Weekly recap: 2023-12-31

Posted by Q McCallum on 2023-12-31

What you see here is the last week’s worth of links and quips I have shared on LinkedIn, from Monday through Sunday.

For now I’ll post the notes as they appeared on LinkedIn, including hashtags and sentence fragments. Over time I might expand on these thoughts as they land here on my blog.

2023/12/26: What’s holding companies back?

According to an Accenture exec, this is what’s holding companies back when it comes to AI:

Most companies are not ready to deploy generative artificial intelligence at scale because they lack strong data infrastructure or the controls needed to make sure the technology is used safely, according to the chief executive of the consultancy Accenture.

(Source: “Accenture chief says most companies not ready for AI rollout” (FT))

I agree with that list. I’m also glad to see that I’m not the only one talking about safety and controls (key elements of risk management) when it comes to AI. Great.

But I also think that list is incomplete.

What’s missing, then?

Given my years of experience consulting in this field, I can tell you that lack of data literacy is a key hurdle in successful AI adoption.

This starts at the top of the org chart. If executives and product owners aren’t aware of what AI really is and what it can(not) achieve, they are more likely to hold unrealistic expectations. If they are not active participants in setting strategy, the company risks releasing needless “AI-infused” products and services. And they’re more likely to encounter problems along the way.

To draw an analogy: you don’t have to turn on the lights when you enter a dark room. But it’s the nicer way of finding out where the sharp corners and pitfalls are located.

Want someone to turn on the lights for your company’s AI adventure? Reach out: https://qethanm.cc/contact/

2023/12/27: What’s in your training data?

What’s in your company’s AI training data?

No, I’m not asking what data you think you have. Nor am I asking what data should be in there. I’m asking how much you have actually checked the data before building AI models on it.

I usually ask this kind of question to confirm that a company has secured the proper rights to use that data. No sense building and deploying a model, only to get sued and have to take it down.

But, well, now there’s another reason to keep an eye on your training data (and the entire data supply chain feeding into it):

Child sex abuse images found in dataset training image generators, report says” (Ars Technica)

2023/12/29: A newsroom shares the workload with AI

Most of the stories about “newspapers and gen AI” focus on lawsuits and new deals for training data. This one’s a little different:

How one of the world’s oldest newspapers is using AI to reinvent journalism” (The Guardian)

Of note:

The AI reporters use an in-house copywriting tool based on the technology ChatGPT, a souped-up chatbot that draws on information gleaned from text on the internet. Reporters input mundane but necessary “trusted content” – such as minutes from a local council planning committee – which the tool turns into concise news reports in the publisher’s style.

I’m on the fence about letting the machines generate the article text. But this group gets credit for separating the responsibilities such that the journalists do what the machines cannot (interview people, visit courtrooms, file FOIA requests).

No matter how this turns out in the long run, they have made some smart moves in designing the experiment.

(As a bonus for NLP/NLU work: since the generated articles will hold the same style, that should simplify parsing and analysis work down the road.)

2023/12/30: Your periodic reminder

Your periodic reminder that a generative AI bot is a system that generates text based on probabilistic patterns in grammar.

It is not a search engine.

(In other words: here’s another one for genAI’s “that’s not how this works” list …)

Michael Cohen used Google’s AI to research legal cases to cite in his appeal. The AI hallucinated them.” (Insider)

2023/12/31: Yet another side door for LLMs

This article points to an interesting side-door exploit in SaaS-style genAI models: fine-tuning a model to bypass some guardrails.

Personal Information Exploit With OpenAI’s ChatGPT Model Raises Privacy Concerns” (NY Times)

This part caught my eye:

Mr. Zhu and his colleagues were not working directly with ChatGPT’s standard public interface, but rather with its application programming interface, or API, which outside programmers can use to interact with GPT-3.5 Turbo. The process they used, called fine-tuning, is intended to allow users to give an L.L.M. more knowledge about a specific area, such as medicine or finance. But as Mr. Zhu and his colleagues found, it can also be used to foil some of the defenses that are built into the tool. Requests that would typically be denied in the ChatGPT interface were accepted.

“They do not have the protections on the fine-tuned data,” Mr. Zhu said.

The researchers’ technique raises important questions around the generated artifacts:

  • A client who has fine-tuned a model could argue: “the fine-tuned model represents my creation and I should be able to use it as I please.”
  • A hosted model provider and SaaS host could argue: “hey you should be responsible in how you use that fine-tuned artifact.”

What complicates the provider’s side is … how to enforce that view? If you think that content moderation of user-generated content (UGC) is hard, what about content moderation for all of your customers’ genAI models – checking both prompt inputs and model outputs for problems?