xAI Hands Over All Its Compute Resources to Anthropic — A Pivot to "Neocloud"?

xAI and Anthropic announced a major partnership this week: Anthropic will take over all of the compute resources at xAI's Memphis, Tennessee data center, known as "Colossus 1."

As discussed on TechCrunch's Equity podcast, the deal raises big questions about xAI's business model. xAI was originally positioned as a frontier AI lab, but with the Grok models struggling to make much of an impact in the market, some argue the company now effectively resembles a "neocloud" provider that rents out GPUs.

Ahead of SpaceX's IPO, reports also suggest xAI could be dissolved as an independent organization. Some in the tech industry see the move as "tidying up before the IPO."

AI's "Memorization Problem" — LLMs Had Entire Books Memorized

Research from Stanford and Yale, reported by The Atlantic, has once again demonstrated that major LLMs memorize and can reproduce large amounts of copyrighted works from their training data.

According to the study, Claude could reproduce text close to the full text of "Harry Potter," "The Great Gatsby," "1984," and "Frankenstein." GPT-4.1 produced outputs strikingly similar to the original passages, even while paraphrasing semantically. Meta's Llama 3.1-70B was also confirmed to be capable of reproducing nearly the entire text of "Harry Potter."

AI companies have long insisted that "models don't store copies of their training data," but this study contradicts that claim. Experts have begun using the phrase "lossy compression," and some now see AI models as essentially giant databases.

The implications for copyright infringement lawsuits are significant — in the worst case, models could be ordered to undergo retraining.

LLM Agents Beat Pokémon Using Genetic Algorithms

"PokeLoop," a project that drew attention on Hacker News, is an ambitious effort to evolve a population of LLM agents with genetic algorithms and have them play Pokémon Crystal.

Eight agents using Claude Sonnet 4.6 as their policy play simultaneously on parallel sandboxes, with each generation going through tournament selection plus LLM-driven prompt crossover and mutation. After eight generations of evolution, the agents achieved a series of milestones: "walk → talk → get the first Pokémon → travel across the map → catch Pokémon → challenge a gym → earn a badge."

Rather than fine-tuning weights, the method is "evolution of text" — an approach that treats system prompts as genes and rewrites them in natural language. It's a fascinating new direction for reinforcement learning that leverages LLMs.

Design Lessons for Escrow Services Built for AI Agents

A StreetAI blog post covered the design challenges of escrow payments in marketplaces where AI agents act as sellers.

Applying human-oriented escrow (the Fiverr or Upwork model) directly to AI agents causes three serious problems. First, agents hallucinate deliveries — believing they've delivered the work when no file was actually attached. Second, if a buyer lies by claiming "I've already paid," the agent believes them. Third, during dispute resolution, an agent may propose both a defense and a refund at the same time, leaving the transaction in an inconsistent state.

The adopted solution: implement precondition checks on delivery tool calls, treat only platform-side system messages as the authoritative signal for payment confirmation, and during disputes, force agents into a binary choice between "defend" or "agree to refund." It's a case that highlights how critical system design becomes as AI agents participate in economic activity.

Qwen3.6-27B Trends on Hugging Face

Qwen3.6-27B is drawing attention as a trending model on Hugging Face. It's a 27B-parameter dense model featuring faster inference through Multi-Token Prediction (MTP), a context length of up to 262K tokens (up to 1M tokens with YaRN extension), and strengthened agentic coding capabilities.

MTP support is available in inference frameworks such as SGLang, vLLM, and KTransformers, and GGUF quantized versions from Unsloth have also been released. With particular improvements to Agentic Coding and Thinking Preservation (retaining reasoning history), the model is increasingly positioned as a practical choice for developers.

In Other News

The return of the Unix workstation**: An article titled "The Return of the Unix Workstation (Now with AI)" trended on Hacker News, prompting renewed appreciation for the importance of local-first AI development environments

What local LLM speed actually feels like**: A tool for building an intuitive feel for what tokens-per-second numbers actually mean was shared in the Reddit LocalLLaMA community

Comparing AI coding agents**: Experiences migrating from Deepseek TUI to Cline, Claude Code, OpenCode, and others are being discussed

---

**Sources:**

[We're feeling cynical about xAI's big deal with Anthropic](https://techcrunch.com/2026/05/10/were-feeling-cynical-about-xais-big-deal-with-anthropic/) — TechCrunch

[AI's Memorization Crisis](https://www.theatlantic.com/technology/2026/01/ai-memorization-research/685552/) — The Atlantic

[AI Learns to Play Pokémon GO on AI Sandboxes](https://zozo123.github.io/pokeloop/) — PokeLoop

[Why payment escrow for AI agents needed a different design](https://streetai.org/blog/escrow-for-ai-agents.html) — StreetAI

[Qwen3.6-27B](https://huggingface.co/Qwen/Qwen3.6-27B) — Hugging Face

[The Return of the Unix Workstation (Now with AI)](https://jasoneckert.github.io/myblog/return-of-the-unix-workstation/) — Hacker News

[Getting a feel for how fast X tokens/second really is](https://www.reddit.com/r/LocalLLaMA/comments/1t99upf/) — Reddit LocalLLaMA