Anthropic filed for IPO. Then asked everyone to stop. DeepSeek raised $7B anyway.
DeepSeek's $7B raise, SpaceX IPO at $135, OpenAI on AWS, ChatGPT Dreaming V3, and 20+ more stories
Must-Read
Anthropic Filed a Confidential Draft IPO Registration — Anthropic confidentially submitted a draft S-1 registration statement to the SEC for a proposed initial public offering. No pricing or share counts are set — the filing remains subject to regulatory review and market conditions. Coming one week after the $965B valuation and $47B revenue run rate, the timing is deliberate: Anthropic is moving toward a public listing while the numbers are at their most impressive.
Anthropic Oceanus Leaked — Next Mythos Model — A checkpoint of Anthropic’s next Mythos-class model, codenamed Oceanus, was made available to red teamers — typically a week before wider launch. The program was paused after an individual in the red team resold API access through a Chinese proxy. Oceanus is reportedly better than Mythos Preview. Whether the leak delays the launch is unknown, but the model is real and imminent.
TLDR; (30-sec)
Anthropic files confidential draft IPO registration with the SEC — no pricing set yet
Anthropic Oceanus leaks — Mythos successor made available to red teams, program paused after Chinese API proxy resale
Anthropic urges global pause in AI development, citing recursive self-improvement risk
DeepSeek raising $7B in first-ever funding round — founder committing $3B of his own money
SpaceX targets $135 per share for IPO — roadshow underway ahead of June 12 Nasdaq debut
OpenAI and Codex reach general availability on AWS — enterprise integration through Bedrock
ChatGPT Dreaming V3 ships — new memory synthesis for longer-horizon continuity
Microsoft launches 7 MAI models with Frontier Tuning + Mayo Clinic healthcare AI collaboration
NVIDIA Nemotron 3 Ultra released — 550B params, most intelligent open-weights US model
Meta keeps delaying Muse Spark release — no planned date for developer access
Anthropic Urges Global Pause in AI Development — Anthropic published a call for a global pause in AI development, claiming AI systems will soon be able to improve themselves without human intervention. The timing is loaded — this comes the same week they filed for an IPO and leaked their next frontier model. Critics are already framing it as competitive positioning: slow everyone else down while you go public. But the technical argument underneath — that recursive self-improvement creates uncontrollable feedback loops — is worth reading on its own.
DeepSeek Raising $7B in Maiden Fundraising — DeepSeek is set to raise roughly 50 billion yuan (~$7B) in its first-ever funding round. Founder Liang Wenfeng committed 20 billion yuan of his own money. Fewer than 10 investors. For a company that made the 75% price cut permanent last week, raising $7B signals this is a long-term play, not a cash-burning sprint.
SpaceX Targets $135 for IPO — SpaceX set its IPO target price at $135 per share, with a June 12 Nasdaq debut planned. This is relevant beyond aerospace: SpaceX is now Anthropic’s compute provider via the $45B deal from two weeks ago, and holds a $60B acquisition option on Cursor. SpaceX going public makes both of those relationships more visible and more scrutinized.
OpenAI and Codex Reach AWS — OpenAI announced general availability of its frontier models and Codex on AWS through Amazon Bedrock. Enterprises can now access OpenAI through their existing AWS security, governance, procurement, and billing workflows. This is OpenAI’s most significant cloud distribution deal — it puts their models inside the enterprise buying motion that AWS already owns.
ChatGPT Dreaming V3 — OpenAI introduced a new memory synthesis system for ChatGPT designed to improve freshness, continuity, and relevance over longer time horizons. Rolling out to Plus and Pro users in the US. The “Dreaming” architecture name suggests the system processes and consolidates memories in the background — similar to how sleep consolidates human memory.
Meta Keeps Delaying New AI Model Release — Meta still hasn’t set a firm date to release its newest AI models to developers, despite earlier indications of a June API launch. The Muse Spark model is reportedly competitive with OpenAI and Anthropic, but hasn’t been evaluated by outside firms. The repeated delays raise questions about how quickly Meta can monetize its massive AI investments.
Quick question — Anthropic dominated this week (IPO, Oceanus, global pause, recursive self-improvement). Is this the right framing, or should we pull the Anthropic stories into a single narrative block? Hit reply.
Build Tips & Engineering
Rethinking Search as Code Generation — Perplexity introduces Search as Code (SaC), letting AI models control the search process directly via an SDK. Instead of fixed search pipelines, models configure tailored search workflows per query. This is a fundamental rethink of how search systems are built — worth reading if you’re building anything that retrieves information.
LLM Hacking Benchmark — $1,500 Experiment — A developer built a vulnerable app and spent $1,500 testing if LLMs could exploit it. GPT-5.5 performed best (7/10 runs), DeepSeek-V4-Pro was runner-up (3/10), and Claude Sonnet 4.6 was the most expensive but only solved 2/10 — though five runs hit the budget limit. Real-world security eval data.
Defending Code Reference Harness — Anthropic open-sources a reference implementation for autonomous vulnerability discovery and remediation with Claude. Build custom vulnerability pipelines based on general best practices.
Preventing AI Inference Theft at Scale — Vercel explains how attackers resell stolen AI inference by exploiting exposed endpoints and why traditional rate limits fall short. Their approach uses BotID analysis to verify every request.
Wall Attention Mechanism — Tilde Research introduces an attention mechanism using persistent “wall” memory tokens that improve long-context reasoning. Designed to help models maintain coherence across very long sequences.
🧬 Model Releases
NVIDIA Nemotron 3 Ultra — 550B parameters (55B active), the most intelligent open-weights model from a US company. Scores 48 on the Artificial Analysis Intelligence Index. Available in NVFP4 quantization for inference performance.
Gemma 4 12B — Google’s latest Gemma model at 12B parameters. Designed for efficient deployment and fine-tuning.
MiniMax M3 — MiniMax launches M3 with MSA architecture, 1M-token context window, native multimodality, and agentic coding support. Weights and technical report coming within 10 days.
Nemotron 3.5 Content Safety — NVIDIA releases a unified model for multimodal, multilingual enterprise safety enforcement with auditable reasoning.
Tools & Product Updates
Codex New Capabilities + 6 Role-Specific Plugins — OpenAI released new Codex capabilities with six industry-specific plugins: data analytics, creative production, sales, product design, equity investing, and investment banking. The shift from general-purpose coding tool to role-specific workflows.
Microsoft 7 MAI Models + Mayo Clinic — Microsoft launched seven MAI models using Frontier Tuning — AI adapting to specific workflows through reinforcement learning. Plus a Mayo Clinic collaboration for healthcare AI, deploying within Mayo first then distributing through Azure Foundry.
Perplexity Hybrid Local-Cloud Inference — Perplexity unveiled a system at Computex 2026 that routes queries between on-device models for lightweight tasks and cloud models for complex reasoning.
Quick Bits
SpaceX Targets $135 for IPO — SpaceX targets $135/share for its IPO, with a June 12 Nasdaq debut planned — could be the largest tech listing of the year.
ChatGPT Dreaming V3 — OpenAI ships background memory synthesis for ChatGPT Plus and Pro users.
Deep Dives
When AI Builds Itself: Recursive Self-Improvement — Anthropic’s Institute published a deep examination of recursive self-improvement — AI systems that can autonomously design and develop their own successors. This is the technical foundation for Anthropic’s global pause call. Whether you agree with the policy conclusion or not, the technical argument deserves a careful read.
GitHub’s Plan for Agents — A 90-minute deep dive with GitHub COO Kyle Daigle on how GitHub is positioning itself for the coding agent era. Covers platform strategy, how agents change developer workflows, and what GitHub looks like when more code is written by AI than humans.
That’s a wrap for this week.
If this was useful, the best thing you can do is share it with someone who’d get value from it too. Forward this email, drop it in your team’s Slack, or just send it to that one friend who’s always asking “what happened in AI this week?” Every share helps us keep this going.
Hit reply if I missed something — I read everything. See you next week.
Stay curious.