Hey hustlers,
Quiet week on the surface. Absolutely chaotic underneath.
Two of the biggest tech companies in the world started trading people for GPUs. Elon's AI just made a case for replacing your entire customer support team. And China dropped its most powerful open model yet, built on chips that have nothing to do with Nvidia.
Let's take a look.
👾 WHAT'S NEW IN AI
1. Meta and Microsoft Are Cutting 20,000 Jobs to Pay for AI
Meta announced it's cutting 10% of its workforce, around 8,000 employees, starting May 20. The memo used the exact words "offset the other investments we're making."
Microsoft, the same day, offered voluntary buyouts to 7% of its US staff, the first time in its 51-year history it has done this. Together, that's over 20,000 potential job cuts in a single Thursday. Amazon, Salesforce, Snap, and Block have made similar moves this year. Over 92,000 tech workers have been laid off in 2026 so far. Amazon, Google, Meta, and Microsoft alone are expected to spend $650 billion on AI capital this year.
Why you should care: This is not a "tech winter." It's a deliberate swap: human headcount out, AI infrastructure in. The companies doing this are not struggling. Meta's stock is near all-time highs. This is a bet that AI can do what those teams were doing, and do it cheaper. Whether that bet pays off or not, the signal is clear. The AI economy is being built on the savings from the human economy.
2. xAI Just Launched a Voice AI That Closes Sales Calls on Its Own
Elon Musk's xAI released Grok Voice Think Fast 1.0, their new flagship voice model available via API. It tops the Tau Voice Bench leaderboard at 67.3%, beating Gemini, GPT Realtime, and every other voice model in the field.
The real flex? xAI revealed it's already powering Starlink's customer support line at +1 (888) GO STARLINK. The results:
70% of support calls resolved with no human involved.
20% of sales calls result in a purchase, with the AI closing the deal.
28 tools used by a single agent across hundreds of workflows.
It handles heavy accents, noisy calls, interruptions, and structured data collection like addresses and account numbers, all without any awkward pauses.
Why you should care: Voice AI just crossed from "cool demo" to "running a real business workflow at scale." If you've ever thought AI can't handle the messiness of actual phone calls, this is the answer. The model reasons in the background while responding, so it sounds natural while doing complex work. At $0.05 per minute, it's also absurdly cheap. Phone-based customer support as we know it has a shelf life now.
3. DeepSeek V4 Dropped and It Runs Without Nvidia
Chinese AI lab DeepSeek released a preview of V4, its most powerful model yet, in two versions: V4 Flash (fast and cheap) and V4 Pro (frontier grade).
It's fully open source, supports a 1 million token context window, meaning you can feed it an entire codebase in one go, and benchmarks put it between GPT-5.2 and GPT-5.4 on reasoning and coding tasks. The most geopolitically significant detail: V4 was built and runs on Huawei Ascend chips, not Nvidia. DeepSeek gave Chinese chipmakers early optimization access and deliberately shut out Western suppliers. Jensen Huang reportedly called this "a disaster" for the US.
Why you should care: A year ago DeepSeek's R1 model wiped a trillion dollars off global stock markets in a single day. V4 probably won't do that because markets have already priced in the reality that China builds competitive AI. But here's what matters: a frontier-class open-source model that runs on domestic chips just became available to every developer on the planet for free. The Nvidia dependency that the entire AI industry was built on just got a serious alternative.
👾 THE GOOD STUFF
🔧 AI Tool: ElevenLabs Voice Studio
If the Grok Voice story got you thinking about voice AI for your own work, ElevenLabs is the most accessible way to experiment with it. Clone a voice, generate realistic speech in 32 languages, and build voice agents without writing a single line of code. Free tier available. Used by creators, podcasters, and enterprises alike.
🐙 GitHub: deepseek-ai/DeepSeek-V4
The official repo for DeepSeek V4, now public on Hugging Face and GitHub.
If you want to run the most powerful open-source AI model in the world locally or via API, start here. The Flash version is surprisingly lightweight for what it can do.
🎬 YouTube: Web Design Just Got Easier Using the Kimi K2.5 Open AI Model
Everyone's talking about DeepSeek V4 but this model quietly dropped and it might be better for design work. Worth watching before you sleep on it.
👾 TO READ
"AI Models Are Now Detecting Harmful Content
From the Inside Out”
📎 https://huggingface.co/papers/2604.18519
Researchers from the University of Toronto built SIREN, a lightweight system that detects harmful content by reading the internal signals of an AI model while it processes a prompt, rather than scanning the final output.
What they built: Instead of checking what the AI says after it's done thinking, SIREN reads what's happening inside the model layer by layer as it generates a response.
It identifies "safety neurons" that light up when harmful content is being processed, before it ever reaches the output.
What they found: SIREN outperforms every other open source content moderation system on the market while using 250 times fewer parameters. It also works in real time while the model is still generating, catching harmful outputs mid-stream before they finish.
Why it's interesting: Every AI product you use right now has some version of a content filter sitting on top of it. SIREN flips that model entirely. Instead of a bouncer at the door, it's a monitor inside the room. In a week where AI safety and misuse are front and center, this is the research that quietly matters most.
🧵 Thread Drop
Twenty thousand jobs. A voice AI closing sales calls. China's most powerful open model yet, running without a single Nvidia chip.
This week wasn't loud. No flashy keynotes, no viral product launches. Just quiet, structural shifts that will look obvious in hindsight.
That's usually the scariest kind.
👾 See you soon 👾
Want to work together?
If you're building something in AI and
think our audiences would vibe with it, let's talk.
