Jason Katz

I've been asked this so many times, I decided to put together a cheat sheet. There is a ton of confusion in the synthetic respondent domain. To start with, not every approach to synthetic data uses LLMs. For example, the model approach used for subgroup boosting is completely different from the LLM approaches being used elsewhere. It has very specific applications and limitations. On the other side, the "Big Data" approaches used by Qualtrics and a few others are also not LLM-driven methods, but heavily leverage past data sources to create a massively parameterized quantitative model that's actually more similar to the models used for image generation. (If you're thinking this creates some severe data privacy/confidentiality issues, you're right - I've covered this elsewhere). We've seen these tools can replicate well-known data such as surveys about Google search activity, which is an easy case, but I haven't seen validation data for difficult cases yet. If you want citations and more details, email me or ask in the comments. If you disagree with anything, please let me know why and send me the validation data. #SyntheticRespondents #CustomerInsights #MarketResearch #DigitalTwins

150

35 Comments

If you read that “Diabolus Ex Machina” post about AI lying to a user, it is worth noting that the problem that AI systems cannot reliably follow links pasted into prompts & instead just make up content is a long-standing one dating back to the original ChatGPT 👇 It should be fixed as it is a common, bad failure mode. Better models can actually go to links, and a better prompt fixes it, but most users don’t know that.

304

107 Comments

Why the Absolute Zero paper should be on your radar Most people still assume better LLMs need better data. This new approach skips datasets, supervision, and labels entirely. It works through runtime feedback alone — and still beats strong baselines. That shift has direct implications for how engineers train and evaluate models. #llms

17

2 Comments

Yext's latest study of 6.8 million AI citations prove that 'brand building' shouldn't be your strategy for AI Search (like I told you). Yext’s new study of 6.8 million AI citations claims 86% come from “brand-controlled” sources — mainly first-party websites (44%) and listings (42%). But this isn’t proof that AI rewards brand power. It’s proof that AI rewards structure. Models like Gemini, ChatGPT, and Perplexity aren’t citing brands because they’re famous, they’re citing them because their data is clean, connected, and verifiable. Forums make up just 2% of citations because they’re unstructured. So when Yext says “brand-controlled,” read it as machine-readable. Visibility in AI doesn’t come from brand awareness, it comes from retrieval readiness. That’s Retrieval-First marketing in practice. https://lnkd.in/e25xaPSf

16

4 Comments

The AI data/labeling space seems kind of bubbly. Foundation labs are willing to spend a LOT for data or labeled data. Startups that provide this data to labs can make a lot of revenue quickly and raise big rounds. There isn't a great accumulating advantage -- lots of startups competing to get data to sell it out to the labs. The labs will have less and less need for general data, and the longer tail of special data will become harder and harder to get.

30

8 Comments

As search engines like Google move toward generative results and AI overviews, structured data has become a key signal of clarity and authority. Schema increases your chances of being cited in AI-generated summaries. In this post I break down what Schema is, why it matters, and how to actually implement to gain an advantage over competitors. Learn more: https://nudgeai.uk

15

6 Comments

What happens when an AI coding agent can read private data, ingest untrusted context, and act on the internet? Last week, Isaac Flath talked me through Simon Willison's "lethal trifecta" formulation, an agent that has - Access to private data, - Exposure to untrusted information, and - The ability to publish / communicate externally. This effectively collapses trust boundaries we usually keep separate in software systems. Much of our conversation focused on context for AI agents and the fact that external sources extend that context. This means external context can become an attack surface: for example, malicious instructions hidden in a white font (or otherwise invisible) on a website your agent is accessing. How do you think about security when using Claude Code, Antigravity, Amp, or Cursor, especially around isolation, permissions, and external context? Loved this chat with Eleanor Berger & Isaac about mastering Agent Skills, Async Agents, and more generally how to become effective at AI-Assisted coding. Check out the full Vanishing Gradients episode for a deeper discussion of agent capabilities, failure modes, and safe usage patterns.

23

4 Comments

AI agents are dominating the 2025 tech landscape with so many options available. This post breaks down the strengths and weaknesses of the some of the widely used agent frameworks so that others can make informed decisions Key Takeaways: ► OpenAI Agents SDK – Best for simplicity and production-readiness, with a low learning curve and built-in primitives like Agents, Handoffs, and Guardrails for structured workflows. ► LangGraph – Ideal for complex, graph-based workflows but has a steeper learning curve. ► CrewAI – Most intuitive for role-based multi-agent systems, but lacks scalability and is still evolving. ► AutoGen – Supports diverse conversation patterns and agent topologies, but documentation is unclear. Link: https://lnkd.in/gr6pVhJ2

13

This looks useful for folks in research or scanning what's coming out on the frontier: Lumi from Google's PAIR research team, a tool for better understanding papers on arxiv. Reading smarter, not harder -- that tagline brings me back to Mortimer Adler's classic 'How to Read a Book'! From their GH repo: Lumi uses AI to help you quickly read and understand arXiv papers. Features include: ✏️ AI-augmented annotations - read summaries at multiple granularities 🔖 Smart highlights - highlight text + ask questions 🖼️ Figure explanations - ask Lumi about images in the paper Demo: https://lnkd.in/ex7erTBb GH repo: https://lnkd.in/emz_Ghcs Medium article: https://lnkd.in/enky8dKC GitHub discussions: https://lnkd.in/efAmiCcp

19

Anthropic just quietly changed how non-technical people will use AI agents. They’ve launched Cowork — a sandboxed, folder-based version of Claude Code built directly into the Claude Desktop app. No CLI. No virtual environments. No “engineer brain” required. You point Claude at a folder → it can read, edit, organize, and reason over files through plain chat. This is important for one reason 👇 Agentic AI is no longer just a developer tool. What started as “AI that writes code” is now becoming: AI that assembles expense reports from receipt photos AI that manages media libraries AI that analyzes conversations, documents, and workflows AI that acts, not just chats Anthropic explicitly warns about risks (prompt injection, file deletion), which tells you something else: We’re entering the phase where AI agents can meaningfully affect real systems, not just generate text. The takeaway: The next wave of AI adoption won’t be about smarter models — it’ll be about safer, simpler interfaces that let agents operate in the real world. This is the same pattern we saw with: spreadsheets → business users no-code tools → operators now: agents → everyone If you’re building products, teams, or workflows around AI, this shift matters. If this kind of practical, real-world AI insight is useful, follow me for more breakdowns on where AI is actually going — not just the hype.

25

How are LLMs trained to "reason"? It's a 4-step process. 1. Train the LLM on datasets. 2. Fine-tune the LLM to generate tokens in chains of thought. 3. Gather human feedback on the responses, and approximate the (responses, scores) into a function itself! It's like finding a function with n points. 4. Do this until the model becomes smart enough (scored by benchmarks). That's it! You now have a Large "Reasoning" Model. The internals are described in the "Reasoning Models" chapter of the InterviewReady course. This is a new addition to the AI engineering course: https://lnkd.in/dwXei6y9 Designed for software engineers looking to transition to AI. #AI #LLMs #Reasoning

53

Concerns Related to "Agentic AI" and its Mitigation: 1. The Alignment Problem 2. Unintended Consequences & Unpredictability 3. Loss of Meaningful Human Control 4. Goal Drift/Instrumental Goals 5. Security Risks & Malicious Use 6. Bias Amplification 7. Accountability Gap 8. Escalation Risks The potential for agentic AI to operate autonomously raises valid and serious concerns about control, safety, and ethics. While the benefits could be enormous, realizing them safely requires careful design, ongoing research into alignment and safety, robust oversight mechanisms, and proactive governance. #DataScience #MachineLearning #AI #ComputerVision #CyberSecurity #Research #Academia #Neuralnetworks #Patternrecognition #Deeplearning #NLP #LLM #Innovation #FutureTech #AdaptiveAI

12

1 Comment

Frontier LLMs get the hype, but small specialized models deliver the results. After reading the WSJ article, I was reminded of a mistake I made early on. We burned through a lot of money running frontier LLMs on large datasets for sampling, thinking bigger automatically meant better. Only later did we learn that a small fine-tuned model could do the same job for a fraction of the cost. That experience changed how we build. Across multiple projects, we’ve consistently seen small task-focused models match or outperform the larger models. Fine-tuned domain models, narrow architectures and small model pipelines have given us better accuracy, lower latency, more privacy and far more predictable costs. Frontier models still have a role in deep reasoning and planning, but in real production environments, it’s the smaller specialized models that quietly do most of the heavy lifting. Curious if others have had a similar “we had to learn it the hard way” moment. #AI #LLM #DataScience #MLOps #AIStrategy #EnterpriseAI

28

5 Comments