Jul 3, 2025
Dear Readers,
Welcome to the 13th edition of Fine-Tuned by Genloop! This edition highlights key observations from Anthropic's mixed legal victory and Cloudflare's new pay-per-crawl system, revealing that data is AI's key competitive moat.
Meanwhile, Meta is rolling out the red carpet for key talent with compensation packages reaching as high as $100M. We hope this helps Meta reclaim ground in the LLM rankings.
We're also proud to share that Genloop is now part of NetApp's 14th Excellerator program! More details below.
So, let's dive in!
🌟 AI Industry Highlights
Anthropic Wins "Fair Use" Ruling for AI Training but Faces Trial for Piracy
A federal judge ruled that training AI models on copyrighted books constitutes fair use, but Anthropic must face trial for using pirated copies from sites like Library Genesis.
Key highlights:
Training Deemed Legal: Judge found AI training on copyrighted works is "transformative" and constitutes fair use, potentially setting precedent for the AI industry
Piracy Exception: Court ruled using pirated copies to build training libraries is not protected, with Anthropic facing potential damages of billions for "stealing" books first
Mixed Precedent: The ruling protects legitimate AI training while establishing that companies must legally source their training data
This ruling may compel AI companies to secure costly licensing agreements instead of relying on freely available pirated material. It also underscores the importance of implementing robust data risk management and privacy controls for enterprises.
Cloudflare Launches Pay-Per-Crawl for AI Content Access
Cloudflare introduced Pay-Per-Crawl, allowing content creators to charge AI crawlers for accessing their content through a programmatic payment system. This highlights how data control is becoming the the ultimate competitive moat.
Key highlights:
Third Option: Provides alternative to blocking AI crawlers entirely or allowing free access, enabling content monetization at internet scale while maintaining data sovereignty
HTTP 402 Integration: Uses forgotten web standard with payment headers, allowing crawlers to either pay configured prices or receive payment-required responses
Publisher Control: Content owners can set flat per-request pricing and choose to allow, charge, or block specific crawlers with Cloudflare handling billing
As data becomes the primary competitive advantage in the AI era, enterprises must strategically control access to their valuable content.
Meta Hires Key OpenAI Researchers for AI Superintelligence Push
Meta has hired four top OpenAI researchers as part of its aggressive talent acquisition for its new AI superintelligence unit, reportedly offering compensation packages up to $100 million. The race for AI supermacy is heating up and we are rooting for more open-weight models to come out of it.
✨ Genloop Updates
Genloop is proud to be part of NetApp’s 14th Excellerator cohort! 🚀
We’ve always believed that GenAI in production must understand and learn from your data to deliver real ROI. General intelligence isn’t enough — true impact comes from domain intelligence. That belief drives our work on use cases like Data-to-Insight, where we help enterprises build business-aware systems that deliver consistent, reliable insights from structured data — not just guesses.
That’s why NetApp is a natural partner in our journey. As a leader in intelligent data infrastructure, their focus on data and security perfectly aligns with our mission: helping enterprises build personalized GenAI systems that learn from their data and deliver insights with precision.
Excited to collaborate with NetApp’s leadership and our fellow cohort members.
Get a sneak peek into what Genloop is building from our founder
Want to know what drives Genloop and where we're heading? Our founder Ayush Gupta recently sat down with Pegasus Accelerator to share his vision for the future of enterprise AI and how we're building personalized GenAI systems that truly understand business context.
Watch here to hear from Ayush Gupta
Research Jam #7: Exploring The Illusion of Thinking
Last week's Research Jam #7 delivered another engaging session as we dove deep into the "The illusion of thinking" paper by Apple.
Watch the full recording here:
Join Research Jam #8: Exploring AceReason-Nemotron 1.1
Research Jam #8 is happening on July 10, where we'll dive into AceReason-Nemotron 1.1 - the top research paper on LLM Research Hub for the week [June 16 - June 20, 2025].
Spots are limited, so register today to secure your place!
📚 Featured Content
Watch: Andrej Karpathy on the Future of Software
Former Tesla AI Director and OpenAI researcher Andrej Karpathy breaks down how we're entering "Software 3.0" - where programming computers in plain English is becoming reality. From comparing current LLMs to 1960s operating systems (hinting we're at the dawn of a computing revolution) to introducing "vibe coding" that makes everyone a programmer, Karpathy's insights reveal why we're witnessing the most significant shift in software development since the internet.
🔬 Research Corner
Check out the top papers of the week on LLM Research Hub. Each week, our AI agents scour the internet for the best research papers, evaluate their relevance, and our experts carefully curate the top selections. Don't forget to follow us to stay up to date with our weekly research curation!
Now, let's deep dive into the top research from the last two weeks:
Chain-of-Experts: Unlocking Sequential Communication in MoE Models
Researchers propose Chain-of-Experts (CoE), a new architecture that enables sequential expert communication within transformer layers, challenging the assumption that MoE experts must operate independently in parallel.
Key findings:
Sequential Processing: CoE processes tokens iteratively across expert chains with dedicated routers at each step, allowing tokens to re-evaluate and select different experts during iterations
Superior Performance: Achieves 6%+ reduced validation loss on math reasoning compared to standard MoE while using 17.6-42% lower memory
New Scaling Dimension: Introduces expert iteration depth as an alternative scaling axis, complementing traditional width and depth approaches
The research suggests expert communication patterns may provide more efficient pathways for scaling language model capacity beyond conventional methods.
Read Our TuesdayPaperThoughts analysis
AceReason-Nemotron 1.1: Optimizing SFT and RL for Mathematical Reasoning
NVIDIA researchers investigate systematic integration of supervised fine-tuning and reinforcement learning to develop mathematical and coding reasoning capabilities in language models.
Key findings:
SFT Scaling Impact: Scaling SFT datasets from 36K to 2.2M samples substantially improves reasoning performance through both prompt scaling and response scaling strategies
RL Training Dynamics: Stronger SFT models yield better RL results, with optimal temperature-adjusted entropy around 0.3 providing effective exploration-exploitation balance
Cross-Domain Benefits: Math-only RL training creates foundation for robust learning, with final 7B model achieving 63.2% on AIME25 and 52.8% on LiveCodeBench v5
The research provides empirical insights into optimizing SFT and RL integration for developing high-performance reasoning models.
Read Our TuesdayPaperThoughts analysis
Looking Forward
For centuries, information was the moat because those who controlled it held power. The internet erased that advantage by giving everyone access to knowledge. Intelligence, the ability to process and use that knowledge, then became the new moat. Now artificial intelligence is spreading that capability to all, so even intelligence is losing its exclusivity. The next competitive edge will lie in how effectively we apply this shared intelligence to solve complex problems, build efficient systems, and push deeper into the unknown.
As we've seen this week— Cloudflare's pay-per-crawl system and Anthropic's legal battles over training data—control over high-quality, domain-specific data is becoming the ultimate differentiator.
GenAI allows enterprises to get beyond generic intelligence while maintaining their data moats. We have been working with many enterprises on data-to-insight solutions, helping functional owners get personalized insights over their data without dependencies and delays. And we have seen it deliver tremendous edge. Those who own and tame their AI to their advantage will be the leaders of tomorrow.
About Genloop
Genloop empowers enterprises to deploy GenAI in production with agents that understand business know-how and processes. We help companies build personalized LLMs that deliver superior performance, control, and simplicity—ideal for use cases like chatting with enterprise databases and transforming click-based workflows into conversational interfaces. Visit genloop.ai, follow us on LinkedIn, or reach out at founder@genloop.ai to learn more.
