AI: Addressing the AI Hallucinations 'Bug'. RTZ #396

...reining in AI errors & 'confabulations' when AI hallucinations are not a 'feature'

The Bigger Picture, Sunday June 23, 2024

Despite all the ebullient, mainstream enthusiasm around this AI Tech Wave and eventual AGI (Artificial General Intelligence), a perennial worry amongst worries remains. And it’s true for all AI stakeholders, enterprises, consumers, investors, and the companies building this stuff, is that AI for now is nowhere as reliable and error-free as traditional computing products and services have been for over fifty years.

In last week’s AI: Bigger Picture, I highlighted how Scaling AI Trust is Job #1, ahead of even Scaling AI itself. In this week’s Bigger Picture, I’d like to revisit one of the key root causes of that lack of Trust, the hallucinations of current LLM AIs. So fixing these ‘bugs’, is the current holy grail for all AI companies and researchers.

As Axios pithily summarized the issue back in February:

“Generative AI is raising the curtain on a new era of software breakdowns rooted in the same creative capabilities that make it powerful.”

“Why it matters: Every novel technology brings bugs, but AI's will be especially thorny and frustrating because they're so different from the ones we're used to.”

“Most AI systems don't operate by commands and instructions — they use "weights" (probabilities) to shape output.”

“Zoom out: Making things up is at the heart of what generative AI does.”

• “Traditional software looks at its code base for the next instruction to execute; generative AI programs "guess" the next word or pixel in a sequence based on the guidelines people give them.”

• “You can tune a model's "temperature" up or down, making its output more or less random. But you can't just stop it from being creative, or it won't do anything at all.”

'“Between the lines: The choices AI models make are often opaque and even their creators don't fully understand how they work.”

• “So when developers try to add "guardrails" to, say, diversify image results or limit political propaganda and hate speech, their interventions have unpredictable results and can backfire.”

• “But not intervening will lead to biased and troublesome results, too.”

For a deeper dive into how AI works differently from traditional software, I’d recommend this other Axios piece from last fall as a refresher. It has deeper links to follow for those interested.

Of course all AI companies are on the the issue of fixing AI hallucinations as a bug of course, even though in some areas, especially creative endeavors, AI hallucinations are a feature not a bug.

But for most areas, progress needs to made to reduce said bugs and errors. Axios this week highlighted some of the latest progress on this front.

“Between the lines: The most advanced chatbots from OpenAI, Meta, Google and others "hallucinate" at rates between 2.5% and 5% when summarizing a document.”

“A new algorithm, along with a dose of humility, might help generative AI mitigate one of its persistent problems: confident but inaccurate answers.

“Why it matters: AI errors are especially risky if people overly rely on chatbots and other tools for medical advice, legal precedents or other high-stakes information.”

• “A new Wired investigation found AI-powered search engine Perplexity churns out inaccurate answers.”

“The big picture: Today's AI models make several kinds of mistakes — some of which may be harder to solve than others, says Sebastian Farquhar, a senior research fellow in the computer science department at the University of Oxford.”

• “But all these errors are often lumped together as "hallucinations" — a term Farquhar and others argue has become useless because it encompasses so many different categories.”

  

“Driving the news: Farquhar and his Oxford colleagues this week reported developing a new method for detecting "arbitrary and incorrect answers," called confabulations, the team writes in Nature. It addresses "the fact that one idea can be expressed in many ways by computing uncertainty at the level of meaning rather than specific sequences of words."

• “The method involves asking a chatbot a question several times — i.e. "Where is the Eiffel Tower?"

• “A separate large language model (LLM) grouped the chatbot's responses — "It's Paris," "Paris," "France's capital Paris," "Rome," "It's Rome," "Berlin" — based on their meaning.”

• “Then they calculated the "semantic entropy" for each group — a measure of the similarity among the responses. If the responses are different — Paris, Rome and Berlin — the model is likely to be confabulating.”

“What they found: The approach can determine whether an answer is a confabulation about 79% of the time — compared to 69% for a detection measure that assesses similarity based on the words in a response, and similar performance by two other methods.”

It’s not a panacea of course, and the piece outlines some of the downsides that have be worked around. Note that these efforts to mitigate AI hallucinations and ‘confabulations’ is separate from addressing the much feared AI existential risks that are being posited and positioned for by AI industry luminaries.

This side of addressing AI hallucinations and ‘confabulations’, is a relatively prosaic and practical effort. And there are technical solutions over time to make more progress. Bugs as in nature are a fact of computer code. And there are ways to address them.

The important point is that at this AI Tech Wave the industry broadly defined, is investing billions in mitigating the bugs, while other figure out how to LLM AI propensities to hallucinate as a feature. More on that in a later piece. Stay tuned.

(NOTE: The discussions here are for information purposes only, and not meant as investment advice at any time. Thanks for joining us here)