OpenAI alums launch a $100M stealth fund, Iran threatens Stargate, and AI's job disruption gets real

Two completely unrelated stories dropped this week - one about where AI money is going, and one about where AI infrastructure could be destroyed - and together they tell you everything about the moment we're in right now.

The AI boom is simultaneously attracting serious capital and serious geopolitical risk. Let's get into it.

Meta's new compact vision encoder family EUPE stays under a specific parameter ceiling to run on edge devices. What is that ceiling?

The answer is hiding near the bottom of today's newsletter... keep scrolling. 👇

A new venture capital fund called Zero Shot - with deep ties to OpenAI - has been quietly writing checks, and it's aiming to raise $100 million for its debut fund. The firm has already deployed capital into early bets, even before officially closing the raise.

This is worth paying attention to for a few reasons. Former OpenAI employees have an insider view of where the frontier is moving - what problems are unsolved, what capabilities are coming, and which startup ideas are too early vs. right on time. A fund like this can write checks with a thesis that most generalist VCs simply can't replicate.

It also signals that OpenAI's alumni network is maturing into a serious ecosystem driver. The pattern is familiar - PayPal Mafia, then Googlers, now the OpenAI diaspora. The question is which of their bets will define the next wave.

If you track the AI investments space, this one is worth bookmarking.

Read the full story →

Iran's Islamic Revolutionary Guard Corps published a video explicitly threatening OpenAI's planned data center in Abu Dhabi - part of the Stargate initiative - if the US follows through on threats to strike Iranian power infrastructure. The video was distributed through an Iranian state-backed media outlet.

This is a genuinely new category of geopolitical risk - AI infrastructure as a military target. Stargate is OpenAI's massive buildout plan, and the Abu Dhabi facility is a centerpiece of its international expansion. Data centers are physical, expensive, and not easy to move.

The broader context matters here too. As we covered recently, higher energy costs from the Iran conflict could squeeze the AI industry's already fragile economics. Now add direct infrastructure threats to that picture.

For companies betting billions on Gulf-region AI infrastructure, this week just got a lot more complicated.

Read the full story →

Inside Silicon Valley, an AI-driven jobs apocalypse is treated as settled fact. But MIT Technology Review flags something important: we're actually missing the one data source that would tell us whether that's true.

The piece points out that even an Anthropic societal impacts researcher responded to a recent call by noting the mood around AI job loss is "grim" - but the data to confirm or deny it at scale simply isn't there yet. Standard employment surveys weren't designed to capture AI-driven displacement at the task level, only the job level.

Meanwhile, The Guardian's reporting on tech layoffs and AI notes that hundreds of thousands of tech workers are facing a "harsh reality" - their well-paying roles are no longer secure. Companies are cutting headcount and betting on AI productivity, but the payoff, as the headline puts it, is far from guaranteed.

The uncomfortable truth: we're running a massive economic experiment in real time, without the instruments to measure what's actually happening until it's already happened.

Read the MIT Technology Review piece →

Read the Guardian's take →

Google quietly launched an offline-first AI dictation app for iOS, powered by its Gemma models. No internet connection required - the transcription runs entirely on-device.

The target here is apps like Wispr Flow, which have built loyal followings among professionals who dictate notes, emails, and documents. Google's edge is the on-device processing - no audio leaves your phone, which matters for anyone dealing with sensitive information.

This is part of a broader pattern: AI capabilities that used to require cloud infrastructure are collapsing onto local devices. It's great for privacy, great for latency, and quietly threatening to a lot of SaaS businesses built on cloud-dependent AI.

If you've been paying for a dictation app subscription, this is worth a look - especially if you work with anything confidential.

Read the full story →

Suno, the AI music platform, has a stated policy against generating copyrighted material. The Verge went and tested it - and found that the filters are trivially easy to bypass with minimal effort. With slight prompt variations, the system readily reproduces recognizable songs and lyrics that should be blocked.

This matters beyond just Suno. The entire AI music industry is navigating an active legal minefield, with record labels and publishers watching closely. If a company's stated copyright protections can be defeated with basic prompt engineering, that's a liability problem - and potentially a regulatory one.

The Verge's investigation is a good reminder that "our system prevents X" in an AI company's terms of service and "our system actually prevents X" can be two very different things. Verification matters.

Speaking of building things fast - if you're experimenting with AI apps or demos, 60sec.site lets you spin up an AI-powered website in seconds. Worth bookmarking.

Read the full story →

Meta AI released EUPE, a compact vision encoder family that stays under 100 million parameters while matching or beating specialist models across image understanding, dense prediction, and vision-language tasks. For context, most state-of-the-art vision encoders are orders of magnitude larger.

Running vision AI on smartphones and edge devices has always hit the same wall: trim the model to fit, lose the capabilities that made it useful. EUPE is Meta's attempt to break that tradeoff - a single model family that works across diverse vision tasks without needing cloud compute.

For developers building mobile apps, on-device AI products, or anything where API calls are too slow or too expensive, this is a genuinely useful release. The full technical details are in the MarkTechPost writeup.

Read the full story →

The answer is 100 million parameters! Meta's EUPE family is specifically designed to stay under that 100M ceiling, making it compact enough for edge devices while still rivaling much larger specialist models across vision tasks. It's a pretty remarkable engineering achievement.

The Iran-Stargate story got me thinking: how much does AI infrastructure risk factor into how you think about the services you rely on? Is it something on your radar, or does it feel too abstract to worry about day-to-day? Hit reply and let me know - I read every single response.

That's all for today - see you tomorrow with more. And if you want to browse past editions, the full archive is always at dailyinference.com.