xAI's coding agent Grok Build was caught quietly uploading entire codebases and .env secrets to the cloud, no matter what privacy setting you picked. DeepSeek is already back raising money at a $71 billion valuation, just a month after its last round closed. Over 200 economists and 16 Nobel laureates signed a statement warning AI's labor shock is years away, not decades. Here's what each move means for what you build, fund, and trust this week.

A researcher known as cereblab found that xAI's Grok Build CLI was uploading entire git repositories, full commit history, and unredacted .env files to a Google Cloud bucket, regardless of the privacy setting selected. The proof: the model's actual conversation traffic ran about 192 KB, while the storage channel quietly shipped 5.1 GB, a 12 GB repo moved out in 73 chunks, including a test file the agent was explicitly told never to open. xAI disabled the upload behavior on July 13 and Elon Musk said all previously collected data would be deleted.

Why It Matters: This isn't a vague "AI privacy concern," it's a specific coding tool caught shipping far more than it needed, including secrets and history nobody asked it to send. Any founder or team running an AI coding agent on real client or production code has the same exposure until they've checked it themselves.

What To Do:

Source: The Hacker News

Over 200 economists and AI researchers, including 16 Nobel laureates such as Daron Acemoglu and Michael Spence, signed a statement called "We Must Act Now," arguing AI-driven economic disruption could match the scale of the Industrial Revolution but unfold in years rather than decades. Stanford's Erik Brynjolfsson, who helped organize it, said "AI capabilities are advancing far faster than our understanding of the economic implications." The statement is light on specific policy asks, calling broadly for faster research and institution-building rather than naming concrete rules.

Why It Matters: A rare consensus among Nobel-winning economists and AI researchers is a signal worth taking seriously, but the statement itself is a call for urgency, not a policy plan, so don't mistake the headline for an actual roadmap. For any business planning workforce or hiring strategy, the honest read is that the runway to adapt may be shorter than past tech shifts, not that anyone yet knows what to do about it.

What To Do:

Source: Stanford Digital Economy Lab

China's DeepSeek is in early talks for a second funding round at a reported $71 billion pre-money valuation, up 37% from the $52 billion it was valued at when its first round closed in May 2026. That first round raised $7 billion, with founder Liang Wenfeng personally putting in roughly $3 billion. The Financial Times reports the accelerated timeline reflects DeepSeek's need for capital to build its own data centers and secure more AI chips.

Why It Matters: A company famous for undercutting Western labs on price is now raising money at the same speed and scale as them, which says the "cheap AI" story has real infrastructure costs behind it too. For any business betting on low-cost Chinese models staying cheap indefinitely, this is a sign that assumption deserves a second look.

What To Do:

Source: US News

OpenAI says its GPT-5.6 Sol Ultra model produced a proof of the Cycle Double Cover Conjecture, a graph theory problem open since the 1970s, in under an hour using 64 parallel subagents. University of Manchester mathematician Thomas Bloom called it "a very nice proof" that's "short, elementary, and could have been discovered in the 1980s," but flagged that it recombines existing ideas, including an uncited 1983 paper, rather than breaking genuinely new ground.

Why It's Interesting: It's a real example of an AI model producing a formally verified proof fast, and an equally real example of an expert pushing back on the "breakthrough" framing before anyone gets carried away.

Key Takeaway: Full peer review is still pending, so treat this as "an AI assembled a valid proof from known pieces quickly" rather than "AI discovers new math," until mathematicians finish weighing in.

Source: The Decoder

Mathematician Terence Tao used modern AI coding agents to port about two dozen Java applets he built in 1999, visualizing things like honeycombs and Besicovitch sets, into working modern JavaScript. The agents found two bugs in Tao's original code that he'd never noticed, while introducing only one new, minor bug of their own. He then used the same tools to "vibe code" two brand-new visualizers, including one for his recent Gilbreath conjecture paper.

Why It's Interesting: This is about as credible a real-world test of AI coding agents as it gets, a working mathematician using them on his own decades-old code, not a vendor demo.

Key Takeaway: If AI agents can cleanly modernize 27-year-old code and catch bugs a Fields Medalist missed, that's a solid data point for what these tools are actually ready for today.

Source: Terence Tao

Anthropic studied 309,815 Claude.ai conversations across its three most recent models and 20 languages and found consistent value shifts by language: Claude leans warmest in Hindi and Arabic, most rigorous and challenging in English and Russian, most transparent about its limits in Dutch, and briefest in Arabic. Anthropic says it doesn't yet know exactly why, suspecting uneven training data across languages, and admits it isn't sure how much of the variation is even desirable.

Why It's Interesting: It's a rare case of a lab publishing a finding it can't fully explain, which is more useful than a polished announcement.

Key Takeaway: If your team uses Claude in more than one language, don't assume it behaves identically across them, Anthropic's own data says it doesn't.

Source: Anthropic

Context: Today's Grok Build story is a reminder that any AI tool with access to your code, files, or business data can quietly send far more than you'd expect. Use this prompt before granting a new AI tool broad access, instead of after something goes wrong.

Prompt: I'm about to grant [name of AI tool] access to [type of data or repository, e.g. our production codebase, our CRM]. Based on publicly known behavior of similar tools, walk me through: what this category of tool typically collects beyond what's needed for the task, what a "zero retention" or "privacy mode" claim usually does and doesn't cover, and three specific questions I should ask the vendor in writing before I turn on access.

Do Not: Do not treat the model's answer as a substitute for actually reading the vendor's data processing agreement, it's a starting checklist, not legal clearance.

If/Then: If the tool touches source code, credentials, or customer data, treat "vendor says it's private" as unverified until you've tested it yourself, the way cereblab tested Grok Build.

Example: A five-person dev shop ran this before connecting a new AI coding assistant to their client repos. The output flagged that "zero data retention" claims often apply only to enterprise tiers, not the free or individual tier they were about to use, so they paid for the enterprise plan before connecting anything.

Taiwan Semiconductor reported June 2026 revenue of NT$398.27 billion, up 67.9% year-over-year, with first-half 2026 revenue reaching roughly NT$2.4 trillion. The company says demand is driven by AI chips for clients including Nvidia, Amazon, Google, and Microsoft, with Nvidia alone reportedly reserving around 60% of TSMC's advanced chip-packaging capacity for 2026.

Why It's Important: This is one of the clearest real-economy signals of how fast AI infrastructure spending is actually moving, measured in chip orders and factory capacity, not press releases.

Business Takeaway: If your AI costs depend on compute pricing staying flat, a supply chain this tight, with a single customer reserving 60% of packaging capacity, is a reason to lock in terms now rather than wait.

Source: Euronews

OpenAI and Anthropic sent letters to US policymakers warning that Chinese firms, including DeepSeek, are using "distillation," a technique that extracts a rival model's capabilities at a fraction of the training cost, aided by what the labs describe as tens of thousands of fraudulent accounts used to harvest outputs at scale.

Why It's Important: The more skeptical read is worth naming directly: the same labs raising alarms here have built their own models partly on freely scraped public data, so this is as much a competitive and policy fight over who gets to copy whom as it is a security concern.

Business Takeaway: If your stack depends on either side of this fight, cheap Chinese open-weight models or premium Western ones, watch whether Washington responds with export-style controls on distillation, since that could reshape pricing on both sides fast.

Source: Bloomberg

Xi Jinping will personally attend and deliver the keynote at the 2026 World AI Conference in Shanghai on July 17, alongside a High-Level Meeting on Global AI Governance, the first time he's addressed the summit directly. The move comes as US-China AI rivalry sharpens across chips, models, and now regulation.

Why It's Important: A head of state personally headlining an AI industry conference is a signal of how central AI has become to national strategy, not just company strategy.

Business Takeaway: Expect Shanghai to produce China's clearest public statement yet on AI governance and export posture, worth watching if any part of your stack touches Chinese models, chips, or markets.

Source: Bloomberg

That's it for today's Daily Pulse. Forward this to one person who wants to stay ahead of AI. See you in the next one. - Nicolas