Suggestions
Curtis Northcutt
CEO & Co-Founder @ Cleanlab. MIT PhD in CS. I build AI companies to empower people. Former Google, Oculus, Amazon, Facebook, Microsoft
Curtis Northcutt is an American computer scientist, entrepreneur, and CEO & Co-Founder of Cleanlab, an AI software company that focuses on improving machine learning model performance by automatically fixing data and label issues in real-world datasets.12
Education and Achievements
Curtis completed his PhD in Computer Science at the Massachusetts Institute of Technology (MIT), where he:
- Invented Cleanlab's algorithms for automatically finding and fixing label issues in datasets1
- Received the MIT Morris Levin Thesis Award3
- Amassed over 1,500 citations for his research2
Professional Experience
Before founding Cleanlab, Curtis gained experience working with several leading AI research groups, including:
- Oculus
- Amazon
- Microsoft
- NASA1
Cleanlab
As CEO and Co-Founder of Cleanlab, Curtis leads a company that:
- Provides enterprise-grade data curation solutions
- Improves the reliability of data-driven systems like LLMs, ML models, and analytics
- Works with various data types, including text, tabular, and image data2
Cleanlab's technology is based on confident learning, a sub-field of machine learning that Curtis invented during his PhD research.2
Recognition
Curtis and Cleanlab have received industry recognition:
- Featured in Forbes and CB Insights as an industry leader in AI
- Listed among the top 50 most innovative firms in AI and GenAI2
Curtis Northcutt's LinkedIn username is cgnorthcutt, as specified in the query.3
Highlights
We're getting to the point where a scrappy solo serial entrepreneur (or any engineering manager) can build a $1B+ company with this stack: AWS, Stripe, TLM. My biggest worry is with all the noise in GenAI, engineering teams and entrepreneurs might miss how cool TLM actually is.
4 takeaways about the Cleanlab TLM :
- TLM doesn't just detect hallucinations, it also EXPLAINS them.
- TLM improves the accuracy of Claude, GPT-4o... out-of-the-box.
- TLM works like any LLM API, but also tells you how trustworthy the LLM is with each response.
- On the HaluEval (Hallucination Evaluation) benchmark dataset, TLM catches 80% of hallucinations with <5% mistakes.
TLM addresses the biggest problem in GenAI: reliability -- the delta between experimental GenAI at your company and profitable automation for your business with RAG and LLMs. -- TLM is free to try and pay-per-token like any LLM API.
TLM makes RAG and GenAI actually work. And Cleanlab is the only one adding data quality and hallucination detection (with explanations) built-in to the LLM and RAG system directly.
Neat video with examples below.
#genai #dataquality #llms #rag #llm #generativeAI #gpt #hallucinations
Anthropic says Claude Sonnet 3.5 beats GPT-4o-- we tested it out in a real-world customer use case. Results below.
We benchmarked GPT-4o vs Sonnet 3.5 on our Banking Task Benchmark (intent recognition for categorizing customer support) using 3 approaches: zero-shot, few-shot, and Cleanlab-curated few shot.
In this task, Anthropic is both more accurate and more affordable. The improvement over GPT-4o is slight but consistent, and with Claude 3.5 Sonnet's per-token prompt cost at almost half the price of GPT-4o, it's cheaper to feed in large prompts and examples.
Both LLMs improved using Cleanlab to curate the few shot examples + reduced costs by removing and correcting problematic few shot examples.
Shout out to Nelson Auner for the analysis.
#genAI #llms #agi #artificialintelligence #machinelearning #datacuration #datacentricAI
