Log of insights and news – Gorkem Turgut (G.T.) Ozer
This is an (almost) complete picture of LLMs, highlighting the underlying math for probability calculations. It’s a great companion for the 64-picture visual summary I curated for a talk earlier this year.
Good reminder that “Life’s most important questions are, for the most part, nothing but probability problems.” Pierre-Simon Laplace in Théorie Analytique Des Probabilités
I once quoted Abraham Maslow in my note “Using predictive modeling as a hammer when the nail needs more thinking“, a problem that has since eased as focus shifts to causal modeling and optimization.
Now, if the problem is extracting handwritten digits from a PDF or an image, what’s the solution? I guess good old reliable OCR? There is nothing generative about extraction, so using an LLM may be moot, if not actively detrimental (due to hallucinations). OCR is widely available through cloud like AWS Textract and Google Cloud Vision, some of which also offer human validation.
Instead, the Reuters team intriguingly chose the shiniest hammer, an LLM, to extract data from handwritten prison temperature logs. They uploaded 20,326 pages of logs to Gemini 2.5 Pro, followed by manual cleaning and merging. One big problem with this approach is the inevitable made-up text (hallucinations), which required the team to hand-code 384 logs.
So why not OCR but an LLM? All I can think of is, LLMs may be useful in extraction when the text is highly unstructured and contextual understanding is needed, neither of which seems to apply here. Surprisingly, the methodological note doesn’t even mention OCR as a solution.
Putting the tool choice aside, though, Reuters asks an important question here: “How hot does it get inside prisons?” and the answer is “Very.” I applaud the effort and data-centric journalism, and I recommend reading the story.
Credit for the image goes to Adolfo Arranz.
How do you get to new ideas when data is always looking back?
In this latest Data Duets post, we discuss a case from United Airlines and share five key lessons from an interview with Patrick Quayle, Senior VP of Global Network Planning.
The post explores how to go beyond historical data by using:
Our Director’s Cut expands the discussion to offer insights for retail merchandising.
Enjoy the read and feel free to leave your comments on LinkedIn here.
[Click title for image]
This is so hilarious I had to share. A major issue with using LLMs is their overly obsequious behavior. They aren’t much help when I’m right; I don’t want to be right, I want to be corrected.
This project uses a Python script to count how often Claude Code says you’re “absolutely right.” The script doesn’t seem to normalize the counts by usage, which might be a good next step.
In January, I wrote a short note based on one of my talks: “How to use LLMs for learning in 2025.” In that note, I differentiated between using LLMs (1) to learn and (2) to do. With the new semester now underway, I’ve checked some usage numbers and read the Ammari et al. (2025) paper on how students use ChatGPT. I was particularly interested in the second RQ: “Which usage patterns correlate with continued or increased reliance on ChatGPT over time?”
An over-reliance on any tool, regardless of what it is, is a potential red flag for persistent learning, especially when the goal is comprehension. For example, understanding derivatives and calculating them using a computer are two distinct learning objectives. If the reliance on a tool substitutes for understanding, long-term implications may not be a net positive.
The article does not really answer the reliance part of the question. It does, however, report some interesting correlations between LLM behavior and student engagement. Notably, when ChatGPT asks for clarifications, provides unintended or inconsistent answers, or communicates its limitations, students are less likely to continue using it.
Plausible, but what these correlations mean for learning and comprehension is unclear. What is the next step after disengagement? Do they switch to another LLM to get a direct answer without having to answer follow-up questions, or do they go back to figuring it out on their own?
Class of 2029, I guess the answer lies with you. Welcome!
A recent New York Times story offers insights.
Using LLMs as a “second pair of eyes” or as a fallible assistant seems to work well. Automation also works effectively when the instructions are clear and the objectives are defined unambiguously. In both cases, human agency remains central.
Use case #15 in the article, “Review medical literature,” reminded me of a study I shared earlier (How do LLMs report scientific text?). The study showed that LLMs systematically exaggerate claims they found in the original text. The user in this case is a medical imaging scientist and is aware of the danger. When a tool isn’t foolproof, the user’s expertise and awareness make all the difference.
The high-demand use cases are quickly scaling into independent businesses with more standardized output, often with LLMs as the core wrapper. I suspect some are marketed as “magic,” and to resist that hype, users will need a combination of expertise and awareness.
If you’re a visual learner looking to deepen your understanding of AI and language models:
I’ve just made the deck from my recent talk, “AI in 64 Pictures,” available. It’s a visual journey through language processing: from word embeddings to RNNs, attention, and transformers.
Understanding AI models better helps us discover more use cases and navigate their limitations. And if you’re looking to dive deeper, you can follow the links in the slides.
Nearly half of all AI agent projects are set to fail (as Gartner predicts here). Why? Unclear business value, inadequate risk controls, and escalating costs.
As I see it, much of this is fueled by hype, leading to existing solutions being relabeled as “Agentic AI” without any rethinking of business processes.
Human creativity is missing in this picture. It’s this creative thinking that should move agent use beyond just automating or augmenting individual tasks with LLMs, leading instead to the redesign of business processes and a vision for how humans and AI can truly complement each other.
The risks and costs are more straightforward to resolve:
– Managers who are most excited about AI agents often do not fully understand the risks and limitations of LLMs. They should invest as much in understanding these models as they do in using them.
– The true cost of scaling proof-of-concept GenAI solutions is often underestimated. This is on selecting the right vendor. Gartner estimates only about 130 of the thousands of agentic AI vendors are real.
Andrew Gelman’s latest “rant” is worth a read. Everybody lies, but why, even when data clearly refutes it?
It’s interesting to think a little bit more and understand why and how people lie, especially when it comes to scientists, medical doctors, and law enforcement officials. Spoiler is, the answer is not always money.
[Click title for image]
Turns out Anthropic literally purchased and scanned millions of print books to train its language model Claude.
The judge finds that the scanning of these purchased books is “fair use” even without licensing, and goes on to say:
Everyone reads texts, too, then writes new texts. They may need to pay for getting their hands on a text in the first instance. But to make anyone pay specifically for the use of a book each time they read it, each time they recall it from memory, each time they later draw upon it when writing new things in new ways would be unthinkable. For centuries, we have read and re-read books. We have admired, memorized, and internalized their sweeping themes, their substantive points, and their stylistic solutions to recurring writing problems.
This is quite interesting, as it draws an analogy between how humans read, comprehend, and write text and how a language model operates.