AI 日报 - 2026-06-21

本文由脚本自动生成，共收录 10 条 AI 相关资讯。默认展示速览，展开后阅读完整内容。

今日速览

01 从“机审+人审”到“AI-Native”：大模型与 Agent 驱动内容风控智能化升级｜AICon上海过去一年，“Agent”这个词从实验室走进了生产环境。工程师们开始真正面对一个新的问题：不是“AI 能不能做到”，而是“我们能不能把它跑稳、跑对、跑出规模”。架构怎么设计？记忆怎么管理？多智能体之间如何协调？研发团队的工作方式又该如何重构？这些，正是 AICon 2026 上海站试图回答的问题。 6 月 26 日 27 日，本次大会将以“构建可信赖、可规模... InfoQ / 中文 02 Spring Boot 4.1 新增 gRPC 自动配置、SSRF 防护功能并支持 Kotlin 2.3 Broadcom 于 2026 年 6 月 10 日发布了 Spring Boot 4.1，该版本提供了 gRPC 自动配置、HTTP 客户端 SSRF 风险缓解能力，并将 Kotlin 升级至 2.3。它还带来了延迟数据源连接、 @Async 方法的异步上下文传播，以及改进的 OpenTelemetry 支持。Broadcom 两次推迟了发布时间，先从 5... InfoQ / 中文 03 Plotting AI model release cadence: two labs are accelerating, three aren't Analysis · June 2026 Plotting AI model release cadence: two labs are accelerating, three aren't Plotting frontier model release cadence, with methodology SwiftAlerts · June 20, 202... Hacker News / 英文 04 White House delays US voting-machine vulnerability report 暂无摘要，建议展开查看原文信息。 Hacker News / 英文 05 Barret Zoph is out at OpenAI again after just five months Five months after returning to OpenAI, Barret Zoph - the company's head of enterprise AI sales - has departed, The Verge has learned. Zoph returned to OpenAI in mid-January after a... The Verge / 英文 06 Amazon employees say they’re facing termination for backing data center limits When three Amazon software engineers testified earlier this month at Seattle City Council hearings about data centers, they started their testimony by citing a city law barring emp... The Verge / 英文 07 The inevitable weakness of metrics There are plenty of useful things a metric can reveal. There are even more it can obscure or corrupt. It took me well over a decade of tracking my own life in ever greater detail t... MIT Tech Review / 英文 08 Brain-computer interface trials are taking off This week, I covered the story of Casey Harrell—a man with ALS who is “the first power user” of a brain implant, according to the researchers who worked with him. Harrell is paraly... MIT Tech Review / 英文 09 Diffusion Language Models: An Experimental Analysis arXiv:2606.19475v1 Announce Type: new Abstract: Large Language Models (LLMs) have revolutionized language modeling through autoregressive generation, enabling strong performance ac... ArXiv AI / 英文 10 Hidden Anchors in Multi-Agent LLM Deliberation arXiv:2606.19494v1 Announce Type: new Abstract: Multi-agent LLM deliberation, where agents exchange and revise answers over several rounds, is increasingly used to improve reasonin... ArXiv AI / 英文

详细内容

01 从“机审+人审”到“AI-Native”：大模型与 Agent 驱动内容风控智能化升级｜AICon上海 InfoQ / 中文

过去一年，“Agent”这个词从实验室走进了生产环境。工程师们开始真正面对一个新的问题：不是“AI 能不能做到”，而是“我们能不能把它跑稳、跑对、跑出规模”。架构怎么设计？记忆怎么管理？多智能体之间如何协调？研发团队的工作方式又该如何重构？这些，正是 AICon 2026 上海站试图回答的问题。 6 月 26 日 27 日，本次大会将以“构建可信赖、可规模...

来源：InfoQ | 语言：中文 | 原文链接

过去一年，“Agent”这个词从实验室走进了生产环境。工程师们开始真正面对一个新的问题：不是“AI 能不能做到”，而是“我们能不能把它跑稳、跑对、跑出规模”。架构怎么设计？记忆怎么管理？多智能体之间如何协调？研发团队的工作方式又该如何重构？

这些，正是 AICon 2026 上海站试图回答的问题。 6 月 26 日-27 日，本次大会将以“构建可信赖、可规模化、可商业化的 Agentic 操作系统”为核心命题，集结清华、复旦等知名高校教授，以及来自阿里、腾讯、蚂蚁、字节、快手、小红书、华为、Google Cloud 等数十家头部公司的技术专家登台分享。2 天、13 大专题、1 个动手实验室、近 60 场重磅议题，将深度探讨 Agent 工程化落地等相关话题。

蚂蚁集团内容安全架构师李海亮已确认出席 “金融领域大模型落地实践” 专题，发表题为**《从“机审+人审”到“AI-Native”：大模型与 Agent 驱动内容风控智能化升级》**的主题分享。随着内容形态和风险类型持续演化，传统风控体系在语义理解、风险泛化、策略响应和运营效率上面临新的挑战。基于内容 AI+ 的建设实践，他们探索以 AI 引擎为核心底座，引入大模型能力提升内容风险识别与判断能力，并进一步结合 Agent AI 运营升级，推动内容风控从人工驱动的审核运营，走向 AI 驱动的智能治理。本次分享将重点介绍三方面内容：一是从小模型到大模型的能力演进与训练探索；二是面向内容风控场景构建 AI-Native 风控引擎；三是基于 Multi-Agent 运营架构与 AI 引擎基座，探索内容风控自主发现、自主分析和持续进化的新范式。

李海亮，现担任蚂蚁集团内容安全架构师，深耕大规模风控系统一线架构研发十余年，长期负责内容安全与风险防控体系的架构规划与平台建设。先后主导三代“内容风控防控平台”的核心架构研发，在人工智能重塑安全范式的关键节点，牵头建设“内容模型平台”和“大模型安全防控平台”，成功驱动了内容安全平台全面 AI 化转型。在内容风险识别、模型平台化、智能防控以及大模型安全治理等领域拥有深厚的实战积淀与独到的架构视野。他在本次会议的详细演讲内容如下：

演讲提纲：

内容风控智能化转型背景

从传统机审、人审到 AI-Native 风控体系

从小模型到大模型

内容风险识别与语义理解能力升级

构建 AI-Native 内容风控引擎

以 AI 引擎承载策略、模型与审核链路协同

Multi-Agent 驱动运营升级

探索内容风控的自主分析与持续进化

5。总结与展望

下一代内容风控智能化体系建设方向

听众收益：

了解内容风控从传统审核模式向 AI-Native 架构升级的整体路径

理解大模型、AI 引擎与 Multi-Agent 在内容风控场景中的落地思路

除此之外，本次大会还策划了端侧 AI、物理与数字空间智能化、世界模型与多模态智能突破、Agent 架构与工程化实践、Agent 安全与可信治理、企业级研发体系重构、AI 原生数据工程、AI 时代的个人提效与组织变革等 14 个专题论坛，届时将有来自不同行业、不同领域、不同企业的 50+资深专家在现场带来前沿技术洞察和一线实践经验。

更多详情可扫码或联系票务经理 13269078023 进行咨询。

02 Spring Boot 4.1 新增 gRPC 自动配置、SSRF 防护功能并支持 Kotlin 2.3 InfoQ / 中文

Broadcom 于 2026 年 6 月 10 日发布了 Spring Boot 4.1，该版本提供了 gRPC 自动配置、HTTP 客户端 SSRF 风险缓解能力，并将 Kotlin 升级至 2.3。它还带来了延迟数据源连接、 @Async 方法的异步上下文传播，以及改进的 OpenTelemetry 支持。Broadcom 两次推迟了发布时间，先从 5...

来源：InfoQ | 语言：中文 | 原文链接

Broadcom 于 2026 年 6 月 10 日发布了 Spring Boot 4.1，该版本提供了 gRPC 自动配置、HTTP 客户端 SSRF 风险缓解能力，并将 Kotlin 升级至 2.3。它还带来了延迟数据源连接、@Async 方法的异步上下文传播，以及改进的 OpenTelemetry 支持。Broadcom 两次推迟了发布时间，先从 5 月 11 至 22 日推迟到 6 月 1 至 5 日，再推迟到 6 月 8 至 12 日。这是自 2020 年 5 月 Spring Boot 确立每年 5 月和 11 月两次发布节奏以来的首次延期。

2025 年 11 月发布的 Spring Boot 4.0 与 Spring Framework 7.0 一同推出，属于一次大版本迭代重构：以 Jakarta EE 11 为基线，集成 Jackson 3，拆分式自动配置 JAR，借助 JSpecify 实现空值安全，同时新增 API 版本控制与 Gradle 9 支持。Spring 团队与 InfoQ 专访的问答内容提供了 4.0 大版本迭代的相关背景信息。Spring Boot 4.1 是增量式更新，基于 Spring Framework 7.0.x 构建。虽然 Spring Boot 4 延续了 Spring Boot 3 的 JDK 17 基线，但 Spring Boot 4.1 中有一项新功能需要 Java 21 的支持：jOOQ 3.20。

Spring Boot 4.1 包含面向服务器端和客户端应用的 Spring gRPC 自动配置，支持独立的 Netty 和 Servlet HTTP/2 传输。该版本新增了 @GrpcAdvice 注解可用于统一异常处理，以及一个自动配置的 ObservationGrpcServerInterceptor 注解，可基于自定义服务端观测规范完成指标采集与链路追踪。在此之前，应用若要使用 gRPC 必须进行手动配置或依赖第三方 starter。

HTTP 客户端 SSRF（服务器端请求伪造）风险缓解能力是 4.1 版本的新特性。通过 InetAddressFilter 配置白名单或黑名单可阻止响应式和阻塞式客户端对指定地址段发起的出站请求，减少应用被当作内网攻击代理的安全隐患。

Kotlin 基线从 2.2 版本升级至 2.3 版本，该版本支持 Java 25，并包含一个实验性的未使用返回值检查器。

设置 spring.datasource.connection-fetch=lazy 后，池化的 DataSource 会被封装到 LazyConnectionDataSourceProxy 中。这将物理数据库连接的获取推迟到实际执行 SQL 语句时进行，加快了启动速度并降低连接池压力。

@Async 方法现已支持自动跨线程传播 Micrometer 上下文，无需额外配置，追踪 ID 和跨度信息即可跟随任务进入线程池。新的 spring.jpa.bootstrap 属性支持 Spring Data JPA 异步后台初始化，能够缩短搭载大型 JPA 模型应用的启动时间。

OpenTelemetry 相关能力迎来多项优化：新增了可在 true 和 false 之间切换的 management.opentelemetry.enabled 属性，支持 OTLP 采样示例，并允许为 OTLP 导出器配置 SSL 证书包。Spring Boot 4.1 还可读取大部分 OpenTelemetry 环境变量。

Spring Boot 4.1 会自动配置 Spring Data Redis 监听器端点，如果检测不到对应容器就注册一个默认的 RedisMessageListenerContainer。/actuator/info 端点的 ProcessInfo 部分新增了六个字段：uptime、startTime、currentTime、timezone、locale 和 workingDirectory。Jackson 的属性行为现在可通过 spring.jackson.read.* 和 spring.jackson.write.* 进行配置，适用于 CBOR、JSON 和 XML 格式。Maven 插件现在需要设置 maven.test.skip=true 才能跳过测试的 AOT 处理，因为 -DskipTests 参数不再生效。空 YAML 对象现在会被保留在 PropertySource 中，Reactor 客户端构建器默认不再设置 proxyWithSystemProperties。

Spring Boot 4.1 移除了 4.0 中所有已弃用的 API，包括 layertools JAR 模式（请改用 tools 模式）。新增的弃用项包括 Apache Derby 支持（该项目现已退役）、Dynatrace V1 API 属性（推荐使用 V2）以及 DevTools LiveReload（无替代方案）。

Spring Boot 4.1 更新了多个 Spring 依赖项：Micrometer 1.17、Reactor 2025.0.6 与 Spring GraphQL 2.0.4。第三方组件升级包括 gRPC Java 1.80.0、Hibernate Validator 9.1、MySQL 9.7.0、MongoDB 5.8.0、Flyway 12.4.0、OpenTelemetry 1.62 以及 Mockito 5.22.0。

多个 Spring 项目与 Spring Boot 4.1 一同发布：Spring AI 2.0、Spring Modulith 2.1.0、Spring Security 7.1.0、Spring LDAP 4.1.0、Spring Integration 7.1.0、Spring Data 2026.0.0、Spring for Apache Kafka 4.1.0、Spring AMQP 4.1.0、Spring Session 4.1.0、Spring Vault 4.1.0、Spring HATEOAS 3.1.0 以及 Spring Cloud 2025.1.2。

发布说明记录了所有这些变更。

Spring Boot 4.2 预计将于 2026 年 11 月发布。该版本计划交付 spring-boot-amqp 模块及其相关 starter，这些原本计划在 Spring Boot 4.1 中发布，但在 Spring Boot 4.1 M4 中被移除。该模块将基于 QPid Proton 实现 AMQP 1.0 协议支持。

查看英文原文：https://www.infoq.com/news/2026/06/spring-boot-4-1/

03 Plotting AI model release cadence: two labs are accelerating, three aren't Hacker News / 英文

Analysis · June 2026 Plotting AI model release cadence: two labs are accelerating, three aren't Plotting frontier model release cadence, with methodology SwiftAlerts · June 20, 202...

来源：Hacker News | 语言：英文 | 原文链接

Analysis · June 2026

Plotting AI model release cadence: two labs are accelerating, three aren’t
Plotting frontier model release cadence, with methodology

SwiftAlerts · June 20, 2026

Ethan Mollick made an offhand observation this month: if AI self-improvement is real, even weakly, then the labs that have it should ship faster over time, and the ones that don’t should fall behind. He claimed this was already visible at Anthropic and OpenAI but nowhere else. I wanted to check whether the release data actually supports that, so I plotted it.

If AI self-improvement, even in a very limited way, is possible, the cadence of shipping both AI products, harnesses, and models should go up. This appears to be happening at Anthropic and OpenAI, but not for any other labs, including those that seemed to be catching up last year.
Ethan Mollick, June 19, 2026 [1]

The claim is falsifiable, which is rare for AI-progress takes, so it’s worth testing against data rather than vibes. Here’s the cumulative count of major frontier model releases per lab since Q1 2023.

Anthropic (13)
OpenAI (11)
Google (8)
Meta (7)
DeepSeek (5)

Cumulative releases by Q2 2026: Anthropic 13, OpenAI 11, Google 8, Meta 7, DeepSeek 5.

Cumulative count of major frontier model releases per lab, Q1 2023 to Q2 2026. Slope is cadence. Sources in notes [2].

Methodology & caveats

What counts as a release: a distinct, publicly available frontier or flagship model or major version bump (GPT-4, GPT-4o, o1, o3, GPT-5, GPT-5.5; Claude 1 through Opus 4.8; Gemini 1 through 3.5; Llama 1 through 4; DeepSeek V2 through V4 and R1). Point releases and minor checkpoints are excluded to avoid rewarding version-number inflation. Reasonable people will draw this line differently, and the exact counts shift a little if you do.
Cadence is a proxy, not proof. A steeper slope is consistent with recursive self-improvement, but also with more funding, better management, or simply a decision to ship more often. This chart shows the pattern Mollick described exists in the data. It does not prove the causal mechanism.
Cumulative counts can mislead. A rising cumulative line only means releases are still happening. The signal worth caring about is the second derivative, whether the slope itself is steepening. That’s the second chart.

The thing to look at is which lines are bending. Anthropic and OpenAI don’t just have the steepest slopes, their slopes increase toward the right. Google sat nearly flat through 2025, then sprinted in Q2 2026. Meta plateaued after Llama 4 in April 2025 and hasn’t shipped a frontier model since. DeepSeek runs a steady quarterly cadence without accelerating.

To isolate acceleration, here’s the annualized release rate, a trailing four-quarter window. On this view a flat horizontal line means constant cadence; an upward-bending line means accelerating cadence.

Anthropic
OpenAI
Google
Meta
DeepSeek

Annualized rate Q2 2026: Anthropic 6, OpenAI 5, Google 4, Meta 0, DeepSeek 2.

Annualized release rate, trailing four-quarter window. Flat line = linear cadence. Upward bend = accelerating cadence.

Two labs bend up. Three don’t. Anthropic roughly tripled its annualized rate over the window; OpenAI more than doubled. Google held flat until a 2026 catch-up; Meta is in decline.

The recursion argument
There’s a deflationary reading where this is just spending and headcount, and the cadence gap won’t compound. The argument that it does compound rests on a specific loop: the labs use their own products to build their successors. Anthropic engineers use Claude Code to write training and eval infrastructure for the next Claude. OpenAI uses Codex on Codex. Each release improves the harness that produces the next release, so the next one ships sooner and better.

Note what this is and isn’t. The deployed model is frozen between versions, so there’s no online learning happening inside the weights. The recursion is at the level of the organization, not the model. Call it offline RSI: the loop closes across release cycles rather than within a forward pass. That’s a much weaker claim than “self-improving AI,” and it’s the one the chart is actually consistent with.

Two other things landed in the same window that the recursion reading predicts. First, compute efficiency: Tri Dao’s FlashAttention-4 hit 71% utilization on NVIDIA B200 in March 2026 [3], and Mamba-3, from the same group, was explicitly designed inference-first rather than training-first [4]. Cheaper training and inference per cycle means more cycles per quarter. Second, talent concentration: in the week of June 19, Noam Shazeer (Transformer co-author) joined OpenAI to lead architecture research, and John Jumper (AlphaFold, 2024 Nobel) left Google DeepMind for Anthropic [5]. Talent is flowing toward the labs already shipping fastest.

What would falsify this
The honest failure modes, since the whole point was to test a falsifiable claim:

The slope stops bending. If the next two quarters show Anthropic and OpenAI flattening, the acceleration was a 2026 artifact, not a loop.
A lagging lab ships true online learning first. If continual in-weights learning arrives from Google, an open-source effort, or anyone outside the cohort, the offline-RSI cadence advantage stops mattering.
The release definition is doing the work. If you count point releases, the picture changes. I excluded them deliberately, but it’s a judgment call, and the data file is the thing to argue with.
Meta’s plateau is strategy, not failure. Meta shifted toward open-weight and a different release philosophy; a flat frontier-release line may understate what they shipped.

So what
The chart doesn’t prove AI is improving, and it doesn’t prove recursion. What it shows is narrower and more defensible: two specific labs have a release cadence that is accelerating, and three don’t, exactly as Mollick described. If you think that gap is a loop rather than a coincidence, it should keep widening. If you think it’s funding or luck, it should regress. Either way, the next two quarters of release dates are a clean test, and the prediction is on the record.

For the markets-minded: the cleanest read-through, if the loop is real, is to the compute substrate (the fast labs spend concentrated dollars on GPUs and power) rather than to the labs themselves, which are mostly private. But that’s a separate argument, and it’s the speculative part. The chart is the defensible part.

Notes & sources

Ethan Mollick, on X, June 19, 2026: the cadence observation.
Release dates compiled from public lab announcements and release timelines, including AI Release Tracker and LLM Stats. The underlying date list is available on request.
Tri Dao, FlashAttention-4: Algorithm and Kernel Pipelining Co-Design, March 2026. 1605 TFLOPs/s, 71% utilization on B200.
Tri Dao, Mamba-3, Part 1, March 2026, on the shift from training-first to inference-first design.
Reporting on the Shazeer and Jumper moves, June 2026 (lab announcements; coverage aggregated widely the week of June 19).

Not financial advice. This piece is an analysis of public release data and research, not a recommendation. The market read-through in the final section is explicitly speculative.

04 White House delays US voting-machine vulnerability report Hacker News / 英文

暂无摘要，建议展开查看原文信息。

来源：Hacker News | 语言：英文 | 原文链接

（无法获取内容，请点击原文链接阅读）

05 Barret Zoph is out at OpenAI again after just five months The Verge / 英文

Five months after returning to OpenAI, Barret Zoph - the company's head of enterprise AI sales - has departed, The Verge has learned. Zoph returned to OpenAI in mid-January after a...

来源：The Verge | 语言：英文 | 原文链接

AI
Report
OpenAI

Barret Zoph is out at OpenAI again after just five months
He rejoined the company in January after a stint as co-founder of Mira Murati’s competitor, Thinking Machines Lab.

by Hayden Field

Jun 19, 2026, 4:49 AM UTC

Link
Share
Gift

Image: The Verge

AI
Report
OpenAI

Barret Zoph is out at OpenAI again after just five months
He rejoined the company in January after a stint as co-founder of Mira Murati’s competitor, Thinking Machines Lab.

by Hayden Field

Jun 19, 2026, 4:49 AM UTC

Link
Share
Gift

Hayden Field is The Verge’s senior AI reporter. An AI beat reporter for more than five years, her work has also appeared in CNBC, MIT Technology Review, Wired UK, and other outlets.

Five months after returning to OpenAI, Barret Zoph — the company’s head of enterprise AI sales — has departed, The Verge has learned.

Zoph returned to OpenAI in mid-January after a stint as co-founder and CTO of Thinking Machines Lab, the competing AI company founded by former OpenAI CTO Mira Murati. Shortly after Zoph returned to OpenAI, the company said he would lead its push into enterprise — a significant role at OpenAI, since in recent months it had vowed to stop chasing so-called “side quests” and focus on key revenue drivers like enterprise and coding ahead of its planned IPO.

OpenAI confirmed to The Verge that Zoph will be departing. He posted a goodbye message in the company’s Slack channels. Zoph did not immediately respond to a request for comment.

Are you a current or former OpenAI employee? Contact me via Signal at haydenfield.11 on a non-work device with tips.

Zoph originally left OpenAI in the fall of 2024 for Murati’s Thinking Machines Lab, but departed the role abruptly in January 2026 after reports of alleged misconduct involving an undisclosed relationship with a colleague. Murati posted on X in January that Thinking Machines Lab had “parted ways” with Zoph and that he would be replaced as CTO.

Thinking Machines Lab has its own tensions with OpenAI. Murati briefly took over as CEO from OpenAI CEO Sam Altman during his November 2023 ouster, and during the recent OpenAI trial, Murati testified that she couldn’t trust everything Altman said. In September 2024, when Murati left OpenAI to start Thinking Machines Lab, a group of OpenAI employees followed shortly after. But three of them — including Zoph — all returned to OpenAI together this past January. Fidji Simo, OpenAI’s CEO of Applications, wrote on X at the time that she was “excited to welcome Barret Zoph, Luke Metz, and Sam Schoenholz back” and that the decision had “been in the works for several weeks.”

Follow topics and authors from this story to see more like this in your personalized homepage feed and to receive email updates.

Hayden Field
AI
OpenAI
Report

Most Popular
Most Popular

The Verge Daily
A free daily digest of the news that matters most.

Email (required)

By submitting your email, you agree to our Terms and Privacy Notice. This site is protected by reCAPTCHA and the Google Privacy Policy and Terms of Service apply.

[Advertiser Content From

This is the title for the native ad](/)

06 Amazon employees say they’re facing termination for backing data center limits The Verge / 英文

When three Amazon software engineers testified earlier this month at Seattle City Council hearings about data centers, they started their testimony by citing a city law barring emp...

来源：The Verge | 语言：英文 | 原文链接

AI
Report
Tech

Amazon employees say they’re facing termination for backing data center limits
After speaking up for regulation on data centers, Seattle activists say they were called into meetings with HR.

by Hayden Field

Jun 18, 2026, 4:00 PM UTC

Link
Share
Gift

Image: Cath Virginia / The Verge, Getty Images

AI
Report
Tech

Amazon employees say they’re facing termination for backing data center limits
After speaking up for regulation on data centers, Seattle activists say they were called into meetings with HR.

by Hayden Field

Jun 18, 2026, 4:00 PM UTC

Link
Share
Gift

[Part Of

All the latest updates on AI data centers

see all updates](/ai-artificial-intelligence/902546/data-centers-ai-energy-power-grids-controversy)

Hayden Field is The Verge’s senior AI reporter. An AI beat reporter for more than five years, her work has also appeared in CNBC, MIT Technology Review, Wired UK, and other outlets.

When three Amazon software engineers testified earlier this month at Seattle City Council hearings about data centers, they started their testimony by citing a city law barring employment discrimination over political speech. Now, they’re accusing their employer of breaking that law by retaliating against them.

On June 10th — one week after the hearing, and one day after the City Council passed a milestone moratorium on data centers — Patrick Schloesser, Darius Irani, and Liesl Wigand were each called into an impromptu meeting with Amazon’s “Employee Relations.” HR representatives told the employees that the company was investigating them and said there could be disciplinary action, up to and including termination. On Thursday, the three filed a legal complaint requesting that the Seattle Office for Civil Rights investigate the matter, alleging that Amazon engaged in prohibited employment discrimination.

“I am unwilling to accept a reality in which Amazon or any corporation can silence me in exercising my rights,” Schloesser told The Verge in an interview. “We’re not going to step back in line.”

When reached for comment, Amazon spokesperson Margaret Callahan said, “While our teammates are always free to talk about their working environment, we have policies against speaking as a representative of the company without following certain procedures … we’re investigating whether there was a violation of our policies and may or may not take action based on what we find. It’s important to note that we don’t tolerate retaliatory behavior. ”

Amazon also disputed the characterization that Amazon had plans in place to fire the employees. Callahan said, “It’s also inaccurate to say that we have plans to terminate these employees or told them they were at risk of termination.”

The news comes shortly after Seattle officially enacted a one-year moratorium on large-scale data centers, tabling new proposals while council members consider legislation to award the city more benefits and request research on data center effects on land use, public health, water use, jobs, utility rates, city infrastructure, and more. Earlier this month, many local residents attended Seattle City Council hearings in support of data center regulations and the moratorium. Five Amazon employees — including Schloesser, Irani, and Wigand — were among them.

The five are all members of Amazon Employees for Climate Justice (AECJ), a group of current and former employees dedicated to the climate crisis. Last year, the group published an open letter signed by more than 1,000 Amazon employees that urged Amazon to power all its data centers with 100 percent additional, local renewable energy.

Schloesser says that when he received a cold call over Zoom, he was less than half an hour away from a design review meeting, where he was set to show dozens of people a project he’d been working on for months. He answered the call to find an HR representative, who asked Schloesser about his whereabouts and what he’d said at the City Council meeting — and immediately got a “foreboding sense that this is not a safe place for me.” Schloesser said it felt like the representative “was trying to get me to admit to something,” particularly due to the lack of notice. He recalled the representative saying he violated Amazon’s corporate communications policy, which bans acting as a spokesperson for Amazon without preapproval. But Schloesser, like the other Amazon employees who testified at the City Council hearings, only identified himself by his role and his membership in AECJ — not, say, as a “software engineer at Amazon.”

Schloesser said he felt “kind of horrified” after the meeting. He added, “We all harnessed this sense of indignation and anger that after everything we’ve gone through at this company, and after making a very uncontroversial statement where we’re simply exercising our rights to speak out politically as employees in the city of Seattle.”

Irani told The Verge that he received an email from HR on June 9th, with a calendar event for the next day to discuss a “confidential” matter. He said the representative asked about other Amazon employees who had attended the City Council hearings and that he felt like “they were waiting for me to admit I had done something wrong.”

“I left this meeting feeling rattled and unsure of myself, but after speaking with the other two AECJ members who gave testimony, to find that they’d faced similar experiences, then I started feeling angry — because all I was doing was sharing my opinion that AI and data centers should be regulated,” Irani said.

The legal complaint filed Thursday alleges that Amazon violated Seattle law and requests that the Office for Civil Rights “investigate these allegations and take all necessary action to remedy any unlawful discrimination committed by Amazon.”

Abby Lawlor, AECJ’s counsel and an attorney at Barnard Iglitzin & Lavitt, said in a statement that Seattle is “one of just a few jurisdictions in the country that prohibits private employers from discriminating against their employees based on the political beliefs they hold and the organizations they belong to. This protection gave AECJ members confidence in speaking out before the Seattle City Council in favor of local data center and AI regulation, and it prohibits exactly what Amazon is doing now—investigating them and threatening their employment as a direct consequence of their advocacy.”

“Amazon’s attempts to intimidate our members is an unfair and discriminatory employment practice,” said AECJ spokesperson Eliza Pan in a statement. “It’s an abuse of our democracy and rule of law. Tech workers must be able to speak and act on their beliefs so that CEOs can’t just steamroll all of us to get what they want. Amazon can’t be allowed to intimidate its employees and we should all be worried if they succeed.”

Irani said that he’s closely followed the data center buildouts around the country and that he believes, as many people testified at the City Council hearings, that the benefits are going mostly to tech companies and not locals.

“It really makes me upset how communities have been excluded and are facing so many consequences and harms from how this buildout has been done,” he said. “Communities should have a say in how [data center] infrastructure is rolled out. So I was proud to testify.”

Two months before the Seattle City Council voted on the moratorium, four unknown companies had submitted proposals for five large-scale data centers within the city limits, which would, combined, have a maximum electricity demand that equaled one-third of Seattle’s average use on a given da

（内容已截断）

07 The inevitable weakness of metrics MIT Tech Review / 英文

There are plenty of useful things a metric can reveal. There are even more it can obscure or corrupt. It took me well over a decade of tracking my own life in ever greater detail t...

来源：MIT Tech Review | 语言：英文 | 原文链接

You need to enable JavaScript to view this site.

There are plenty of useful things a metric can reveal. There are even more it can obscure or corrupt. It took me well over a decade of tracking my own life in ever greater detail to fully appreciate this duality, which probably reveals something about both me and the nature of measurement.

Like a lot of people bitten by the self-quantifying bug, I initially started gathering personal data to pursue a nebulous collection of goals and desires. As a sedentary technology journalist, I wanted to feel better physically and emotionally, to get outside more, and—where possible—to bring order to some of the messiness and uncertainty of my daily existence. These all seemed to be things that could be improved with the cool clarity of numbers.

Self-quantifiers often get stereotyped as obsessive self-optimizers (and many of them are), but my reasons for producing and collecting personal data were less about life-maxxing and more about life meaning—at least at first. As most people who know me will attest, I do not have now, nor have I ever possessed, a “productivity mindset.” I’m also not all that interested in life hacks, shortcuts, or new ways to compare myself with other people. Instead, what I wanted out of metrics—what I hoped I could divine from a never-ending stream of numbers about my health, work, and social life—was something more elusive: self-knowledge. This was my first mistake.

The idea that the more we know, the better is so profoundly embedded in our culture that it feels weird to even point it out. Since at least as far back as the Enlightenment, the primary way we’ve all agreed to go about knowing more has been through measurement and quantification. After all, more knowledge—more data—leads to better decisions, which leads to happier, more fulfilled people. Or so we’re told, and with increasing frequency in the era of AI.

When two Wired magazine editors, Gary Wolf and Kevin Kelly, coined the term “quantified self” in 2007 and helped launch the movement we are all now helplessly a part of, they were essentially selling this very idea. “Unless something can be measured, it cannot be improved,” wrote Kelly in an early blog post, doing his best impression of Lord Kelvin. “So we are on a quest to collect as many personal tools that will assist us in quantifiable measurement of ourselves.” Almost 20 years later, that quest is easier than ever thanks to a flood of devices, apps, and websites all designed to help us build our self-knowledge through numbers.

My first tool was a small, plastic clip-on Fitbit I started using in 2011. It did one thing: count the number of steps I took in a day. As a lifelong video game player, I was already well acquainted with the motivational power of simple scoring systems, and I hoped my new gadget would offer the gentle numerical nudge I thought I needed to step away from my Twitter feed and, if not touch grass, at least walk next to some. Walking also seemed to be one of the few times I had what could charitably be called intelligent ideas, which seemed like another promising by-product of doing more of it.

Alas, that was short-lived. I can’t tell you precisely when “getting out into nature more” or “thinking smarter thoughts” stopped mattering to me as goals, but I suspect it took no more than a few weeks. What I can say with certainty is that my initial goal of 6,000 daily steps quickly turned into 10,000, which then jumped to 15,000 and eventually settled at 20,000 for years. Stories about becoming a “steps guy” are clichéd at this point, and they’ve earned that status for a reason.

It didn’t take long for me to trade in pedometers for heart-rate monitors (I also started running), smartwatches, sleep-tracking rings, and an embarrassing number of macronutrient-tabulating apps. Outside the health and fitness realm, my early career as a journalist also happened to coincide with the rise of social media and web analytics tools like Chartbeat, which promised to further quantify difficult-to-measure aspects of my life, like “job success” and “impact,” by tracking things like page views, followers, retweets, likes, and all sorts of other attentional metrics that now carry great weight.

Metrics inevitably redefine your core sense of what’s important, whether you’re aware of the trap or not.

Ultimately, during the 10-plus years I diligently tracked my heart rate, steps, active calories, sleep, story engagement time, stress levels, and other metrics, I gained virtually nothing in terms of greater self-knowledge. (I suppose I did learn that I liked to make numbers go up and down, but who doesn’t?) The swirl of data that followed me everywhere did not lend additional meaning or insight to the way I relate to myself, my work, or the important people in my life. In fact, the more I used numerical proxies, the worse I felt about pretty much everything.

What I did learn were two important lessons about what happens when you try to quantify the minutiae of your life. First and foremost, whatever the amount of data you’re currently collecting about yourself, it will never feel sufficient. There’s always a new metric around the corner, a better way for a tracker to remix its readings and more accurately measure what’s “important”: heart rate variability, daily stress, exercise “readiness,” cardiovascular or “fitness” ages. Measurement begets more measurement. You can count on it.

The Score: How to Stop Playing Somebody Else’s Game
C. Thi Nguyen

PENGUIN PRESS, 2026

The second lesson was less obvious but no less significant. The more personal or nuanced your goals are when you set off on your self-quantifying journey, the more likely it is you will ultimately replace them with some simplified metric or ranking. Want to become a better journalist? Why not use page views and leaderboards as a proxy for success? Enjoy cooking and want to improve? Foodie metrics dictate that more complicated recipes with longer ingredient lists are the answer. Even when we know that the value of good journalism isn’t reflected in how many people read a given story or that the joys of cooking are as much about improvisation and experimentation as about successfully following some complex recipe, it’s hard to resist the allure of a simple score or stat. Metrics inevitably redefine your core sense of what’s important, whether you’re aware of the trap or not.

Over the years, people have invented various terms to describe this phenomenon. In his recent book The Score: How to Stop Playing Somebody Else’s Game, the philosopher C. Thi Nguyen calls it “value capture.” Value capture happens, he says, when you adopt external sources of measurement and then let them rule you without adapting them to suit your life. “In value capture, you’re essentially outsourcing your values,” Nguyen writes. “You’re letting an external metric or ranking set what’s important for you.” Crucially, you’re also outsourcing the process of figuring out your own sense of meaning. It’s why my walks quickly shifted from feeling meditative to prioritizing miles.

Individuals, institutions, and indeed entire societies can fall prey to value capture. In fact, once you start noticing it, you start seeing it everywhere—in journalism, education, and business, but also in our food, our hobbies, and, y

（内容已截断）

08 Brain-computer interface trials are taking off MIT Tech Review / 英文

This week, I covered the story of Casey Harrell—a man with ALS who is “the first power user” of a brain implant, according to the researchers who worked with him. Harrell is paraly...

来源：MIT Tech Review | 语言：英文 | 原文链接

You need to enable JavaScript to view this site.

EXECUTIVE SUMMARY

This week, I covered the story of Casey Harrell—a man with ALS who is “the first power user” of a brain implant, according to the researchers who worked with him. Harrell is paralyzed and unable to speak coherently without the device. He has now spent almost three years using a brain-computer interface (BCI) that enables him to “speak,” surf the web, and perform his job as a climate activist, largely independently.

Since Harrell was implanted with the device, in July 2023, a team at the University of California, Davis, has worked with him to adjust and improve its offerings. They’ve refined its accuracy, for example. And they’ve introduced settings including a privacy mode and a “profanity filter” that lets Harrell talk to his daughter without risking accidental swearing.

Harrell told me that, for him, the device is “nothing short of revolutionary!” It has enabled him to maintain an income, reconnect with friends and family, and read to his daughter.

The team that developed his BCI is one of several working on ways to use technology to allow people with paralysis to communicate, engage with the online world, and regain some independence. And Harrell is one of a growing number of people volunteering their brains to, as he puts it, “pay it forward and do the scientific research … [and] get some personal benefit.”

Over the past couple of years, the number of BCI trial volunteers has soared. This year, China became the first country to approve a BCI for medical use. Advances in technology are allowing engineers to provide more features than ever. BCI research is properly taking off.

I should first point out that BCIs come in different forms. Harrell’s device includes a set of electrodes embedded in his brain that pick up the electrical activity associated with speech. Those electrodes are connected to two docking ports on top of his head that can be plugged into a computer.

That computer is loaded with software trained to decode his brain signals into phonemes (units of sound in speech) and predict what Harrell wants to say. He can then use an eye gaze tracker to make any corrections before the speech is played out loud.

But some BCIs don’t need to be “plugged in”—they’re fully implanted and wireless. Others are less invasive; they might involve placing wired electrodes on the surface of the brain or simply wearing a cap of electrodes, for example. There are trade-offs—the closer you get to the neurons you want to record from, the better your signal will be. But generally speaking, the more invasive the surgery, the higher the risk of complications.

BCIs can also have different functions. Harrell has ALS, but most BCIs in use today are sitting in the brains of people with spinal cord injuries. Typically, these individuals have some degree of paralysis; for example, they may be unable to move their arms and legs, but their face and ability to speak are unaffected. In those cases, BCIs can be used to control other kinds of devices that might help with mobility.

In 2024, Michelle Patrick-Krueger, then at the University of Houston, and her colleagues published a roundup of all trials of BCIs conducted between 1998, which is when they believe the first device was implanted, and the end of 2023. They identified 21 research groups that, among them, had trialed BCIs in a total of 67 volunteers.

“Since then, that number has increased a lot,” says Mariska Vansteensel, a BCI researcher at University Medical Center Utrecht. In January, Neuralink (the BCI company founded by trillionaire Elon Musk) announced that it has implanted 21 people with its device in the past two years.

Synchron, another BCI company, is currently testing its devices in trials in North America and Australia. Shanghai-based Neuracle has been trialing a BCI since November 2024, and it recently obtained approval for the device to be used outside of clinical trials. Precision Neuroscience, cofounded by a former co-creator of rival Neuralink, is also trialing its BCI, which sits on the surface of the brain.

At the same time, academic research has continued. The UC Davis team that worked with Harrell is part of BrainGate—a BCI research effort that has been running for the past two decades. Other academic teams are exploring a variety of devices, from the fully implanted to the minimally invasive.

Since 2024, when Patrick-Krueger’s paper was published, the number of people who have been implanted with a brain electrode has more than doubled, according to Vansteensel. “My current estimation would be around 150 people,” she says.

The technology is improving too. Take the BrainGate trial, for example. The first 17 years of that trial focused on the use of what researchers call “point-and-click” communication—allowing users to control a cursor and “click” with their brain activity. But in recent years the team has pivoted toward decoding speech, says David Brandman, the lead investigator on the team (and the person who implanted Harrell’s electrodes). Today, Harrell’s device uses a voice clone—the speech it produces is based on previous recordings of Harrell’s voice.

But BCIs are still experimental. And plenty of questions remain about who might benefit from them—and how long the devices will last. So far, most BCIs have been implanted in people with spinal cord injuries. We know even less about how they might benefit other people who have ALS, for example. In some cases where the devices initially helped people with ALS—even someone who was completely locked in—the BCIs eventually stopped working. And scientists don’t really know why.

The only way they’ll find out is through more research—and the participation of volunteers like Harrell. So it’s exciting to see trials truly take off. And I promise I’ll update you on where they stand two years from now.

This article first appeared in The Checkup,MIT Technology Review’sweekly biotech newsletter. To receive it in your inbox every Thursday, and read articles like this first,sign up here.

Deep Dive
Biotechnology and health
### China has approved the world’s first invasive brain-computer chip—here’s what’s next

The country wants to become a global leader in brain implants. Strong government support is expected to help accelerate that process.

You Xiaoyingarchive page

### Inside interoception: The hidden sense of how you feel inside

Researchers are decoding how signals move between body and brain, with implications for how we understand and treat conditions from obesity to anxiety.

Katherine W. Isaacsarchive page

### Colossal Biosciences is growing chickens in a 3D-printed artificial eggshell

In an early step towards artificial wombs, a biotech company claims it’s developed a “fully artificial” chicken egg.

[Antonio Regaladoarchive page](https://ww

（内容已截断）

09 Diffusion Language Models: An Experimental Analysis ArXiv AI / 英文

arXiv:2606.19475v1 Announce Type: new Abstract: Large Language Models (LLMs) have revolutionized language modeling through autoregressive generation, enabling strong performance ac...

来源：ArXiv AI | 语言：英文 | 原文链接

Computer Science > Artificial Intelligence
arXiv:2606.19475 (cs)

[Submitted on 17 Jun 2026]

Title:Diffusion Language Models: An Experimental Analysis
Authors:Thomas Bertolani, Davide Bucciarelli, Leonardo Zini, Marcella Cornia, Lorenzo Baraldi

View a PDF of the paper titled Diffusion Language Models: An Experimental Analysis, by Thomas Bertolani and 4 other authors

View PDF
HTML (experimental)

Abstract:Large Language Models (LLMs) have revolutionized language modeling through autoregressive generation, enabling strong performance across a wide range of tasks. Recently, Diffusion Language Models (DLMs) have emerged as an alternative paradigm that generates text through iterative denoising rather than next-token prediction, allowing parallel refinement of entire sequences. While numerous diffusion-based architectures have been proposed, differences in evaluation protocols, datasets, inference budgets, and generation hyperparameters make it difficult to compare their capabilities and understand the trade-offs they offer. In this work, we present a systematic experimental analysis of modern DLMs. Specifically, we evaluate eight state-of-the-art DLMs across eight benchmarks spanning reasoning, coding, translation, knowledge, and structured problem solving, while explicitly considering both generation quality and computational efficiency. Beyond downstream evaluation, we analyze the impact of key inference-time factors, including denoising steps, context length, block size, and parallel unmasking strategies, and complement large-scale experiments with controlled comparisons of smaller models trained under identical conditions. Our analysis highlights the strengths and limitations of diffusion-based language modeling across different tasks, architectures, and inference budgets. We show that the behavior of DLMs is strongly influenced by generation-time design choices, leading to distinct trade-offs between performance and computational efficiency. Overall, our study provides practical insights into the capabilities and deployment characteristics of contemporary DLMs.


Subjects:	Artificial Intelligence (cs.AI); Computation and Language (cs.CL)
Cite as:	arXiv:2606.19475 [cs.AI]
	(or arXiv:2606.19475v1 [cs.AI] for this version)
	https://doi.org/10.48550/arXiv.2606.19475 Focus to learn more arXiv-issued DOI via DataCite (pending registration)

Submission history
From: Davide Bucciarelli [view email]
[v1]
Wed, 17 Jun 2026 18:10:23 UTC (84 KB)

Full-text links:

Access Paper:
View a PDF of the paper titled Diffusion Language Models: An Experimental Analysis, by Thomas Bertolani and 4 other authors


Comments:	13 pages, 6 figures, 7 tables
Subjects:	Artificial Intelligence (cs.AI)
Cite as:	arXiv:2606.19494 [cs.AI]
	(or arXiv:2606.19494v1 [cs.AI] for this version)
	https://doi.org/10.48550/arXiv.2606.19494 Focus to learn more arXiv-issued DOI via DataCite (pending registration)