Apr 3, 2025
Welcome to Edition 7 of Fine-Tuned by Genloop – your go-to guide for the latest in LLM customization. In this edition, we're covering major research breakthroughs from Google to DeepSeek, exploring the viral topic of the week - GPT-4o's new image generation capabilities.
Plus, don't miss our expanded Research Corner featuring both our weekly top papers roundup and deep dives into Gemma 3 and DAPO's fully open-source reinforcement learning system.
We're thrilled by the amazing response to our first Research Jam on Meta's SWE-RL paper! 🎉
Missed it? No worries - catch the recording and join us for Research Jam #2 on "Transformers without Normalization" coming up on April 10th. (More details in the Genloop updates below!)
🌟 AI Industry Highlights
OpenAI Integrates Advanced Image Generation Directly into GPT-4o
We are sure, by now you must have heard about and viewed Ghibli-style images a hundred times if not more. So, let's highlight the elephant in the room - OpenAI introduced an impressive new image generation capability directly into GPT-4o, moving beyond their previous DALL-E models.
Key highlights:
Enhanced Functionality: The new system excels at creating practical visuals like diagrams, menus, and UI mockups, with remarkable accuracy when rendering text and following detailed instructions
Creative Explosion: Users have quickly pushed the boundaries by creating Studio Ghibli-style Lord of the Rings videos, perfect wine glasses (previously challenging for AI), South Park character transformations, transparent background assets, and even anatomy diagrams of animated characters
Image is part of the Ghibli style remake of LOTR
Moderation Changes: OpenAI has notably relaxed some content policies, now allowing generation of public figures and certain previously restricted content while maintaining safeguards.
This release represents a significant leap forward in making image generation a practical tool for visual communication, though questions remain about its potential impact on creative industries and content moderation standards.
Google Unveils Gemini 2.5 - Their most intelligent AI model
Google DeepMind has introduced Gemini 2.5, describing it as their most intelligent AI model to date. The first release in this series, Gemini 2.5 Pro Experimental, claims the top position on the LMArena leaderboard by a significant margin.
Key findings:
Thinking Approach: Designed as a "thinking model," Gemini 2.5 reasons through problems before responding, resulting in improved accuracy and performance
Benchmark Leadership: Achieves state-of-the-art results across reasoning, mathematics, and science benchmarks, including an 18.8% score on Humanity's Last Exam
Coding Improvements: Shows substantial improvements in code generation, with the ability to create complex applications like games from simple prompts
Gemini 2.5 Pro maintains a 1 million token context window (with 2 million coming soon) and is currently available in Google AI Studio and the Gemini app for Advanced users. It will be available on Vertex AI in the coming weeks with pricing details to be announced soon.
DeepSeek Releases V3-0324 with Major Performance Improvements
DeepSeek has released an updated version of their V3 model with significant enhancements across key capabilities. The new DeepSeek-V3-0324 model shows substantial benchmark improvements while maintaining the same open-source MIT license.
Key findings:
Enhanced Reasoning: Dramatic benchmark improvements, including +19.8 points on AIME (39.6 → 59.4) and +9.3 on GPQA (59.1 → 68.4)
Improved Coding: Better front-end web development with more executable code and visually appealing web pages, with LiveCodeBench scores increasing by 10 points
Chinese Content Optimization: Enhanced writing quality with better multi-turn interactive features and improved translation capabilities
DeepSeek-V3-0324 maintains the same model structure as its predecessor, making it easy to run locally with support for function calling, JSON output, and FIM completion. With performance approaching or exceeding larger models while being significantly faster and more cost-effective, this release represents a major advancement in open-source LLM development.
✨ Genloop Updates: Research Jam Recap & Our Next Deep Dive into Transformers
Thanks for the warm response to Genloop Research Jam #1! It was fantastic to deep dive into Meta's SWE-RL paper together.
In our session, we explored two key opportunities: 👉 Using better reward functions beyond simple sequence matching on patches 👉 Enhancing performance through improved agent scaffolding
Missed the session? No worries, we've got you covered! Check out the recording below to get up to speed:
We're excited to announce Research Jam #2 happening on April 10, where we'll dive into Transformers without Normalization - the top research paper on LLM Research Hub for the week of March 3rd, 2025.
Spots are limited, so register today to secure your place in what promises to be another insightful discussion!
Register for Research Jam: https://lu.ma/5apxsz6o
📚 Featured Blog Posts
We stumbled upon a fascinating read that showcases how the AI landscape is evolving:
11x: The AI Sales Startup Using Questionable Tactics to Appear Successful
A TechCrunch investigation reveals how a16z and Benchmark-backed AI sales automation startup 11x has been misrepresenting its customer base and financial health while raising significant venture capital. The heavily funded startup, which creates AI bots for sales outreach, has been displaying logos of companies like ZoomInfo and Airtable on its website and marketing materials despite these companies confirming they were never actual customers.
Cursor's Decline: User's Frustration with the AI Coding Assistant
A long-time Cursor user shares the common disappointment with the popular AI coding tool, citing deteriorating performance, reduced context windows, and questionable pricing strategies. The user points to the introduction of Claude 3.7 Sonnet as "the beginning of the end," with subsequent updates making the tool progressively worse while pushing users toward more expensive options.
🔬 Research Corner
Check out the Top 3 papers of the Week [March 24-28 2025] suggested by Genloop's LLM Research Hub - where AI-powered filtering meets expert human curation to deliver the most impactful research across multiple sources.
Checkout the post https://www.linkedin.com/posts/genloop-ai_top-3-papers-of-the-week-march-24-28-2025-activity-7312624774478737408-5Qsb.
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:
Gemma 3 Technical Report
This week's research spotlight features Google's "Gemma 3 Technical Report," detailing their latest family of open-weight models ranging from 1 to 27 billion parameters.
Key highlights:
Multimodal Capabilities: Integrates image understanding through a tailored SigLIP vision encoder using a "Pan and Scan" approach for flexible handling of varying image resolutions
Efficient Long-Context Processing: Combines local and global attention layers to process up to 128K tokens without the typical memory explosion issues
Superior Performance: The refined post-training recipe delivers exceptional results, with the 4B-parameter model performing on par with its 27B Gemma 2 predecessor
Our internal testing shows Gemma 3 dramatically outperforming GPT-4o in document processing (90% vs 10% accuracy), showcasing the rapid advancement of open-weight models.
Read our TuesdayPaperThoughts analysis
DAPO - Fully Open-Source LLM Reinforcement Learning System
ByteDance's "DAPO - Fully Open-Source LLM Reinforcement Learning System" represents a landmark contribution to the AI community.
Key highlights:
True Open-Source Approach: Unlike proprietary systems that conceal training details, DAPO provides a completely transparent framework including code, algorithm, and dataset
Superior Mathematical Reasoning: Using Qwen2.5-32B, the system achieves 50 points on AIME 2024, outperforming DeepSeek's R1 (47 points) with 50% fewer training steps
Technical Innovations: Tackles critical RL training challenges through Clip-Higher for diversity, token-level policy gradient loss for long-chain reasoning, and overlong reward shaping for stability
This breakthrough democratizes advanced LLM training techniques, empowering the broader community to build powerful reasoning models for their applications.
Read our TuesdayPaperThoughts analysis
Looking Forward
We're witnessing remarkable progress across the LLM landscape this week - from OpenAI's groundbreaking image generation capabilities to Google's Gemini 2.5 "thinking model" and DeepSeek's impressively efficient open-source alternatives. The boundaries between text, image, and reasoning continue to blur as these models become increasingly powerful and accessible.
If you'd like to dive deeper into such advancements, join our Research Jam #2 on April 10. Register here to secure your spot before they fill up!
About Genloop
Genloop delivers customized LLMs that provide unmatched cost, control, simplicity, and performance for production enterprise applications. Please visit genloop.ai, catch us on Linkedin, or email founder@genloop.ai for more details.