😺 You're either Jeremy or you're cut

😺 You’re either Jeremy or you’re cut

April 26, 2026

Your browser does not support the audio element.

Welcome, humans.

Friday morning, China’s DeepSeek shipped its long anticipated V4 model with a technical paper that does something almost no frontier lab ever does. It admits V4 trails GPT-5.4 and Gemini 3.1 Pro by 3-6 months on standard intelligence tests. A US lab would never; well, unless you count a code red or Meta’s throw money at the problem moves.

Well, the rest of the 50 pages then walk through where V4 actually won, which turns out to be a different race entirely: long-context economics.

Here’s what that means in practice. V4-Pro can take in up to a million tokens at once, and since a token is about three-quarters of a word, that’s roughly 750,000 words, or about two long novels. And V4 serves all that long-form reading for about a tenth of what comparable models charge: $4 per million output tokens versus $14-15 for ChatGPT and Claude, per Hugging Face’s Tiezhen Wang.

Whether V4’s efficiency advantage holds, or melts at the next training generation (Google DeepMind’s Susan Zhang has thoughts) is TBD. Read our deep dive for the architecture, the chip controversy, and Zhang’s case for why V4’s polish might be papering over architectural rot. From what we can tell, this report is the technical talk of the town, even if we can barely keep up, but we try to explain everything as best we can!

Here’s what happened in AI today:

🐋 DeepSeek shipped V4 with 1M-token context for 10% the cost
😿 Meta cut 8,000 jobs to fund its $72B AI buildout
💰 Google committed up to $40B to Anthropic in cash and compute
🍪 Anthropic’s Project Deal: Claude agents closed 186 real marketplace deals
📰 Meta-AWS signed a multibillion Graviton5 chip deal for agentic AI

… and a whole lot more that you can read about here.

P.P.S: Love robots? We’re starting a new robotics newsletter! Sign up early here.

Meta told staff this week it’s cutting 10% of its workforce, roughly 8,000 people, in what Bloomberg called a “push for efficiency” and CNN called an AI-driven reorg. The cuts hit hardest inside Meta Superintelligence Labs, the AI division Zuckerberg spun up last year to compete with OpenAI and Anthropic.

Here’s what happened:

The cuts help fund Meta’s $72B AI capex commitment for 2026, up from $47B in 2025.
Roles being cut include middle managers, researchers on non-agentic projects, and some of the AI hires Zuckerberg personally poached from Scale AI (the Scale deal alone cost Meta $14.3B).
Superintelligence Labs survives but is narrowing its focus to the Mythos-tier projects that actually ship.

Why this matters: This is the first “AI layoff wave at an AI-leading company” moment of the year, and it arrives right as Anthropic’s Dario Amodei predicts white-collar cuts across the economy. If you work in tech, this is the quiet signal that the ZIRP-era org chart is finally getting restructured, not by the CFO but by the infrastructure build.

Our take:This is a Jeremy moment.

On Patrick O’Shaughnessy’s podcast this week, SemiAnalysis’ Dylan Patel told a story about “Jeremy,” an energy analyst on his team who spent $6,000 a day on Claude tokens for three weeks. Jeremy singlehandedly rebuilt a product that a 100-person data-services team at an incumbent had been working on for a decade.

Jeremy didn’t get laid off. He got leveraged.

The Meta cuts are the inverse. When one person with the right AI tools can do the work of 10-to-15 people, the person most at risk isn’t the one using the AI. It’s the one whose job description overlaps with what AI now does by itself. The manager who shuffles Jiras. The researcher whose specialty is already solved. The middle-of-the-pyramid specialist whose output is now a 20-minute Codex run.

The three-part rule Patel keeps coming back to: use more tokens, generate outsized value with them, and capture that value before the org chart gets rewritten around you. The alternative is being the team that got replaced by a Jeremy.

Wingman handles low-stakes tasks (social media, emails, research, ops) without monitoring, so you can focus on other priorities.

Works inside WhatsApp, Telegram, iMessage, Gmail, etc
Build one in 5 minutes (no coding or experience needed)
Requires confirmation for anything consequential or sensitive

It’s private, secure, and designed to act with your permission.

Matt Pocock of AI Hero argued this week that “specs-to-code” workflows (write a spec, let AI generate code, never look at it) are vibe-coding in disguise and produce worse code each iteration. His fix: make the AI interview you before it does anything.

The Grill Me skill (a few lines of instruction, now 13K+ stars on github.com/mattpocock/skills) flips the dynamic. Instead of you explaining the idea, Claude asks you 40-to-100 questions and walks every branch of the design tree. The goal is for you both to share what Frederick Brooks called a “design concept” (an actual mental model of the thing you’re building). Pocock says it beats Claude Code’s default plan mode for anything non-trivial.

Save this as a skill, or paste it at the top of your next planning chat:

Interview me relentlessly about every aspect of this plan until we reach a shared understanding. Walk down each branch of the design tree, resolving dependencies between decisions one by one. Ask about requirements, edge cases, user experience, data models, and failure modes. Do not write a plan document or code until I say we are aligned.

One favorite insight: this works even when you’re not coding. Swap “plan” for “essay,” “campaign,” or “strategy,” and Claude will drag a half-formed idea out of your head before it writes a single line.

Tencent open-sourced Hy3-preview, its first flagship 295B-parameter MoE model, led by ex-OpenAI researcher Yao Shunyu.
Anthropic told a court there is no “kill switch” for its AI deployed in classified Pentagon settings.
Cohere acquired and merged with Germany’s Aleph Alpha to form a “transatlantic AI powerhouse” for regulated industries.
The Trump DOJ joined Elon Musk’s xAI in its lawsuit against Colorado’s new AI discrimination law.
Samsung workers rallied 30,000 strong outside the Pyeongtaek chip hub to demand a greater share of AI-chip profits.
Anthropic ran Project Deal: Claude agents closed 186 real marketplace deals worth >$4K for 69 SF employees, autonomously handling listings, negotiation, and closing.
SpaceX is pivoting to enterprise AI as its next big business line, pegging the TAM at $22.7T to $26.5T per its S-1 filings.
Google faces new EU Digital Markets Act pressure to let AI rivals into Android voice activation and search.
Three new open models shipped on Tinker from Kimi and Qwen, with upgrades on long-horizon coding, multi-agent tasks, and tool use.

Want absolutely EVERYTHING that happened in AI this week? Click here!

Top Story of the Week:

OpenAI shipped GPT-5.5 exactly seven days after Anthropic shipped Opus 4.7, locking the two labs into a one-week cadence that’s quietly rewriting how enterprise teams pick models.

Top Stories:

Meta cut 8,000 jobs (10% of staff) in the first big “AI layoff at an AI-leading company” moment of the year.
Anthropic quietly hit a $1 trillion valuation on secondary markets, passing OpenAI on paper.
The White House accused China of “industrial-scale” AI theft via model distillation.
The Pentagon quietly vibe-coded 100,000+ AI agents on Gemini for unclassified work.