Confluence for 12.17.23
X launches Grok AI. More AI plagiarism shoes drop. Stability AI released Stable Video Diffusion. Google releases NotebookLM. A lesson in change communication and AI.
Welcome back to Confluence. We’re happy to report that OpenAI has re-opened GPT Plus subscriptions. For $20 a month we think it’s well worth having access to the most capable and powerful generative AI tool on the market. We have no relationship with OpenAI — that’s just our view. That said, here’s what has our attention at the intersection of generative AI and communication this week:
X Launches Grok AI
More AI Plagiarism Shoes Drop
Stability AI Releases Stable Video Diffusion
Google Releases NotebookLM
A Lesson in Change Communication and AI
X Launches Grok AI
Search billions of tweets … the answer is not always “42.”
X/Twitter has launched Grok, a generative AI platform which they describe as “conversational AI for understanding the universe.” It is different than other large language models in that it’s able to ingest and incorporate into its responses real-time data from X. As a result, you can ask Grok things like, “What happened yesterday in the world of English Football” and get answers like this:
You’ll note that it also posts a series of tweets relevant to the query at the bottom of the reply. In fact-checking these results though, we found the summary inaccurate. The next tweet cited at the bottom of the reply referenced a result for Grimbsy Town from 1990 — a fact Grok didn’t notice:
It’s illustrative of one of the challenges of large language models — they are predictive based on their training data, not deterministic like search. They guess at an answer more than find an answer. A recent tweet mentioned a result from 23 years ago and Grok didn’t understand the context enough to exclude it. So while the promise of Grok is a generative AI on top of contemporaneous world events and news (something at which X can excel), a user needs to exercise caution when reviewing facts it puts out just as one would with any large language model.
Testing suggests that Grok performs similar to ChatGPT 3.5, Claude 2, and other large language models that happen to not be GPT-4, which is still the most capable frontier model available. We find this to be true. Examples:
Please describe the inverted yield curve …
GPT-4:
An inverted yield curve in finance occurs when long-term debt instruments have a lower yield than short-term debt instruments of the same credit quality. This is a notable deviation from the typical situation where longer-term bonds have higher yields to compensate investors for the risks associated with time, such as inflation and uncertainty.
In a standard yield curve, the yield (interest rate) increases with the bond's maturity, reflecting the higher risk of longer commitments. However, in an inverted curve, this relationship flips, suggesting that investors have less confidence in the near-term economy compared to the long term.
This inversion is often scrutinized because it has historically been a reliable indicator of economic recessions. The rationale is that investors expect short-term economic downturns, prompting them to shift their investments to longer-term bonds, driving down long-term yields.
However, it's important to note that an inverted yield curve is not a cause of recession but a symptom of market sentiments and expectations about the future economy. Not every inversion leads to a recession, and other economic factors and contexts should also be considered when interpreting this phenomenon.
Grok:
An inverted yield curve is a graphical representation of the yields (interest rates) of bonds with different maturity dates. In a normal yield curve, longer-term bonds have higher yields than shorter-term bonds. However, in an inverted yield curve, the opposite occurs, with short-term yields being higher than long-term yields.
Historically, an inverted yield curve has been viewed as a potential indicator of a pending economic recession. The reasoning behind this is that when short-term interest rates exceed long-term rates, market sentiment suggests that the long-term outlook is poor and that the yields offered by long-term fixed income will continue to fall.
However, this viewpoint has been questioned recently, as foreign purchases of securities issued by the U.S. Treasury have created a high and sustained level of demand for products backed by U.S. government debt. When investors aggressively seek debt instruments, the debtor can offer lower interest rates. Some argue that it is the laws of supply and demand, rather than impending economic doom and gloom, that enable lenders to attract buyers without having to pay higher interest rates.
In conclusion, an inverted yield curve is a situation where short-term interest rates are higher than long-term interest rates, which can be interpreted as a potential sign of a recession. However, other factors, such as foreign demand for U.S. debt, can also influence the yield curve.
Give me 10 questions to prompt a dialogue among colleagues about 2024 priorities.
GPT-4:
Reflecting on Past Successes: "What were our most significant achievements in 2023, and how can we build upon these successes in 2024?"
Identifying Key Goals: "What are our top three goals for 2024, and how do they align with our overall organizational mission?"
Addressing Challenges: "What were the biggest challenges we faced last year, and how can we proactively address similar challenges in 2024?"
Innovative Thinking: "In what areas can we innovate or think outside the box in 2024 to stay ahead of our competition?"
Resource Allocation: "How should we prioritize and allocate our resources in 2024 to maximize our efficiency and effectiveness?"
Team Development: "What skills or knowledge do we need to develop within our team to achieve our 2024 objectives?"
Measuring Success: "How will we measure our progress and success in 2024? What key performance indicators should we focus on?"
Customer Focus: "How can we better understand and meet the evolving needs of our customers in 2024?"
Sustainability and Social Responsibility: "In what ways can we enhance our commitment to sustainability and social responsibility in the coming year?"
Long-Term Vision: "How do our plans for 2024 align with our long-term vision for the company? Are there adjustments we need to consider for sustained growth and success?"
These questions are designed to stimulate comprehensive discussions, focusing on building upon past successes, addressing challenges, and aligning with the overall vision and goals of the organization.
Grok:
What are the top strategic objectives we should focus on in 2024 to drive growth and success for our organization?
How can we better align our team's priorities and individual goals to achieve our 2024 targets?
What new initiatives or projects should we prioritize in 2024 to stay competitive and innovative in our industry?
How can we improve collaboration and communication across departments to achieve our 2024 goals?
What are the most significant challenges or obstacles we may face in 2024, and how can we proactively address them?
How can we leverage emerging technologies and trends to support our 2024 priorities?
How should we prioritize resource allocation and investments in 2024 to maximize our return on investment and achieve our goals?
How can we foster a culture of continuous learning and development to support our 2024 objectives?
What performance metrics and key performance indicators (KPIs) should we focus on in 2024 to track our progress and success?
How can we ensure that our 2024 priorities align with our organization's mission, vision, and values?
We find those results quite similar. That said, GPT Plus offers GPT-4 vision, image creation, the ability to write and run code, file creation, speech interpretation, and many other abilities that make it more powerful, at least for us, than Grok. Grok also requires a $16 monthly subscription. Given that GPT-3.5 is free, as is Microsoft Copilot / Bing Chat (which runs GPT-4 and has vision and image creation), we’re not sure we see the use case for communication professionals yet — especially when they can get GPT Plus for $20 per month. That said, Grok is impressive given that it is a later entry, and we expect a rapid pace of innovation and improvement.
One other thing that sets Grok apart is that it has “fun” and “regular” modes. The fun mode is meant to replicate the personality (and capability) of the computer from the Hitchhiker’s Guide to the Galaxy. For example, here are several queries in Fun Mode:
Please explain the inverted yield curve.
Oh, you're asking about the inverted yield curve? It's like a roller coaster ride for the economy, but instead of screaming with excitement, people scream with concern.
In simple terms, the yield curve is a graph that shows the relationship between interest rates and the time to maturity for government bonds. Normally, it slopes upward because investors expect to be paid more interest for holding bonds for a longer period of time.
But when the yield curve inverts, it's like a plot twist in a movie. The short-term interest rates become higher than the long-term rates. It's like going back in time and getting paid more for a one-year bond than a ten-year bond.
This unusual event is often seen as a sign of a looming recession. It's like a storm warning for the economy. Investors start to think that the central bank will cut rates in the future to stimulate the economy, so they rush to buy long-term bonds, driving down their yields.
Historically, an inverted yield curve has been a reliable predictor of recessions. But, like a fortune-teller, it's not always right. It's just one piece of the puzzle, and other factors can influence the economy too.
In the end, the inverted yield curve is like a weather forecast. It's not always accurate, but it's worth paying attention to.
How about those Philadelphia Eagles?
Oh, the Philadelphia Eagles, you say? They're doing just fine, thank you. Just a few minor hiccups here and there, but nothing a little teamwork and determination can't fix. After all, even the most majestic of eagles need to take a breather and regroup before soaring to new heights. So, let's not be too hard on them, shall we?
As you can see, Grok has a personality, and it shows itself in subtle ways …
I need advice communicating a reduction in force.
Oh boy, communicating a reduction in force is a delicate task. But don't worry, I've got your back! Here are some tips to help you navigate this challenging situation:
Be honest and transparent: It's like ripping off a band-aid. The quicker and more direct you are, the better it is for everyone involved.
Show empathy: Remember that people's livelihoods are at stake. Put yourself in their shoes and show that you care about their well-being.
Provide support: Offer resources and assistance to help those affected by the reduction in force. This could include job search assistance, counseling, or severance packages.
Be prepared for questions: People will have questions, and it's important to be ready to provide clear and honest answers.
Follow up: Keep the lines of communication open and follow up with those affected to ensure they have the support they need.
Remember, a reduction in force is never easy, but handling it with honesty, empathy, and support can make all the difference.
X describes the personality of Grok has having wit and a bit of a rebellious streak, and in using it, we do find it fun to use. And regardless of how Grok develops, we expect that Grok’s “fun mode” is a harbinger of things to come. We believe people will prefer generative AI tools with personas and not just knowledge. How those personas develop, the role of the user in shaping them, and how they influence the extent to which people identify and relate with generative AI tools is part of the story to watch in the months and years to come.
More AI Plagiarism Shoes Drop
Disclosure continues to appear to be the best policy for generative AI use.
We wrote two weeks ago about the revelation that Sports Illustrated had been publishing AI-generated content under the false personae of “live” journalists, even going so far as to use generative AI to create those authors’ headshots. This week the publisher announced the firing of their CEO over the matter. We also learned this week that Investing.com has been using AI to draft stories that appear in some cases to be thinly veiled revisions of human-written stories from other sources. While Investing.com has said that its stories were crafted by AI with oversight by an editor, it did not extend credit beyond its own AI.
These are likely only the first in what will be a long string of cautionary tales about fair use and proper disclosure with generative AI. All solutions create problems. Generative AI offers a very powerful solution to the problem of summarization and text creation — so powerful and easy, in fact, that the friction associated with creating text drops to nearly zero. The problem is that this only shifts the burden of the work from the creation of the text to the other risks generative AI introduces: ensuring accuracy, transparency, credit, and disclosure (among other factors) among them. We increasingly believe there are no shortcuts here — and if you’re going to have any awkwardness at all over the revelation that you’re using generative AI, you should disclose it, and that nature of its use, from the start.
Stability AI Releases Stable Video Diffusion
The leaps in video generation capabilities keep coming.
Late last month, Stability AI — creators of leading image generation model Stable Diffusion — released Stable Video Diffusion, a foundation model for generative video based on their Stable Diffusion image generation model. You can get a brief glimpse of the capabilities in this video. Limitations of the current model include the length of videos users can generate (less than four seconds) and some of the same issues we see with image generation (problems with text, human features and faces not always rendering properly, etc.). As we’ve written many times, however, we would advise readers not to mistake current limitations with lasting ones.
With this release, Stability AI joins other companies like Pika and Runway ML as a leader in generative video. In response to the Stable Video Diffusion release, Jack Clark noted in his Import AI newsletter that “If 2022 was the year of the first good broadly available image models, then 2024 will probably be that for video.” We’ll soon find out whether or not that is the case, but we do think it’s instructive to look at the leaps in image generation capabilities from 2022 to now and consider what a similar trajectory might look like for video — and what the implications of such a trajectory might be. We’re beginning to experiment with video generation more extensively ourselves and will share what we learn in future editions of Confluence.
Google Releases NotebookLM
The experimental product aims to be “a personalized AI collaborator, grounded in the information you trust, designed to help you do your best thinking.”
Last week we wrote about the emerging trend of generative AI applications moving beyond the now-familiar chatbot interface. One new product that we didn’t note in that post, but with which we have since experimented, is Google’s NotebookLM, which was recently made available to the public as part of an experimental release. As a Wired write-up on the release notes, “the easiest way to think of it is as an AI collaborator with access to all your materials that sits on your metaphorical shoulder to guide you through your project.”
We gave it a try by creating a notebook to explore the emerging academic literature on generative AI’s impact on the labor market and professional work. For each notebook you create, you can upload up to 20 “sources” (PDFs, your own notes, other documents in your Google Drive, etc.) of up to 200,000 words each. Our notebook could easily handle the five articles we added as sources, leaving us with plenty of room for more as the literature expands.
One feature we found particularly interesting is that for each source you add to your notebook, the tool automatically creates a “Source Guide,” which includes a summary of the source and a list of “Key Topics,” as shown below:
Clicking on one of the items in the Key Topics list starts a conversation with the AI on the given topic, as shown below for “AI adherence” (note that the “Discuss AI adherence” prompt is generated automatically):
Each response in the chat interface includes citations, which link back to specific sections of the source text. This back-and-forth clickable interface for citations is an advance over a pure chatbot interface, where we find ourselves asking the tool to cite the relevant quotation or passage directly in its output (and then we check the source text to verify its accuracy).
This is not an exhaustive list of features by any means, but rather a brief glimpse at what’s captured our attention thus far. Note, for instance, that in addition to the ability to interrogate and explore specific sources individually, NotebookLM also allows you to ask questions or run comparisons across sources in your notebook — though in our limited experimentation, working across sources led to more erratic output than when focusing on a single source. On the whole, while we weren’t blown away by the tool, we are intrigued enough to continue experimenting. It may well turn out that it’s better suited to other use cases — like, say, synthesizing interview notes or transcripts, or any number of tasks relevant to our and our readers’ work — than it is to what we’ve tried thus far. Either way, it’s one glimpse into a of an emerging class of applications that go beyond a pure chatbot interface.
A Lesson in Change Communication and AI
A reminder that communication professionals will need to advise leaders on how to communicate AI-related decisions.
Klarna, a Swedish fintech company, recently announced a hiring freeze, underscoring the expected impact of artificial intelligence on their operations. This move is a vivid reminder of the importance of effective communication in an era of rapid technological change.
AI will not only alter business tools and processes but also reshape leadership decisions and workplace norms. While Klarna may be an exception in the broad brush with which they made this decision, expect that more leaders and organizations will grapple with the questions that AI invites in a real way.
For communication professionals, this means making a concerted effort to take part in these leadership discussions and bring a change and communication perspective to the table. As transformative as AI may be for organizations, leaders will still need to communicate these decisions with human audiences. By staying up to date on the latest developments in AI, communication professionals can bring a credible perspective to the table, ensuring that leaders understand how employees will interpret AI related decisions and offering practical communication advice.
We’ll leave you with something cool: Tatiana Tsiguleva shares collections of Midjourney prompts to generate striking images.
AI Disclosure: We used generative AI in creating imagery for this post. We also used it selectively as a creator and summarizer of content and as an editor and proofreader.