The Growing Pains of Generative AI

Love it or hate it, there is no denying that ChatGPT has thrown open the floodgates and opened our eyes to the possibilities of generative AI.

Crucially, top tech firms around the world are now pouring vast amounts of resources into further pushing the envelope of possibilities, while others are racing to incorporate AI capabilities into their products.

As with any new technology, growth pains abound amidst spurts of rapid progress. We look at some upcoming developments, as well as some ways it is stumbling.

What’s happening

As the most advanced generative AI model today, GPT-4 impresses with its incredible ability to string words together. But our fascination with putting it through exams or explaining difficult concepts detracts from the fact that generative AI models are not designed as repositories of knowledge.

As it is, the performance of today’s AI models is constrained by their lack of knowledge, notes entrepreneur Dan Shipper. According to him, large-language models (LLM) are really reasoning engines and not knowledge databases – a point he says OpenAI CEO Sam Altman also spoke about.

This explains OpenAI’s ongoing efforts to work with third-party providers on ChatGPT plugins. With the ability to access the right information to ground its answers, perhaps the sky’s the limit. Indeed, Shipper suggests that private repositories of knowledge are the new oil, including personal knowledge databases.

In other news, AI models are also growing smaller. Last week, we wrote about Dolly, released by Databricks as an open-source model with capabilities that are “close” to that produced by ChatGPT. Designed as a “cheap-to-build” LLM, Dolly can be trained on a single machine with "low-end" machine learning specifications of just 8x Nvidia A100 40GB GPUs (The A100 does have an MSRP of USD15,000 each).

And researchers are scrambling to make LLMs even smaller. In a video posted on LinkedIn recently, four chatbots were shown running on the same computer. It turns out that an optimized version of LLaMA – a ChatGPT-like model by Meta, could already be loaded and run with just 2GB of RAM.

It is not hard to envision additional fine-tuning reducing the requirements further. Perhaps we might have our own ChatGPT running off our smartphones within the next two to three years.

Perils abound

Of course, it is not all a bed of roses. In South Korea, using ChatGPT resulted in at least three data leaks that were recorded by Samsung Semiconductor – and over a period of just 20 days.

As reported by Tom’s Hardware, a Samsung Semiconductor employee submitted the code of a proprietary program to ChatGPT to fix errors. This effectively disclosed the source code of a top-secret application to OpenAI, an external organization.

In the second instance, confidential test patterns meant to identify defective chips were shared; in the third, the transcript of an internal call was used to make a presentation. It is no wonder that South Korean firms are reportedly developing guidelines for the proper use of ChatGPT and similar services to prevent misuse.

Finally, put guidelines around the kind of content that you can ask ChatGPT, and there is no shortage of people who seek to break it. Indeed, an entire community has sprung up around “jailbreaking” ChatGPT from a benign AI model into providing hate speech, produce malware, or advice it would normally not divulge.

Yes, even OpenAI CEO talked about it.

The future of AI

For now, there is an arms race in generative AI, even as governments mull legislation.

Already, AI startup Anthropic says it is working on producing a next-generation AI that is 10 times more capable than today’s most powerful AI, with plans to spend a billion dollars over the next 18 months training it.

“These models could begin to automate large portions of the economy,” read a pitch deck for its Series C fundraising round. “We believe that companies that train the best 2025/26 models will be too far ahead for anyone to catch up in subsequent cycles.”

While some are sanguine about the future of AI, others are far less optimistic. For AI researcher Connor Leahy, future models of AI could potentially achieve a “godlike-level” of super-intelligence that could well break out of whatever controls are in place.

“These are black-box neural networks. Who knows what’s inside of them? We don’t know what they’re thinking and we do not know how they work,” said Leahy. “If you keep building smarter systems, at some point you will have a system that can and will break out.”

And this dystopian future – as outlined by various commentators – could see the AI running a secret network of servers in the cloud, amassing financial resources via cryptocurrency mining or hacking of cryptocurrency exchanges, and blackmailing (or paying) humans for tasks it might not be able to do by itself.

What will the future bring? I suppose only time will tell.

Paul Mah is the editor of DSAITrends. A former system administrator, programmer, and IT lecturer, he enjoys writing both code and prose. You can reach him at [email protected].​

Image credit: bestdesigns