Confluence for 6.23.24

Anthropic releases Claude 3.5 Sonnet. Shaping Claude's character. Lessons on how AI knowledge spreads in organizations. AI's impact on the freelance market.

Midjourney prompt: a new creation blooms, in the style of William Larkin

Welcome to Confluence. Here’s what has our attention this week at the intersection of generative AI and corporate communication:

• Anthropic Releases Claude 3.5 Sonnet

• Shaping Claude’s Character

• Lessons on How AI Knowledge Spreads in Organizations

• AI’s Impact on the Freelance Market

Anthropic Releases Claude 3.5 Sonnet

There’s a new leading model, and you can access it for free.

Anthropic has once again shaken up the AI landscape with the unexpected release of Claude 3.5 Sonnet, which represents a step forward in both performance and accessibility. According to benchmarks released by Anthropic, Claude 3.5 Sonnet outperforms its predecessor, Claude 3 Opus, and OpenAI’s GPT-4o across nearly every standard. We’ve sensed this in our own initial testing. Claude 3.5 Sonnet feels smarter than GPT-4o and Claude 3 Opus while retaining the human-like prose (more on that later in this edition) that has differentiated Claude since the release of the Claude 3 family of models.

Beyond the fact that we now have a new leader in the competition for the most capable LLM, there are two other things worth noting about Claude 3.5 Sonnet.

First, it’s available for free. This move mirrors OpenAI's strategy (OpenAI made its leading model, GPT-4o, available for free in ChatGPT last month) and has far-reaching implications. By lowering the barriers to access to such advanced AI capabilities, Anthropic joins OpenAI in narrowing the “widening gulf” we've previously observed between free and paid AI tools. Before these recent moves, free users of ChatGPT and Claude were limited to less sophisticated models; that’s no longer the case.

The second major development comes in the form of Artifacts, a new feature that changes how users interact with Claude. Artifacts introduces a split-screen interface that transforms the AI interaction from a simple chat into a collaborative workspace. While half of the screen maintains the familiar chat format, the other half serves as a dynamic canvas where Claude showcases its capabilities in real-time. Below is a video shared by Anthropic that shows how Artifacts works, and that we’ve found representative of our experience.

This interface creates new possibilities for direct collaboration between humans and AI. Users can now provide immediate feedback on Claude's output or make their own edits, creating a more interactive and productive way of working with AI. Our initial experiments with Artifacts have generally been playful (experimenting with recreating retro arcade games, for example), but Claude has excelled at these tasks with impressive speed and accuracy. Artifacts offers a glimpse into a future of even more capable AI tools — ones that can see complex tasks through to completion rather than simply responding to isolated prompts. This shift from reactive to proactive AI assistance opens up new possibilities for problem-solving and creative tasks across fields.

The release of Claude 3.5 Sonnet is yet another reminder of the breakneck speed at which AI capabilities are advancing. And there’s likely more to come soon: Anthropic has already hinted at the future release of Claude 3.5 Opus, while industry watchers anticipate a new iteration from OpenAI (possibly GPT-5) in a similar timeframe.

Shaping Claude’s Character

There’s more to AI utility than sheer “horsepower.”

The release of Claude 3.5 Sonnet isn’t the only development at Anthropic that has our attention this week. We’ve also been reflecting on Anthropic Research’s recent publication titled “Claude’s Character,” which explains the approach Anthropic took to shape the “character” of Claude, beginning with the Claude 3 family of models earlier this year.

When we first wrote about Claude 3 Opus in March, we remarked that

For whatever reason — and it’s hard to pin down — Claude 3 Opus just feels better than the other models. Its responses seem to say more with less, its tone more calm and less pedantic, its responses more, well, human. Some are calling it “warmer” than the other models, and we find that true as well, but it also just seems to have better (for lack of a better term) common sense.

Something about the Claude 3+ family of models — including Claude 3.5 Sonnet, as we note earlier in this edition — feels more human, more mature, more natural. With the publication of “Claude’s Character,” we now know much more about why that is. The piece reveals that

Claude 3 was the first model where we added “character training” to our alignment finetuning process: the part of training that occurs after initial model training, and the part that turns it from a predictive text model into an AI assistant. The goal of character training is to make Claude begin to have more nuanced, richer traits like curiosity, open-mindedness, and thoughtfulness.

The rest of the piece explains the considerations that went into shaping Claude’s character and the methodology Anthropic used to do it. It’s a relatively short read and worth the few minutes to read in its entirety.

We think this notion of character is important for a few reasons. First, our experience with the Claude 3+ models over the past several months suggests that efforts to intentionally shape the character of an AI model simply work — and make a noticeable difference in the user experience. GPT-4o and Gemini may be just as powerful as Claude 3 Opus or Claude 3.5 Sonnet, but the extent to which Claude differentiates itself in utility and perceived value may largely be due to Claude’s character training.

Second, we expect this to be an emerging dimension on which the leading models compete. At least in the near term, the leading models are likely to share similar technical architectures and be trained on similar data, so innovations in finetuning like Anthropic’s character training may lead to the biggest differentiators in performance and utility. And as the developers of leading models experiment with the character of their models, their decisions will matter, as they will shape how these models navigate thorny topics like values, bias, and more. As the Anthropic piece notes:

Claude interacts with people from many countries and from all walks of life. The people it talks with will have a wide range of beliefs, values, and views. Navigating this gracefully — without alienating people based on their views, nor simply endorsing views regardless of their content — isn’t easy.

The most important implication for our readers, though, is simply for utility. In his book Co-Intelligence, Ethan Mollick’s third principle for interacting with AI is to “Treat AI like a person (but tell it what kind of person it is).” It’s one of the foundational pieces of advice we most often give clients: think of the AI like a colleague, an expert, a friend, or whatever the right type of person is for a given scenario — but always think of it, and interact with it, like a person. Claude’s character — which was immediately obvious to us when we began interacting with Claude 3 Opus — makes that much easier. It’s easier to think of an AI like a person when it feels like one, and now we know the approach Anthropic took to make that happen. We expect to see much more of this in future models, both from Anthropic and others.

Lessons on How AI Knowledge Spreads in Organizations

A new study highlights the importance of understanding how generative AI works before teaching others how to use it.

The team of researchers behind “Navigating the Jagged Technological Frontier” has released a new working paper that follows up on their groundbreaking study. Their follow-up research, titled “Don’t Expect Juniors to Teach Senior Professionals to Use Generative AI: Emerging Technology Risks and Novice AI Risk Mitigation Tactics,” addresses a common expectation in organizations: when new technology emerges, younger, more junior employees often end up teaching older, more senior employees how to use it. But does this hold true for generative AI?

To investigate this question, the researchers re-interviewed 78 participants from their original study to understand how they might approach teaching senior staff about generative AI tools like ChatGPT. While the study itself gets quite technical in parts, it shows that junior talent may fail to appropriately educate senior talent for two reasons:

1. Junior employees lack a deep enough understanding of how generative AI works, and thus

2. Recommendations from junior employees for using generative AI and mitigating its potential risks deviate in meaningful ways from what experts would advise.

Putting aside the age question for a moment (which, while interesting, doesn’t provide a clear direction for what we should do), we believe that it’s essential for teams using generative AI to have a foundational understanding of how the class of technologies works. And fortunately, it’s a challenge that’s reasonably simple to address within teams.

When we lead sessions with clients, we always start with a discussion of the fundamentals of how large language models (LLMs) work. This includes exploring how their underlying mechanisms affect which tasks they’re well-suited for and which they aren’t. We find value in having a real conversation about this, but even if a facilitated conversation isn’t feasible, there are plenty of resources available to help you understand what’s happening when you prompt ChatGPT, Claude, or another chatbot¹. In fact, you can also ask these models themselves about how they work if you’re looking for a quick overview. See below what Claude 3.5 Sonnet has to say:

User: I lead a corporate communications team at a mid-sized consumer products organization. We expect generative AI could affect our work in meaningful ways and want to make sure we have a foundational understanding of how this technology works. Can you explain in simple terms how generative AI works, focusing on large language models?

Claude 3.5 Sonnet: I’d be happy to explain how generative AI and large language models work in simple terms. Let me provide a concise overview, and I can elaborate on any specific areas you'd like to know more about.

Generative AI, particularly large language models (LLMs), work by processing and learning from vast amounts of text data. Here's a basic breakdown:

1. Training: LLMs are trained on enormous datasets of text from the internet, books, and other sources. They learn patterns, relationships between words, and contextual information.

2. Pattern Recognition: Through this training, the model learns to recognize complex patterns in language, including grammar, context, and even some basic facts about the world.

3. Prediction: When given a prompt or question, the model uses what it has learned to predict the most likely next words or sequences of words that would follow.

4. Generation: Based on these predictions, the model generates text that is contextually relevant and coherent.

5. Fine-tuning: Models can be further trained on specific datasets to specialize in certain tasks or domains.

The key aspect is that these models don't truly “understand” language in the way humans do. They're essentially very sophisticated pattern-matching systems that can produce human-like text based on statistical probabilities learned from their training data.

For corporate communications, this technology could potentially assist in drafting press releases, creating content for various channels, or even helping with crisis communication scenarios.

Would you like me to elaborate on any specific aspect of how generative AI works, or discuss potential applications in corporate communications?

When we have this conversation with clients, it typically takes only a few minutes of discussion for people to grasp these central points. Generative AI can feel like a black box to those who haven't invested the time to learn about it or experiment with it in a meaningful way. But as soon as you pull back the curtain, even a bit, the tools themselves become much less intimidating and people can start to hone their own sense of how and when to use them.

If you and your team don’t yet have an understanding of how generative AI works, now is the time to change that. The longer teams put off learning more about how these models work (and, perhaps more importantly, how they don’t) the greater the risk of missteps.

We recognize that diving into the technical details can be daunting, especially for those with non-technical backgrounds, but it’s absolutely necessary. It’s going to become an expectation that communication teams are using this technology, and that means we need to think intentionally about how to leverage generative AI in our work today.

For what it’s worth, those of us working on Confluence have academic backgrounds in subjects like Anthropology, English, French, and Political Science. We’re not exactly software engineers. Yet we’ve found that taking the time to understand the basics of how LLMs function has been invaluable in guiding how we bring generative AI to our work.

AI’s Impact on the Freelance Market

Is freelance work going away — or evolving?

A recent story from The Wall Street Journal, “AI Doesn’t Kill Jobs? Tell That to Freelancers,” highlights AI’s impact on freelancers, particularly in creative fields. It’s a worthwhile read, with some valuable insights for professionals across industries on how AI is reshaping the work landscape. One study cited in the article found that since the debut of ChatGPT and other generative AI models, the number of freelance jobs posted on platforms like Upwork and Fiverr has dropped by as much as 21% in areas where AI excels like data coding and writing. The story isn’t all doom and gloom, though. Other freelancers report an uptick in work, particularly for tasks like refining AI-generated content or tackling more complex, specialized tasks that AI can’t handle effectively.

Consider the task of rewriting job descriptions. Glowforge co-founder Dan Shapiro developed this detailed prompt that uses AI to streamline the process, making it more efficient and effective. This is an excellent example that illustrates how AI can augment human capabilities, allowing professionals to focus on higher-value tasks while improving routine processes.

As AI capabilities advance, professionals — and not just freelancers — will need to stay informed and adapt accordingly. As always, the key is to view AI as a collaborator rather than a competitor, finding ways to intelligently integrate it into workflows to boost productivity and creativity — without inviting reputational risk. Human oversight remains critical, especially when dealing with sensitive information or client-facing content.

The rapid evolution of AI tools is reshaping the professional landscape in real-time. Used well, the technology is powerful tool to augment human capabilities and drive innovation across various fields. The future of work is evolving, and those who engage with AI purposefully and ethically will be best-positioned to thrive.

---

We’ll leave you with something cool: Luma’s Dream Machine, a new AI tool that turns images and text into short videos.

---

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.

1. We often recommend “Generative AI Exists Because of the Transformer” from the Financial Times and Andrej Karpathy’s “Intro to Large Language Models”