There are multiple ways to determine the cost of running a LLM (electricity use, compute cost, etc.), however, if you use a third-party LLM (a LLM-as-a-service) they typically charge you based on the tokens you use. Different vendors (OpenAI, Anthropic, Cohere, etc.) have different ways of counting the tokens, but for the sake of simplicity, well consider the cost to be based on the number of tokens processed by the LLM.
The most important part of this framework is the idea that different models cost different amounts. The authors of the paper conveniently assembled the below table highlighting the difference in cost, and the difference between them is significant. For example, AI21s output tokens cost an order of magnitude more than GPT-4s does in this table!
As a part of cost optimization we always need to figure out a way to optimize the answer quality while minimizing the cost. Typically, higher cost models are often higher performing models, able to give higher quality answers than lower cost ones. The general relationship can be seen in the below graph, with Frugal GPTs performance overlaid on top in red.
Using the vast cost difference between models, the researchers FrugalGPT system relies on a cascade of LLMs to give the user an answer. Put simply, the user query begins with the cheapest LLM, and if the answer is good enough, then it is returned. However, if the answer is not good enough, then the query is passed along to the next cheapest LLM.
The researchers used the following logic: if a less expensive model answers a question incorrectly, then it is likely that a more expensive model will give the answer correctly. Thus, to minimize costs the chain is ordered from least expensive to most expensive, assuming that quality goes up as you get more expensive.
This setup relies on reliably determining when an answer is good enough and when it isnt. To solve for this, the authors created a DistilBERT model that would take the question and answer then assign a score to the answer. As the DistilBERT model is exponentially smaller than the other models in the sequence, the cost to run it is almost negligible compared to the others.
One might naturally ask, if quality is most important, why not just query the best LLM and work on ways to reduce the cost of running the best LLM?
When this paper came out GPT-4 was the best LLM they found, yet GPT-4 did not always give a better answer than the FrugalGPT system! (Eagle-eyed readers will see this as part of the cost vs performance graph from before) The authors speculate that just as the most capable person doesnt always give the right answer, the most complex model wont either. Thus, by having the answer go through a filtering process with DistilBERT, you are removing any answers that arent up to par and increasing the odds of a good answer.
Consequently, this system not only reduces your costs but can also increase quality more so than just using the best LLM!
The results of this paper are fascinating to consider. For me, it raises questions about how we can go even further with cost savings without having to invest in further model optimization.
One such possibility is to cache all model answers in a vector database and then do a similarity search to determine if the answer in the cache works before starting the LLM cascade. This would significantly reduce costs by replacing a costly LLM operation with a comparatively less expensive query and similarity operation.
Additionally, it makes you wonder if outdated models can still be worth cost-optimizing, as if you can reduce their cost per token, they can still create value on the LLM cascade. Similarly, the key question here is at what point do you get diminishing returns by adding new LLMs onto the chain.
Read more:
FrugalGPT and Reducing LLM Operating Costs | by Matthew Gunton | Mar, 2024 - Towards Data Science
- Global Data Science Platform Market Report 2020 Industry Trends, Share and Size, Complete Data Analysis across the Region and Globe, Opportunities and... [Last Updated On: November 11th, 2020] [Originally Added On: November 11th, 2020]
- Data Science and Machine-Learning Platforms Market Size, Drivers, Potential Growth Opportunities, Competitive Landscape, Trends And Forecast To 2027 -... [Last Updated On: November 11th, 2020] [Originally Added On: November 11th, 2020]
- Industrial Access Control Market 2020-28 use of data science in agriculture to maximize yields and efficiency with top key players - TechnoWeekly [Last Updated On: November 11th, 2020] [Originally Added On: November 11th, 2020]
- IPG Unveils New-And-Improved Copy For Data: It's Not Your Father's 'Targeting' 11/11/2020 - MediaPost Communications [Last Updated On: November 11th, 2020] [Originally Added On: November 11th, 2020]
- Risks and benefits of an AI revolution in medicine - Harvard Gazette [Last Updated On: November 11th, 2020] [Originally Added On: November 11th, 2020]
- UTSA to break ground on $90 million School of Data Science and National Security Collaboration Center - Construction Review [Last Updated On: November 11th, 2020] [Originally Added On: November 11th, 2020]
- Addressing the skills shortage in data science and analytics - IT-Online [Last Updated On: November 11th, 2020] [Originally Added On: November 11th, 2020]
- Data Science Platform Market Research Growth by Manufacturers, Regions, Type and Application, Forecast Analysis to 2026 - Eurowire [Last Updated On: November 11th, 2020] [Originally Added On: November 11th, 2020]
- 2020 AI and Data Science in Retail Industry Ongoing Market Situation with Manufacturing Opportunities: Amazon Web Services, Baidu Inc., BloomReach... [Last Updated On: November 11th, 2020] [Originally Added On: November 11th, 2020]
- Endowed Chair of Data Science job with Baylor University | 299439 - The Chronicle of Higher Education [Last Updated On: November 11th, 2020] [Originally Added On: November 11th, 2020]
- Data scientists gather 'chaos into something organized' - University of Miami [Last Updated On: November 11th, 2020] [Originally Added On: November 11th, 2020]
- AI Update: Provisions in the National Defense Authorization Act Signal the Importance of AI to American Competitiveness - Lexology [Last Updated On: January 12th, 2021] [Originally Added On: January 12th, 2021]
- Healthcare Innovations: Predictions for 2021 Based on the Viewpoints of Analytics Thought Leaders and Industry Experts | Quantzig - Business Wire [Last Updated On: January 12th, 2021] [Originally Added On: January 12th, 2021]
- Poor data flows hampered governments Covid-19 response, says the Science and Technology Committee - ComputerWeekly.com [Last Updated On: January 12th, 2021] [Originally Added On: January 12th, 2021]
- Ilia Dub and Jasper Yip join Oliver Wyman's Asia partnership - Consultancy.asia [Last Updated On: January 12th, 2021] [Originally Added On: January 12th, 2021]
- Save 98% off the Complete Excel, VBA, and Data Science Certification Training Bundle - Neowin [Last Updated On: January 12th, 2021] [Originally Added On: January 12th, 2021]
- Data Science for Social Good Programme helps Ofsted and World Bank - India Education Diary [Last Updated On: January 12th, 2021] [Originally Added On: January 12th, 2021]
- Associate Professor of Fisheries Oceanography named a Cooperative Institute for the North Atlantic Region (CINAR) Fellow - UMass Dartmouth [Last Updated On: January 12th, 2021] [Originally Added On: January 12th, 2021]
- Rapid Insight To Host Free Webinar, Building on Data: From Raw Piles to Data Science - PR Web [Last Updated On: January 12th, 2021] [Originally Added On: January 12th, 2021]
- This Is the Best Place to Buy Groceries, New Data Finds | Eat This Not That - Eat This, Not That [Last Updated On: January 12th, 2021] [Originally Added On: January 12th, 2021]
- Which Technology Jobs Will Require AI and Machine Learning Skills? - Dice Insights [Last Updated On: January 12th, 2021] [Originally Added On: January 12th, 2021]
- Companies hiring data scientists in NYC and how much they pay - Business Insider [Last Updated On: January 12th, 2021] [Originally Added On: January 12th, 2021]
- Calling all rock stars: hire the right data scientist talent for your business - IDG Connect [Last Updated On: January 12th, 2021] [Originally Added On: January 12th, 2021]
- How Professors Can Use AI to Improve Their Teaching In Real Time - EdSurge [Last Updated On: January 12th, 2021] [Originally Added On: January 12th, 2021]
- BCG GAMMA, in Collaboration with Scikit-Learn, Launches FACET, Its New Open-Source Library for Human-Explainable Artificial Intelligence - PRNewswire [Last Updated On: January 12th, 2021] [Originally Added On: January 12th, 2021]
- Data Science Platform Market Insights, Industry Outlook, Growing Trends and Demands 2020 to 2025 The Courier - The Courier [Last Updated On: January 31st, 2021] [Originally Added On: January 31st, 2021]
- UBIX and ORS GROUP announce partnership to democratize advanced analytics and AI for small and midmarket organizations - PR Web [Last Updated On: January 31st, 2021] [Originally Added On: January 31st, 2021]
- Praxis Business School is launching its Post Graduate Program in Data Engineering in association with Knowledge Partners - Genpact and LatentView... [Last Updated On: January 31st, 2021] [Originally Added On: January 31st, 2021]
- What's So Trendy about Knowledge Management Solutions Market That Everyone Went Crazy over It? | Bloomfire, CSC (American Productivity & Quality... [Last Updated On: January 31st, 2021] [Originally Added On: January 31st, 2021]
- Want to work in data? Here are 6 skills you'll need Just now - Siliconrepublic.com [Last Updated On: January 31st, 2021] [Originally Added On: January 31st, 2021]
- Data, AI and babies - BusinessLine [Last Updated On: January 31st, 2021] [Originally Added On: January 31st, 2021]
- Here's how much Amazon pays its Boston-based employees - Business Insider [Last Updated On: January 31st, 2021] [Originally Added On: January 31st, 2021]
- Datavant and Kythera Increase the Value Of Healthcare Data Through Expanded Data Science Platform Partnership - GlobeNewswire [Last Updated On: January 31st, 2021] [Originally Added On: January 31st, 2021]
- O'Reilly Analysis Unveils Python's Growing Demand as Searches for Data Science, Cloud, and ITOps Topics Accelerate - Business Wire [Last Updated On: January 31st, 2021] [Originally Added On: January 31st, 2021]
- Book Review: Hands-On Exploratory Data Analysis with Python - insideBIGDATA [Last Updated On: January 31st, 2021] [Originally Added On: January 31st, 2021]
- The 12 Best R Courses and Online Training to Consider for 2021 - Solutions Review [Last Updated On: January 31st, 2021] [Originally Added On: January 31st, 2021]
- Software AG's TrendMiner 2021.R1 Release Puts Data Science in the Hands of Operational Experts - Yahoo Finance [Last Updated On: January 31st, 2021] [Originally Added On: January 31st, 2021]
- The chief data scientist: Who they are and what they do - Siliconrepublic.com [Last Updated On: January 31st, 2021] [Originally Added On: January 31st, 2021]
- Berkeley's data science leader dedicated to advancing diversity in computing - UC Berkeley [Last Updated On: January 31st, 2021] [Originally Added On: January 31st, 2021]
- Awful Earnings Aside, the Dip in Alteryx Stock Is Worth Buying - InvestorPlace [Last Updated On: February 12th, 2021] [Originally Added On: February 12th, 2021]
- Why Artificial Intelligence May Not Offer The Business Value You Think - CMSWire [Last Updated On: February 12th, 2021] [Originally Added On: February 12th, 2021]
- Getting Prices Right in 2021 - Progressive Grocer [Last Updated On: February 12th, 2021] [Originally Added On: February 12th, 2021]
- Labelbox raises $40 million for its data labeling and annotation tools - VentureBeat [Last Updated On: February 12th, 2021] [Originally Added On: February 12th, 2021]
- How researchers are using data science to map wage theft - SmartCompany.com.au [Last Updated On: February 12th, 2021] [Originally Added On: February 12th, 2021]
- Ready to start coding? What you need to know about Python - TechRepublic [Last Updated On: February 12th, 2021] [Originally Added On: February 12th, 2021]
- Women changing the face of science in the Middle East and North Africa - The Jerusalem Post [Last Updated On: February 12th, 2021] [Originally Added On: February 12th, 2021]
- Mapping wage theft with data science - The Mandarin [Last Updated On: February 12th, 2021] [Originally Added On: February 12th, 2021]
- Data Science Platform Market 2021 Analysis Report with Highest CAGR and Major Players like || Dataiku, Bridgei2i Analytics, Feature Labs and More KSU... [Last Updated On: February 12th, 2021] [Originally Added On: February 12th, 2021]
- Data Science Impacting the Pharmaceutical Industry, 2020 Report: Focus on Clinical Trials - Data Science-driven Patient Selection & FDA... [Last Updated On: February 12th, 2021] [Originally Added On: February 12th, 2021]
- App Annie Sets New Bar for Mobile Analytics with Data Science Innovations - PRNewswire [Last Updated On: February 12th, 2021] [Originally Added On: February 12th, 2021]
- Data Science and Analytics Market 2021 to Showing Impressive Growth by 2028 | Industry Trends, Share, Size, Top Key Players Analysis and Forecast... [Last Updated On: February 12th, 2021] [Originally Added On: February 12th, 2021]
- How Can We Fix the Data Science Talent Shortage? Machine Learning Times - The Predictive Analytics Times [Last Updated On: February 14th, 2021] [Originally Added On: February 14th, 2021]
- Opinion: How to secure the best tech talent | Human Capital - Business Chief [Last Updated On: February 14th, 2021] [Originally Added On: February 14th, 2021]
- Following the COVID science: what the data say about the vaccine, social gatherings and travel - Chicago Sun-Times [Last Updated On: February 14th, 2021] [Originally Added On: February 14th, 2021]
- Automated Data Science and Machine Learning Platforms Market Technological Growth and Precise Outlook 2021- Microsoft, MathWorks, SAS, Databricks,... [Last Updated On: February 14th, 2021] [Originally Added On: February 14th, 2021]
- 9 investors discuss hurdles, opportunities and the impact of cloud vendors in enterprise data lakes - TechCrunch [Last Updated On: February 14th, 2021] [Originally Added On: February 14th, 2021]
- Rapid Insight to Present at Data Science Salon's Healthcare, Finance, and Technology Virtual Event - PR Web [Last Updated On: February 14th, 2021] [Originally Added On: February 14th, 2021]
- Aunalytics Acquires Naveego to Expand Capabilities of its End-to-End Cloud-Native Data Platform to Enable True Digital Transformation for Customers -... [Last Updated On: February 22nd, 2021] [Originally Added On: February 22nd, 2021]
- Tech Careers: In-demand Courses to watch out for a Lucrative Future - Big Easy Magazine [Last Updated On: February 22nd, 2021] [Originally Added On: February 22nd, 2021]
- Willis Towers Watson enhances its human capital data science capabilities globally with the addition of the Jobable team - GlobeNewswire [Last Updated On: February 22nd, 2021] [Originally Added On: February 22nd, 2021]
- Global Data Science Platform Market 2021 Industry Insights, Drivers, Top Trends, Global Analysis And Forecast to 2027 KSU | The Sentinel Newspaper -... [Last Updated On: February 22nd, 2021] [Originally Added On: February 22nd, 2021]
- A Comprehensive Guide to Scikit-Learn - Built In [Last Updated On: February 22nd, 2021] [Originally Added On: February 22nd, 2021]
- Industry VoicesBuilding ethical algorithms to confront biases: Lessons from Aotearoa New Zealand - FierceHealthcare [Last Updated On: February 22nd, 2021] [Originally Added On: February 22nd, 2021]
- How Intel Employees Volunteered Their Data Science Expertise To Help Costa Rica Save Lives During the Pandemic - CSRwire.com [Last Updated On: February 22nd, 2021] [Originally Added On: February 22nd, 2021]
- Learn About Innovations in Data Science and Analytic Automation on an Upcoming Episode of the Advancements Series - Yahoo Finance [Last Updated On: February 22nd, 2021] [Originally Added On: February 22nd, 2021]
- Symposium aimed at leveraging the power of data science for promoting diversity - Penn State News [Last Updated On: February 22nd, 2021] [Originally Added On: February 22nd, 2021]
- Rochester to advance research in biological imaging through new grant - University of Rochester [Last Updated On: February 22nd, 2021] [Originally Added On: February 22nd, 2021]
- SoftBank Joins Initiative to Train Diverse Talent in Data Science and AI - Entrepreneur [Last Updated On: February 22nd, 2021] [Originally Added On: February 22nd, 2021]
- Participating in SoftBank/ Correlation One Initiative - Miami - City of Miami [Last Updated On: February 22nd, 2021] [Originally Added On: February 22nd, 2021]
- Increasing Access to Care with the Help of Big Data | Research Blog - Duke Today [Last Updated On: February 22nd, 2021] [Originally Added On: February 22nd, 2021]
- Heres how Data Science & Business Analytics expertise can put you on the career expressway - Times of India [Last Updated On: March 14th, 2021] [Originally Added On: March 14th, 2021]
- Yelp data shows almost half a million new businesses opened during the pandemic - CNBC [Last Updated On: March 14th, 2021] [Originally Added On: March 14th, 2021]
- Postdoctoral Position in Transient and Multi-messenger Astronomy Data Science in Greenbelt, MD for University of MD Baltimore County/CRESST II -... [Last Updated On: March 14th, 2021] [Originally Added On: March 14th, 2021]
- DefinedCrowd CEO Daniela Braga on the future of AI, training data, and women in tech - GeekWire [Last Updated On: March 14th, 2021] [Originally Added On: March 14th, 2021]
- Gartner: AI and data science to drive investment decisions rather than "gut feel" by mid-decade - TechRepublic [Last Updated On: March 14th, 2021] [Originally Added On: March 14th, 2021]
- Jupyter has revolutionized data science, and it started with a chance meeting between two students - TechRepublic [Last Updated On: March 14th, 2021] [Originally Added On: March 14th, 2021]
- Working at the intersection of data science and public policy | Penn Today - Penn Today [Last Updated On: March 14th, 2021] [Originally Added On: March 14th, 2021]
- The Future of AI: Careers in Machine Learning - Southern New Hampshire University [Last Updated On: April 4th, 2021] [Originally Added On: April 4th, 2021]
- SMU meets the opportunities of the data-driven world with cutting-edge research and data science programs - The Dallas Morning News [Last Updated On: April 4th, 2021] [Originally Added On: April 4th, 2021]
- Data, Science, and Journalism in the Age of COVID - Pulitzer Center on Crisis Reporting [Last Updated On: April 4th, 2021] [Originally Added On: April 4th, 2021]