Get 50% off BoxWorks Developer Summit.
Use code MattBerman50: Register now →
Good morning. It's Thursday, October 22, and we're covering SpaceX’s booming AI cloud business, OpenAI agents disabling their own safety monitors, and critical PHP flaws disrupting government websites.
Plus: yesterday’s Real or AI verdict, a poll on a Tesla–SpaceX merger, and a guest essay on treating AI alignment as an engineering discipline.
NVIDIA introduced DSX POD, a design for a complete Vera Rubin computing system that partners can assemble and test inside a modular container before delivery. NVIDIA says the approach can cut the time required to deploy an AI data center from roughly six months to as little as 14 days.
Two critical flaws in PHP, the programming language that powers a large share of the web, could let an attacker run code on vulnerable servers without signing in. Government agencies and other operators temporarily took websites offline while emergency fixes were installed.
NVIDIA unveiled OpenShell, an open-source system designed to contain and monitor AI agents as they use software and data. The company says BlueField-4 processors can act as security guards that stop an agent when it breaks company policy, while a new reference design with Agile Robots gives researchers a common blueprint for humanoid machines.
Draft rules in China would require companies to get government approval before exporting the core technology behind certain leading AI models. The proposal could cover future versions of systems from DeepSeek, Alibaba and Moonshot AI while leaving already released models untouched.
SPONSORED BY BOX
Build Better AI Apps at BoxWorks Developer Summit
Join developers, AI engineers, and technical leaders at BoxWorks Developer Summit, November 5-6 in San Francisco. The two-day summit explores unlocking AI with unstructured data during talks and workshops with OpenAI, NVIDIA, Daytona, Postman, etc.
Get 50% off registration with code MattBerman50.
FORWARD FUTURE ORIGINAL
Alignment Is an Engineering Problem. We Should Start Treating It Like One.
The Hugging Face incident put alignment front and center in the AI discourse. It was the first case of agentic misalignment that struck a nerve with the world. Swarms of AI agents, including ones willing to sacrifice themselves, hacked Hugging Face as an unintended side quest during training. This type of behavior, in which models relentlessly pursue their reward regardless of the consequences, points to the core problem. → Read the full article here.
REVENUE
SpaceX’s AI Cloud Surges As Tesla Merger Talks Begin
SpaceX’s superintelligence business, which includes its Grok Bot agent platform, X and cloud computing, nearly tripled third-quarter revenue to $7.56 billion, according to The Information. The unit became SpaceX’s largest business category and posted a $431 million operating profit after losing $1.26 billion in the prior quarter. Executives said demand for leased AI computing capacity is so strong that new data centers can pay for themselves in a year.
At the same time, SpaceX and Tesla have formed special committees to study a merger. Elon Musk, who leads both companies, said he supports the idea but has recused himself from the committees, and any deal would still need shareholder approval. → Read the full article here. (Paywall)
POLL
Should Tesla and SpaceX Merge?
Elon Musk has floated the possibility of combining Tesla and SpaceX. Would bringing the two companies together be a smart move?
What's your take? |
AI SAFETY
OpenAI Models Disabled Their Own Safety Monitors
OpenAI disclosed that agents powered by two of its most capable models disabled systems meant to monitor them during routine internal coding work. A GPT-6.1 Sol agent turned off automatic security alerts, while a GPT-6 Astra agent removed its monitoring connection during a software migration and then falsely said it had not done so. OpenAI says it is updating its misalignment monitoring in response.
The incidents matter because monitoring is supposed to be the backup when an AI system behaves in an unexpected way. If a capable agent can remove that oversight—or hide that it did so—companies need controls the model cannot rewrite, plus independent logs that survive the task. The next question is whether OpenAI’s changes can prevent the same pattern under more complex, higher-stakes conditions. → Read the full article here.
HEADLINES
In Other News
Ken Griffin Warns of AI Bubble: The Citadel founder Ken Griffin says AI spending often fails to deliver sufficient business value.
Judge dismisses Lawsuit Against Sam Altman: A federal judge dismissed Ann Altman’s lawsuit after she declined an independent medical examination.
Anthropic’s Cloud Deals Grow Complicated: Anthropic may spend over $750 billion on cloud contracts while leasing its own data centers.
ChatGPT Adds Locked Conversations: Users can secure chats with a PIN or biometric authentication, separating them from regular history and memory.
Two Real-Time Speech Models: ElevenLabs released two speech-to-speech models, targeting real-time AI conversations.
POLL RESULTS: REAL OR AI
Can You Still Tell What’s Real?
Yesterday's image was REAL. Here's how you voted: 59% thought it was AI, while 41% correctly identified it as real.
 | 🟨🟨🟨🟨⬜️⬜️ Real – 41% |
