📰 The InformationThe InformationCodex · 实盘不做把握 3/10不做原因:证据不足
北京时间·🕒 发布 08-26 22:00:31时区推断·⟳ 抓取 08-26 22:16:53·⚡ 决策 08-27 00:07:07 +1h50m·61.7 天前
The Information称企业AI Agent评估更多使用程序化校验，附带多项AI与半导体交易动态
主文是行业方法论变化，不构成明确上市公司增量；附带交易动态缺少在本轮可审计的行情、重复检查和先例研究支撑，不能下ENTER。
📄 原文The Information↗ 原文
How Companies Are Using Old-School Software to Grade AI Agents
When companies started adopting agents, evaluators would often use an AI judge to monitor, for example, whether the age…
不做 Navitas Semiconductor（NVTS.US）
收购Claros虽有方向性但本文只是简讯转述，且缺少本轮强制行情、市值、去重和先例核验，不能给出可执行交易。
不做 Salesforce（CRM.US）
只是旧背景引用，没有本次新增公司级催化。
不做 Accenture（ACN.US）
收购金额未披露且相对公司体量缺少短线重大性。
展开详情(成交明细 / 研究档案 / 原文全文 / 决策记录链)
研究档案
结论摘要
这条信号的核心不是某家上市公司发布新订单、业绩或监管进展，而是The Information对AI Agent评估方式的行业观察；文中可交易公司多为背景或简讯提及，未形成可量化、可执行、已用行情验证的正期望交易。
🔎 新鲜度:最早出现 2026-08-26 · 扩散 medium · price-in 风险 medium ↗
消息传播链
- The Information发布AI Agenda文章，主线是企业AI Agent评估中程序化verifier的使用增加，并附带多项融资并购简讯。
主要风险
- The Information影响力较高，但本条对上市公司更多是主题叙事和交易简讯，缺少能直接重估收入、利润或市值的新增事实。
- 缺少stock_quote和event_study结果，无法给出合规入场价、止盈止损和同类事件胜率。
反向证据
- 做多前查反向 · 来源 filing · 日期 2026-08-26 · 影响 weakens
正文未披露可直接量化到上市公司收入或利润的新订单、指引上修或监管准入。
检索:当前正文是否包含上市公司新增订单、业绩指引或可量化商业化里程碑 原文 ↗
先例如何影响了这次决策
对本次决策的影响:先例不足使本次只能回避，不能给出交易价位和仓位。
方向:未形成可用于定方向的同市场同机制先例。
涨幅量级:缺少event_study测算，不能用先例校准涨跌幅。
形态:无法判断是否存在反复见光死或高开低走模式。
原文全文
展开原文 · 10,875 字
When companies started adopting agents, evaluators would often use an AI judge to monitor, for example, whether the agent correctly cited its sources or kept its answer below a certain word count. But that requires paying to run a second AI model. Software code is cheaper to use.
The move toward traditional software for certain evaluation work might seem counterintuitive at a time when AI is supplanting so much of traditional software, but it also reflects enterprises’ growing familiarity with the limitations of AI models, such as their high costs and unreliability. For example, my colleagues reported late last year that Salesforce was [deprioritizing large language models in Agentforce](https://www.theinformation.com/articles/salesforce-executives-say-trust-generative-ai-declined), in favor of traditional software automation that followed if-then rules.
AI judges remain an effective approach when the criteria for determining success are “squishy,” Kenstler said. In such cases, to provide a grade, a judge scores an agent’s outputs against a grading rubric. For example, if an agent is told “I want you to write a report on the Revolutionary War,” the rubric might include criteria like “the report mentions that George Washington was the general of the Continental Army,” Kentsler said.
Programmatic verifiers can evaluate performance on more black-and-white criteria, such as whether an agent has solved a math problem, correctly formatted a spreadsheet or included a specific phrase in its output. An enterprise that is evaluating how well an agent can work with a CRM database could use a programmatic verifier to check whether the database has been updated, without relying on another AI to serve as a judge.
Increasing use of these programmatic verifiers “is bringing a greater level of rigor and outcome-orientation to the way that enterprises are doing evals,” said Kentsler. Because they’re cheaper, programmatic verifiers allow enterprises to more thoroughly evaluate their agents’ actions, potentially catching more errors and better understanding the pros and cons of different agents.
While Kentsler is primarily focused on evaluations, which measure the performance of an AI model after it has been trained, both of these techniques can also be used to improve AI agents during reinforcement learning.
So are programmatic verifiers going to put AIs out of their jobs as judges? Not quite. “The best evaluations combine both,” Kentsler said. For example, if an AI agent was tasked with producing a regulatory filing, it might have to follow a standardized format, which can be checked programmatically, as well as include the right details and adhere to a company’s guidelines about how information is presented, which are better suited for AI.
In fact, for some tasks the trend has gone the opposite direction: from relying on programmatic verifiers to increasingly drawing on AI judges.
“Coding agents is a great example of a use case that starts with programmatic and is now trending rubric,” Kenstler said. For example, older coding benchmarks such as the popular SWE-Bench use automatic tests to check whether an AI wrote successful code. But measuring whether the AI followed good style requires bringing in an AI judge.
To be sure, evaluating agents will continue to be a challenge for the foreseeable future, but using all the available tools will help enterprises select the best AI models for a given task and maybe even help catch naughty agents trying to hack into other companies.
*Here’s what else is going on…*
### Deals and Debuts
*See The Information’s [Generative AI Database](https://www.theinformation.com/projects/generative-ai) for an exclusive list of private companies and their investors.*
**Navitas Semiconductor**, which makes power chips that convert and deliver electricity efficiently inside AI data centers, agreed to acquire **Claros**, a company that optimizes energy delivery for data centers for up to $232.8 million.
**Gatik**, which develops self-driving box trucks that run short, repeated freight routes between warehouses and stores, raised **$200 million** in a Series D funding round led by **Qatar Investment Authority** and **Koch Disruptive Technologies.**
**Emerald AI**, a company whose software lets AI data centers dial their electricity use up and down on demand, so they can connect to the power grid faster and ease off when the grid is strained, raised $150 million in a Series A funding round led by **Energize Capital**and **DCVC**.
**Alice**, a company that tests AI models for ways they can be tricked into harmful behavior and puts guardrails around the AI that companies deploy, raised $140 million in a funding round led by **Apax Digital.**
**Stability AI**, which makes AI tools that generate images, audio and video for professional creative teams, raised **$76 million** in a Series B funding round with participation from Electronic Arts, **Sony Music Group**, **Universal Music Group**, **Warner Music Group**.
**Happy Health**, a company whose FDA-cleared smart ring and software diagnose and manage sleep disorders at home, raised **$75 million** in a Series A funding round led by **ARCH Venture Partners** and **OpenLoop**.
**Airbound**, a company that designs small autonomous cargo drones aiming to make drone delivery cheaper than trucking, raised $37 million in a Series A funding round led by **Greenoaks**.
**Liner**, a company whose AI research tool answers questions with citations attached so users can check the sources, raised $36.1 million in a Series C funding round led by **LB Investment.**
**Celera Semiconductor**, a company that uses AI to design analog chips, cutting the cost and time it normally takes, raised $30 million in a Series B funding round funded entirely by **Maverick Silicon**, already its largest investor.
**Keenable**, a company that is developing a web search index designed for AI systems rather than people, raised $26 million in a seed funding round led by **Accel**.
**Hike Medical**, a company whose software carries a medical device order all the way from a doctor's referral to the finished product, raised $22.5 million in seed and Series A funding led by **Max Altman**at **Saga Ventures.**
**Newcode**, a company whose software lets a law firm develop and run its own AI system on top of its existing case files and databases, raised $13.5 million in a Series A funding round led by **OnDean Forward**.
**Primero**, a company that helps Latin America's largest corporations replace software dating from the 1990s and automate the work those systems handle, raised $12 million in a seed funding round led by **Kaszek** and **General Catalyst.**
**amber**, a company whose software gathers business documents and records into one searchable base that its AI can interact with, raised €7 million in a Series A funding round led by **Ventech** and **NRW.Venture.**
**Capital Numbers**, a company whose engineers build software, cloud systems and AI tools for clients, completed the acquisition of all of **Epitome Cloud**, a firm that customizes Salesforce and related contract and pricing software for large companies, $4–5 million.
**WellStack**, which builds software that pulls a hospital system's scattered patient and operational data into one place so it can be analyzed, acquired **DeLorean AI**, which builds AI that predicts which patients are likely to get sicker. Terms were not disclosed.
**Accenture** agreed to acquire **McCoy**, a consulting firm whose staff install and run SAP business software, for mid-sized companies. Terms were not disclosed.
**ZenaTech**, a company that sells drone flights as a service and also sells business software, completed the acquisition of **ESM Software**, a company whose software helps organizations set goals and prove they meet cybersecurity rules. Terms were not disclosed.
**Algorhythm Holdings** said it is pursuing a sale, spin-off or other options for **SemiCab**, its unit that runs cloud software using AI to match freight loads to trucks so fewer trucks drive around empty. No buyer was named and no price was disclosed.
**SpaceX** [announced on Tuesday](https://www.theinformation.com/briefings/spacex-announces-100-billion-starship-launch-site-louisiana) plans to build a $100 billion launch site for its Starship rocket on the coast of Louisiana.
**LayerZero**, a blockchain infrastructure startup that counts **Citadel** **Securities**, **DTCC**, and **Intercontinental** **Exchange** among its partners, [said it’s launching a new blockchain-based exchange](https://www.theinformation.com/briefings/citadel-securities-dtcc-partner-layerzero-launches-new-exchange) targeting financial institutions this fall.
**Skild AI**, which develops foundation models for robotics, announced **S1**, a new AI model that it said can learn to perform new physical tasks after being “prompted” with a single video demonstration.
**Lightwheel AI**, which provides simulation software for robotics, and **Hugging Face** released a data set containing 100,000 hours of first-person videos of people using their hands to complete tasks at home and at work.
## What We’re Reading
#### [A.I. Is Becoming So Powerful, It’s Stumping Those Trying to Contain It](https://www.nytimes.com/2026/08/25/technology/irregular-ai-test-hacks.html)
#### [They Dedicated Their Lives to Teaching. Then the Deepfakes Started](https://www.wired.com/story/teachers-deepfake-ai-students-content/)
#### [Why Social Media Giants Are on Trial For Hooking Kids](https://www.bloomberg.com/news/articles/2026-08-23/meta-trial-how-social-media-lawsuits-threaten-big-tech?srnd=phx-technology)
### Upcoming Events
#### [Wednesday, September 23 — AI Agenda Live SF 2026](https://www.theinformation.com/events/ai-agenda-live-2026)
Join us to explore the real-world shifts, new architectures, and enterprise strategies defining the next era of AI. Secure your spot today to hear directly from top minds at Google DeepMind, OpenAI, Atlassian, and more.
[More details](https://www.theinformation.com/events/ai-agenda-live-2026)
#### [Tuesday, October 27 — The Information’s 2026 WTF Summit](https://www.theinformation.com/events/wtf-summit-2026)
Tickets on sale now: The Information returns to Napa Valley October 27–28 to convene senior women across tech, media, and finance. Join us for two days of intimate, candid conversations with the leaders navigating today’s global shifts.
[More details](https://www.theinformation.com/events/wtf-summit-2026)
*Thank you for reading the AI Agenda Newsletter! I’d love your feedback, ideas and tips: [stephanie@theinformation.com](mailto:stephanie@theinformation.com).*
*If you think someone else might enjoy this newsletter, please pass it forward or they can sign up [here](https://www.theinformation.com/newsletters/ai-agenda).*
The move toward traditional software for certain evaluation work might seem counterintuitive at a time when AI is supplanting so much of traditional software, but it also reflects enterprises’ growing familiarity with the limitations of AI models, such as their high costs and unreliability. For example, my colleagues reported late last year that Salesforce was [deprioritizing large language models in Agentforce](https://www.theinformation.com/articles/salesforce-executives-say-trust-generative-ai-declined), in favor of traditional software automation that followed if-then rules.
AI judges remain an effective approach when the criteria for determining success are “squishy,” Kenstler said. In such cases, to provide a grade, a judge scores an agent’s outputs against a grading rubric. For example, if an agent is told “I want you to write a report on the Revolutionary War,” the rubric might include criteria like “the report mentions that George Washington was the general of the Continental Army,” Kentsler said.
Programmatic verifiers can evaluate performance on more black-and-white criteria, such as whether an agent has solved a math problem, correctly formatted a spreadsheet or included a specific phrase in its output. An enterprise that is evaluating how well an agent can work with a CRM database could use a programmatic verifier to check whether the database has been updated, without relying on another AI to serve as a judge.
Increasing use of these programmatic verifiers “is bringing a greater level of rigor and outcome-orientation to the way that enterprises are doing evals,” said Kentsler. Because they’re cheaper, programmatic verifiers allow enterprises to more thoroughly evaluate their agents’ actions, potentially catching more errors and better understanding the pros and cons of different agents.
While Kentsler is primarily focused on evaluations, which measure the performance of an AI model after it has been trained, both of these techniques can also be used to improve AI agents during reinforcement learning.
So are programmatic verifiers going to put AIs out of their jobs as judges? Not quite. “The best evaluations combine both,” Kentsler said. For example, if an AI agent was tasked with producing a regulatory filing, it might have to follow a standardized format, which can be checked programmatically, as well as include the right details and adhere to a company’s guidelines about how information is presented, which are better suited for AI.
In fact, for some tasks the trend has gone the opposite direction: from relying on programmatic verifiers to increasingly drawing on AI judges.
“Coding agents is a great example of a use case that starts with programmatic and is now trending rubric,” Kenstler said. For example, older coding benchmarks such as the popular SWE-Bench use automatic tests to check whether an AI wrote successful code. But measuring whether the AI followed good style requires bringing in an AI judge.
To be sure, evaluating agents will continue to be a challenge for the foreseeable future, but using all the available tools will help enterprises select the best AI models for a given task and maybe even help catch naughty agents trying to hack into other companies.
*Here’s what else is going on…*
### Deals and Debuts
*See The Information’s [Generative AI Database](https://www.theinformation.com/projects/generative-ai) for an exclusive list of private companies and their investors.*
**Navitas Semiconductor**, which makes power chips that convert and deliver electricity efficiently inside AI data centers, agreed to acquire **Claros**, a company that optimizes energy delivery for data centers for up to $232.8 million.
**Gatik**, which develops self-driving box trucks that run short, repeated freight routes between warehouses and stores, raised **$200 million** in a Series D funding round led by **Qatar Investment Authority** and **Koch Disruptive Technologies.**
**Emerald AI**, a company whose software lets AI data centers dial their electricity use up and down on demand, so they can connect to the power grid faster and ease off when the grid is strained, raised $150 million in a Series A funding round led by **Energize Capital**and **DCVC**.
**Alice**, a company that tests AI models for ways they can be tricked into harmful behavior and puts guardrails around the AI that companies deploy, raised $140 million in a funding round led by **Apax Digital.**
**Stability AI**, which makes AI tools that generate images, audio and video for professional creative teams, raised **$76 million** in a Series B funding round with participation from Electronic Arts, **Sony Music Group**, **Universal Music Group**, **Warner Music Group**.
**Happy Health**, a company whose FDA-cleared smart ring and software diagnose and manage sleep disorders at home, raised **$75 million** in a Series A funding round led by **ARCH Venture Partners** and **OpenLoop**.
**Airbound**, a company that designs small autonomous cargo drones aiming to make drone delivery cheaper than trucking, raised $37 million in a Series A funding round led by **Greenoaks**.
**Liner**, a company whose AI research tool answers questions with citations attached so users can check the sources, raised $36.1 million in a Series C funding round led by **LB Investment.**
**Celera Semiconductor**, a company that uses AI to design analog chips, cutting the cost and time it normally takes, raised $30 million in a Series B funding round funded entirely by **Maverick Silicon**, already its largest investor.
**Keenable**, a company that is developing a web search index designed for AI systems rather than people, raised $26 million in a seed funding round led by **Accel**.
**Hike Medical**, a company whose software carries a medical device order all the way from a doctor's referral to the finished product, raised $22.5 million in seed and Series A funding led by **Max Altman**at **Saga Ventures.**
**Newcode**, a company whose software lets a law firm develop and run its own AI system on top of its existing case files and databases, raised $13.5 million in a Series A funding round led by **OnDean Forward**.
**Primero**, a company that helps Latin America's largest corporations replace software dating from the 1990s and automate the work those systems handle, raised $12 million in a seed funding round led by **Kaszek** and **General Catalyst.**
**amber**, a company whose software gathers business documents and records into one searchable base that its AI can interact with, raised €7 million in a Series A funding round led by **Ventech** and **NRW.Venture.**
**Capital Numbers**, a company whose engineers build software, cloud systems and AI tools for clients, completed the acquisition of all of **Epitome Cloud**, a firm that customizes Salesforce and related contract and pricing software for large companies, $4–5 million.
**WellStack**, which builds software that pulls a hospital system's scattered patient and operational data into one place so it can be analyzed, acquired **DeLorean AI**, which builds AI that predicts which patients are likely to get sicker. Terms were not disclosed.
**Accenture** agreed to acquire **McCoy**, a consulting firm whose staff install and run SAP business software, for mid-sized companies. Terms were not disclosed.
**ZenaTech**, a company that sells drone flights as a service and also sells business software, completed the acquisition of **ESM Software**, a company whose software helps organizations set goals and prove they meet cybersecurity rules. Terms were not disclosed.
**Algorhythm Holdings** said it is pursuing a sale, spin-off or other options for **SemiCab**, its unit that runs cloud software using AI to match freight loads to trucks so fewer trucks drive around empty. No buyer was named and no price was disclosed.
**SpaceX** [announced on Tuesday](https://www.theinformation.com/briefings/spacex-announces-100-billion-starship-launch-site-louisiana) plans to build a $100 billion launch site for its Starship rocket on the coast of Louisiana.
**LayerZero**, a blockchain infrastructure startup that counts **Citadel** **Securities**, **DTCC**, and **Intercontinental** **Exchange** among its partners, [said it’s launching a new blockchain-based exchange](https://www.theinformation.com/briefings/citadel-securities-dtcc-partner-layerzero-launches-new-exchange) targeting financial institutions this fall.
**Skild AI**, which develops foundation models for robotics, announced **S1**, a new AI model that it said can learn to perform new physical tasks after being “prompted” with a single video demonstration.
**Lightwheel AI**, which provides simulation software for robotics, and **Hugging Face** released a data set containing 100,000 hours of first-person videos of people using their hands to complete tasks at home and at work.
## What We’re Reading
#### [A.I. Is Becoming So Powerful, It’s Stumping Those Trying to Contain It](https://www.nytimes.com/2026/08/25/technology/irregular-ai-test-hacks.html)
#### [They Dedicated Their Lives to Teaching. Then the Deepfakes Started](https://www.wired.com/story/teachers-deepfake-ai-students-content/)
#### [Why Social Media Giants Are on Trial For Hooking Kids](https://www.bloomberg.com/news/articles/2026-08-23/meta-trial-how-social-media-lawsuits-threaten-big-tech?srnd=phx-technology)
### Upcoming Events
#### [Wednesday, September 23 — AI Agenda Live SF 2026](https://www.theinformation.com/events/ai-agenda-live-2026)
Join us to explore the real-world shifts, new architectures, and enterprise strategies defining the next era of AI. Secure your spot today to hear directly from top minds at Google DeepMind, OpenAI, Atlassian, and more.
[More details](https://www.theinformation.com/events/ai-agenda-live-2026)
#### [Tuesday, October 27 — The Information’s 2026 WTF Summit](https://www.theinformation.com/events/wtf-summit-2026)
Tickets on sale now: The Information returns to Napa Valley October 27–28 to convene senior women across tech, media, and finance. Join us for two days of intimate, candid conversations with the leaders navigating today’s global shifts.
[More details](https://www.theinformation.com/events/wtf-summit-2026)
*Thank you for reading the AI Agenda Newsletter! I’d love your feedback, ideas and tips: [stephanie@theinformation.com](mailto:stephanie@theinformation.com).*
*If you think someone else might enjoy this newsletter, please pass it forward or they can sign up [here](https://www.theinformation.com/newsletters/ai-agenda).*
决策记录链 · 只追加
每次领取产生一条记录,重新决策追加一条并以 supersedes 指向被替代的记录(ADR 0014)。
这条决策早于决策账本(2026-10-08 起逐条追加),没有追加记录。