Welcome back to the dAIgest! If you're new here, the first issue catches you up on the big picture. This week, we're covering the stories from roughly the last two weeks of April into early May.

1. DeepSeek V4 Is Here. Finally.

If you had been reading the internal version of this newsletter (hi, Sandstorm team), you know I've been tracking DeepSeek V4 since February, because it was meant to release mid-February. Then March. Then "any day now" for all of April. I started calling it the "Waiting for Godot" of AI models.

Well, Godot showed up.

DeepSeek officially launched V4 in preview on April 24, releasing two models: V4-Pro and V4-Flash. Both are open-source under the MIT license, meaning anyone can download, run, and modify them for free.

1. The specs are enormous. Both models support a 1 million token context window, meaning they can process the equivalent of several novels at once. V4-Pro is now the largest open-weight model ever released.
2. The performance is competitive with the best closed models. On formal math benchmarks, V4 scored a perfect 120/120 on Putnam-2025 proofs. On coding and reasoning benchmarks, it's competitive with GPT-5.4 and approaching Claude Opus territory, though it still trails the top closed-source models by roughly 3-6 months of development.
3. The price is absurd (in a good way!) V4-Flash costs $0.14 per million input tokens and $0.28 per million output tokens. For context, that's cheaper than OpenAI's smallest model. V4-Pro, the flagship, is still a fraction of what Anthropic and OpenAI charge for their top-tier models.
4. It runs on Chinese chips. V4 is optimized for Huawei's Ascend chips rather than Nvidia GPUs. It's the first frontier model designed to run independently of American hardware, a deliberate break from the industry norm and a significant moment in the US-China AI competition.

Why it matters: If you're a developer or tinker with AI, V4-Flash is worth trying: it's absurdly cheap and genuinely capable. For everyone else, the takeaway is simpler: the gap between free/open AI and expensive/closed AI continues to shrink. The companies charging premium prices for their models need to prove the premium is worth it.

2. SpaceX Might Buy Cursor for $60 Billion

Elon Musk's SpaceX announced a deal with Cursor, the wildly popular AI coding tool, that gives SpaceX the option to acquire the company for $60 billion later this year (or pay $10 billion just for the collaboration). For context: Cursor was valued at $2.5 billion in January 2025. By November, it was $29.3 billion. Now we're talking $60 billion.

The deal makes strategic sense for SpaceX, which is targeting a $1.75 trillion IPO in June - potentially the largest in history. Adding an AI coding company to the portfolio makes SpaceX look more like an AI conglomerate and less like "just" a rocket company. SpaceX already absorbed Musk's xAI startup in February.

TechCrunch reported that Cursor was literally hours away from closing a $2 billion funding round when the SpaceX deal landed. The startup was running parallel processes which, apparently, is common in Silicon Valley.

Why it matters: AI coding tools are now valued like entire tech companies were a decade ago. If you write code for a living (or manage people who do), understanding tools like Cursor isn't optional anymore.

3. Hackers Stole 40,000 Voices (and the IDs to Match)

On April 4, the hacking group Lapsus$ posted 4 terabytes of data stolen from Mercor, an AI contractor platform that supplies training data to OpenAI, Anthropic, and Meta. The stolen data includes studio-quality voice recordings, averaging 2-5 minutes each, from over 40,000 contractors, paired with their government-issued ID scans. Five lawsuits were filed within ten days.

High-quality voice cloning typically requires roughly 15 seconds of clean audio; these recordings exceed that threshold by a factor of 8 to 20, and they come bundled with verified identity documents for the same people.

The attack vector was surprisingly sophisticated: hackers compromised LiteLLM, an open-source AI gateway downloaded 95 million times per month, injecting malicious code into its software. Mercor was one of the companies that unknowingly ingested the compromised software.

The real-world implications are already documented: synthetic voice attacks against insurance call centers increased 475% year-over-year in 2025, and the FBI logged $2.3 billion in elder fraud losses in 2026, with emergency impersonation calls as the fastest-growing category.

Why it matters: Your voice is becoming a security credential, and unlike a password, you can't change it once it's stolen. If you've ever recorded voice samples for any AI training platform, this breach is a wake-up call. More broadly, set a verbal codeword with family members for emergencies. If you get a panicked call from someone who sounds exactly like your kid or parent asking for money, that codeword is your verification.

Quick Hits

• Gartner: successful AI organizations invest 4x more in data foundations. AI success is mostly a data quality and change management problem, not a model selection problem.
• Google Gemma 4 released. An open-weight model that runs on just 8GB of VRAM, meaning it can run on a decent laptop.
• UK announced plans for an AI hardware strategy to reduce reliance on foreign chip supply chains.

One Thing to Think About

A Chinese lab just released a model that's competitive with the most expensive AI systems on Earth for a fraction of the price, with open weights anyone can download, running on Chinese-made chips instead of American ones. Three years ago, this would have been unthinkable. A year ago, it was aspirational. Today, it's just... a product you can use.

The implications cascade in every direction. If frontier-quality AI is available for free (or nearly free), what happens to the business models of companies charging $200/month for access? If the best open-source model runs on non-American hardware, what happens to US chip export controls as a tool of geopolitical leverage? If anyone can download and run a model this powerful locally, what happens to the centralized AI platforms that currently gatekeep access? None of these questions have clean answers yet. But DeepSeek V4 makes them a lot more urgent.

The boring, but important, takeaway? The AI models are converging. The gap between the best and the rest is shrinking fast, which means the differentiator increasingly isn't which model you use: it's how well you implement it. That's a more interesting problem, and it's one that us smelly humans are still uniquely equipped to solve.

See you next week!