Skip to content

AI-ML-Notes.md

Latest commit

 

History

History
299 lines (246 loc) · 42.3 KB

AI-ML-Notes.md

File metadata and controls

299 lines (246 loc) · 42.3 KB

Random Artificial Intelligence & Machine Learning Notes

There is an AI theme applied to virtually anything at work. My experience says that signals simple hype (or executive echo chamber) but my reading also suggests that this hum of AI-everything is ignored at our peril. Maybe you have analogous sensations? If not, maybe think some more...
Bill Gates said that AI is "almost like getting free white-collar workers" (Jan/Feb 2025 AARP Bulletin, p. 33).

This is largely-neglected alpha interest work. After percolating in this format for a while, I'll figure out how to better prioritize this reading and content. You can do better than to invest your time here!

In fact, you might want to start at Hari Sekhon's excellent "AI - Artificial Intelligence"
and if you need something more specific "Awesome ChatGPT Prompts" or just https://prompts.chat/

AI Data Center Resource Demands

An analysis from: Prunelia Stuart Senior Program Project Manager | Hyperscale Data Centers | Tech, Construction & Change Management Expert | CAPEX Optimisation | Delivering Cross-Functional Projects at Scale

What Does It Really Cost To Build An Ai Data Center?
Twenty years ago, a “big” data center drew 20 megawatts. Today, hyperscalers are sketching out multi-gigawatt campuses. OpenAI’s Stargate and Meta’s Hyperion? Each could reach 5 GW, enough to power 4–5 million US homes, or all the homes in Greater London combined.

Here’s the rough math:

 • Facility shell: ~$1B per 100 MW → $10B per 1 GW  
 • Compute (GPUs, servers, interconnect): ~$20B per 1 GW  
 • Add dedicated generation (gas/renewables): +$3-4B

So, 1 GW ≈ $30B.
Scale that to 5 GW: $100-150B, depending on density, hardware mix, and power strategy.

And this is where the story sharpens: the real cost is compute.

 • A single AI rack can clear $500K-$3M
 • Nvidia’s GB200/H100-class systems: $2.6M-$3M per NVL72 rack, trending higher for next-gen (𝘏𝘚𝘉𝘊 𝘢𝘭𝘴𝘰 𝘧𝘰𝘳𝘦𝘤𝘢𝘴𝘵𝘴 𝘕𝘷𝘪𝘥𝘪𝘢’𝘴 𝘧𝘶𝘵𝘶𝘳𝘦 𝘝𝘦𝘳𝘢 𝘙𝘶𝘣𝘪𝘯 𝘕𝘝𝘓144 𝘸𝘪𝘭𝘭 𝘣𝘦 𝘱𝘳𝘪𝘤𝘦𝘥 𝘢𝘵 $3.2 𝘮𝘪𝘭𝘭𝘪𝘰𝘯 𝘢𝘯𝘥 𝘵𝘩𝘦 𝘙𝘶𝘣𝘪𝘯 𝘜𝘭𝘵𝘳𝘢 𝘕𝘝𝘓576 𝘸𝘪𝘭𝘭 𝘤𝘰𝘴𝘵 $8.8 𝘮𝘪𝘭𝘭𝘪𝘰𝘯)
 • AI servers draw 10x more power than traditional cloud gear

Zoom out and the picture is staggering:

 • McKinsey projects $6.7T in global data center CAPEX by 2030, with $5.2T for AI-capable sites  
 • Microsoft, Alphabet, Amazon, Meta alone will spend >$300B on capex in 2025  
 • One TechCrunch analysis: a single leading AI DC in 6 years could cost $200B and draw 9 GW

The next 5 years?

 • OpenAI’s Stargate: $100B deployment now, $500B over four years  
 • Meta’s Hyperion: 2 GW online by 2030, scaling to 5 GW  
 • Abilene, TX: an 875-acre “AI factory” campus planned at 1.2 GW  
 • Start Campus, Portugal: a 1.2GW data center campus across six buildings by 2030

WWhat it means for leaders:

 • Power is key, requiring 24/7 clean firm power to secure compute  
 • Chips and interconnects drive timelines as much as construction  
 • Design for modularity and density  
 • Finance meets physics: every gigawatt is a $30B+ bet intertwined with grid, water, workforce, and regulation.

Mapping out AI buildouts requires measuring assumptions against these kinds of magnitudes, as AI data centers are becoming nationwide infrastructure.

References:
𝘈𝘴𝘵𝘦𝘳𝘪𝘴𝘬 (𝘧𝘪𝘷𝘦-𝘎𝘞 𝘋𝘊 𝘮𝘢𝘵𝘩) | 𝘔𝘤𝘒𝘪𝘯𝘴𝘦𝘺 ($6.7𝘛 𝘣𝘺 2030) | 𝘍𝘛/𝘎𝘢𝘳𝘵𝘯𝘦𝘳 (2025 𝘋𝘊 𝘴𝘱𝘦𝘯𝘥) | 𝘖𝘱𝘦𝘯𝘈𝘐/𝘚𝘵𝘢𝘳𝘨𝘢𝘵𝘦 | 𝘛𝘦𝘤𝘩𝘊𝘳𝘶𝘯𝘤𝘩 (𝘔𝘦𝘵𝘢 𝘏𝘺𝘱𝘦𝘳𝘪𝘰𝘯; $200𝘉/9-𝘎𝘞 𝘰𝘶𝘵𝘭𝘰𝘰𝘬) | 𝘋𝘢𝘵𝘢𝘊𝘦𝘯𝘵𝘦𝘳𝘋𝘺𝘯𝘢𝘮𝘪𝘤𝘴 (5-𝘎𝘞 𝘥𝘦𝘤𝘬) | 𝘋𝘢𝘵𝘢𝘊𝘦𝘯𝘵𝘦𝘳𝘔𝘢𝘨𝘢𝘻𝘪𝘯𝘦 (𝘈𝘣𝘪𝘭𝘦𝘯𝘦 1.2-𝘎𝘞) | 𝘓𝘶𝘮𝘦𝘯𝘢𝘭𝘵𝘢 & 𝘊𝘺𝘧𝘶𝘵𝘶𝘳𝘦 (𝘳𝘢𝘤𝘬/𝘈𝘐 𝘴𝘦𝘳𝘷𝘦𝘳 𝘤𝘰𝘴𝘵𝘴) | 𝘋𝘪𝘨𝘪𝘵𝘪𝘮𝘦𝘴, 𝘏𝘚𝘉𝘊 & 𝘉𝘢𝘳𝘳𝘰𝘯’𝘴 (𝘕𝘝𝘓576 / 𝘕𝘝𝘓72+ 𝘱𝘳𝘪𝘤𝘪𝘯𝘨)

Again, this "What Does It Really Cost To Build An Ai Data Center?" section is not my work, but is an analysis posted on LinkedIn by Prunelia Stuart Senior Program Project Manager | Hyperscale Data Centers

Resource Consumption:

  • "As they race to capitalize on a craze for generative AI, leading tech developers including Microsoft, OpenAI and Google have acknowledged that growing demand for their AI tools carries hefty costs, from expensive semiconductors to an increase in water consumption." ... "In its latest environmental report, Microsoft disclosed that its global water consumption spiked 34% from 2021 to 2022 (to nearly 1.7 billion gallons, or more than 2,500 Olympic-sized swimming pools), a sharp increase compared to previous years that outside researchers tie to its AI research." ... "Google reported a 20% growth in water use in the same period, which Ren also largely attributes to its AI work. Google’s spike wasn’t uniform -- it was steady in Oregon where its water use has attracted public attention, while doubling outside Las Vegas. It was also thirsty in Iowa, drawing more potable water to its Council Bluffs data centers than anywhere else." ... "“It’s fair to say the majority of the growth is due to AI,” including “its heavy investment in generative AI and partnership with OpenAI,” said Shaolei Ren, a researcher at the University of California, Riverside who has been trying to calculate the environmental impact of generative AI products such as ChatGPT. In a paper due to be published later this year, Ren’s team estimates ChatGPT gulps up 500 milliliters of water (close to what’s in a 16-ounce water bottle) every time you ask it a series of between 5 to 50 prompts or questions. The range varies depending on where its servers are located and the season. The estimate includes indirect water usage that the companies don’t measure — such as to cool power plants that supply the data centers with electricity." ...In late May, 2023 "Microsoft’s president, Brad Smith, disclosed that it had built its “advanced AI supercomputing data center” in Iowa, exclusively to enable OpenAI to train what has become its fourth-generation model, GPT-4. The model now powers premium versions of ChatGPT and some of Microsoft’s own products and has accelerated a debate about containing AI’s societal risks." ..."describing it as a “single system” with more than 285,000 cores of conventional semiconductors, and 10,000 graphics processors — a kind of chip that’s become crucial to AI workloads. Experts have said it can make sense to “pretrain” an AI model at a single location because of the large amounts of data that need to be transferred between computing cores." "In July 2022, the month before OpenAI says it completed its training of GPT-4, Microsoft pumped in about 11.5 million gallons of water to its cluster of Iowa data centers, according to the West Des Moines Water Works. That amounted to about 6% of all the water used in the district, which also supplies drinking water to the city’s residents." https://apnews.com/article/chatgpt-gpt4-iowa-ai-water-consumption-microsoft-f551fde98083d17a7e8d904f8be822c4 and https://query.prod.cms.rt.microsoft.com/cms/api/am/binary/RW15mgm

Ian Frisch wrote in the 20 Sept 2025 NYT Deal Book:

The big bump in data center activity has been linked to distorted residential power readings across the country. And according to the International Energy Agency, a 100-megawatt data center, which uses water to cool servers, consumes roughly two million liters of water per day, equivalent to 6,500 households. This puts strain on water supply for nearby residential communities, a majority of which, according to Bloomberg News, are already facing high levels of water stress.

“I think we’re in that era right now with A.I. models where it’s just who can make the bigger and better one,” said Vijay Gadepally, a senior scientist at the Lincoln Laboratory Supercomputing Center at the Massachusetts Institute of Technology. “But we haven’t actually stopped to think about, Well, OK, is this actually worth it?”
https://www.nytimes.com/2025/09/20/business/dealbook/data-centers-ai.html

For an up-to-date list of "all" large language models see: https://llmmodels.org/

Where do we encounter AI/ML?

  • Games
  • Investing
  • Logistics systems
  • Medical diagnosis
  • Search
  • Traffic management
  • Autonomous driving
  • Language translation
  • Chatbots (general)
  • Social media (chatbots & logic used to drive platform profits)
  • Interactive personal assistance
  • Interactive tutoring
  • Consumer/customer/consumption suggestions, advice, and support (and product/service reviews)
  • Customer support
  • Application development
  • Science analyses and "proposals"
  • Crime (criminals use available tools, AI assistance in fraud, cybercrime and more)

Required Workflow(s):

Starter list outline below:

  • Data Collection (Acquire New Data as Needed)
  • Resist hostile inputs (Mission Integrity, Trust-Building and Transparency)
  • Understanding the Data
  • Data Governance (Mission Integrity, Trust-Building and Transparency)
    • Copyright/trademark/servicemark/etc. management
    • "Documenting" the data (at many levels) (Trust-Building and Transparency)
    • Legal review and approval (Mission Integrity, Trust-Building and Transparency)
  • Understanding the Target Business (including model performance metrics and business KPIs)
  • Feature Engineering & Data Preparation (Analysis, Cleaning, Repair; resist hostile inputs)
  • Model Training (resist hostile inputs)
    • Quality testing, quality controls and reporting (Mission Integrity and Trust-Building)
  • Model Evaluation
    • Quality testing, quality controls and reporting (Mission Integrity and Trust-Building)
  • Model Deployment (resist hostile inputs)
    • Quality testing, quality controls and reporting (Mission Integrity and Trust-Building)
    • Vulnerability testing and reporting (Mission Integrity and Trust-Building)
  • Monitor Deployed AI Model Quality (Mission Integrity and Trust-Building)
    • Vulnerability testing and reporting (Mission Integrity and Trust-Building)
  • Monitor Deployed AI Model Fairness (Mission Integrity and Trust-Building)
  • Monitor Deployed AI Model Explainability (Trust-Building and Transparency)

List above began with a model IBM AI lifecycle

Summer 2025: Five leading AI models

  • OpenAI’s ChatGPT
  • Anthropic’s Claude
  • X/xAI’s Grok (owned by Elon Musk)
  • Google’s Gemini
  • Perplexity.

Maybe Leave This Page Now and Read "This" for Context!

Yes. Go read this now...
"AI Is a Lot of Work."
As the technology becomes ubiquitous, a vast tasker underclass is emerging — and not going anywhere.
By Josh Dzieza, June 20, 2023. New York Magazine.
https://nymag.com/intelligencer/article/ai-artificial-intelligence-humans-technology-business-factory.html

Resources:

How AI is impacting 700 professions — and might impact yours. By Youyou Zhou. 28 July 2025. read it in the Washington Post

Natural Language Processing and Common Sense

Types of Reasoning:

  • Human intuition and instinct: Its processes are unconscious, effortless, fast, associative, automatic-pilot, slow-learning. Its content is conceptual, temporal (for example before and after) and can be evoked by language. D.Kahneman1
  • Rational thinking: Its processes are require effort, and are slow, logical, governed by rules, often serial, sometimes flexible, and indecisive. Its content is conceptual, temporal and can be evoked by language. D.Kahneman1

Timeline Ideas:

  • 2025 :List of LLM Providers -- https://docs.litellm.ai/docs/providers
  • 2024: Feb 8th Google changed the name of its chatbot from Bard to Gemini and introduced a smartphone app Gemini -- like a talking digital assistant and a conversational chatbot.
  • 2023: Google releases a conversational AI service called "Bard" (or go directly to https://bard.google.com/) to compte with ChatGPT. ...There are many competitors: Bing AI (or go directly to https://www.bing.com/new) (research), Claud/Claude 2 (writing), Google Docs AI (Notes), BlueWillow (images), ChatGPT+ (research, writing, search), Notion AI (Notes), Midjourney (images), Canva AI (art/design), Le Chat (Mistral AI, free access at https://chat.mistral.ai)
  • 2022: OpenAI released ChatGPT - an a LLM conversational AI service that accepts inputs in plain language and spits out human-friendly, content relevant results.
  • 2021: Language Model for Dialogue Applications (or LaMDA for short)
  • 2020: OpenAI GPT-3 a language model that generates text using pre-trained algorithms.
  • 2019:
  • 2018: 'Human-level performance' on reading comprehension (limited data set & limited definition of comprehension) -- Alibaba language processing AI beats top humans at a Stanford University reading and comprehension test, scoring 82.44 against 82.30 on 100,000 questions.
  • 2017:
  • 2016:
  • 2015:
    • 'Super-human performance' on image captioning
    • 'Super-human performance' on object recognition
  • 2014:
  • 2013: Robot HRP-2, built by SCHAFT Inc. of Japan, a subsidiary of Google, defeats 15 teams to win DARPA's Robotics Challenge Trials by completing disaster response tasks, including driving a vehicle, walking over debris, climbing a ladder, removing debris, walking through doors, cutting a wall, closing valves, and connecting a hose.
  • 2012:
  • 2011: IBM's Watson wins Jeopardy. Apple introduces Siri.
  • 2010:
  • 2005: Computer scientist Sebastian Thrun and a Stanford Artificial Intelligence Laboratory team build a driverless car (Stanley) that was the first autonomous vehicle to complete a DARPA Grand Challenge 132-mile course in the Mojave Desert.
  • 1980: The "neocognitron", which introduced the two basic types of layers in Convolutional neural networks (CNNs) -- convolutional layers and downsampling layers -- was introduced by Kunihiko Fukushima in 1980.
  • 1955: "A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence. August 31, 1955."

Timeline Notes:

Some Definitions:

"AI guidance, terms added to AP Stylebook." Aug. 17, 2023, by Nicole Meir. https://blog.ap.org/ai-guidance-terms-added-to-ap-stylebook

Guidance on how to cover artificial intelligence and 10 key AI terms were added today to the AP Stylebook, to help journalists accurately explain the potential, inherent risks and varying effects of AI and generative AI models.

  • Algorithmic bias "A term used to describe the negative impacts of AI tools that draw from large datasets that are skewed by historical or selection bias." (from the AP Stylebook)

  • Common Sense "It's the unspoken, implicit knowledge that you and I have. It's so obvious that we often don't talk about it." Interview with Yejin Choi

  • Common Sense is the basic level of 'practical' knowledge and reasoning concerning everyday situations and events that are commonly shared among most people. It is essential for humans to live and interact with each other in a reasonable and safe way. Yejin Choi at Common Sense is a difficult challenge and integration into AI implementations still suffer from:

  • Generative artificial intelligence (generative AI) "Generative AI is the technology to create new content by utilizing existing text, audio files, or images. With generative AI, computers detect the underlying pattern related to the input and produce similar content. This is in contrast to most other AI techniques where the AI model attempts to solve a problem which has a single answer (e.g. a classification or prediction problem)." (from the AIMultiple) Various techniques include (but are not limited to):

    • Transformers: "Transformers, such as GPT-3, LaMDA, Wu-Dao and ChatGPT imitate cognitive attention and differentially measure the significance of the input data parts. They are trained to understand the language or image, learn some classification tasks and generate texts or images from massive datasets." (from the AIMultiple)
    • Generative adversarial networks (GANs): "GANs are two neural networks: a generator and a discriminator that pit against each other to find equilibrium between the two networks: The generator network is responsible for generating new data or content resembling the source data. The discriminator network is in charge of differentiating between the source and the generated data in order to recognize what is closer to the original data." (from the AIMultiple)
    • Variational auto-encoders: "The encoder encodes the input into compressed code while the decoder reproduces the initial information from this code. If chosen and trained correctly, this compressed representation stores the input data distribution in a much smaller dimensional representation." (from the AIMultiple)
    • Hallucinate: When a chatbot fabricates information or emits definitive statements on uncertain fact sets from Cade Metz in the NYT.
      Always consider "the models’ tendency to generate inaccurate responses to queries." (from the AP Stylebook)

  • Machine learning (ML) is often viewed as a subset of Artificial Intelligence. ML employes previously collected data to predict outcomes. ML models may depend upon direct human inputs (training or supervision), or not, depending on their algorithms and their level of training maturity. ML enables software (and supporting infrastructure) to build upon training/experience and improvise suggestions or results. At some point, modern implementations involve AI systems helping humans oversee other AI" to enable economical ML/AI at scale.

  • Natural language processing (NLP) is a growing field within artificial intelligence. The fundamental goal of NLP is to program computers capable of human-level understanding of natural language. Common NLP applications include personal assistants and chatbots, automatic translation, question answering, sentiment analysis and summarization. Among the main challenges of NLP research is that human language is often ambiguous and underspecified. A person processing language relies heavily on their commonsense knowledge and reasoning abilities to resolve these ambiguities and complete missing information. Machine learning based NLP models, on the other hand, lack this commonsense and often make absurd mistakes. From Dr. Vered Shwartz in her 2022 NLP course description

  • Convolutional neural network (CNN)"In deep learning, a convolutional neural network (CNN, or ConvNet) is a class of artificial neural network (ANN), most commonly applied to analyze visual imagery. CNNs are also known as Shift Invariant or Space Invariant Artificial Neural Networks (SIANN), based on the shared-weight architecture of the convolution kernels or filters that slide along input features and provide translation-equivariant responses known as feature maps. Counter-intuitively, most convolutional neural networks are not invariant to translation, due to the downsampling operation they apply to the input. They have applications in image and video recognition, recommender systems, image classification, image segmentation, medical image analysis, natural language processing, brain–computer interfaces, and financial time series."

  • Generative adversarial networks (GANs) and reinforcement learning endow deep networks with the ability to produce artificial content such as fake images that pass for the real thing. "GANs consist of two interlocked components—a generator, responsible for creating realistic content, and a discriminator, tasked with distinguishing the output of the generator from naturally occurring content. The two learn from each other, becoming better and better at their respective tasks over time" AI100Report_MT_10, page 12.

  • Retrieval-Augmented Generation (RAG), focusing on high-dimensional embeddings crucial for machine learning... What is this?

  • Recurrent neural network (RNN) "is a class of artificial neural networks where connections between nodes can create a cycle, allowing output from some nodes to affect subsequent input to the same nodes. This allows it to exhibit temporal dynamic behavior. Derived from feedforward neural networks, RNNs can use their internal state (memory) to process variable length sequences of inputs. This makes them applicable to tasks such as unsegmented, connected handwriting recognition[4] or speech recognition."

  • Large Language Model (LLM)

  • Reinforcement Learning from Human Feedback (RLHF) https://en.wikipedia.org/wiki/Reinforcement_learning_from_human_feedback

  • Sensibleness and Specificity Average (SSA)

  • Training data "refers to the vast corpus of information used to train large language models. This data can include text, images, audio, or video. Generative models learn patterns from this data, enabling them to generate new content matching the input data’s complexity, style, and structure." (from AIMultiple) "Explaining the term, and advising to consider the specific information training data may contain as machine learning models learn to find patterns from these datasets. The types of training data used in different AI tools can vary widely." (from the AP Stylebook)

  • NIST Taxonomy of Attacks, Defenses, and Consequences in Adversarial Machine Learning: NIST Taxonomy of Attacks, Defenses, and Consequences in Adversarial Machine Learning: (from: https://nvlpubs.nist.gov/nistpubs/ir/2019/NIST.IR.8269-draft.pdf accessed 2024-05-08)

Some References:

"The emergence of products fueled by generative artificial intelligence (AI) such as ChatGPT will usher in a new era in the platform liability wars. Previous waves of new communication technologies—from websites and chat rooms to social media apps and video sharing services—have been shielded from legal liability for content posted on their platforms, enabling these digital services to rise to prominence. But with products like ChatGPT, critics of that legal framework are likely to get what they have long wished for: a regulatory model that makes tech platforms responsible for online content."

Anyone who's tried ChatGPT, Microsoft's Bing chatbot or Google's Bard will have quickly learned that they have a tendency to fabricate information and confidently present it as fact. These systems, built on what's known as large language models, also emulate the cultural biases they've learned from being trained upon huge troves of what people have written online.

From Los Angeles to Colorado and throughout Oregon, as child welfare agencies use or consider tools similar to the one in Allegheny County, Pennsylvania, an Associated Press review has identified a number of concerns about the technology, including questions about its reliability and its potential to harden racial disparities in the child welfare system. Related issues have already torpedoed some jurisdictions' plans to use predictive models, such as the tool notably dropped by the state of Illinois. According to new research from a Carnegie Mellon University team obtained exclusively by AP, Allegheny's algorithm in its first years of operation showed a pattern of flagging a disproportionate number of Black children for a "mandatory" neglect investigation, when compared with white children. The independent researchers, who received data from the county, also found that social workers disagreed with the risk scores the algorithm produced about one-third of the time.

Any output from a generative AI tool should be treated as unvetted source material. AP staff must apply their editorial judgment and AP’s sourcing standards when considering any information for publication.
Generative AI makes it even easier for people to intentionally spread mis- and disinformation through altered words, photos, video or audio, including content that may have no signs of alteration, appearing realistic and authentic. To avoid using such content inadvertently, journalists should exercise the same caution and skepticism they would normally, including trying to identify the source of the original content, doing a reverse image search to help verify an image’s origin, and checking for reports with similar content from trusted media.

ToDo: Experiment with the following to build additional context:

"Seven new no-cost generative AI training courses to advance your cloud career." May 18, 2023 https://cloud.google.com/blog/topics/training-certifications/new-google-cloud-generative-ai-training-resources

"Generative AI Data in 2023: Importance & 7 Methods" Updated on July 10, 2023 https://research.aimultiple.com/generative-ai-data/

"Introduction to Generative AI." 2 hr 33 min Learning Path 3 Modules. https://learn.microsoft.com/en-us/training/paths/introduction-generative-ai/

"Generative Design & Generative AI: Definition, 10 Use Cases, Challenges." By Cem Dilmegani https://research.aimultiple.com/generative-design/

"Generative Adversarial Networks (GAN) & Synthetic Data [2023]." By Cem Dilmegani, Updated on December 22, 2022 https://research.aimultiple.com/gan-synthetic-data/

"Generative AI: 7 Steps to Enterprise GenAI Growth in 2023." By Cem Dilmegani, Updated on July 20, 2023 https://research.aimultiple.com/generative-ai/

"Mistral AI releases new model to rival GPT-4 and its own chat assistant." By Romain Dillet, February 26, 2024 https://techcrunch.com/2024/02/26/mistral-ai-releases-new-model...chat-assistant/](https://techcrunch.com/2024/02/26/mistral-ai-releases-new-model-to-rival-gpt-4-and-its-own-chat-assistant/?utm_source=tldrnewsletter)

https://apnews.com/hub/generative-ai

Random

ImageNet (zero-shot): SOTA, surpassing OpenAI CLIP (https://openai.com/blog/clip/).
LAMA (factual and commonsense knowledge): Surpassed AutoPrompt (https://arxiv.org/abs/2010.15980).
LAMBADA (cloze tasks): Surpassed Microsoft Turing NLG (https://www.microsoft.com/en-us/research/blog/turing-nlg-a-17-billion-parameter-language-model-by-microsoft/).
SuperGLUE (few-shot): SOTA, surpassing OpenAI GPT-3 (https://arxiv.org/abs/2005.14165).
UC Merced Land Use (zero-shot): SOTA, surpassing OpenAI CLIP (https://openai.com/blog/clip/).
MS COCO (text generation diagram): Surpassed OpenAI DALL·E (https://openai.com/blog/dall-e/).
MS COCO (English graphic retrieval): Surpassed OpenAI CLIP and Google ALIGN (https://ai.googleblog.com/2021/05/align-scaling-up-visual-and-vision.html).
MS COCO (multilingual graphic retrieval): Surpassed UC² (best multilingual and multimodal pre-trained model) (https://arxiv.org/pdf/2104.00332.pdf).