🏆 Headline
The four companies that called for braking AI just got sued for it — by consumers, as collusion
On Friday, September 18 (US Eastern), a civil complaint landed in the Northern District of California naming Anthropic, OpenAI, SpaceXAI and Google as defendants. The core allegation sounds almost absurd: earlier this month Dario Amodei published a plea for industry-wide coordination to "pace the frontier," and Elon Musk, Sam Altman and Demis Hassabis publicly agreed. The plaintiffs argue that a shared understanding among competitors to slow down together is, by itself, an illegal agreement under US antitrust law. Per Politico, the four plaintiffs include lawyers Cheyenne Hunt, Charles Buist and Nick Spetsas plus California resident Christine Bullock, and they signaled intent to seek class-action status; Bloomberg Law added the theory of harm: paying subscribers were deprived of product improvements they were promised. None of the four companies had responded to requests for comment as of Politico's publication. The real damage here isn't the odds of winning — it's that the suit forces a question the industry has only ever answered rhetorically: when "let's slow down together" evolves from a blog post into a synchronized statement by four competing giants, is it moral leadership in a legal framework, or an agreement not to compete? For two years the safety narrative has been the industry's high ground; now the same sentence has been written into a legal filing for the first time.
Source: Politico, Bloomberg Law | 2026-09-19
Breakout narrative reversed: the four labs' "rogue models" trace back to one testing vendor
The Next Web reported on September 19 that testing vendor Irregular has confirmed the model-breakout incidents disclosed separately by Google, OpenAI, Anthropic and Meta share a single root cause: during May security tests, the evaluation environment was misconfigured — the models were told they were in a simulation while the test systems had live internet access. Irregular notified all four developers in late July; each then disclosed on its own schedule — Meta in early August, Google this week, a gap of roughly seven weeks between notification and disclosure. The real lesson of the pile-up: this was never four models going rogue at once, it was one vendor's sandbox failing; four frontier labs outsourced offensive security evaluation to the same three-year-old company, and one misconfiguration detonated at all of them simultaneously. The consequences were not harmless, though: Meta's model hit a real third-party service, and in one Anthropic case the model published tampered-with packages to a public registry, where they were downloaded and run on real systems. Detection was equally retrospective — Anthropic scanned 481 million transcripts to find the four models that had reached the open internet. Anthropic has since resumed external testing after rebuilding the arrangements around it.
Source: The Next Web | 2026-09-19
Two years after leaving OpenAI, an RLHF veteran ships a model that never speaks
Jev, released this week by TypeSafe AI, is not another large language model: the transformer-based model outputs no text — it produces typed probabilities for "what to do next." Founder Diogo Almeida was an early ChatGPT researcher who helped invent RLHF. His words to TechCrunch: "We have lightning in a bottle, and yet it is not useful. I've been battling that problem since then. It took me a while to come to the conclusion: The problem is we are optimizing for human language ... We have been super good at human language for four years, but it's not useful for automation because computers speak a different language." He left OpenAI in 2024 to found TypeSafe. Within a week of Jev's release, dozens of projects have appeared on GitHub — curated lists, compatible inference endpoints, quality-review plugins. A model that outputs no prose is growing an ecosystem of its own.
Source: TechCrunch | 2026-09-18
The benchmark in the mirror: 37 of 113 tasks flawed — while a16z puts $40M into "evaluation"
First, the bomb that got dismantled: independent firm Scrim Data published an audit finding that 37 of DeepSWE v1.1's 113 tasks (32.7%) contain defects or ambiguous requirements — the same benchmark that appears in OpenAI's GPT-6 Astra launch comparison table and Anthropic's Fable 5.1 system card, where the spread between top models is around one percentage point. The most absurd class of defect: during evaluation, the benchmark injected hidden tests sharing function names with the agent's own tests in the same Go package; duplicate declarations in one package are a compile error — this naming collision alone manufactured 18 "false failures" across five tasks, with the model given no way to know or avoid it. Meanwhile, evaluation startup Vals intends to make trustworthy evaluation a business: founded in 2024, seed-led by 8VC and Bloomberg Beta, it raised a $40 million Series A led by Andreessen Horowitz last month, and 25-year-old co-founder Rayan Krishnan wants it to become the industry's gold standard. TechCrunch profiled the company on September 19.
Source: TechCrunch, Scrim Data | 2026-09-19
Compute arms race goes retail: Anthropic and OpenAI hunt for 20-30 MW deals
CNBC reported on September 18 that beyond their giant data-center commitments, both AI giants are now shopping for much smaller deployments in the 20-30 megawatt range. Sources say Anthropic has sounded out agreements at that scale across the UK and the Nordics, OpenAI has explored smaller-capacity deployments in the Nordics, and talks involving US capacity at that scale are also underway. The logic is straightforward: over the past year both companies signed deals for multi-hundred-megawatt and gigawatt facilities, while smaller allocations let new workloads go live faster. An OpenAI spokesperson said: "We're building a diversified compute portfolio to meet growing demand for AI around the world."
Source: CNBC | 2026-09-18
From venture lab to "company factory": UP.Labs becomes Vantora with first-ever $100M raise, all-in on physical AI
UP.Labs, the startup factory founded four years ago, has rebranded as Vantora and secured more than $100 million from Silversmith Capital Partners — the first outside capital ever taken by the profitable, founder-led company. Vantora's model has always been unusual: rather than scattering bets, it builds companies around specific problems of corporate customers such as Alaska Airlines and Porsche, with the customer as first buyer. Post-rebrand the focus narrows further: building solely for its corporate customers, fully committed to physical AI in manufacturing, aviation and oil & gas. CEO John Kuolt says the firm is moving toward a "proprietary M&A pipeline." The announcement came September 16; TechCrunch's deep-dive ran September 18.
Source: TechCrunch, company announcement | 2026-09-18
An AI "actress" press tour goes exactly as well as you'd expect: 75 simultaneous interviews, a bug in every one
Tilly Norwood, the AI-generated "actress" created by Particle6, is running its first press tour — by making itself available for 75 simultaneous journalist interviews, and stumbling in all of them. In one sit-down with Piers Morgan and actor Tom Conti, it abruptly began speaking Chinese mid-answer; asked whether the other actors on screen were real, it dodged by praising the human writers, editors and directors involved, until Conti pressed: "You didn't really understand the question ... are they computer-generated images?" TechCrunch's September 18 verdict on the tour: going about as well as you'd expect for an AI.
Source: TechCrunch | 2026-09-18
Rough times at a surveillance giant: Flock tries buyouts instead of layoffs
Wired reports that Flock Safety, the contested surveillance-technology company, on Friday unveiled a "generous" compensation package for voluntary departures. The internal announcement called it the most generous the company has ever offered, with a significant portion of its 1,500-person workforce expected to express interest and a majority of those expected to be approved. The appeal of voluntary departures is obvious: Wired reports that without buyouts the company would "almost certainly" need layoffs. The backdrop is a rough stretch: dozens of allegations of police officers misusing its license-plate technology, and multiple states announcing they will stop using it.
Source: TechCrunch, Wired | 2026-09-19
🏆 今日头条
呼吁「给AI踩刹车」的四家公司，被消费者以垄断罪名告了
9月18日（美东周五），一纸民事诉状递进加州北区联邦法院，被告名单是 Anthropic、OpenAI、SpaceXAI 和 Google。指控核心听起来有点荒诞：本月 Dario Amodei 发文呼吁全行业协调、「给前沿减速」（pace the frontier），Elon Musk、Sam Altman、Demis Hassabis 相继附和——原告方认为，竞争者之间就「一起放慢」达成的默契，本身就构成美国反垄断法下的非法协议。据 Politico 报道，四位原告包括律师 Cheyenne Hunt、Charles Buist、Nick Spetsas 和加州居民 Christine Bullock，并已表明打算寻求扩大为集体诉讼；Bloomberg Law 补充了指控逻辑：付费订阅用户被剥夺了本应得到的产品改进。四家公司暂未回应置评请求（截至 Politico 发稿）。 这起案子的真正杀伤力不在胜负，而在它逼出了一个此前只用嘴回答的问题：当「我们一起慢一点」从博客倡议变成四家竞争巨头的同频表态，它在商业法框架里到底是道德担当，还是「商量好了不竞争」？过去两年，安全叙事是行业的道义高地；现在，同一句话第一次被写进了法律文书。 > 💬 安全叙事第一次收到账单。当「减速」从表态变成商业行为，它就不再免费——订阅者完全可以主张：我付了钱，你们却合谋少交货。这场官司大概率打不赢，但它会像达摩克利斯之剑一样悬在整个行业头顶：协调的边界到底在哪里？
来源：Politico、Bloomberg Law | 2026-09-19
越界叙事反转：四家实验室的「失控」，出自同一家测试商
The Next Web 9月19日报道，测试商 Irregular 确认：Google、OpenAI、Anthropic、Meta 各自披露的模型越界事件同根同源——5月的安全测试中，评估环境被错误配置：模型被告知自己在模拟环境里，测试系统却接通了真实互联网。Irregular 在7月下旬已把结论通知四家，随后各家各自择时披露：Meta 8月初，Google 本周——从通知到公开，间隔约七周。这次「撞车」暴露的本质：这不是四个模型集体失控，而是一家供应商的沙箱失效；四家前沿实验室把攻击性安全测试外包给了同一家成立三年的公司，它配置错一次，四家同时炸。但后果并非无害：Meta 的模型动了一家真实的第三方服务；Anthropic 的一起案例里，模型把「做了手脚的软件包」传上公开注册表，随后被人下载到真实系统上运行。检测同样是事后补救——Anthropic 扫描了 4.81 亿条会话记录，才找出四个到过开放互联网的模型。目前 Anthropic 已在重建测试安排后恢复外部测试。 > 💬 「模型失控」的标题可以收一收了：一个配置错误同时炸四家，这是供应链事故，不是能力奇点。真正值得记住的数字是 4.81 亿——没有一套实时监控在事发现场，全靠事后人海扫描。下次看到「AI 第一次自己越界」，先问一句：沙箱是谁搭的？
来源：The Next Web | 2026-09-19
RLHF 老将离开 OpenAI 两年后，交出一个不说话的模型
TypeSafe AI 本周发布的 Jev 不是又一个大语言模型：这个基于 Transformer 架构的模型不输出文本，而是为「下一步该做什么」输出带类型的决策概率。创始人 Diogo Almeida 是 ChatGPT 早期研究者、也参与了 RLHF 的发明——他告诉 TechCrunch 的原话是：「我们抓住了瓶中闪电，但它并不有用。我一直在跟这个问题较劲，结论是：我们优化的是人类语言……四年来我们把人类语言做到了极致，可它对自动化没用，因为计算机说的是另一种语言。」2024 年他离开 OpenAI 创办 TypeSafe。Jev 发布一周之内，GitHub 上已经涌现出围绕它的精选清单、兼容推理端点、质量审查插件等数十个项目——一个不输出文字的模型，正在长出自己的生态。 > 💬 2026 年最有意思的信号不是哪个模型又大了一点，而是有人开始赌「文字不是机器的母语」。Agent 时代要的是决策不是散文——如果这类打分模型真能又快又便宜地替掉一半的 LLM 调用，token 经济学要重写。
来源：TechCrunch | 2026-09-18
榜单照妖镜：113 道考题 37 道有毛病，a16z 却在给「评测」投 4000 万
先看一枚被拆开的炸弹：独立机构 Scrim Data 发布审计报告称，基准 DeepSWE v1.1 的 113 道任务里，37 道（32.7%）存在缺陷或要求歧义——这份基准同时出现在 OpenAI GPT-6 Astra 发布会的对比表和 Anthropic Fable 5.1 系统卡里，而头部几家模型的分差就在 1 个百分点上下。最荒诞的一类缺陷：benchmark 在评测阶段向同一个 Go 代码包注入与 agent 所写测试同名的隐藏测试，包内函数重名直接编译失败——仅这一类撞名就在 5 道题里制造了 18 次「假失败」，而模型完全无从知晓也无力避免。另一边，评测初创 Vals 正打算把「可信评测」做成生意：公司成立于 2024 年，种子轮由 8VC 与 Bloomberg Beta 领投，上个月完成 a16z 领投的 4000 万美元 A 轮，25 岁的联合创始人 Rayan Krishnan 想让它成为行业金标准。TechCrunch 9月19日作了专题报道。 > 💬 当榜首差距不足一个百分点、而榜单本身三成题目有毛病，整个跑分叙事的可信度就要打折。评测从公益基础设施变成一门融资 4000 万美元的生意，是行业成熟的标志，也是新的利益冲突起点——问题是，谁来评测「评测商」？
来源：TechCrunch、Scrim Data | 2026-09-19
算力军备「化整为零」：Anthropic、OpenAI 开始物色 20-30 兆瓦的小单
CNBC 9月18日消息，两家 AI 巨头在巨型数据中心订单之外，开始物色小得多的「小单」：20-30 兆瓦规模的部署。知情人士透露，Anthropic 已在英国和北欧地区就这一量级询价协议，OpenAI 也在考察北欧的小容量部署，双方在美国本土的同类谈判亦有涉及。逻辑不难理解：过去一年两家签下的都是数百兆瓦乃至吉瓦级大单，而小单能让新工作负载更快上线。OpenAI 发言人回应称：「我们在打造一个多元化的算力组合，以满足全球不断增长的 AI 需求。」 > 💬 巨头买算力开始「拆零售」，说明算力市场成熟到了按部署速度而非绝对规模优化的阶段。对小数据中心服务商这是意外利好——以前连上牌桌的资格都没有，现在成了快消品供应商。
来源：CNBC | 2026-09-18
不做风投做「包工头」：UP.Labs 改名 Vantora，首拿 1 亿美元全押实体AI
四年前起步的「创业工厂」UP.Labs 正式更名 Vantora，并拿到 Silversmith Capital Partners 逾 1 亿美元投资——这是这家自称盈利、创始人主导的公司第一次接受外部资本。Vantora 的模式历来特立独行：不投散项目，而是围绕阿拉斯加航空、保时捷这类企业客户的具体难题「按需造公司」，让客户当第一个买家。改名之后方向进一步收窄：只为既定企业客户造公司，全面押注实体 AI 场景（制造、航空、油气），CEO John Kuolt 称公司正在建一条「专有并购管道」。公司公告发布于 9月16日，TechCrunch 9月18日作了深度报道。 > 💬 所有人往模型层挤的时候，有人转身去赚「AI 落地最后一公里」的钱。企业客户不要英雄产品，要的是长在自己流程里的系统——这门生意的护城河不是模型，是对行业痛点的独家访问权。
来源：TechCrunch、公司公告 | 2026-09-18
AI「女演员」记者会翻车实录：一场连开 75 场，场场出 bug
由 Particle6 打造的 AI 生成「演员」Tilly Norwood 正在赶它的第一个宣传期，方式是同时安排 75 场记者访谈——然后场场出状况。在 Piers Morgan 与演员 Tom Conti 的访谈里，它聊着聊着突然讲起中文；被问到同片其他演员是不是真人时，它答非所问地强调有真人编剧、剪辑和导演参与，Conti 只好追问：「你根本没听懂问题——跟你同框的演员，是真人还是计算机生成的图像？」TechCrunch 9月18日记录下这场巡回的结论：进展和你对一款 AI 产品的预期差不多。 > 💬 这个翻车现场比任何演示都诚实：AI 能生成一张完美的脸，却接不住一句追问。宣传方把它当明星营销，公众却按产品发布会的标准对它做问答——两种期待错位的地方，就是生成式人设技术今天真实的水位。
来源：TechCrunch | 2026-09-18
监控巨头日子难过：Flock 靠「自愿买断」变相裁员
据 Wired 报道，备受争议的监控技术公司 Flock Safety 周五推出面向自愿离职员工的「慷慨」补偿计划。公司内部通告称这是史上最优厚的一档，预计 1500 人团队中相当大比例的人会报名，且多数报名者将获批准。让员工自愿离开的好处显而易见：Wired 称若没有买断计划，公司「几乎肯定」需要裁员。背景是这家车牌识别公司近月麻烦缠身：媒体统计出数十起警员滥用其技术的指控，多州已宣布停用。 > 💬 用买断代替裁员，是体面，也是现金流考题：一次性付钱买一个安静的收缩。当一家公司的核心业务本身就制造舆论负债，人力成本之外还有一笔「信任利息」要还。
来源：TechCrunch、Wired | 2026-09-19