, and climate resilience.

The report categorizes blue tech into distinct clusters based on impact and market commitment. AI, machine learning, robotics, and industrial automation are identified as "transformational technologies," boasting over 200 scaleups and $20 billion in total investment. In contrast, high-potential disruptive technologies like clean propulsion, sustainable fuels, and supply chain tech currently have lower density but are seen as promising future growth areas.

Ports are emerging as critical hubs for testing and scaling these innovations. Acting as mini "blue economies," ports like Barcelona, Rotterdam, and Los Angeles are implementing open innovation tools to integrate energy, logistics, and data systems. Initiatives such as BlueTechPort at the Port of Barcelona exemplify how port authorities are collaborating with startups and researchers to drive the radical evolution of one of the world's largest economic sectors.

출처:Crunchbase News

리서치

Tesla Cybercab's $299 Monthly Fee Makes It Costlier Than Model 3 After Five Years; Musk Calls Price Too High

중대
2026-11-19 18:47 GMT+8

Tesla's Cybercab, priced at $29,990, is the company's first vehicle under $30,000, but its mandatory $299-a-month Autonomy Subscription significantly increases long-term ownership costs. According to Business Insider calculations, by the 45th month, the Cybercab becomes more expensive than a base Model 3 equipped with a $99-a-month FSD (Supervised) subscription. By the fifth year, the Cybercab's total cost reaches $47,631—$3,160 more than the Model 3 and $160 more than a larger Model Y with the same monthly FSD subscription.

CEO Elon Musk acknowledged the pricing issue in a response to tech reviewer Marques Brownlee (MKBHD) on X. He stated that the $299 monthly fee was "too high" and should align closer to the cost of Tesla FSD and connectivity for other models, which is approximately $115 per month. This comment suggests Tesla may adjust its subscription strategy to mitigate consumer resistance to high recurring costs.

The subscription covers not only autonomous driving software but also roadside assistance and certain maintenance services, such as tire and windshield replacements. However, for average consumers, the Cybercab remains geofenced to small areas in Texas and faces an ongoing safety audit by the National Highway Traffic Safety Administration (NHTSA). While Tesla plans to begin delivering personal-use units to invited customers in December, the tension between high fixed monthly expenses and limited operational scope may slow widespread adoption.

출처:Business Insider

리서치

Appian CEO Calkins: AI Has Not Touched Serious Business, Announces Launch of Enterprise Control Plane

중대
2026-11-19 19:42 GMT+8

Calkins' core judgment is that the AI revolution is "strange" because, unlike any previous technology revolution, it did not begin with users who had the most pressing needs. The first people to buy mobile phones were precisely those who needed mobile devices the most, and the first people to use electric lights illuminated the streets that needed them most. However, AI is exactly the opposite—it started with low-value scenarios such as personal productivity and minor cost savings, while large organizations that truly possess the most valuable high-frequency processes and the lowest tolerance for error—banks, pharmaceutical companies, and the public sector—have barely adopted AI.

Calkins believes the reason is straightforward: serious organizations will never trust unfiltered probabilistic AI. These organizations have spent decades standardizing, integrating, and governing complex operations to achieve the repeatability, reliability, and control required for large-scale work. They are not going to undermine all that effort simply because someone promises that AI has great potential.

His prescription is this: AI is not a standalone technology; its uncertainty requires deterministic software infrastructure to constrain it. Calkins proposes four capabilities that must be wrapped around AI models to allow them to participate in serious enterprise work: deterministic process orchestration (to maintain boundaries), distributed data fabric (to keep them grounded), governance infrastructure at the execution level (to monitor and constrain behavior), and the freedom to choose how to execute based on the type of task (frontier models, small models, RPA bots, or humans).

Based on this, Appian announced its official entry into the control plane market and launched its product in beta. Calkins said, "The control plane is an emerging term; some companies are using it or similar terminology. We are announcing today that we are entering this market; we have a product, launching it in beta today."

The core architectural principle of the control plane is that agents should not call other agents directly but should go through a deterministic layer. Calkins stated, "We believe agents should call a deterministic layer, which then calls an agent, which then calls a deterministic layer." This means enterprises can limit AI's discretion to individual tasks without surrendering control over the entire operation.

Calkins further pointed out that the true value of the control plane lies not just in tracking and constraining, but in generating a dataset: recording when each agent was called, at which stage of the process, and whether the entire sequence led to success. With this data, organizations can optimize how agents are deployed—comparing their costs, reliability, and contribution to overall process outcomes—to determine whether a specific task should really be handed over to AI.

While the reporting author Ian Thomas affirmed this argument, he raised a notable counterpoint: Calkins' centralized coordination architecture bears structural similarity to the hub-and-spoke model of early Enterprise Application Integration (EAI), merely shifting the application from application systems to operating models. As agent capabilities continue to enhance, the balance between centralized control and distributed agency may need to be readjusted. Organizations may eventually find themselves needing to move beyond the incremental starting point described by Calkins.

출처:diginomica

리서치

US Optical Communications Firm Coherent Corp. Board Authorizes $2 Billion Share Repurchase Program

중대
검증됨 2026-11-19 19:58 GMT+8

Coherent Corp. (NYSE: COHR), a US-based optical communications and laser technology company, announced on November 18, 2026, that its Board of Directors authorized a share repurchase program allowing the company to buy back up to $2.0 billion of its common stock. The repurchases can be made through open market transactions, negotiated purchases, block transactions, or derivative transactions, excluding fees and commissions.

According to an 8-K filing with the U.S. Securities and Exchange Commission (SEC), the program has no expiration date, does not obligate the company to repurchase any specific number of shares, and may be suspended or discontinued at any time. This authorization reflects management's approach to capital allocation and shareholder returns, though actual execution will depend on future market conditions and corporate decisions.

출처:sec.gov

리서치

BlackLine CEO Owen Ryan: The Bottleneck for Financial AI Adoption Is Governance, Not Model Capability

중대
2026-11-19

The attitude toward AI agents in the finance and accounting sectors is contradictory, according to BlackLine CEO Owen Ryan: agents are arriving quickly, and customers want them, but adoption in critical workflows remains very slow. This is not an issue of capability, but of trust.

Ryan describes a specific adoption status: most of the agents co-developed with clients have not yet touched the books. What companies are doing is running parallel systems for multiple quarters to see if the results are consistent. This means financial AI is currently in a cautious experimental phase—the models are good enough, but organizations need to accumulate sufficient confidence in parallel environments before allowing agents to directly write to production financial data.

Why is finance so different? Ryan’s explanation is the irreversibility of consequences: in marketing, a wrong draft can be rewritten; in engineering, bugs are caught in code reviews. But in finance, a wrong number reaches auditors, the board, and even the market. This consequence structure dictates that financial organizations move much more cautiously than other functions from AI assistance, to agents completing tasks, to agents running critical workflows end-to-end.

Gartner data supports this judgment: only 13% of organizations believe they currently possess the correct AI governance. Ryan says this aligns perfectly with feedback from their clients and partners—the problem is not that AI doesn't work, but that organizational governance maturity is insufficient to rely on it in accuracy-critical scenarios like accounting.

So what breaks this deadlock? Ryan believes it is a control layer: encompassing policies, observability, agent registry, audit trails, and risk scoring. This is precisely the function targeted by BlackLine’s Finance Control Console, which serves as the governance layer of its Agentic Financial Operations Platform.

In terms of the competitive landscape, OpenAI launched a ChatGPT plugin for the accounting sector this week, recalling earlier interpretations that Anthropic would replace lawyers. Ryan and BlackLine CTO Jeremy Ung respond with cooperation rather than confrontation: OpenAI models will connect into BlackLine’s workflows; BlackLine has also listed on Anthropic’s Claude Marketplace, where products like Verity accruals are already powered by Claude’s reasoning capabilities.

Ung adds that through Model Context Protocol (MCP) integration, agents can dynamically discover BlackLine’s capabilities; customers do not need to wait for BlackLine’s educational materials to connect Claude or OpenAI and execute workflows. diginomica journalist Stuart Lauchlan summarizes this relationship as 'co-opetition,' common in the tech industry: customers use models they already know and trust, but bring financial data into BlackLine’s controlled environment. Ryan’s own statement is: 'We now have greater collaboration with the largest AI labs. Customers will be able to use models they already know and trust, then bring their financial data into our control environment.'

출처:diginomica

리서치

Community Inference Tools Bring 176-Billion-Parameter Model to Desktop, Achieving ~200 Tokens per Second in Testing

중대
2026-11-19 17:20 GMT+8

A 176-billion-parameter model running at 200 tokens per second on a desktop PC—this sounds like marketing hype, but Christian Winkler, Professor of NLP at the Nuremberg University of Applied Sciences (TH Nürnberg), verified this figure using his own machine: 128GB RAM plus an Nvidia RTX 5090, running Alibaba's Qwen3.8-Flash-Next via the community inference tool Basalt, a derivative of Strata, actually achieved approximately 200 tokens/s. His assessment: the perceived speed feels faster than some commercial services.

This is possible due to two key layers of model architecture design. Qwen3.8-Flash-Next has a total of 176B parameters, but 51B are so-called N-Gram parameters that can be offloaded to CPU memory or even SSDs, without needing to reside permanently in GPU VRAM. The remaining 122B use a Mixture-of-Experts (MoE) structure, activating only 6B during each inference run. This means only 6 billion parameters are truly involved in computation, similar to a small model, yet the model possesses the knowledge capacity of 122B parameters seen during training. This design decouples 'how large the model is' from 'how much is computed during inference.'

The community tools address the other half of the problem: how to efficiently schedule these parameters scattered across different storage tiers. Strata allows running large models on relatively small graphics cards, provided the system has sufficient memory. It can even run larger models like DeepSeek-V4.1-Flash and GLM-5.3-Flash. Basalt is a fork of Strata, specifically optimized for further improving inference speeds for Qwen3.8-Flash-Next. Winkler noted that these tools largely come from so-called 'vibe coding'—developers using AI assistance for rapid prototyping—and the project maturity and long-term maintainability remain uncertain.

The cost is quality. Winkler explicitly stated that local inference quality does not reach the level of frontier large language models. This represents a trade-off in capability tiers: you gain speed, privacy, and fixed costs, but lose the most complex reasoning abilities and breadth of knowledge. For everyday tasks like writing assistance, code completion, and information summarization, this tier may already be sufficient; for scenarios requiring deep reasoning or precise knowledge retrieval, commercial APIs still lead.

Notable comparisons include other new models mentioned in the same article—Moonshot's Kimi K3.1-Pro has 2.8 trillion parameters, and the so-called Anthropic leaked checkpoint numbat has 14 trillion parameters—these still require datacenter-level hardware. What truly brings large models out of server rooms is not making models smaller, but designing architectures that are 'large in storage, small in computation,' and then using software to execute the separation of storage and computation effectively. Qwen3.8-Flash-Next is currently the model on this path with publicly available weights that can be directly invoked by community tools.

출처:Developer

리서치
다음 →
로그인 / 회원가입
독서리서치레이더투자 프레임워크

Beyond these technical terms, the author cites the experience of Brex CFO Erica Dorfman in the Capital One acquisition case to illustrate conditions worth fighting for beyond price. Dorfman stated that both Brex and Capital One agreed that maintaining operational independence post-acquisition—including its own brand, product roadmap, and leadership team—is 'as important to us as any dollar figure.' Brex CEO Pedro Franceschi continued to serve as CEO after the deal closed, and team retention and continuity with customers were listed as core negotiation conditions.

The author also mentions an often-overlooked practical suggestion: before signing the exclusivity clause, confirm that the buyer has conducted sufficient due diligence to feel confident moving forward. Otherwise, the greatest risk to the seller during the exclusivity period is not that terms are lowered, but that the deal falls through entirely—during which time the seller cannot find alternative buyers and may discover the buyer ultimately abandons the deal.

In summary, the LOI stage represents the peak of the seller's negotiation leverage in the entire M&A process. Once the exclusivity clause is signed, the seller's bargaining power enters an irreversible downward channel. It is far more effective to clarify definitions, boundaries, and key conditions within this window than to attempt renegotiation at the definitive agreement stage. As the author says: 'Don’t neglect the LOI because it’s non-binding on price. It is one of the most important stages of the process.'

It should be noted that this is BlackLine’s own perspective. Positioning governance as the sole unlocking condition conveniently places its product at a key position in the value chain. Gartner’s 13% figure is a secondary citation, with methodology and sample size not independently verified. However, regardless of motive, if the specific adoption status of running parallel systems without touching the books is true, it indeed indicates that the commercialization timeline for financial AI depends on the maturity of compliance infrastructure, rather than the next leap in model capability.