MLWhiz Weekly Recsys/ML/GenAI Newsletter # 7 - The week Karpathy Joined Anthropic
The week Andrej Karpathy picked his side, and everyone else picked theirs.
Hey, Rahul here! 👋 Each week, I publish long-form ML+AI posts covering ML, AI, and System design for MLwhiz. Paid subscribers also get how-to guides with full code walkthroughs. I publish occasional extra articles. If you’d like to become a paid subscriber, here’s a button for that:
🏆 Story of the Week: Karpathy Joins Anthropic
I’ve been following Karpathy’s career since his Stanford days. So when he joined Anthropic on Monday, I stopped what I was doing and read the announcement twice.
Think about his resume for a second. Co-founded OpenAI. Ran all of Tesla’s AI. Coined “vibe coding.” Taught more people about neural networks on YouTube than most universities.
He could have gone anywhere. He picked Anthropic.
His role will be leading a new team under pre-training lead Nick Joseph, focused on using Claude to accelerate pre-training research. Using the model to make the model better.
Here’s what I find interesting. Every major career move Karpathy has made has tracked where the field was shifting. OpenAI in 2015 was deep learning going mainstream. Tesla in 2017 was AI at an industrial scale. Back to OpenAI in 2023 was LLMs. Now Anthropic.
And he’s not the only one. In the last four months, the CTOs of Workday, You.com, and Instagram all left their companies to take IC research roles at Anthropic. Senior technical leaders voluntarily giving up executive positions to work on Claude.
When that many people make the same bet, it tells you something.
The recursive angle is the interesting technical bet. If Claude can meaningfully speed up the research that produces its next version, each generation makes the next one cheaper or better to build. That advantage compounds with each iteration. I think this is what Karpathy is actually excited about.
This hire caps a week where Anthropic also acquired Stainless (the SDK company powering OpenAI’s, Google’s, and Cloudflare’s API libraries), announced a $200M partnership with the Gates Foundation, and is reportedly in talks for a $30-50B raise at ~$950B valuation with an IPO as early as October.
The pace at which talent, infrastructure, and distribution are pulling toward Anthropic right now is something I haven’t seen from any AI lab before.
🤖 Models That Dropped This Week
Cerebras IPO (May 15): This one is not a model per se, but the AI chip market got its second public company, and the market went wild. Priced at $185, opened at $385, and closed day one at $311. $66B valuation, $5.5B raised. That 108% first-day pop tells you institutional investors think AI chip demand is structurally undersupplied and Nvidia alone can’t fill the gap.
Cerebras makes wafer-scale engines: entire silicon wafers as single chips, 4 trillion transistors on one piece of silicon. At $66B, it’s worth more than AMD’s AI-specific business.
Qwen 3.7 (May 19): Spotted on Qwen Chat, open weights expected any day.
🧠 Papers That Matter
SID-MLP: MLPs Beat Transformers for Generative Recommendation (Snap): I love when a paper makes you question an assumption everyone took for granted. Everyone uses beam search for token prediction, while the prediction difficulty in Semantic ID decoding drops sharply after the first token. So most of the Transformer’s capacity is wasted on easy subsequent tokens. Snap shows MLPs match or beat Transformer decoders at a fraction of the latency.
HyDRA: Model Routing Saves 54% in GitHub Copilot (Microsoft) : A lightweight ModernBERT encoder routes queries across a pool of LLMs based on four dimensions: reasoning, code generation, debugging, and tool-use. 54.1% cost savings at identical quality. Already deployed in Copilot’s VS Code Chat. We are in the cost saving zone and we are going to see many such papers going forward.
MARS: Agentic RecSys with Hierarchical Memory (Meta) : This one is for the RecSys folks. It treats recommendation as a partially observable planning problem with three-tier memory: episodic (what happened), semantic (what it means), and strategic (what to do about it). It’s a completely different way to formulate the problem. If your RecSys team is thinking about incorporating agents into the recommendation loop, start here.
📝 Some Good Reads
Pinterest: An Engineer’s Guide to Testing AI Agent Skills in Production : One of the first engineering blog posts from a major platform on how to systematically test AI agents in production recommendation systems. If you’re deploying agents at scale and making it up as you go, this gives you a real framework.
⚡Quick Hits
Amazon kills Rufus, launches Alexa for Shopping : agent that comparison-shops across the web and buys from competitors. The “Buy for Me” feature on competitor sites is Amazon betting that owning the agent layer is worth more than owning every transaction.
Anthropic reportedly in talks for $30-50B raise at ~$950B valuation : IPO as early as October 2026.
Anthropic and Gates Foundation announce $200M partnership : four-year commitment targeting polio, HPV, eclampsia, and African language AI.
Anthropic acquired Stainless : the SDK company that builds OpenAI’s, Google’s, and Cloudflare’s API libraries. Anthropic now owns the company that generates its competitors’ SDKs.
Meta reassigns 7,000 employees to AI-focused roles : the largest internal redeployment in Big Tech’s AI pivot. Total AI headcount redeployment exceeds 15,000 in 2026, on top of 8,000 layoffs.
Karpathy at Anthropic changes the signal for where serious AI automation research is heading. I work with small business owners on practical AI for business every week, and this foundational work upstream shapes what becomes usable six months from now. Do you think this accelerates or widens the gap between enterprise AI and small business AI?