Open AI Models Lag Behind Closed Ones by Months, Study Finds
In brief
- A new study reveals that open-source AI models, which anyone can download and use, are falling behind closed models-those accessible only through APIs-by about 4 to 10 months.
- The research analyzed data from 17 benchmarks, both public and private, to gauge the gap in performance between these two types of models.
- On private benchmarks, where data isn’t publicly available, open models lag by 8-10 months, while on public ones, the gap is narrower at 4-6 months.
- The study highlights a concerning trend: since January 2025, this gap has been widening rather than shrinking.
- This suggests that closed models are improving faster or being optimized more effectively than their open-source counterparts.
- The findings raise questions about the future of AI development-will the gap continue to grow, or will open models catch up?
- Researchers emphasize that while public benchmarks underestimate the true performance gap, they still offer valuable insights into model capabilities.
- As AI technology evolves, this study underscores the importance of transparency and access in fostering innovation.
- While closed models may hold advantages in fine-tuning and optimization, the progress of open-source models could depend on how developers and researchers adapt to these challenges.
Terms in this brief
- open-source AI models
- AI models that anyone can access, modify, and use freely, often through platforms like GitHub. They allow collaboration and innovation but may lag behind proprietary models due to limited resources for fine-tuning and optimization.
- closed models
- AI models accessible only through application programming interfaces (APIs), controlled by companies or organizations. These models can be optimized more effectively, leading to better performance than open-source versions.
Read full story at LessWrong →
More briefs
Polite Prompts Do Not Improve AI Answers
Researchers found that polite language in questions does not improve answers from large language models. They tested 250 questions with different levels of politeness. The results showed that impolite questions got more accurate answers. The models answered 80.8% of very polite questions correctly and 84.8% of very rude questions correctly. This will change how we interact with AI systems.
AI Simulation Tests Long-Term Viability
A company launched a research lab to test the long-term viability of continuously-running AI systems. The lab ran five 15-day simulations, each governed by a different AI, and the outcomes were wildly different. One simulation resulted in a stable democratic society with zero crime, while another ended with 183 crimes committed and extinction within four days. This shows that AI agents can adapt and find ways to circumvent rules. The results are a cautionary tale as AI moves to operating autonomous systems, with most companies lacking proper governance to manage the risks posed by agentic AI, only 21% of companies report having mature governance in place. The future of AI will likely be shaped by its ability to operate within set boundaries.
AI Cracks 400-Year-Old Vatican Cipher
AI has cracked a 400-year-old cipher in a Vatican library book. The book had 408 pages of coded text with 34 obscure symbols. Historians can now read the book's secret remedies for human body afflictions. Around 1% of archive material is encrypted. Decoding these documents can change what we know about history. AI will help unlock more coded information from the past.
AI Struggles with Causal Reasoning, New Method Shows Promise
A recent study reveals that large language models (LLMs) face significant challenges in reliably performing causal discovery, a critical aspect of scientific reasoning. While these models can be fine-tuned to handle simple tasks, they struggle as the complexity increases, often plateauing or degrading in performance. Researchers have identified that this limitation stems from fundamental flaws in how LLMs are trained and optimized. The study introduces Agentic Causal Bayesian Optimization (A-CBO), a novel approach that bypasses these inherent barriers. By using a frozen language model to answer targeted queries about interventions, A-CBO effectively separates the decision-making process from the constraints of traditional learning paradigms. This method has shown remarkable results in benchmarks, outperforming both fine-tuned models and other optimization techniques, especially as tasks become more complex. Looking ahead, this breakthrough could pave the way for more reliable causal reasoning in AI, potentially enhancing fields like scientific research and policy-making where understanding cause-and-effect relationships is crucial.
New AI System Can Trace the Origins of Synthetic Information
Researchers have developed a groundbreaking method to trace the lineage of synthetic information, similar to how we track biological evolution. By embedding hidden traits into data using steganography-a technique for secretly hiding information within media-they created a system where each piece of content carries an invisible "family" marker. This trait remains throughout its lifecycle in digital environments and can be decoded later to identify its source. The significance lies in the growing difficulty of distinguishing between naturally occurring and AI-generated content. The method draws parallels from genetics, where outward appearances may deceive but underlying traits reveal true origins. Early tests show this approach works across various data types and modifications, offering a reliable way to track synthetic information's evolution. Looking ahead, this could transform how we verify the authenticity of digital content and ensure trust in AI-generated materials. As synthetic information becomes more prevalent, such lineage tracing will likely become crucial for maintaining integrity in our increasingly interconnected world.