Cassie Kozyrkov

Kozyr

Speaking and advisory services. Request form: makecassietalk.com

Kozyr was renamed from Data Scientific.

Chief Executive Officer

Federal Reserve Bank of New York

Served a full two year term as Innovation Advisor and Councillor at the Federal Reserve Bank of New York, with a focus on the impact of AI on the financial markets and global policy.

Innovation Advisory Council Member

Google

Decision Intelligence is an interdisciplinary field concerned with all aspects of decision-making.  It combines Data Science (statistics, machine learning, AI, analytics) with the Behavioral Sciences (psychology, neuroscience, economics, and the managerial sciences).

I serve as a thought leader in this new discipline, guiding Google and the world in making safe, effective use of AI and data.

Chief Decision Scientist

My mission is to democratize artificial intelligence and guide business leaders in harnessing the transformative power of machine learning.

Chief Data Scientist, Office of the CTO

Youngest ever member of Google's Office of the CTO, specializing in AI and data science.

I'm also a coauthor of Google's Technical Director training program to teach others how to succeed in the role.

Technical Director, AI and Data

My roles are: statistician, machine learning expert, mentor, data scientist, evangelist, decision scientist, and consultant.  I bridge the gap between leaders and quantitative analysts / researchers.

As Google's statistics and machine learning sensei:
* I've advised 200+ teams at Google on their analytics processes.
* My ideas have sparked 400+ new projects.
* I've trained 20,000+ analysts/leaders/engineers in rigorous data-driven decision-making and/or using machine learning to transform their work.

Largest projects in: Search Advertising, Google Maps, Machine Intelligence, Infrastructure, YouTube

Statistician

Self-Employed

Data science consulting and analytics services for clients from government organizations, private sector firms, and academia.

Data Science Consultant

North Carolina State University

Statistics Lecturer

Statistical Computing, Advanced Topics in Statistics: Big Data, Special Topics in Statistics: Data Mining, Applied Time Series Analysis, Applied Multivariate Data Analysis, Advanced Probability and Measure Theory, SAMSI Computational Advertising, Statistical Methods II, Statistical Consulting, Statistical Theory I, Statistical Theory II, Linear Models and Variance Components, Stochastic Processes, Doctoral Research, Supervised Doctoral Teaching, R Programming

Mathematical Statistics PhD Student

Duke University

Psychology & neuroscience with specialization in transcranial magnetic stimulation (TMS) methods and the science of decision-making.

Neuroeconomics PhD Candidate

Center for Cognitive and Social Neuroscience at the University of Chicago

Coordinated lab research projects and team effort. Designed and ran experiments.

Project Manager

Charitable contributions and decision-making research with Professors Steve Levitt and John List.

Research Assistant - Prof. John List

Cassie is Google’s former Chief Decision Scientist and a sought-after AI advisor, working with senior leaders at places like Lenovo, NASA, and Gucci. She is currently CEO of Data Scientific and is known for founding the field of Decision Intelligence.

Cassie is a beloved personality in the data and AI community, followed by over half a million tech professionals. She is also a renowned keynote speaker and writer. ( If you'd like her to speak at your event, here's her booking form: makecassietalk.com )

In almost 10 years at Google, Cassie personally trained over 20,000 Googlers in data-driven decision-making and AI and has helped over 500 projects implement decision intelligence best practices. Before that, she served in Google's Office of the CTO as Chief Data Scientist. The rest of her 20 years of experience was split between data science, consulting, lecturing, and academia.

If you've had a good chuckle while learning about AI, statistics, or decision-making, chances are you've encountered her writing, which has reached millions of readers. She has been honored as a LinkedIn Top Voice for 5 years and as a #1 Artificial Intelligence writer on Medium for 7 years.

Speaking and advisory services. Request form: makecassietalk.com

86% of enterprises plan to invest more in agentic AI.
6% trust agents to run core processes.

Let's talk about that gap!

Harvard Business Review just released some findings about tech decision-makers (n=603):
🤖 86% expect agentic AI investment to grow in 2 years
🤖 6% fully trust agents with end-to-end processes
🤖 ~25% say they're "ready" on infrastructure, data, security, and governance

That lack of trust is a good thing, in my opinion. Trusting complex probabilitistic systems without re-archtiecting complex processes is a recipe for kablooies. If anything it should give you hope. The trouble is that the redesign is difficult except in narrow use cases.

The model isn't the bottleneck anymore.
The plumbing and the people are.

Fragmented apps and data → agents can't get reliable context 
Legacy processes → never designed for autonomous execution 
Security and compliance → no playbook for "agents at the core" 
Employees → unclear if agents are tools or replacements

The report points to enterprise orchestration as the missing layer—connecting systems, standardizing actions, enforcing governance so agents behave predictably instead of becoming shadow IT.

Old question: "Which LLM should we use?"
New questions:
Which processes are we willing to redesign around agents?
How do we orchestrate data, tools, and guardrails so we can trust them with real work?

In 3 years, the gap won't be "who has agents."
It'll be who rebuilt their processes so agents can actually run them.

Where is your team right now?
A) Still experimenting at the edges (help desks, copilots) 
B) Re-architecting workflows so agents own real outcomes 
C) Somewhere in between, trying to figure out where to start
D) Waiting and watching others stub their toes

Drop your letter below. Your comments make my (digital) world 'round.

P.S. On a completely different note,  my Decision-Making with ChatGPT course is running on demand right now with a live Q&A with me next week. Come ask me anything! decisiongptcourse.com

CEO, Google's first Chief Decision Scientist, AI Adviser, Decision Strategist, Keynote Speaker (makecassietalk.com), LinkedIn Top Voice

Agentic HBR

Been wondering when my next cohort will be? Surprise! The time is NOW. If you missed the previous run or have been waiting to join, enrollment is now live:

 👉 https://lnkd.in/eG7EBtTv

The last cohorts made something clear: leaders aren’t looking for more information. They’re looking for clearer judgment. ChatGPT helps - but only when you use it in the right parts of the thinking process.

In this cohort, we’ll continue the work of turning AI into a genuine thinking partner instead of a shortcut. You’ll learn how to:

• Use ChatGPT to reduce effort without losing rigor
• Avoid the trap that turns “AI-assisted” into lazy thinking
 • Spot cognitive biases before they distort outcomes
 • Separate luck from decision quality
 • Build a decision scribe so you finally learn from your choices
 • Operate with confident judgment instead of drowning in advice

You’ll get 30+ on-demand videos (available as soon as you enroll), a 2-hour live Q&A with me (at 1 PM ET Dec 18), prerecorded Q&A from earlier cohorts, templates, frameworks, and a global community.

If you're working on sharpening your thinking in an AI world, Cohort 3 is now open.
 👉 https://lnkd.in/e-RaxVJJ

If this resonates, a repost or subscribing to decision.substack.com helps the course reach the leaders who need it.

#DecisionIntelligence #AILeadership #StrategicThinking #DecisionMaking #AI

Vincent Granville

BondingAI

GenAI platform based on xLLM, the game-changing technology that started the LLM 2.0 revolution. Boosting productivity, automating tasks, delivering ROI at scale to Enterprise customers. Without GPU, hallucinations, security issues, or latency. With zero weight, real-time fine tuning, exhaustive but concise results, customizable relevancy scores, and a modern UI - not just a prompt box.

Chief AI Architect, Co-Founder & Investor

NVIDIA

Advocate for products and services pertaining to AI, generative AI and LLMs. Featuring curated Nvidia content in my GenAI newsletter (200k subscribers) and other media, presenting at selected Nvidia events. Deep understanding of the AI landscape and technology, building ground-breaking architecture known as LLM 2.0 / xLLM, with our startup BondingAI.io participating in the Nvidia Inception program.

AI Advocate & Startup Inception Program

GenAItechLab.com

Now part of BondingAI.io. Building better, faster, and less expensive GenAI products with explainable AI, such as NoGAN (data synthesizer). Key product:  xLLM, a fast in-memory agentic multi-LLM system built from scratch, with zero weight, no training, no latency, no hallucination, fully contextual from the ground up, with real-time fine-tuning. For scientific or enterprise search, as well as clustering and predictions.  Also, builder of the best model evaluation metric on the market.

Founder, Lead AI Architect

AB InBev

Helping Fortune 100 company InBev implement my xLLM technology as an alternative to OpenAI and LangChain to boost relevancy (hallucination-free), reduce latency, and increase speed by several orders of magnitude, while increasing security via sub-LLMs accessible to authorized users only. Offering a better solution for search and cataloging, to help users navigate their corporate repository. With a focus on concise but exhaustive results (via augmentation) and specialized sub-LLMs. Local implementation without external API calls.

GenAI & LLM Consultant

MLtechniques.com

AI optimization research lab, developing state-of-the-art technology with a focus on high quality and fast algorithms, thus cost effective. Our GenAI Web APIs are accessible on GenAItechLab.com. Covering synthetic tabular data, artificial DNA sequences, time series & geospatial data generation, agent-based modeling, synthetic graphs, and customized LLMs.

Chief AI Scientist

Data Science Central

The company was acquired in 2020.  I am still a contributing author.

Executive Data Scientist, Co-Founder, Managing Partner

Microsoft

Developed a strategy to collect the right data and detect minuscule change points in time series associated with KPI's (when the changes occur, their amplitudes and directions). Goal: assessing the impact of events influencing core metrics. Example: changing the definition of a website user resulting in shift in user count, or changing some features on Bing.com resulting in increase / decrease in pageviews. These were blind tests: I did not know when the website changes were made. But when I detected a change, it was matched against known events that took place at the time flagged by my algorithm as "change point". I designed and then helped automate this algorithm.

Machine Learning Consultant

LookSmart

Optimizing advertising campaigns, Ad arbitrage. Projects and technology include:

• Created list of top commercial keywords accounting for 85% of all pay-per-click revenue
• Leveraged Google API's to automatically find millions of high yield keywords 
• Automated taxonomy creation (NLP)  
• Click fraud detection (click jacking and Botnet detection)
• Internet traffic quality scoring using hidden decision trees 
• Automation of keyword bidding 
• Reverse engineering Google keyword pricing algorithms, for ad arbitrage

Chief Scientist

eBay

Keyword pricing optimization. Improving keyword bid pricing algorithm to update bids on millions of keywords purchased on Google AdWords. Automated daily updates. Setting bids for large tail of keywords with no conversion history, based on aggregation methods (NLP). Detection of new keywords.

Adometry (acquired by Google)

Built scoring system to measure traffic quality associated to keywords and publishers, for keyword ad networks, search engines, or advertisers (the customers). Scores were used for fraud detection, removal of bad publishers and keyword pricing. Detected major Botnets. Developed proofs of concept, attracting Looksmart and eBay as customers.

Chief AI Scientist, GenAItechLab.com

Full-stack entrepreneur with multimillion-dollar exit. Co-Founder at BondingAI.io. Author (Wiley, Elsevier). Creator of open-source Python libraries / Web APIs such as xLLM and GenAI-Evaluation. Builder of high performance in-memory agentic multi-LLMs for professional users and enterprise, with real-time finetuning, self-tuning, no weight, no training, no latency, no hallucinations, no GPU. Made from scratch, leading to replicable results, leveraging explainable AI, adopted by Fortune 100. With a focus on delivering concise, exhaustive, relevant, and in-depth search results, references, and links. 

Vincent is a leading expert, mathematician, patent owner, former post-doc at Cambridge University, former VC-funded executive, with 20+ years of corporate experience including CNET, NBC, Visa, Wells Fargo, Microsoft, eBay.  Vincent also founded and co-founded a few startups, including one with a successful exit (Data Science Central acquired by Tech Target). He is also a top influencer for brands such as Nvidia, SingleStore, and The Linux Foundation.

Privately funded GenAI and ML Optimization research lab, creating and disseminating open source intellectual property. Developing explainable and generative AI, synthetic data, computer vision, NLP and generic ML technologies. Check out our certifications, Python libraries, eBooks, articles, courses and podcasts, including our best seller "Synthetic Data and Generative  AI" (Elsevier). Visit our company website at MLtechniques.com.

Some LLMs don't have parameters in the classical sense. Our parameters are not billions of DNN weights, but instead dozens of core intuitive parameters attached to the various components. It makes fine-tuning different, easier, and faster. See https://lnkd.in/g6KPESC6

Chief AI Architect, Investor & Co-Founder

The best fine-tuning guide you'll find on arXiv this year.

Covers:
> NLP basics
> PEFT/LoRA/QLoRA techniques
> Mixture of Experts
> Seven-stage fine-tuning pipeline 
▸ 5-min daily newsletter for developers to keep up with AI: https://lnkd.in/gUB_3cWq 

▸ Source: https://t.co/x0ihiqfPFD

Processing LLMs/RAG Prompts 100x Faster with Optimized JSON https://lnkd.in/gX3AWmGi

JSON is great to handle input sources (Enterprise corpuses) used in LLMs and RAG. While well structured, it can be slow. You probably heard about TOON as an alternative, or my nested hashes that work much faster with JSON-like input optimized for speed, without vector database. In my case the architecture is Python-native, ideal for sparse tables such as token correlation tables, and similar to a light version of Berkeley DB. Yet cost-efficient without deep nesting that slows down processing. 

In this live session, you will see real use cases based on similar technology in production environment. This is not about my xLLM technology, but instead a different framework also adopted by companies small and large..

➡️ Presentation Overview

Modern applications rely heavily on semi-structured JSON data, but querying large JSON arrays can quickly become a performance bottleneck. In this session, Spandana will walk through how SingleStore’s new multi-value hash indexes make it possible to query JSON arrays with lightning-fast lookups. She’ll cover how this indexing innovation works under the hood, show real-world performance benchmarks, and demonstrate how developers can apply it to speed up analytical and operational workloads without restructuring data.

➡️ What You’ll Learn:

• How multi-value hash indexes improve query performance on JSON data
• Best practices for indexing nested and array fields
• A live demo of real-world JSON queries running 100x faster

Recording will be available to participants who are unable to attend the live event. 

➡️ Register at https://lnkd.in/gX3AWmGi

#JSON #LLMs #RAG

Wait till I publish my new technique predicting response with 95% R-squared on tabular data with 10k data points, each with 1000 features. Not a DNN, nothing to do with trees either, or with any kind of regression for that matter. 

You won't see it in any journal or conference as I publish on my blog only, at https://lnkd.in/gQ7_A9tx. Probably in a week or two. It's been fully tested at this point.

PhD in AI, author of 📖 The Hundred-Page Language Models Book and 📖 The Hundred-Page Machine Learning Book

This paper really is groundbreaking. It solves a long-standing embarrassment in machine learning: despite all the hype around deep learning, traditional tree-based methods (XGBoost, CatBoost, random forests, etc) have dominated tabular data—the most common data format in real-world applications—for two decades. Deep learning conquered images, text, and games, but spreadsheets remained stubbornly resistant.

This paper's (published in Nature by the way) main contribution is a foundation model that finally beats tree-based methods convincingly on small-to-medium datasets, and does so very fast. TabPFN in 2.8 seconds outperforms CatBoost tuned for 4 hours—a 5,000× speedup. That's not incremental; it's a different regime entirely.

The training approach is also fundamentally different. GPT trains on internet text; CLIP trains on image-caption pairs. TabPFN trains on entirely synthetic data—over 100 million artificial datasets generated from causal graphs.

TabPFN generates training data by randomly constructing directed acyclic graphs where each edge applies a random transformation (using neural networks, decision trees, discretization, or noise), then pushes random noise through the root nodes and lets it propagate through the graph—the intermediate values at various nodes become features, one becomes the target, and post-processing adds realistic messiness like missing values and outliers. By training on millions of these synthetic datasets with very different structures, the model learns general prediction strategies without ever seeing real data.

The inference mechanism is also unusual. Rather than finetuning or prompting, TabPFN performs both "training" and prediction in a single forward pass. You feed it your labeled training data and unlabeled test points together, and it outputs predictions immediately. There's no gradient descent at inference time—the model has learned how to learn from examples during pretraining.

The architecture respects tabular structure with two-way attention (across features within a row, then across samples within a column), unlike standard transformers that treat everything as a flat sequence.

So, the transformer has basically learned to do supervised learning.

Talk to the paper on ChapterPal: https://lnkd.in/e_rdrePu

Download the PDF: https://lnkd.in/ekrDRXZY

Python Code and Material from the Book "Synthetic Data and Generative AI"  -  GitHub Repository

Open article: GitHub - VincentGranville/Main: Main folder. Material related to my books on synthetic data and generative AI. Also contains documents blending components from several folders, or covering topics spanning across multiple folders.. by mltblog.com

GitHub - VincentGranville/Main: Main folder. Material related to my books on synthetic data and generative AI. Also contains documents blending components from several folders, or covering topics spanning across multiple folders..

Wolfram has been doing that for over 20 years. Last time I manually computed an integral manually was over 15 years ago. Same with differential equations and bunch of mechanical, mundane math computations no matter how advanced.

🤖 Generative AI Lead @ AWS ☁️ (150k+) | Startup Advisor | Public Speaker | AI Outsider | Founder Thinkfluencer AI

Holly s****

Do you know how long did it take to me to solve this integral I just wrote in a piece of paper?

2 secs by using this prompt: re-generate the image including the solution of the integral step by step.

I know, you must believed actually I solved the problem by myself but I didn’t… AI did it for me.

Then of course I asked for the detailed explanation with the logic on how to solve it. The goal is not to have the solution but to understand how to do it.

We (humans) usually have very short memory but this would have been crazy only 2 years ago…So for the ones who believe this is just another trend, trust me, this is not.

Imagine how a kid in the middle of Africa, without resources, can learn now mathematics. And take this to every single aspect of our lives.

Love it!!!!

#ai #genai

For the best data synthesizer on the market, check out my own. Free, much faster, much better quality. For tabular synthetic data. It also comes with the best evaluation metric in that context. See https://lnkd.in/gSG3wWKC

👋 Need synthetic data? You’re in luck.

Our team at NVIDIA just open sourced NeMo Data Designer, a general framework for generating high-quality synthetic data from scratch or based on your own seed data.

Data Designer has been my primary focus since my time at Gretel before we joined NVIDIA. I'm beyond thrilled that we are giving this library back to the community – I can't wait to see what you build with it ❤️

Highlights
- 🙌 Truly Open: Apache 2.0 license
- 🎨 UX: Designed to make SDG fun and flexible!
- 🤖 Model Agnostic: BYO model endpoints
- 🎲 Versatile: Multiple generation strategies
- 🔌 Extensible: Customize with plugins  

Get started here
- 📦 Install: pip install data-designer
- 💻 Code: https://lnkd.in/eJdAG7qG
- 📗 Docs: https://lnkd.in/eUAXnjxf

Huge shoutout to our amazing team at NVIDIA! Special thanks to Nabin Mulepati, Michael Knepper, Andre M., Kirit Thadaka, Anastasia Nesterenko, and Eric Tramel for diving deep into the weeds to help make this happen 🙏

Feel free to reach out if you have any questions!

Happy Building 🚀

#OpenSource #GenerativeAI #SyntheticData #NVIDIA

taplio

The Battle

Cassie Kozyrkov

Vincent Granville

Followers

Posts

Engagements

Likes

Performance ScoreCassie Kozyrkov's LinkedIn Performance

Performance ScoreVincent Granville's LinkedIn Performance

Ready to grow your LinkedIn presence?

taplio

The Battle

Cassie Kozyrkov

Vincent Granville

Followers

Posts

Engagements

Likes

Performance Score.css-goub6s{font-size:var(--chakra-fontSizes-md);font-weight:500;color:var(--chakra-colors-gray-600);display:block;margin-top:var(--chakra-space-1);}Cassie Kozyrkov's LinkedIn Performance

Performance ScoreVincent Granville's LinkedIn Performance

Ready to grow your LinkedIn presence?

Performance ScoreCassie Kozyrkov's LinkedIn Performance