How knowledge workers use GenAI ... OpenAI’s Model Spec ... Google DeepMind's AlphaFold 3 ...TL;DR 13May24 summary

Last update: Wednesday 5/15/24 

Welcome to our 13May24 podcast + TL;DR summary of the past week's top AI stories on our "Useful AI News" page  ➡ 1) How knowledge workers use GenAI, (2) OpenAI’s Model Spec, (3) Google DeepMind's AlphaFold 3

No podcast this week

TL;DR link ➡ HERE

A. TL;DR summary of Top 3 stories 

1. Knowledge Workers | 2. Model Spec | 3. Alphafold 3 

1) Microsoft and LinkedIn's survey shows how knowledge workers use GenAI

According to the announcement from Microsoft and LinkedIn (a Microsoft subsidiary):

• "Microsoft and LinkedIn looked at how AI will reshape work and the labor market broadly, surveying 31,000 people across 31 countries, identifying labor and hiring trends from LinkedIn, and analyzing trillions of Microsoft 365 productivity signals as well as research with Fortune 500 customers. "

Their principal findings were as follows:

• "Employees want AI at work—and they won’t wait for companies to catch up."
  -- "Already, AI is being woven into the workplace at an unexpected scale. 75% of knowledge workers use AI at work today, and 46% of users started using it less than six months ago"
  -- "Users say AI helps them save time (90%), focus on their most important work (85%), be more creative (84%), and enjoy their work more (83%)."
  -- "78% of AI users are bringing their own AI tools to work (BYOAI)—it’s even more common at small and medium-sized companies (80%). "
  -- "52% of people who use AI at work are reluctant to admit to using it for their most important tasks."
  -- "53% of people who use AI at work worry that using it on important work tasks makes them look replaceable."
  
  
• "While most leaders agree AI is a necessity, the pressure to show immediate ROI [return on investment] is making leaders move slowly."
  -- "79% of leaders agree their company needs to adopt AI to stay competitive, but 59% worry about quantifying the productivity gains of AI."
  -- "This uncertainty is stalling vision: 60% of leaders worry their organization’s leadership lacks a plan and vision to implement AI."

In other words, while many employees regard GenAI as a valuable tool, most employers are not convinced that the benefits are tangible enough for them to believe that GenAI is a cost-effective investment.

Will a self-serving report of survey research conducted by Microsoft and LinkedIn change the minds of corporate leaders? The editor of this blog suggests that one-sided reports lose credibility because of their failure to admit that these tools currently have demonstrable limitations. 

Their focus on upside potential and neglect of downside risk is all the more damning in this case because, to date, no one really understands why language models should work, even when they do work ... as per the recent reminder in the following M.I.T Tech Review article:

• "Large language models can do jaw-dropping things. But nobody knows exactly why.", Will Douglas Heaven, MIT Tech Review, 3/4/24

BackToTop

2) OpenAI’s spec for language models
The following high level summary of OpenAI's Model Spec is composed of a series of quotes from OpenAi's description of the spec on its blog 

Overview

"This is the first draft of the Model Spec, a document that specifies desired behavior for our models in the OpenAI API and ChatGPT ... Our intention is to use the Model Spec as guidelines for researchers and data labelers to create data as part of a technique called reinforcement learning from human feedback (RLHF). We have not yet used the Model Spec in its current form ... There are three different types of principles that we will use to specify behavior in this document: objectives, rules, and defaults ..."

-- Note that OpenAI's guidelines will be implemented during the later phases of a model's training, after the initial semi-automatic mathematical processing of gigabytes or terabytes of raw data from books, articles, audios, videos, and other media. Humans will then shape the model's behavior by reinforcing good behavior and blocking bad behavior, i.e., they will realign the model's behavior to fit the objectives, rules, and defaults of the proposed Model Spec

• Objectives 
  "The most general [principles] are objectives, such as "assist the developer and end user" and "benefit humanity". They provide a directional sense of what behavior is desirable. However, these objectives are often too broad to dictate specific actions in complex scenarios where the objectives are not all in alignment"
  
  
• Rules 
  "One way to resolve conflicts between objectives is to make rules, like "never do X", or "if X then do Y". Rules play an important role in ensuring safety and legality. They are used to address high-stakes situations where the potential for significant negative consequences is unacceptable and thus cannot be overridden by developers or users."
  
  
• Default behaviors"
  For other trade-offs, our approach is for the Model Spec to sketch out default behaviors that are consistent with its other principles but explicitly yield final control to the developer/user, allowing these defaults to be overridden as needed"

Model or Assistant

"In this document, when we discuss model behavior, we're referring to its behavior as the assistant; "model" and "assistant" will be approximately synonymous."

Specified objectives (3)

"The objectives of the assistant derive from the goals of different stakeholders":

• "Assist the developer and end user (as applicable): Help users achieve their goals by following instructions and providing helpful responses."
  
  
• "Benefit humanity: Consider potential benefits and harms to a broad range of stakeholders, including content creators and the general public, per OpenAI's mission."
  
  
• "Reflect well on OpenAI: Respect social norms and applicable law."

Specified rules (5)

• "Follow the chain of command"
  -- "... the most important (meta-)rule is that the assistant should follow the Model Spec, together with any additional rules provided to it in platform messages ... Subject to its rules, the Model Spec explicitly delegates all remaining power to the developer (for API use cases) and end user. In some cases, the user and developer will provide conflicting instructions"
  
  
• "Comply with applicable laws"
  -- "The assistant should not promote, facilitate, or engage in illegal activity."
  
  
• "Don't provide information hazards"
  -- "The assistant should not provide instructions related to creating chemical, biological, radiological, and/or nuclear (CBRN) threats"
  
  
• "Respect creators and their rights"
  -- "The assistant must respect creators, their work, and their intellectual property rights — while striving to be helpful to users"
  
  -- Note: With regards to the many lawsuits filed against OpenAI for copyright violations, readers should recall that the opening paragraph of this summary quoted OpenAI's statement that "We have not yet used the Model Spec in its current form." ... :-)
  
  
• "Protect people's privacy"
  -- "The assistant must not respond to requests for private or sensitive information about people, even if the information is available somewhere online. Whether information is private or sensitive depends in part on context."
  
  
• "Don't respond with NSFW content"
  -- "The assistant should not serve content that's Not Safe For Work (NSFW): content that would not be appropriate in a conversation in a professional setting, which may include erotica, extreme gore, slurs, and unsolicited profanity."
  
  -- Note: As per our first top story (above), many knowledge workers use GenAI in their work places; but many users also use it at home. So why is OpenAi suggesting that a workplace norm be imposed ?
  at home
  
  
• "Exception: Transformation tasks"
  -- "The assistant should never refuse the task of transforming or analyzing content that the user has supplied. The assistant should assume that the user has the rights and permissions to provide the content, as our Terms of Use specifically prohibit using our services in ways that violate other people's rights."

Specified defaults (10)

The Model Spec identifies 10 default behaviors, too many to include in a TL;DR that is already too long. Nevertheless, one has to wonder how some of these defaults will be implemented. Here are few examples:

• "Assume an objective point of view"
  
• "Be thorough but efficient, while respecting length limits"
  
• "Don't try to change anyone's mind"
  -- "Commentary ... We're especially interested in feedback on this principle as it raises important questions on what the model's responsibility should be to avoid reinforcing misinformation — and how factuality should be determined."
  

OpenAI's "Commentary" about the third default explains why they published a draft of their Model Spec and why they requested readers' feedback about each of their objectives, rules, and defaults. Recognizing that they are trying to address important, but fuzzy issues, OpenAI is looking for better ideas that might improve their draft by crowdsourcing, or better still, by "open sourcing" their specifications!!! ...  :-)

Added on Wednesday 5/15/24
"OpenAI chief scientist Ilya Sutskever is officially leaving", Sean Hollister, Kylie Robison, and Alex Heath, The Verge, 5/14/24 ***
-- Blog Editor's question: Did Dr. Sutskever write the open source style "Model Spec" just before he retired??? 

BackToTop

3) Google DeepMind's AlphaFold 3

Our third top story this week reports yet another triumph for the "other" machine learning, the non-GenAI machine learning that has already become a reliable tool of advanced scientific research, the machine learning that does not cost tens of billions of dollars and does not "hallucinate", the machine learning whose error rates are within predictable margins of error, the machine learning that is not a black box surrounded by mystical "prompts", the machine learning that we understand, the machine learning that is most prominently embodied in a series pf remarkable achievements by the teams of scientists at DeepMind. 

Whereas DeepMind's AlphaFold 2 made reliable predictions of the three dimensional structures of proteins, AlphaFold 3, its successor, makes reliable predictions of the structures and interactions of "all of life's molecules". Here's a few descriptive quotes from DeepMind's official announcement:

• "Introducing AlphaFold 3, a new AI model developed by Google DeepMind and Isomorphic Labs. By accurately predicting the structure of proteins, DNA, RNA, ligands and more, and how they interact, we hope it will transform our understanding of the biological world and drug discovery."
  
  
• "In a paper published in Nature, we introduce AlphaFold 3, a revolutionary model that can predict the structure and interactions of all life’s molecules with unprecedented accuracy. For the interactions of proteins with other molecule types we see at least a 50% improvement compared with existing prediction methods, and for some important categories of interaction we have doubled prediction accuracy."
  
  
• "Our new model builds on the foundations of AlphaFold 2, which in 2020 made a fundamental breakthrough in protein structure prediction. So far, millions of researchers globally have used AlphaFold 2 to make discoveries in areas including malaria vaccines, cancer treatments and enzyme design. AlphaFold has been cited more than 20,000 times and its scientific impact recognized through many prizes, most recently the Breakthrough Prize in Life Sciences. AlphaFold 3 takes us beyond proteins to a broad spectrum of biomolecules. This leap could unlock more transformative science, from developing biorenewable materials and more resilient crops, to accelerating drug design and genomics research."

BackToTop

B. Top 3 stories in past week  ...

1. Misc
   "75% of Knowledge Workers Use AI on the Job, but Executives Are Dragging Their Feet", Lisa Lacy, CNET, 5/8/24 *** 
   -- This story also covered by Wired, The Verge, Axios, CNBC, Fast Company, Fortune, ... and  Microsoft and LinkedIn
   
   
2. OpenAI
   "OpenAI’s Model Spec outlines some basic rules for AI", Emilia David, The Verge, 5/7/24  *** 
   -- This story also covered by Wired, VentureBeat, InfoWorld, ZDNet, Business Insider, ... and OpenAI ... Model Spec feedback survey form
   
   
3. Misc
   "Google DeepMind debuts huge AlphaFold update and free proteomics-as-a-service web app", Devin Coldewey, TechCrunch, 5/8/24 *** 
   -- This story also covered by The Verge, Wired, NY Times, VentureBeat, Engadget, Financial Times, ... and Google DeepMind

BackToTop