Confluence for 5.19.24
OpenAI reveals GPT-4o. The coming wave.
Welcome to Confluence. It was a big week in the evolution of generative AI, with OpenAI announcing a new model (and a mode of generative AI / human interaction, really) on Monday and Google announcing a host of new generative AI models and tools on Tuesday. We have a roundup of the OpenAI announcements here, and we’ll have more on Google next week. And because both pieces this week are relatively long, we only have two items for today — but it’s plenty to go with your coffee (or lunch), and certainly, plenty to think about. With that said, here’s what has our attention at the intersection of generative AI and corporate communication:
OpenAI Reveals GPT-4o
The Coming Wave
OpenAI Reveals GPT-4o
The latest model from OpenAI represents a leap in speed and functionality, though not necessarily intelligence.
On Monday, OpenAI answered the question we raised last week about what they would show next — and it’s GPT-4o. The “o” in GPT-4o stands for “omni,” a nod to the fact that this new model is truly multi-modal, capable of processing text, audio, images, and videos, almost in real time.
While we’ll have to wait a bit longer to see the full breadth of GPT-4o’s multi-modal features in action (OpenAI didn’t commit to a firm release date), we do have access to its text-based capabilities through ChatGPT and saw enough through OpenAI’s demos to begin to form our perspective on what this will mean for communication professionals and other users of generative AI. There are several things we think you should know, and we have some early thoughts.
First and perhaps most important (more on this below), GPT-4o will be free to use for all ChatGPT users, not just those with Plus, Team, or Enterprise accounts. While there will be some form of cap on how much free users can use GPT-4o, it will still give people one more way — and the easiest way so far — to access a frontier model for free. Until now, the gap between the paid and free versions of most AI tools had been widening. GPT-4o represents an important reversal of that direction. We expect many of those who have been underwhelmed by generative AI when using GPT-3.5 to gain a new appreciation for what AI can do once they have access to GPT-4o. If GPT-3.5 is a bicycle, GPT-4o is a sports car, and millions are about to drive it for the first time.
Second, and to that point, GPT-4o is noticeably faster than GPT-4, even with just text. Responses start nearly instantly, are faster overall, and are often longer. GPT-4o isn’t necessarily more intelligent than GPT-4, but it produces more output more quickly than before. Below is an example of what GPT-4o produces compared to GPT-4 with the same prompt.
GPT-4
You are an expert in qualitative analysis. What types of analysis could you run on a set of open-ended survey comments from an all company town hall?
Analyzing open-ended survey comments from an all-company town hall can provide rich insights into employee sentiments, concerns, and suggestions. Here are several types of qualitative analysis methods that could be effectively applied:
Thematic Analysis: This involves identifying patterns or themes within qualitative data. You would code the data by segmenting comments into meaningful groups and then examining these for recurring themes or ideas. This method helps in understanding the dominant sentiments or topics discussed during the town hall.
Content Analysis: This more systematic technique focuses on the presence of certain words, phrases, or concepts within the data. You can quantify and analyze the presence, meanings, and relationships of such words and concepts. This could be particularly useful if you're looking to measure the frequency of specific issues or topics mentioned, such as "work-life balance" or "company culture".
Discourse Analysis: This approach is used to analyze the ways in which language is used in the comments to understand more about the social context. For instance, you could analyze how employees discuss leadership or policy changes, focusing on the language used to express agreement, dissent, or neutrality.
Grounded Theory: This method involves developing a theory based on the data collected. By coding the data and looking for emerging patterns, you can form a hypothesis about the underlying structures or dynamics within the company culture as expressed in the town hall comments.
Narrative Analysis: This approach focuses on the stories within the comments, looking at how individuals describe their experiences and the sequence of events that emerge from their narratives. This can be particularly illuminative for understanding personal employee experiences and the impact of company policies on individual workers.
Sentiment Analysis: While traditionally quantitative, sentiment analysis can be adapted for qualitative data by manually categorizing comments into sentiments like positive, negative, or neutral. This can help gauge the overall mood or employee satisfaction within the company.
Each of these methods can be used independently or in combination to provide a comprehensive understanding of the qualitative data from your all-company town hall, depending on your specific objectives and the nature of the data collected.
GPT-4o
You are an expert in qualitative analysis. What types of analysis could you run on a set of open-ended survey comments from an all company town hall?
Conducting qualitative analysis on open-ended survey comments from a company town hall involves several methods to ensure comprehensive and insightful findings. Here are key types of analysis you could run:
1. Thematic Analysis
Description: Identify and analyze recurring themes or patterns within the data.
Steps:
Familiarization with the data.
Generating initial codes.
Searching for themes.
Reviewing themes.
Defining and naming themes.
Producing the report.
2. Content Analysis
Description: Quantify and analyze the presence of certain words, phrases, or concepts.
Steps:
Coding the text into manageable content categories.
Analyzing the frequency and relationships among words or concepts.
Interpreting the context within which these words or phrases appear.
3. Narrative Analysis
Description: Focus on the stories and accounts within the comments to understand how people make sense of events and actions.
Steps:
Identify the narrative structure (e.g., setting, characters, plot).
Analyze how stories are constructed and the meaning they convey.
Explore the functions of narratives in the context of the organization.
4. Grounded Theory
Description: Develop a theory grounded in the data through a systematic methodology.
Steps:
Open coding: Identify concepts and categories in the data.
Axial coding: Identify relationships between categories.
Selective coding: Integrate and refine the theory.
5. Discourse Analysis
Description: Examine how language is used in the comments to understand social and organizational constructs.
Steps:
Identify discursive practices and patterns.
Analyze the power dynamics, identities, and social roles constructed through language.
Contextualize findings within broader social and organizational contexts.
6. Framework Analysis
Description: Use a structured approach to sort and interpret data within a pre-existing framework or one developed from the data.
Steps:
Familiarization with the data.
Identifying a thematic framework.
Indexing: Applying the thematic framework to the data.
Charting: Rearranging the data according to the framework.
Mapping and interpretation.
7. Sentiment Analysis
Description: Determine the emotional tone of the comments.
Steps:
Use coding or software tools to classify comments as positive, negative, or neutral.
Analyze the distribution of sentiments.
Interpret the overall mood and implications for the organization.
8. Comparative Analysis
Description: Compare different groups' responses (e.g., departments, job roles).
Steps:
Segment the data by relevant categories.
Compare themes, sentiments, or narratives across groups.
Interpret differences and their implications.
9. Constant Comparative Method
Description: Compare data continuously to refine categories and identify variations.
Steps:
Collect and code data simultaneously.
Constantly compare new data with existing codes and categories.
Refine and adjust categories throughout the analysis process.
10. Lexical Analysis
Description: Analyze the language use and vocabulary within the comments.
Steps:
Identify keywords and phrases.
Analyze word frequency and co-occurrence patterns.
Interpret lexical choices in context.
Implementation Tips:
Triangulation: Use multiple methods to cross-verify findings.
Software Tools: Consider using qualitative analysis software (e.g., NVivo, MAXQDA, ATLAS.ti) for coding and analyzing data.
Intercoder Reliability: Ensure reliability by having multiple analysts code the data and compare results.
By employing a combination of these methods, you can gain a deep understanding of the feedback provided during the company town hall, uncovering actionable insights and informing decision-making processes.
Not dramatically different in terms of quality, but GPT-4o gave us four additional types of analysis compared to GPT-4. This difference persisted in other tests we ran, with GPT-4o consistently simply producing more content more quickly in comparison to GPT-4.
Third, the combination of multi-modality and speed will make GPT-4o feel very different compared to working with GPT-4 (or other frontier models). Even without direct access to the multi-modal features of GPT-4o, the demos are impressive enough to convince us. This difference will be most obvious when using voice.
When you use ChatGPT's voice mode today, it relies on a three-step process: first, it transcribes your voice to text, then it writes a response to that text, and finally it converts the response back to audio. While the quality of the spoken interaction is remarkable, this process results in noticeable lag and an inability to interpret tone or cadence. While this approach still has utility (as we've discussed), it has its limits. Now watch this:
The pace of that conversation is much closer to what you’d experience talking to another person. In addition — and we think this is big — GPT-4o is able to read the nonverbal elements of the human side of the conversation: tone, pacing, volume, and expressions via the camera if you engage it (expect solicitations for a thousand speech coaching AIs to flood your spam folder soon). Translations will likely never be the same, either:
GPT-4o opens up many other possibilities thanks to its new vision capabilities. If you wish, it will be able to “see” things on your desktop or screen in real time, as this segment from the announcement event shows:
GPT-4o’s image creation also got a boost. It’s now much better at creating words, fonts, and many other elements of the image creation process that GPT-4 used to fumble, as shown in this GPT-4o-generated image:
There are many more examples of what GPT-4o can do here, and we think it’s worth looking at them all. The bottom line is that GPT-4o’s speed and true multi-modality, combined with the sheer capabilities of a GPT-4 class model, will create real opportunities to “bring AI to every table” in ways that we simply haven’t seen before.
The Coming Wave
Get ahead of shadow generative AI use before it gets ahead of you.
As we noted in the previous section, one of the most significant parts of OpenAI’s event last Monday was the announcement that GPT-4o will be available to everyone, for free. For the first time, people everywhere will have unfettered access to the most powerful generative AI tool in the world via their computers and mobile devices. While there have been ways to get to GPT-4 for free in the past (through Microsoft’s Copilot, for one), those avenues always involved some friction, and always limited the full capabilities of GPT-4. Going forward, this will no longer be the case. Anyone with a cell phone will be able to have the most powerful artificial intelligence the world has ever known just one icon tap away, for free.
What does this mean? Given that generative pre-trained transformers like ChatGPT are general purpose technologies, we don’t think anyone really knows. General purpose technologies have proven to be powerful in both their social and economic effects, and very difficult to predict in terms of what those effects will be. And while we are not economists or technologists, we’ve been advising clients for more than 30 years, and so in some ways we feel we’ve seen at least scenes of this movie before.
We can quickly think of four possible analogues: email, intranet search, web pages, and instant messaging. For all four, most people first used these technologies in their personal lives before they used them at work (with some exceptions for employees in technically progressive organizations or organizations in certain sectors like technology, financial services, or defense). Many of us of a certain … ahem … “experience level” … had email via MSN.com or Yahoo! Mail before we did at work. Many of us had used AOL.com or had created our own web pages with GeoCities or in our own internet sandboxes before we had intranets at the office. Many of us could search for sources with AltaVista, Yahoo!, or the new kid on the block, Google, before we could search our intranets. And many of us had text messaging and micro-blogging tools like Twitter before we had direct messages or Slack at work.
The bottom line is that in our experience there are times when the retail sector of technology adoption leads the commercial. And as people begin to use technology in their personal lives, we’ve seen that a few things tend to happen. One is that people discover all sorts of day-to-day ways that a technology fits into and adds value to their lives (something we and others call “mundane utility”). Another is that they become increasingly reliant on that technology as that utility shapes their lives. And yet another is that they begin to carry the expectations they have for that utility into work.
But since they can’t use that technology at work, they experience a perceived friction of ineffectiveness or inefficiency. With prior technologies like email or the internet, where employees couldn’t easily remove that friction by creating workarounds with their own technology, people generally had to live with it until the organization caught up technically. But with direct messaging, you saw something different — people setting up group chats via text on their phones, or using applications like WhatsApp outside of the corporate technical environment, to communicate with colleagues and among teams. We expect this will be the case with generative AI. Thanks to OpenAI’s announcement this week, now millions of employees will be able to simply (and without any personal expense) use something like GPT-4o in the shadows. They will become, in essence, secret cyborgs, to quote Ethan Mollick.
While this might be great for them and their effectiveness, it’s a terrible situation for the organization in terms of governance and security. Generative AI introduces all sorts of risks along with its benefits — the protection of confidential information and questions of source authenticity looming large among them. As generative AI has an uneven frontier of abilities, being scary good at some things (like brainstorming ideas or critiquing work or generating amazing amounts of content in a short period of time) and quite poor at others (like citing facts), there will be a significant risk of employees creating reputational or operational risk by using generative AI to create work without peer review (a risk we and others call “falling asleep at the wheel”).
So far, we’ve seen our clients typically take one of three approaches to the question of generative AI at work. One is to do nothing and instead ignore generative AI — which we still see a surprising number of organizations doing. We feel this is short-sighted: generative AI is here to stay. A second is to keep everything locked down for employees while any number of committees and working groups explore the matter, working toward some institutional approach. This invites secret cyborg behavior. A third we call “let ‘er rip,” which is to give employees full freedom to explore generative AI use. This approach welcomes risk.
We don’t think any of these are really the best answer given the likelihood that within a year many, many more people will probably be using very powerful generative AI daily in their personal lives. So, what to do? In broad strokes, this is what we’d suggest:
First, engage with the right leader in your own chain of command and get approval to begin a governed, principled exploration of use cases for generative AI for you and your team, department, or function. This could be as simple as getting the green light for your group to start a hands-on exploration of possible use cases for GPT-4o. Establishing a set of principles to govern this use in concert with stakeholders in cybersecurity and legal should be part of the arrangement.
The very next step is to get a set of principles in place, in concert with the right governance stakeholders, that can govern your team’s initial uses of generative AI. We’ve put the ones we established more than a year ago in our own firm in the footnotes in case they are a source of inspiration (note that we review them quarterly).1
With those principles in place, have a conversation with your team, department, or function in which you make these principles explicit. Nobody should have “not knowing” as an excuse for being outside the decided principles.
Establish a place where your team, function, or department can have a dedicated and ongoing conversation about generative AI and to share use cases, learnings, successes, and failures. We do this in a dedicated channel in Microsoft Teams, but there are lots of ways to get there.
Offer some form of training in generative AI. There are many free resources online. You can offer some reading resources. Past editions of Confluence offer many readings that we recommend, if you’re looking for somewhere to start. But do something to help educate people before they begin use, especially regarding what generative AI does well and does poorly. Again, there is plenty about this online that you can easily source.
As a group, start using whichever tool you’ve decided to use and start the conversation. Ideally, get approval to use a tool like ChatGPT Team or Enterprise, or Claude Pro for Teams, that brings security provisions with it. Keep your stakeholders in the know generally, and every quarter do an overall assessment of how things are going and if anything in the principles or approach needs to change. Expect that to happen, because the technology is changing quickly.
We think this approach affords a way to begin getting ready for generative AI, and to bring more of the mundane utility within the walls of the organization in a governed and principled way, while the organization works through the larger process of technical governance and implementation. But as people begin to get more than 10 hours under their belts using these tools at home, they will quickly realize how much mundane utility things like GPT-4o afford — and they are not going to just leave that on the table when they start work every day. Without permission, they’ll seek that utility in the shadows, with all the risk that might bring.
And really, why wouldn’t you want them to come into the daylight? Why should someone spend 40 minutes reformatting the headings and bullets on a set of feedback when GPT-4o can do it in 30 seconds? Why would anyone need to spend an hour looking for duplicates in a large Excel file when GPT-4o can find them and cleanse that data in less than a minute? Why would you want someone to spend two hours reading a long policy document, thinking of common questions it could raise, and drafting an FAQ when they could instead spend one hour quality-checking and improving the FAQ that Claude 3 Opus drafted in five minutes? There are just three examples … there are untold more. But once people get a taste of them at home, they’re not going to avoid them at work. Of that, you can be certain. And we think it’s better to ride a wave than fight one.
We’ll leave you with something cool: Alie Jules’ growing library of Midjourney style reference codes.
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.Our current principles for the use of generative AI:
Rather than be subject to the advancement of this technology we will seek to understand it and embrace its use on our terms and in ways consistent with our values. Primary among these are relationships, excellence, making our own people and our clients better, and fulfillment in our work.
Overall, always be deliberate and thoughtful in the use of generative AI and do so with the intention of bringing greater value to clients. When doing so:
Honor client confidentiality agreements and redact all information than can identify organizations and individuals, including those in our firm, prior to submitting queries to an AI tool unless that tool offers full data security (e.g., it is part of our Microsoft Office suite).
Submit no CRA | Admired Leadership propriety content to AI tools unless the tool offers full data security (e.g., it is part of our Microsoft Office suite).
Fact-check all AI output (names, dates, citations, etc.).
Subject all AI-generated content to human quality review and revision prior to distribution outside the firm.
Prioritize excellence over efficiency. While these tools can and should save us time, we should not overestimate the time they will save nor trade doing it well for doing it fast.
Prioritize skill development over efficiency. In cases where AI ability overlaps tasks that are part of a colleague's developmental path (proofreading, editing, thematic analysis, etc.), we should avoid the use of AI and keep those opportunities as part of our developmental method as they teach core skills beyond that of the task itself (attention to detail, writing skillfulness, attention to context, etc.).
Once a colleague has reached a high level of proficiency in a specific skill, consider using AI on overlapping tasks as a means of freeing that talent to add higher levels of value.
If it would be awkward relationally or if it would introduce reputational risk for a third party to learn we had used generative AI, we should disclose its use. Always err on the side of transparency in all judgement calls.