Confluence for 1.26.2025
Five frontiers. A free reasoning model. OpenAI and partners announce $500-billion Stargate Project.
Welcome to Confluence. Here’s what has our attention this week at the intersection of generative AI and corporate communication:
Five Frontiers
DeepSeek’s Free Reasoning Model
OpenAI and Partners Announce $500-Billion Stargate Project
Five Frontiers
Checking in on five areas we’re watching in 2025 — and getting ahead of their implications for corporate communication.
In a conversation late last year, a client asked where we expected to see big steps forward in generative AI in 2025 and beyond. Off the top of our heads, we listed five areas:
Increasing ubiquity and pervasiveness
Improved memory and personalization
Expanded multimodal capabilities
The arrival and utility of agentic capabilities
Growing context windows
This was not intended to be a comprehensive list. As we’ll share in a bit, there are other important areas we’re watching as well. But we are closely tracking these five areas, and in the past six weeks or so we’ve seen major announcements or advances pertaining to each of them.
So, for today’s edition of Confluence, we thought it would be valuable to provide a quick summary of the recent progress in each of these areas with the goal of catching readers up on these developments — and getting ahead of their implications for corporate communication. Below, we provide a brief overview of each of these areas, some signs of recent progress, and Claude’s take on the implications for corporate communications.
Increasing Ubiquity and Pervasiveness
We expect generative AI to continue to become more ubiquitous and pervasive in the software and devices we use every day. Three mutually-reinforcing factors are driving our expectations here: sustained or increased investment in generative AI across the technology sector, decreasing costs of both development and deployment, and the emergence of powerful small models capable of running locally on devices like a smartphone or personal computer.
Signs of progress:
Apple is continuing to add AI features to its operating system and devices, despite lukewarm reception of the initial rollout of Apple Intelligence. This will affect billions of Apple devices worldwide.
Meta is planning to increase its investment in AI by approximately 70% in 2025. Over three billion people use Meta’s core software products — Facebook, Instagram, WhatsApp, and Messenger — on a daily basis, and generative AI is likely to be increasingly integrated into each.
Microsoft is now including Copilot in the Microsoft Personal and Family Licenses. A year ago, we wrote that we expect Copilot to eventually become part of Microsoft’s standard offering, and this certainly appears to be a step in that direction.
Claude’s take on the implications for corporate communication:
The ubiquity of generative AI creates both urgency and complexity for corporate communicators. As these tools become standard features in the devices and software that stakeholders use daily, the question shifts from whether organizations will engage with AI to how they’ll do so effectively and responsibly. Communications teams need clear policies about AI use across channels — from social media responses to employee communications — that reflect this new reality. They also need to prepare for stakeholders who increasingly expect AI-enabled personalization and responsiveness in their interactions with companies. While some organizations might be tempted to wait for the technology landscape to settle, the pervasiveness of these tools suggests that waiting too long could leave them notably behind stakeholder expectations and competitor capabilities.
Improved Memory and Personalization
While generative AI has made remarkable strides in many areas, true memory and personalization remain more aspirational than achieved reality. Today’s AI can maintain context within a single conversation and, in some cases, reference past interactions through careful engineering. But we’re still far from AI systems that genuinely “know” us — that remember our preferences, understand our context, and build on our shared history the way humans do naturally. The progress we’re seeing suggests we’re moving in that direction, but the gap between current capabilities and truly personalized AI interactions remains substantial.
Signs of progress:
Google Research recently announced a new family of architectures called Titans, designed specifically to optimize memory.
In a recent interview at the World Economic Forum in Davos, Anthropic CEO Dario Amodei noted that memory is “very important” to Anthropic’s broader vision and that some form of memory is “coming soon” to Claude.
OpenAI continues to refine ChatGPT’s memory capability, which originally launch in 2024 (and is now available to all ChatGPT users for free).
Claude’s take on the implications for corporate communication:
For corporate communications teams, improved AI memory could be transformative for daily workflow and efficiency. Instead of repeatedly briefing an AI system on brand voice, style preferences, and organizational context, communicators will be able to build on accumulated knowledge across projects and campaigns. This persistent memory will streamline everything from content creation to media monitoring to stakeholder tracking. The practical impact will be significant: faster onboarding of new team members, more consistent outputs, and reduced time spent providing context and background. Beyond these internal benefits, as people grow accustomed to AI systems that remember their preferences and past interactions, they'll expect similar continuity from organizations – raising the bar for how companies maintain relationships with their stakeholders across multiple touchpoints.
Expanded Multimodal Capabilities
When ChatGPT first launched in November of 2022, it could only work in text, processing a user’s text input and producing text output of its own. Since then, the most widely-used generative AI products — most notably ChatGPT and Gemini — have incorporated the ability to work in other modalities, including image, audio, and video. Recent developments point to models and tools that are able to seamlessly work across modes in real time, and this integration is accelerating.
Signs of progress:
OpenAI recently added video integration to ChatGPT’s Advanced Voice Mode, allowing you to share real-time video from your device with ChatGPT during a conversation.
Google’s Gemini can now “simultaneously process multiple visual streams in real time.”
Video generation models, like OpenAI’s Sora and Google’s Veo 2, are increasingly integrated into the ChatGPT and Gemini user experience.
Claude’s take on the implications for corporate communication:
For corporate communications teams, expanded multimodal capabilities dramatically reshape both what's possible and what's expected. The ability to seamlessly convert and adapt content across formats — turning blog posts into videos, press releases into social media assets, or earnings calls into shareable clips — will fundamentally change content creation workflows. This evolution demands new thinking about how to maintain consistent brand voice and messaging across an expanding range of formats. Communications teams will need to develop new skills and processes to effectively orchestrate these capabilities, while being mindful that greater creative possibility doesn't always mean greater communication effectiveness. The key will be leveraging multimodal AI to enhance, rather than complicate, core messaging strategies.
Arrival and Utility of Agentic Capabilities
“Agent” has been perhaps the buzziest of AI buzzwords over the past several months, and we’ve written about AI agents and their implications quite a bit here in Confluence (most recently at the end of December). There’s no agreed-upon definition of AI agents, but our working definition is that the concept represents “AI that has agency: goals, the ability to plan, and the ability to make decisions and take actions in service of those goals and plans.” Recent developments suggest that the buzz is justified, and that 2025 may indeed be the year where AI agents start to make a mark.
Signs of progress:
This past Friday, OpenAI introduced Operator, “a research preview of an agent that can use its own browser to perform tasks for you.” We have access to Operator and are putting it through its paces. We’ll share our impressions in the coming weeks.
In Dario Amodei’s aforementioned interview at Davos, Amodei discusses Anthropic’s vision of “virtual collaborators” and their goal of building “a model that is able to do anything on a computer screen that a virtual human could do.” This comes on top of Claude’s Computer Use capabilities, which Anthropic announced in October and we wrote about here.
Microsoft has reorganized its development team around AI, with CEO Satya Nadella noting that “… our AI platform and tools will come together to create agents, and these agents will come together to change every SaaS application category, and building custom applications will be driven by software (i.e., ‘service as software’).”
Claude’s take on the implications for corporate communication:
The emergence of AI agents marks a fundamental shift from AI as a tool to AI as a collaborative partner in communications work. This evolution will enable communications teams to delegate entire workflows — from media monitoring to content distribution to stakeholder tracking — to AI agents that can work independently toward defined goals. While this promises significant efficiency gains, it also introduces new challenges around oversight and quality control. Communications leaders will need to develop frameworks for effectively delegating to AI agents while maintaining appropriate human judgment and strategic direction. The key will be identifying which tasks truly benefit from autonomous execution versus those that require ongoing human involvement. Most importantly, teams will need to resist the temptation to automate simply because they can, focusing instead on where agentic capabilities genuinely enhance the quality and effectiveness of communications work.
Growing Context Windows
You can think of a model’s context window as its working memory. It’s a measure of how much information these models can process at once. When ChatGPT launched in late 2022, it could handle roughly 3,000 words in a single interaction. Today, GPT-4o, the model that powers the free version of ChatGPT, can handle roughly 96,000 words, a 32x increase. Other models, like Google’s Gemini, have significantly larger context windows than even that. This exponential growth in context capacity enables AI to maintain longer conversations, analyze larger documents, and tackle more complex tasks while maintaining coherence and accuracy. There are real technical barriers to overcome to maintain this trajectory, but the trend line is clear.
Signs of progress:
Google’s Titans family of architectures, which we referenced earlier, represents a novel approach to increasing context windows, with promise to scale context windows to two million tokens (roughly 1.5 million words) while maintaining performance. This is roughly twice the size of the largest context windows available today.
Researchers are actively exploring novel ways to expand context windows even further.
Claude’s take on the implications for corporate communication:
Expanded context windows fundamentally change how communications teams can leverage their organization's knowledge base and history. Rather than analyzing documents piecemeal, teams can process entire archives of press releases, annual reports, and media coverage simultaneously — enabling deeper pattern recognition and more consistent messaging over time. This capability becomes particularly powerful when combined with memory and personalization: imagine being able to analyze years of stakeholder interactions alongside current market conditions to craft more nuanced communication strategies. For day-to-day work, larger context windows mean communications teams can trust AI systems with more complex tasks, from drafting comprehensive crisis communications plans to synthesizing extensive stakeholder feedback. The challenge will be developing new workflows that take full advantage of this expanded capability while maintaining strategic focus.
So, there you have it. The above is by no means a comprehensive summary of every recent development of note, but it is a snapshot on where we’re seeing progress in key areas and what continued progress might mean for the work of corporate communications. This was all prompted by a simple question someone asked us a few months ago. If someone asked us that question again today, we’d add a sixth item to the list of areas we’re watching: the rapid advances in generative AI “reasoning” models. More on that below.
A Free Reasoning Model
DeepSeek-R1 gives us our first free and open source reasoning model.
We haven’t talked about DeepSeek — a Chinese AI lab noted for releasing open source models that, while capable, are more notable for their stated low cost to train — in past editions of Confluence, but they’ve recently achieved something that has caught our (and many others’) attention.
This week, DeepSeek released DeepSeek-R1 — the first major “reasoning” model to come from a lab other than OpenAI. We’ve been writing about this class of models since September when OpenAI announced o1-preview, but the most important thing to understand about them is that they spend time thinking before they respond, unlike other LLMs. While DeepSeek-R1 doesn’t quite reach the levels of o1 or o1 Pro based on benchmarks or anecdotal evidence, it does have one important edge — it’s free.
Every person with an internet connection can now access a reasoning model that is near the frontier of generative AI capabilities for free. And if you haven’t spent time working with o1-preview, o1, or o1 Pro, it’s worth spending time with DeepSeek-R1 to get a feel for how this class of models works. A quick note, you’ll want to make sure that you click the “Deep Think (R1)” button before submitting your prompt to make sure you’re working with R1.
One thing we’ve noticed is that DeepSeek-R1 is more transparent about how it “thinks” before responding to your prompt. For example, below is an example of our initial prompt and DeepSeek-R1’s “thoughts” in italics. Note that these are DeepSeek-R1’s thoughts before responding and producing an output — not the output itself.
I’m the head of communication for a company planning to divest a sizeable business unit that makes and sells furniture. It’s successful, but far from our core business of running hotels. I need help mapping out a complete communication strategy that addresses all our key stakeholders (including leadership, employees, the media, local government representatives, key customers and suppliers, etc.). I’ll need key messaging, a strategy, a plan that aligns with that strategy, and supporting content. Before you get started, ask me any clarifying questions that would be helpful.
Okay, let’s start by understanding the user’s needs. They’re the head of communication for a company divesting a furniture business unit. The main goal is to create a comprehensive communication strategy addressing all key stakeholders. The user wants key messaging, a strategy, a plan, and supporting content.
First, I need to gather more information to tailor the strategy effectively. Let me think about what questions would clarify the situation.
1. **Reason for Divestment:** Why is the company selling the furniture unit? Is it strategic realignment, financial reasons, or something else? The reason will shape the messaging, especially around the narrative of focusing on core business.
2. **Timeline:** When is the divestment happening? Knowing if it’s imminent or in a few months affects the communication plan's urgency and phases.
3. **Impact on Employees:** Will there be layoffs, transfers, or changes? Employee concerns are critical, so the messaging must address job security and transitions.
4. **Buyer Information:** Is there a buyer identified? If yes, how will they be involved in communications? If not, how transparent can the company be?
5. **Customer and Supplier Commitments:** How will existing obligations be handled? Assurances here prevent panic and maintain trust.
6. **Local Government Relations:** Are there regulatory or community impacts? This affects messaging to government stakeholders and possibly the media.
7. **Previous Communication:** Have there been prior announcements? Consistency is key to avoid mixed messages.
8. **Leadership Alignment:** Are all leaders on the same page? Internal alignment ensures consistent external messaging.
9. **Reputation Risks:** Any potential backlash or concerns? Anticipating these helps in crafting proactive messages.
10. **Cultural Sensitivities:** Any regional or cultural factors? Important for tailoring messages to different audiences.
Each answer will influence the strategy. For example, if the divestment is due to strategic focus, the core message would be about strengthening the hotel business. If there's a buyer, joint communications might be needed. Employee impact details will dictate the level of support messaging required.
I also need to consider the channels for each stakeholder. Employees might need town halls and emails, while media requires press releases and briefings. Local governments might need direct meetings. Suppliers and customers would need personalized communications.
Supporting content like FAQs, press releases, talking points, and social media drafts will depend on the specifics of the situation. Ensuring all content aligns with the key messages and strategy is crucial.
Potential pitfalls include inconsistent messaging across stakeholders, not addressing employee concerns adequately, or failing to manage media narratives. The plan must be phased, starting with internal communications before going public, to control the information flow.
I should structure the strategy around key phases: pre-announcement, announcement, post-announcement, and ongoing management. Each phase will have tailored actions for different stakeholders. The messaging hierarchy should start with the core reason, then address each stakeholder's concerns, ensuring empathy and clarity.
Overall, the strategy needs to be transparent, timely, and tailored, with a focus on maintaining trust and minimizing disruption. The supporting content should reinforce these elements across all channels and touchpoints.
When faced with an ask and little context, DeepSeek-R1 does a remarkable job of thinking through many of same (though not all) questions and concerns we would have. Similar to when we wrote about o1 Pro, it’s remarkable to see how capable these models are out of the box. And we expect that with more context and a particular perspective, they can become even more capable, more useful.
What else does DeepSeek-R1 signal? Simply put, there’s no slowing down. Because it’s open source, it’s easier for others to see how a reasoning model like it works and thus easier to emulate. The fact that DeepSeek was able to develop R1 (and do so at a purported low cost) means that other reasoning models from other labs will not be far behind. Even if the current leaders at OpenAI and Anthropic stay the leaders for the foreseeable future, we expect there to be more labs creating more models that originate from more countries that get quite close to the true frontier of capabilities.
OpenAI and Partners Announce $500-Billion Stargate Project
The major investment reflects a new blueprint for AI expansion and positions AI infrastructure as a national priority.
In a press conference at the White House this week, OpenAI and several key partners unveiled the Stargate Project. According to OpenAI’s website announcement, Stargate is “a new company which intends to invest $500 billion over the next four years building new AI infrastructure for OpenAI in the United States. We will begin deploying $100 billion immediately.” The initiative brings together SoftBank, Oracle, and MGX as funding partners, alongside Arm, Microsoft, and NVIDIA as technology partners.
The initial site buildout is already underway in Abilene Texas, with more “campuses” to come. Oracle, NVIDIA, and OpenAI will manage the computing system’s build-out and operations. As for Microsoft, OpenAI says that it “will continue to increase its consumption of Azure as OpenAI continues its work with Microsoft with this additional compute to train leading models and deliver great products and services.”
The announcement’s timing in the first week of the new presidential term positions AI infrastructure development as a national priority — to say nothing of the investment of $500 billion over four years, which represents a scale that moves beyond typical corporate investment into territory previously reserved for major national initiatives.
The Stargate project likely marks the beginning of a new phase in AI development, where the computing requirements for next-generation AI systems demand infrastructure investment at national-project scale. The involvement of SoftBank, traditional tech giants like Oracle and Microsoft, and state funds like MGX suggests a new model of financing and operating AI infrastructure is emerging — one that combines technical expertise, massive capital, and strategic partnerships in novel ways.
One theme in our first two items in this edition of Confluence is that there are no signs of generative AI progress or investment slowing down. In fact, the reverse rings more true. The announcement of Project Stargate is another data point — and a massive one, at that — that suggests that we should expect the rapid pace of progress to continue in 2025.
We’ll leave you with something cool: The Winter X Games are using AI to help judge snowboarding competitions.
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.