Suggestions
Harsha Varun
Data Science | Gen AI Solutioning & Delivery @ Accenture
Harsha Varun is a Senior Manager in AI Delivery at Accenture, having transitioned from BRIDGEi2i. With over nine years of experience in data analytics and machine learning, he specializes in AI consulting and delivery. He holds a degree from the Birla Institute of Technology and Science. His LinkedIn profile indicates a strong background in leveraging AI technologies for business solutions, reflecting his expertise in the field.123
Highlights
Will #LLMbias impact our decision making ability?
There's a lot of research these days on the issue of data bias in LLMs and its impact on their performance. One study that particularly caught my attention is David Razado’s research on the political bias of AI. Here are a few key takeaways:
- Most LLM chat models lean slightly left or progressive on the political spectrum. Surprisingly, Elon’s #Grok, which was built to address this bias leans most to the left
- Base models appear more neutral; however, fine-tuned or RL-tuned models show bias, indicating that bias might be introduced during post-training (through fine-tuning or RL)
- It's very easy to bias an LLM by fine-tuning it with limited data. The author fine-tuned GPT on heavily biased data to create left-wing and right-wing variants, which then exhibited strong bias
Is this intentional by LLM providers? I don't think so, but a few factors should be considered:
- The left wing might be more vocal or produce more content online, which could create bias during the pre-training stage
- AI-generated content is increasingly prevalent, which could further bias the models
- A lot of data used for RL/fine-tuning is synthetically generated from GPT, and this bias might, in turn, influence the tuning of other LLMs
This indicates that AI reasoning is still fractured—it's largely repeating the data it's fed. This isn't just a problem for LLMs; it's a broader issue in AI that is influencing human behaviour and election preferences already.
This raises a critical question: how will this bias impact business applications? Most of the knowledge in the industry remains with few people, and relying solely on this knowledge can lead to repeated unoptimal decisions. As decision scientists, it's important for us to be aware of this bias and work to avoid it by ensuring that the data we use to tune models remains as neutral as possible. This requires a fusion of business knowledge with traditional statistics, ML, and LLM techniques, a return to our roots in my view.
There's a great video by Sabine Hossenfelder on this, recommend to check it out: https://lnkd.in/gib2xRmq
What are your thoughts?
Link to the research: https://lnkd.in/gH9CDBnK
#ArtificialIntelligence #ai #aibias #llm #openai #data
Mindsearch is a Gen AI framework for web search and is a new player in the block. What sets it apart? It's opensource.
Mindsearch is a multi-agentic framework designed to mimic human search behavior. It decomposes complex questions into smaller search queries, creates a search plan as a graph, executes the search, and iteratively enhances the search graph based on results. They claim to outperform Perplexity and ChatGPT in web search capabilities.
I've been testing it over the past few days, and the results have been impressive. The real beauty of this framework is its customization potential. You can integrate any Gen AI model, embed domain intelligence using RAG or few-shot examples, and achieve even greater accuracy.
Want to try it out? The live version is in Chinese, but you can clone the branch and update the model here: 🔗 GitHub: InternLM/MindSearch 🔗 Live Version: Mindsearch on OpenXLab
As open-source frameworks and accessible models are now competing with their paid counterparts, the barrier to entry in AI is lowering, and this will further accelerate innovation. It's inspiring to see such a collaborative developer community.
#AI #OpenSource #WebSearch #Innovation #Mindsearch #GenAI
