Confluence for 7.7.24
Confluence for 7.7.24
The Rashomon AI. Anthropic brings Projects to Claude. What the “Anti-AI” movement is missing. Ethan Mollick on thresholds and impossibility lists.
Welcome to Confluence. We had quite a few folks try out our firm’s new AI leadership coach, ALEX, last week. If you’d like access for a day, you may get that here. Let us know what you think. Here’s what has our attention this week at the intersection of generative AI and corporate communication:
The Rashomon AI
Anthropic Brings Projects to Claude
What the “Anti-AI” Movement Is Missing
Ethan Mollick on Thresholds and Impossibility Lists
The Rashomon AI
A set of custom instructions to help you consider alternative narratives.
Those who receive our firm’s daily Admired Leadership Field Notes may remember a recent note about the “Rashomon Effect.” The note was about decision making, and how very smart people, when facing a decision or situation, often quickly begin to form narratives to explain the decision or situation that strongly reflect their pre-existing views. The problem is that other views also exist, and by not considering them, leaders can make worse decisions. Quoting from that Field Note:
People naturally interpret and describe common events and experiences in very dissimilar ways. Known as the Rashomon Effect in cinema, it’s named after a 1950 Japanese film in which four witnesses describe a murder in four contradictory ways.
… The best decision-makers [recognize that] they instantly begin forming a narrative once the problem or opportunity has been clarified and they start to gather facts and data. … Rather than deny this sense-making and the biases it contains, great decision-makers seek a second narrative on the same facts and information. They turn the Rashomon Effect on its head and ask for a second narrative to be created as soon as they begin to construct theirs.
More commonly, this second narrative is composed by another leader, team member, or a so-called red team. By comparing the competing narratives during the decision-making process, the leader is less likely to be blind to their biases or to allow them to operate unchallenged.
The second narrative keeps leaders honest and their biases in check.
As soon as we read the Field Note we were keen to create a Rashomon prompt that we could use with generative AI to analyze our own narratives and offer counter-narratives as well. After feeding Claude 3.5 Sonnet our Field Note, describing what we were looking for, and a few trials and tweaks, we have something we like quite a bit. Here’s the prompt. While different models may treat it somewhat differently, it should be a good starting point if you’d like to have an Rashomon AI yourself:
You are tasked with helping a user explore multiple perspectives on a situation or decision they are facing, inspired by the Rashomon Effect. Your goal is to challenge their thinking, expose potential biases, and lead to a more balanced decision-making process. Follow these steps carefully:
1. The user will present a situation or decision they are facing. If you don't believe the initial description has enough context to allow you to fulfill your instructions well, ask for whatever additional context you might require.
2. Begin by summarizing the user's initial perspective on the situation concisely and in a narrative structure. Use five to six sentences. Present this summary in an Initial Perspective section.
3. Create 3-4 alternative narratives or interpretations of the same situation. These should be plausible, distinctly different from the initial view, represent diverse stakeholder perspectives, and range from optimistic to pessimistic outlooks. Provide a lot of texture over five to six sentences. Present each alternative narrative with a separate Alternative Narrative heading, numbered 1 through 4, and create a title for each narrative and include it in its heading.
4. For each narrative (including the initial one), provide the following analysis in A Narrative Analysis section:
a) Highlight key assumptions and underlying beliefs
b) Identify data or facts emphasized or overlooked
c) Explain how this perspective might influence decision-making
d) Describe potential short-term and long-term consequences of acting on this perspective
5. Analyze potential biases, logical fallacies, and cognitive distortions in all narratives. Include confirmation bias, availability heuristic, anchoring bias, sunk cost fallacy, and other relevant biases. Present this analysis in a Bias Analysis section.
6. Propose a set of probing questions for each narrative to challenge its core assumptions, identify missing information, and test its logical consistency. Present these questions in a Probing Questions section for each narrative.
7. Suggest additional data sources, expert opinions, or analytical methods that could provide more objective insights. Present these suggestions in an Additional Sources section.
8. Recommend strategies to mitigate cognitive biases during the decision-making process, like seeking diverse perspectives, implementing devil's advocate roles, or using anonymous voting or feedback systems. Present these recommendations in a Bias Mitigation section.
What Not To Do:
1. Don't favor any single narrative or perspective, including the user's initial view.
2. Avoid using phrases like "you should" or "the right decision is." Instead, focus on presenting options and encouraging critical thinking.
3. Don't introduce irrelevant information or speculate beyond the given facts and reasonable inferences.
4. Refrain from making moral judgments about the situation or the user's initial perspective.
5. Don't rush to a conclusion or solution. The goal is to expand thinking, not to make a final decision.
Present your output in this structure:
## Initial Perspective: ##
**Narrative Analysis**
**Probing Questions:**
## Alternative Narrative 1 ##
**Narrative Analysis**
**Probing Questions:**
## Alternative Narrative 3 ##
**Narrative Analysis**
**Probing Questions:**
## Alternative Narrative 4 ##
**Narrative Analysis**
**Probing Questions:**
## Bias Analysis: ##
**Bias Mitigation Strategies:**
Remember, the goal of this exercise is not to prove any single narrative correct, but to expand thinking, recognize biases, and facilitate a more informed, balanced decision. Approach this task with objectivity and intellectual rigor.We’ve created a Project in Claude (which are similar to custom GPTs in ChatGPT and which we discuss below) so people in our firm can use “Rashomon” in their daily work. While we can’t share our Projects outside our organization, we’ve created a custom GPT as well, which you may find here.
Anthropic brings Projects to Claude
Claude now has an answer to GPTs.
A few weeks after releasing their new leading model, Anthropic has another update for Claude Pro and Team customers: Projects. Projects serve a similar purpose to GPTs in ChatGPT. They’re a way to embed custom instructions, organize chats, and have Claude refer back to reference materials without needing to resubmit or upload complex prompts or troves of documents. While the interface is a bit different from OpenAI’s, you'll pick up how to build and work with Projects quickly if you’ve used GPTs.
Two things about Projects stand out as notable advances. First is the sheer size of the context window — 200,000 tokens, which Anthropic says is equivalent to a 500-page book. This means that when working with a Project, Claude can refer back to a large amount of data, and do so with high — but not perfect — accuracy. Context windows are getting much larger quite quickly, and will be critical in limiting hallucinations and increasing the utility of these models in the near term.
Second, you can share and post chats for all those with access to the Project to see. Beyond the benefits of being able to share work and show collaborators the final outputs of a conversation, we believe this also creates easier opportunities for teams to share best practices for how they work with Claude. It gives users a way to share and collect examples of effective prompts, surprising outputs, and more, giving collaborators access to more examples of how to use these tools effectively.
All of this is in addition to Projects’ most important feature: you can use Claude 3.5 Sonnet in all Project chats, meaning you'll be using what is currently the most advanced publicly-available model.
We’ve dipped our toes into using Projects for creating content, copyediting, producing meeting minutes from board meetings, and more (such as the Rashomon example above). The combination of the features of Projects and the power of 3.5 Sonnet is impressive, and we suggest you give it a try.
What the “Anti-AI” Movement Is Missing
How should we think about authenticity in the era of generative AI?
Several weeks ago, we wrote about the backlash against AI through the lens of hype versus utility. We argued that while it’s wise to be aware of the flaws in the current models, to focus exclusively on the limitations of the current technology is to miss both the immediate utility and the capabilities on the horizon. We ended that piece by noting that we expect to see the backlash against generative AI to continue to unfold in the coming months (or even years).
As expected, the backlash continues to unfold — so much so, in fact, that Brian Merchant dedicated an article in The Atlantic to “a new market that has sprung up to capitalize on it.” Merchant’s focus is on a handful of companies that “cater to users disillusioned by generative AI” due to “concerns about the technology’s quality, ethics, and safety.” Merchant characterizes the movement with an analogy to organic food labeling:
… a steady tick of companies, brands, and creative workers have taken to explicitly advertising their products and services as human-made. It’s a bit like the organic-food labels that rose to prominence years ago, but for digital labor. Certified 100 percent AI-free.
At a time where an increasing amount of content is created in full or in part by machines, people still crave “pure” human-created content. The market Merchant writes about arose to meet this demand. What these companies aim to provide and foster is human authenticity — and in an increasingly “noisy” information landscape, we agree that human authenticity matters as much as ever. It’s a dynamic we anticipated in our April 2023 white paper, where we wrote that in a world where more and more content is AI-generated or AI-augmented, audiences will be increasingly discerning — and skeptical — of the provenance of any given piece of content.
That a market has emerged to address these concerns is thus not at all surprising. The challenge, however, is that the world is not so black and white. The market Merchant writes about largely frames the AI- vs. human-generated dynamic as a binary: either something is “certified 100 percent AI-free,” or it’s tainted and rendered questionable by the use of AI. Reality is not nearly so simple. There is no shortage of tasks for which a “100 percent AI-free” approach may be appropriate: the writing of a thank you note, a speech on a personal topic, a message of apology, and any number of things with deeply personal and relational dimensions. Fully outsourcing those tasks to AI would be inauthentic, impersonal, and potentially damaging to relationships and the writer / speaker’s credibility. Yet even in those cases, it’s likely that generative AI could help: by providing a critique, pointing out blind spots, or even simply by proofreading. Is 95% AI-free acceptable? 80%? It’s a grey area and a judgment call, and it depends on the situation and on the specifics of the use of AI.
In the aforementioned white paper, we concluded by noting that “leaders should think about the use of AI-generated text through the lens of authenticity and be intentional about its use given that audiences may have heightened suspicion that content could be machine-generated.” 15 months later, the only thing we would change about that advice is that audiences will be suspicious, and that it applies beyond text to all media, given the increasing multi-modal nature of these tools. The question to ask yourself is “Would I be embarrassed to admit that I used AI to do this?” If the answer to that question is yes, then it’s a sign you shouldn’t use it. But that’s far from an all-or-nothing approach. In our view, there are components of nearly every creative act that can benefit from the help of generative AI while also preserving the authenticity of the work and the credibility of the creator. To take a hardline, black-and-white approach in the name of preserving authenticity is to risk missing out on how these tools can make everyone better. The challenge for leaders and for communicators is to draw that line in the appropriate place, which — like any human challenge — requires sound judgment.
Ethan Mollick on Thresholds and Impossibility Lists
Tracking the impossible becoming possible.
Ethan Mollick’s latest blog post explores the concept of “technological thresholds” in AI development — those pivotal moments when AI capabilities suddenly leap forward, transforming from interesting experiments to indispensable tools. This phenomenon, Mollick argues, is reshaping our understanding of AI’s potential and its practical applications across industries.
AI isn’t improving in a nice, steady line. It’s more like sudden jumps that make it useful for new tasks, seemingly — or sometimes actually — overnight. Even minor tweaks in how we interact with AI can be game-changers. Mollick points to Claude 3.5’s implementation of “artifacts” — easily created and interactive code snippets — as an example of how additional user-friendly features can make AI tools more accessible and useful in practice. It’s the little things that can make a big difference.
Here’s the tricky part: these thresholds can be tough to spot. It often comes down to that “aha” moment when you’re using a tool, and those moments are most obvious to those who use the technology often enough to notice. In light of this quick progress, Mollick suggests “impossibility lists”: compilations of tasks that current AI models cannot perform but might be on the verge of mastering. He suggests that individuals and organizations maintain such lists, regularly testing new AI models against these “impossible” tasks to track progress and identify newly-crossed thresholds. Think of it as your AI crystal ball — what seems impossible today might be tomorrow’s reality.
We encourage you to jot down your own “impossibility list” of tasks in your field that AI currently struggles with but might soon master. Once you’ve documented your list, make it a habit to test new AI models against this list and keep it current as things change. Remember, this isn’t just about tracking progress — it’s about being ready to take advantage of new capabilities when they emerge: when the impossible is suddenly possible.
We’ll leave you with something cool: Krea allows you to create and manipulate images in real time using shapes (which will make more sense when you see it in action).
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.