In partnership with
> SPOTLIGHT
TODAY’s 3 MOST IMPORTANT
Anthropic sued the Pentagon after being labeled a "supply chain risk" when negotiations over Claude's military deployment terms fell apart. OpenAI immediately signed a similar deal with the DoD. While the two labs fought publicly, Google quietly locked in a contract to deploy AI agents across the Pentagon's 3-million-person workforce. Claude paid subscribers grew 200% YoY, Gemini 258%, but neither lab is profitable. Alphabet sits on $400B in annual revenue and can afford to wait.
Anthropic launched the Anthropic Institute today, led by co-founder Jack Clark, consolidating three research teams covering red teaming, societal impacts, and economic research. At the same time, the company is opening a DC office and tripling its policy team, with focus on chip export controls and a federal AI regulatory framework. For the first time, Anthropic is openly betting on shaping the rules, not just playing by them.
Karpathy (pre-Director of Tesla AI, founding team OpenAI) ran an autoresearch agent for 2 days across 700 experiments to optimize NanoChat. The agent found 20 real improvements, all additive, all transferable to larger models, cutting "Time to GPT-2" from 2.02 to 1.80 hours. It caught mistakes Karpathy had missed after years of manual tuning: attention too diffuse, AdamW betas misconfigured, weight decay unscheduled. This isn't AI writing code. It's AI optimizing the process of training AI. Karpathy calls it the "final boss battle" for frontier labs. Round 1 just landed better than expected.
> SIGNAL HEADLINES
Capture the shift
The legal AI startup closed a $550M Series D, now serving 800 law firms and growing from 40 to 400 employees in a year. Rival Harvey is already at $8B and reportedly raising at $11B.
→ Two AI legaltech startups are racing valuations in parallel, while Anthropic's new legal plugin for Claude already sent traditional legal software stocks sliding.
Aaru runs thousands of AI agents simulating consumer behavior for McDonald's, A24, Bayer, and EY. EY used it to fully replace some global surveys. The three founders dropped out of Dartmouth and Harvard within weeks of arriving.
→ When AI agents simulate human responses more accurately than actual humans, a multi-billion dollar market research industry starts asking hard questions about its own existence.
Nvidia is backing the former OpenAI CTO's new frontier AI lab in a multi-year partnership, with plans to deploy at least 1 gigawatt of GPU capacity for training and inference.
→ Jensen Huang just pulled back from OpenAI and Anthropic last week. Now Nvidia is placing a direct bet on Murati. That's a deliberate signal about where Nvidia sees the next frontier lab standing.
> PRESENTED BY THE CODE
Tech moves fast, but you're still playing catch-up?
That's exactly why 200K+ engineers working at Google, Meta, and Apple read The Code twice a week.
Here's what you get:
Curated tech news that shapes your career - Filtered from thousands of sources so you know what's coming 6 months early.
Practical resources you can use immediately - Real tutorials and tools that solve actual engineering problems.
Research papers and insights decoded - We break down complex tech so you understand what matters.
All delivered twice a week in just 2 short emails.
> TRY THIS TODAY
Turns any website into clean data for AI
Today I came across a pretty interesting tool: Firecrawl.
You give it a URL, and it crawls the entire website and returns clean, structured data (markdown or JSON) that AI can actually use.
It solves a common problem: when you scrape websites, you usually end up with messy HTML full of ads, navigation menus, and cookie banners.
Firecrawl cleans all of that and turns webpages into LLM-ready data.
Useful for things like:
Turning documentation sites into a RAG knowledge base
Giving AI agents the ability to read websites
Crawling sites for research or training data
Firecrawl is open source, offers SDKs for Python, Node, Go, and Rust, and integrates with tools like LangChain, LlamaIndex, CrewAI, and Dify.
> WORTH READING
Analysis & Thesis
This piece breaks down the growing debate between open weights and open training in AI. Many companies claim their models are open because they release the weights, but they often keep the training data and pipeline private.
The author argues that true transparency and reproducibility in AI research require going beyond open weights to fully open training, including datasets, methods, and infrastructure.
Nvidia and Hugging Face are collaborating on an initiative to create high-quality open datasets for AI training. As models grow larger, the quality of training data becomes increasingly important.
The project aims to clean, standardize, and release open datasets so researchers and developers can build stronger AI models without relying solely on proprietary data from large tech companies.
Found this useful?
👉 Forward it to someone trying to keep up with AI.
👉 Read online: techzip.beehiiv.com
Tech.zip Newsletter
| Zipping what truly matters in the AI era.
