Presented by
Good morning, AI enthusiasts. The company that has spent years arguing that AI development should slow down just filed to speed up its path to Wall Street. Anthropic confidentially submitted a draft S-1 to the SEC, beginning the formal process that could make it the most valuable AI lab ever to go public.
The Economist asked this week whether the stock market can even swallow a company like Anthropic. It's a fair question. Once quarterly earnings pressure and fiduciary duty to shareholders are in the picture, the space for slow, expensive safety research gets a lot smaller. Or maybe not. We're about to find out.
In today's recap:
Anthropic submits confidential IPO filing to the SEC
Ex-DeepMind team raises $50M to build discovery AI
NYC startup trades free cleaning for robot training data
Build search-powered agents with Perplexity Search as Code
4 new AI tools, prompts, and more
ANTHROPIC
Anthropic files for IPO with the SEC
Recaply: Anthropic just submitted a confidential draft S-1 registration statement to the U.S. Securities and Exchange Commission, giving the company the option to pursue an initial public offering after the SEC completes its review, with no share count or price set yet.
Key details:
The confidential S-1 filing lets Anthropic test regulatory and market waters without committing to a public offering; the company can choose to proceed or withdraw once the SEC review is complete, and no announcement constitutes an offer to sell securities under Rule 135 of the Securities Act.
Anthropic's most recent Series H funding round raised $65B at a $965B post-money valuation, making it one of the most valuable pre-IPO AI companies ever to reach the SEC filing stage.
The Economist published a companion piece asking whether the stock market can absorb Anthropic, SpaceX, and OpenAI simultaneously, gathering 299 points and 518 comments on Hacker News on the same day as the filing news.
The proposed IPO's timing and final terms depend on SEC review completion and market conditions, with the company stating no specific timeline.
Why it matters: There's been a consistent argument that Anthropic's safety mission and the commercial pressures of being a public company are incompatible. Anthropic doesn't agree, and $965B in implied valuation suggests investors don't either. What's interesting is that the company has structured itself as a Public Benefit Corporation from the beginning, which legally requires it to balance shareholder returns against its stated mission. Whether that structure holds up under earnings calls is the experiment we're all about to watch.
PRESENTED BY DEEL
Your first HR system, implemented right
Rolling out your first HR tool? Get a step-by-step guide to avoid common mistakes, drive adoption, and build a scalable HR foundation.
AI RESEARCH
Ex-DeepMind team raises $50M to build discovery AI agents
Inherent
Recaply: Inherent just closed a $50M seed round to build a London-based research lab focused on AI agents that discover new scientific knowledge, with backing from Index Ventures, Radical, and NVIDIA's venture capital arm NVentures.
Key details:
Inherent is built around "recursive self-improvement at the collective level," where AI agents and the research institution itself continuously improve together, with the lab designed to embrace what it calls AI-driven aesthetic discovery in the sciences.
The $50M seed was led by Index Ventures and Radical VC, with participation from NVentures (NVIDIA's VC arm), Metaplanet, Macroscopic, Mythos Ventures, Charlie Songhurst, Dwarkesh Patel, Thom Wolf, and Max Jaderberg, and is advised by Matthew Clifford of Entrepreneur First.
Inherent is structured as a Public Benefit Corporation headquartered in London, the same legal form as Anthropic, signaling a deliberate alignment between commercial goals and scientific impact from day one.
The lab is currently hiring for open roles at inherentlabs.ai, with the full research vision published alongside the launch announcement.
Why it matters: NVIDIA's VC arm co-investing is the single most meaningful signal here. NVentures doesn't back labs without a strong thesis on future compute demand. If Inherent's "machine-driven scientific inquiry" thesis proves out, the return profile looks less like a traditional AI startup and more like scientific infrastructure: recurring demand for large-scale compute across every domain the agents touch, from biology to materials science to physics.
GUIDES
Build search-powered AI agents with Perplexity Search as Code
Perplexity
Recaply: In this tutorial, you will learn how to apply Perplexity's Search as Code (SaC) architecture to give your AI agent direct control over retrieval, letting it compose custom search pipelines for each task instead of calling a fixed monolithic endpoint.
Step-by-step:
Get access to the Perplexity API at perplexity.ai and install the Agentic Search SDK in your project. Set your API key as an environment variable. The SDK exposes individual search building blocks, including retrieval, ranking, filtering, fanout, deduplication, and rendering, as callable Python functions rather than a single query endpoint.
Define the search context for your agent's task. Instead of sending a single string query, pass the full task description so the model can reason about what retrieval strategy fits. The key shift from traditional RAG is that the model decides how to search, not just what to search for.
Let the model generate Python code that composes the search primitives for the specific task. For simple lookups, this might be a handful of high-level calls. For complex research tasks, the code can include conditional execution, parallel fanouts across multiple query variants, and intermediate deduplication before the results reach model context.
Execute the generated code inside a secure sandbox. SaC runs each script in an isolated environment, so the model can issue hundreds of retrieval calls without polluting its own context with noisy intermediate state. Only the final, processed result set is returned to the model for reasoning.
Consume the compact result set in your agent's context and iterate. Because the pipeline was assembled specifically for the query, the retrieved content is more precise than what a fixed endpoint returns. Refine by asking the model to adjust the pipeline, prompting: "The ranking missed recent sources, add a recency filter to the fanout step."
Pro tip: For research workflows that need multiple angles covered in parallel, use SaC's fanout primitive to run query variants simultaneously rather than sequentially through model turns. This matches how Perplexity Computer handles tasks internally, cutting latency dramatically on information-dense requests.
TOGETHER WITH HUBSPOT
Want to get the most out of ChatGPT?
ChatGPT is a superpower if you know how to use it correctly.
Discover how HubSpot's guide to AI can elevate both your productivity and creativity to get more things done.
Learn to automate tasks, enhance decision-making, and foster innovation with the power of AI.
MICROAGI
NYC startup trades free apartment cleaning for robot training data
Shift
Recaply: MicroAGI just launched Shift in New York City, a service offering free professional home cleaning in exchange for first-person video footage recorded by cleaners during each job, with the footage used to train household robots.
Key details:
Vetted Shift operators wear recording devices during jobs; the footage is anonymized before processing and sold to robotics companies and AI labs as training data for domestic AI systems, with the data value covering the full cost of the cleaning service.
The company says it already works with 10,000+ businesses and households across 15+ countries, and the NYC launch is the starting point for a planned expansion to handymen, repairs, and errands globally.
The launch announcement asks users to comment "shift" for early access, and the startup's model is explicit: "You get a spotless apartment. We get training data. Everyone wins," according to the shiftapp.nyc homepage.
Shift is live for free cleanings in NYC today, with additional services and cities coming soon but no specific timeline given.
Why it matters: This business model works only if first-person home video is genuinely worth more than the cleaning cost. For household robot training, egocentric video of real domestic environments is among the rarest and most valuable data types in the world. The privacy debate will be persistent, and "anything personal is anonymized" without a clear audit trail will not satisfy everyone. But if Shift navigates that, it's a template that scales to almost any service where human physical labor generates training signal.
TOOLS
Trending AI Tools
🤖 Qwen3.7-Plus - Alibaba's latest proprietary reasoning model for agentic tasks, now available on the API
⚙️ Mellum2 - JetBrains' open-source 12B mixture-of-experts model (Apache 2.0) built for routing, Q&A, and sub-agent workflows in production systems
🧠 Koji - Brilliant's AI tutor for kids that coaches through math and coding without giving direct answers, by CEO Sue Khim
⚙️ OpenLogi - Open-source Logitech device manager for macOS (2.4k stars) for adjusting DPI, button mapping, and device settings without official software
NEWS
What Matters in AI Right Now?
OpenAI opened its frontier models and Codex on Amazon Bedrock, giving AWS customers a path to deploy OpenAI capabilities within their existing security, procurement, and governance workflows. Codex, used by more than 5 million developers weekly, is now available in both Commercial and GovCloud regions.
Perplexity introduced Search as Code (SaC), a new architecture that exposes its search stack as programmable SDK primitives rather than a fixed endpoint, letting models generate Python code to assemble bespoke retrieval pipelines for each query. Inside Perplexity Computer, single tasks already invoke hundreds of retrieval operations within minutes using this approach.
Google shipped Gmail, Drive, and Sheets integrations inside AI Studio Build, letting developers connect live workspace data to vibe-coded apps without leaving the platform. Full public sharing for built apps is coming soon; early access is live at ai.studio/build.
Alphabet announced an $80B equity raise including a $10B private placement to Berkshire Hathaway and a $30B concurrent public offering, as it targets $180-190B in annual capital spending on AI infrastructure. Berkshire more than tripled its Alphabet stake just last month before this transaction.
NVIDIA launched its open-source Agent Toolkit at GTC Taipei, anchored by Nemotron 3 Ultra, a 550B mixture-of-experts model delivering 5x faster inference and up to 30% lower cost for long-running agentic workloads. Cadence, Siemens, CrowdStrike, and Palantir are among the first enterprise partners.p
An OpenAI model disproved the Erdos unit distance conjecture, a geometry problem first posed in 1946, producing an infinite family of constructions that achieve a polynomial improvement over the prevailing bound. The proof was verified by a group of external mathematicians who also wrote a companion paper on its significance.
Intel is targeting a new AI data center chip by year end, positioning itself to compete in inference hardware as Nvidia and AMD consolidate their positions across the data center market.
🧡 Enjoyed this issue?
🤝 Recommend our newsletter or leave a feedback.
How'd you like today's newsletter?
Cheers, Jason
