Some thoughts on the latest GenAI news

There is a lot of news lately on OpenAI losing large amounts of money in its operations and how opensource LLMs match performance with commercial LLMs.? This is double bad news for companies such as OpenAI.? Bad news does not stop here.? The Fortune 500 companies are disillusioned with ROI on this huge promise of GenAI models. Enterprises are not seeing large-scale transformation or cost savings out of these models.? On the contrary, I saw a study report claiming that overall productivity has come down due to new skills learning time and the need to double-check all answers from LLMs due to the fear of hallucinations and wrong answers. One headline screamed about a $1 Trillion drop in USA tech stocks including NVIDIA-type AI infrastructure companies due to worries about AI investments not delivering business value as promised. With all this worrying news, some experts are even more concerned about another AI winter like 1980s when AI Expert systems research just slowed down to a halt for almost 20 years. ?

Experts like Gary Marcus have been advocating reducing investments in LLM-scaled next versions and looking at other AGI techniques.? They argue that more scaling will not improve reliability but may even reduce reliability as current public domain data is flooded with fake websites, fake news, fake articles, fake images, etc. generated by current GenAI technologies.?

At @AiThoughts.Org , our view is more moderate.

Let us list some positive points first.

·??????? We can confidently say that the better the prompt, the better the GenAI Response.? We have seen lots of examples where a properly scripted prompt provides excellent responses.

·??????? We can also see promises in the RAG technologies.?? By using RAG and properly slicing the domain knowledge documents, we can expect more relevant answers to our questions.

·??????? AI Agent building technologies are available today with good capabilities to Plan, set workflows and execute actions to get value out of LLM + RAG Domain + Enterprise systems.? Agents will deliver business value and will be better adapted by business users.

Let us also list some good ideas which is not getting traction which may help to realize better value from GenAI technologies.

·??????? We heard about the Small Language Model based on a limited set of training data sets based on published books, published scientific papers, reputed journals, news papers, etc.? I do not see much excitement in this space yet.? This may improve the reliability of the responses and reduce hallucinations.??

·??????? We also heard about watermarking generated content with crypto technologies which will allow LLMs to pre-check if an input is original content or generated by another LLM. This got traction when copyright issue lawsuits and the risk of deep fakes were discussed.? But lately not much traction on these.

·??????? Explainability research on pre-Gen AI models came to an end after the release of GenAI models.? It is as if the industry accepted the fact that it is not possible to explain decisions or responses from such a gigantic model and any attempt to retrace the steps to understand the reasoning will not work.? However, enterprises and government organizations need to explain why a decision was taken and also subject to audit every decision affecting customers and citizens.?

?

?What is our recommendation?

·??????? Let us spend money on production-scale enterprise use cases using LML ++ technologies i.e AI agents technologies. ?Human in the Loop in every phase is OK.? Human-in-the-loop is the best way to get industry use cases and no need to look for a no-human solution. This means some systems integration and API work, let it be. We need to get Fortune 500/1000 enterprises to start seeing visible benefits from AI technology in 2024-25.

·??????? Let us spend money on ?AI Ethics and risk management research including some form of explainability.

·??????? Let us spend money on Cyber Security.? Enterprises will be very concerned about AI agents interfacing with their ERP systems inside the firewall.

·??????? Finally let us make AI work for enterprises.?

More later.

L Ravichandran