Narahara Chari D.

"... Hallucinations need not be mysterious—they originate simply as errors in binary classification. If incorrect statements cannot be distinguished from facts, then hallucinations in pretrained language models will arise through natural statistical pressures. We then argue that hallucinations persist due to the way most evaluations are graded—language models are optimized to be good test-takers, and guessing when uncertain improves test performance... " https://lnkd.in/gC8r-J7c

6

If you haven't heard about the Scholarly Graph yet, you should check it out! Super fascinating stuff. EBSCO's linked data experts are passionate about making research more connected — linking over 900 million scholarly entities and 3.8 billion relationships to uncover hidden connections and trace the story of scholarship. It's starting to power Citation Discovery features in EBSCO Discovery Service and EBSCOhost — and that is just the beginning! Learn more: https://ow.ly/cu8Z30sQe1I

7

AGI Is Infeasible. Instead, Pursue Superhuman Adaptable Intelligence On a recent episode of the Dr. Data Show, my co-host Luba Gloukhova and I tackled a new paper authored by AI luminary Yann LeCun alongside other researchers. We had been tipped off by another co-author of the paper, AI researcher Philippe Wyder, who reached out on social media to say the paper related to our prior episode since it “shows a path out of our muddled AGI discourse by embracing specialization.” The paper proposes a much-needed pivot for the industry: Because the ever-popular notion of artificial general intelligence offers only a hazy, overzealous goal, these researchers have proposed a new North Star they call “superhuman adaptable intelligence.” Defining Superhuman Adaptable Intelligence What is SAI? The authors’ definition breaks down into two parts. First, SAI is capable of adapting to exceed humans at any task humans can do – but crucially, it espouses tackling one task at a time. Second, SAI can also adapt to useful tasks outside the human domain. SAI establishes a new goal for the industry. Instead of trying to build a singular, all-capable solution that can do everything a human can do, which is the goal represented by AGI, this framework champions a return to specialized, narrow AI. These researchers advocate leveraging massive amounts of data through self-supervised learning, as for example LLMs do, but they also advise that such tech then be adapted to solve specific problems. One of the most consequential insights highlighted by the authors is... Read the full article, my latest in Forbes: https://lnkd.in/gQqaExTs

11

2 Comments

Exciting insights from the Nvidia partners event at JPM led by Jensen Huang (reposting now because original post was mysteriously deleted): - Patrick Collison, The visionary Founder and CEO of Stripe and co-founder of Arc Insitute (they have developed Evo (1, 2) genomic foundation models) - Christina Zorn, Chief Administrator Officer of Mayo Clinic, aims to enhance patient care by harnessing their vast repository pf 20 Billion images and clinical data sets - Jacob Thaysen, PhD, CEO of Illumina, is steering the company from their current status as a sequencing leader to a pioneering digital biology platform - Ari Bousbib, CEO of IQVIA, wants to develop AI agents to extract insights from a remarkable 64 petabytes of exclusive healthcare data Lots of great promise for AI in improving lives of patients! #NvidiaEvent #healthcareinnovation #AIfor healthcare

115

6 Comments

Why Do Large Language Models Hallucinate? A new paper from OpenAI & Georgia Tech unpacks how and why today’s LMs sometimes confidently generate plausible sounding but false statements (“hallucinations”). The authors present a clean, statistical view of the problem and argue that it’s baked in, unless we rethink the way we train, evaluate, and reward these systems. 🔍 Core ideas 1. Hallucination arises from statistical “errors in classification” The paper draws an analogy: a language model’s job is akin to a binary classification task distinguishing “valid” (true or well-grounded) outputs from “error” (hallucinations). Even if your training data were perfect, the very nature of probability estimation means there’s risk of misclassification, especially when facts are rare or unique. 2. Some facts are inherently hard to learn (or “singleton facts”) If a particular fact appears only once in your training data (or not at all), there’s little statistical signal to learn from. The model might “guess” plausible but incorrect variants. The authors show that the rate of hallucination necessarily scales with how many facts are “singletons.” 3. Post-training (fine-tuning, RLHF, etc.) doesn’t eliminate hallucinations and can even exacerbate them Why? Because most evaluation benchmarks and leaderboards reward confident answers and penalize uncertainty. If your model can hedge (“I don’t know”), that’s often scored worse than making a wrong but plausible guess. The incentives align with bluffing. 4. The evaluation bias is a major culprit The paper argues that the dominant metrics in the field (accuracy, pass rate, strict “correct vs incorrect”) systematically discourage models from saying “I don’t know” or expressing appropriate uncertainty. That, in turn, pushes models toward overconfidence and hallucination. 5. A better path: integrate confidence targets into benchmarks Rather than relying on binary scoring, the authors propose explicitly specifying confidence thresholds and penalties. That way, a model that declines to answer when uncertain (or hedges) is not universally penalised and we encourage more trustworthy behaviour. ✨ Why this matters • For users: it explains why LMs sometimes lie, even when they seem “very sure.” • For builders: it shows that minimising hallucinations isn’t just about architecture or more data, we also need to rethink how we judge models. • For the field: by aligning incentives (training + evaluation), we may shift toward AI that’s more transparent about uncertainty, instead of always overpromising. #LanguageModels #TrustworthyAI #Hallucinations

16

2 Comments

Agentic AI represents the next major leap in artificial intelligence. In this article, I emphasize that it's no longer just about automating tasks or generating content—it's about building autonomous agents that can think, adapt, and act independently to achieve goals. Personally, I find it exciting to work with systems that don’t merely follow instructions but actively determine the best way to accomplish tasks, helping businesses become faster, smarter, and more efficient. #AgenticAI #LLMs #Governance #Challenges #Takeda https://lnkd.in/eVVtqTQ2

34

1 Comment