AI “will alter our experience as reasoning beings and
permanently change our relationship with reality.”
(Kissinger/Schmidt/Huttenlocher; The Age of AI and our Human Future; 2022
Since fall of 2022 with the first emergence of OpenAI's DALI-e and ChatGPT the "threat" of AI has moved from sci-fi to an active discussion in popular media, policy forums and even a topic of serious concern among technical experts. The spring letter(s) from multiple AI researchers and resignation of one of Google's leading AI experts are indications of this. A useful "talking head" documentary is "The A.I. Dilemma" (by the folks that brought you The Social Dilemma)
I've played a bit with DALI-e, Microsoft's variant of ChatGPT, Google's BARD, Google's augmented search, Claude2 ("constitutional AI"), Sudowrite (to create books, lyrics), VoiceMod ("sings" to canned audio tracks), with interesting results. I've posted a few of these (democracy, peer review), and also OpEd or two to our local paper. Below is a recent snapshot of a PPT presentation I'm developing to facilitate discussion of these topics in this very rapidly changing environment.
(a related list of resources can be found <here>.)
There are a few threats I suggest are very real, and they don't include a Skynet Terminator appearing from the future, or even HAL (2001) taking over the ship (or planet).
Generative AI is being used to create fake news sites & content--- these can span from being jokes, to serious extensions of political/propaganda efforts to undermine targeted nation states (not just the U.S.)
AI is currently used to amplify fear and outrage in social media and other online channels. This is driving social disruption, and in some cases genocide (see my new class on the Dystopian Spirals)
and driving our youth crazy (see Surgeon General's report)The current AI tools serve objectives of their creators/managers to the extent these folks know what is happening, are willing to intervene, and have reasonably moral intentions. All three of these assumptions have been proven optimistic. AI's given the directive to increase the sale of shoes may accomplish this in ways that are highly successful, and given the economic incentives of advertising driven major corporations (think Google, Facebook, and others) this makes sense, and sells shoes. The same success can be anticipated in selling political candidates, riots, legislative efforts, hate attacks, etc. It is unclear that humans can win at the game of free-will any more than we can at chess.
Geoffrey Hinton, ex-Google AI guru, concurs with many of these issues but sees one more as an existential threat. This is the possibility that an AI given some objective will "realize" (not as a conscious entity necessarily) that it could be more successful if it acquired more power --- essentially identifying a "sub goal"of control as a way to advance it's assigned goal (sell shoes, candidates, genocide, whatever.) A LinkedIn discussion between Hinton and Andrew Ng indicates their perception that this generation of AI's "understand the world". I would disagree, since Chatbots seem to be willing to "groom" 13 year old girls and inform teens how to cover up the smell of pot and alcohol for a "great party", their understanding is problematic at best. (36 minute 2024 presentation by Hinton outlines concepts and issues.)
If you are concerned about students/kids/etc. using chatbots to create fake essays, etc. Here is a suggestion: (1) give them an assignment to have a chatbot create an essay. (2) Then use "strike though" to indicate the misleading/incorrect parts, (3) use "highlight" to indicate additions made by the student and (4) have them end the assignment with a paragraph or two of their own describing their experience.
Resource list:
https://sites.google.com/site/jimisaak/Home/stem-4-all/ai---social-media-resources
Chatbot and Image Creation Platforms (click down arrow to see descriptions/links)
Hands On Generative AI (GenAI) 27 March 2024 notes for Spring OLLI Class "Fun with AI"
Generative AI tools have blossomed (or spread like weeds) since 2022. Some of these are available for "free" to the public. Trying these out can help you gain some sense of the pros and cons. Below are some thoughts about using and trying these, and also a list of a few of the most visible ones that you can access. Google lists (today) in response to "New free AI tools", about three billion hits.
"Free" is important to understand. A very successful business model (think Google, Facebook, Tic-tok, etc.) is to collect personal information about users and sell access to those users to advertisers that are informed by that personal data.
A reasonable assumption about any "Free" online service is that this is the model providing revenue for the provider. The GenAI tools are no different. [Although, March 2024 articles indicate that Google and Microsoft are unclear how to monetize GenAI.]
Not surprising then is that most ask you to "register" to use the tool. This typically includes a username and password (and the creation of "cookies" on your computer to track your activities.) Some, Claude2 in particular, seek to verify your identity with two-step confirmation using email, cell phone or both. If you are already "known" to the provider, which is often the case with Google/GMAIL (BARD now Gemini) and Meta/Facebook (Llama2) you may not need to register further.
In any case, before we try to test these tools, you may want to register in advance and avoid that delay.
Also, be aware that GenAI responses are not necessarily accurate, and they certainly lack "common sense" so be skeptical about any generated content and verify any aspects that are important to your objectives.
Here are some of the most visible tools in the press. The cost of creating large language models is significant in both terms of compute cycles and volumes of data to analyze, so there is some overlap between players. [We will probably start with BING since it has both chat and image options.]
ChatBots (text generation)
Alphabet/Google "Gemini nee Bard" is a chatbot you can query directly if you have a Google account. This now (2024) includes the "Gemini" version which most recently is not producing images (likely copyright concerns), but does include a level of web search (and conversely Google is now providing some level of AI integration into web search and other online Google tools like Docs, etc.)
[Google turned off image generation with Gemini when it appeared to have reverse bias, generating persons of color in unexpected situations.] https://gemini.google.com/ [requires a Google account such as Gmail]
Meta/Facebook/Llama2 -- this can be directly accessed by going to www.Llama2.ai [not working 4/12/24] (.ai is the national domain for Anguilla, a boost to their GDP) Facebook has chosen to release it's Large Language Model as "open source", which means that variations will be sprouting everywhere. They allow this for both research and commercial use. https://llama.meta.com/llama2/ has more info about Llama2. [Meta allows registered individuals to download open source variations of Llama2, some of which are available via the web site: HuggingFace https://huggingface.co/chat/ ]
OpenAI: ChatGPT, DALL-E with significant funding from Microsoft who has incorporated advanced versions of these into their BING/Chat and Image Create tools. https://openai.com/ (DALL-E is now presented in the context of ChatGPT3 ) -- Registration required for DALL-E access https://chat.openai.com/ currently has fee access to ChatGPT 3.5
[Training data ended in 2022, according to ChatGPT 3.5] As of Feb 2024, the DALL-E image generation is limited to paid subscribers, however the Microsoft "Bing, Chat, Image Create" options are free via Microsoft's offerings.
OpenAI has also exposed (not released for public use) Sora, a video generation tool that is impressive. https://openai.com/sora
Microsoft "BING" search engine has top level menu options for "Chat" (a recent, augmented version of Chat GPT) and "Image Create" (ditto DALL-E). This currently is one of the few "free" image generation systems.
BING now incorporates "Copilot", a variation on ChatGPT BING/Copilot will only create images for registered users.
https://www.bing.com/chat
Inflection AI corp, PI: https://pi.ai/onboarding -- fairly new offering (onboarding is the registration for new users)
Login seems to now prefer Google, Facebook or some other account, no longer using their own login. (April 224)
[A key inflection founder just went to work for Microsoft in March 2024]
Sudowrite will help you write text articles, stories, poems, even short stories or novels .. free trial, monthly fee if serious.
https://www.sudowrite.com/
Tools suggested by Pi.ai:
YouChat, Poe, Jasper, Quillbot, and Vertex AI
https://you.com/?chatMode=default "signin to unlock free premium modes"
https://poe.com/login login to Talk to ChatGPT, GPT-4, Claude 3 Opus, DALLE 3, and millions of others - all on Poe. [Poe is likely trying to build a revenue stream off work of others, so perhaps not a first mover]
https://www.jasper.ai/ focused on writing marketing "copy" [free trial available]
https://quillbot.com/ paraphrasing tool
Vertex is a Google tool, now "enhanced by Gemini" -- entry way to a paid "pro" set of tools
Search
http://Perplexity.ai a quite new "search" AI system created by folks who have left OpenAI and Meta ...
BING Copilot
no doubt there is a Google variation here as well
Image Generation (in addition to BING/image generator)
Stable Diffusion allows for up to 10 "free" images per day, so worth considering your prompts in advance.
https://stablediffusionweb.com/ (is the online version vs download)
NightCafe Creator (https://creator.nightcafe.studio/studio) registration required (emails challenge events regularly)
Picarts (https://picsart.com/)
Midjourney (https://www.midjourney.com/) not yet available for free generation (can sign up for Beta perhaps)
Dream by WOMBO (https://www.w.ai/ ) login to reduce popup disruptions [but required anyway]
Suggested by (Google) Gemini:
There are several freely available AI image generators, such as the one from Microsoft called Designer
https://www.microsoft.com/en-us/edge/features/image-creator?form=MA13FJ.
Websites like Zapier have reviews of the best AI image generators in 2024, which can help you choose one that meets your needs
https://zapier.com/blog/automate-ai-images/.
A quite different application: Synthetic Memories
Tools for education
http://Playlab.ai ... AI-powered tools built by a community of educators; Toni, a writing TA ...
may be limited to teachers & their students (I've requested access)
Tools we are not trying to track at this point include
video generation:
Synthesia https://www.synthesia.io/ text to video for "presentations" -- could be OLLI previews! (fee to unlock features)
http://Runwaymi.com ... Gen2 fee is 5 cents/second generated video .. offers free trial
Sora (OpenAI) (Diffusion model plus transformer model)