Confluence for 2.16.2025

OpenAI updates Chat GPT-4o. Google offers infinite chat memory. Perplexity brings Deep Research to all. New data on generative AI in the workplace.

Midjourney prompt: The discovery of a well of infinite knowledge, in the style of Alphonse Mucha

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

• OpenAI Updates Chat GPT-4o

• Google Offers Infinite Chat Memory

• Perplexity Brings Deep Research to All

• New Data on Generative AI in the Workplace

OpenAI Updates Chat GPT-4o

But it’s up to us to figure out how.

One thing that got our attention just yesterday is that OpenAI updated GPT-4o (the version used by the vast majority of users), and we — and just about everyone else who use it — have noticed it’s much better. There’s no change log posted, so we can’t say how or why, but it seems smarter, and more collegial. Many are saying it’s more like Claude, which will be good news as we and many others have preferred Claude for some time.

It also seems really, really good at web search. We had no problem with questions that required both accuracy and currency, like last night when we asked “Who’s winning the US Canada game” and quickly got:

This morning we asked for today’s weather forecast for Daytona, FL and any key news about the Daytona 500 from the past 24 hours, and GPT-4o gave us an accurate and helpful readout on both. For years, many have considered voice agents like Siri and Alexa the leading edge of information assistants — that technology is about to reach a whole new level thanks to large language models. While tools like ChatGPT can still invent facts or get details wrong, those hallucinations are far less frequent. And as these models get better at reasoning, their ability to make good choices about what to include when answering queries that require an internet search will only improve.

If nothing else, your news clipping service is in serious trouble.

Google Offers Infinite Chat Memory

We may stand on another threshold of increased utility.

One of the things that limits the helpfulness of large language models is that while they remember what’s happened within a chat, they don’t really remember what’s happened across all your chats. There are some ways around this — ChatGPT selectively remembers some facts from your conversations, and Claude (and our leadership AI, ALEX) allow you to add preference information that they remember in every chat— but it’s just not the same as having real memory of past conversations.

While most regular users have gotten used to this, in reality, it’s a big limitation. Imagine having a colleague who immediately forgets your past conversations in every new conversation — that would be … suboptimal.

But now it seems Google has changed all that, as they have brought chat memory to Gemini Advanced users. We have not yet put this through its paces, but we will. If it works as advertised, this is a big deal — not just for Gemini users, but for all users of large language models as we presume it’s a technical ability that will diffuse across the industry. Imagine having a colleague who immediately remembers, with superhuman accuracy, your past conversations in every new conversation — that would be … helpful, to say the least.

Perplexity Brings Deep Research to All

A new development in the latest AI threshold

Last week, we wrote about OpenAI’s Deep Research capability, describing it as a threshold moment in generative AI that stunned by its ability to produce research analyst-level work in minutes. Now comes another noteworthy development in AI research capabilities — not so much a new breakthrough as the democratization of the capabilities we discussed last week.

Perplexity announced that it has launched its own Deep Research feature, and while the capability may not match OpenAI’s in raw power, it comes with a crucial difference: it’s free for all users, with unlimited access for Pro subscribers. The system functions similarly to OpenAI’s offering: you submit a query, and it conducts autonomous research by performing dozens of searches, reading hundreds of sources, and synthesizing the information into a report, typically in under three minutes.

Our initial testing suggests that while Perplexity’s reasoning capabilities may not quite match OpenAI’s model, its speed and accessibility make it a remarkable tool. Perplexity claims impressive benchmark scores – including 21.1% accuracy on Humanity’s Last Exam – suggesting robust performance across a wide range of subjects (Open AI’s Deep Research scores a 26.6%).

What makes this development particularly interesting isn’t just the technology itself, but rather its positioning in the market. While OpenAI’s paid Deep Research tool represents the cutting edge of what’s possible with AI research, Perplexity’s offering makes similar functionality accessible to everyone, without the $200/month pricetag. OpenAI’s Sam Altman has announced plans to offer limited Deep Research access on the free tier (2 uses per month, compared to 10 for Plus users), but Perplexity is already delivering broader access now. Similar to Google’s infinite chat memory we wrote about above, this move is another powerful AI capability reaching the masses. This democratization of access means the threshold we discussed last week isn’t just approaching — it’s here, and it’s becoming available to everyone.

We suggest you start exploring Perplexity’s Deep Research now. To get started, visit perplexity.ai and select “Deep Research” from the drop-down menu:

Even if you’re not ready to commit to OpenAI’s $200 monthly subscription, you can begin understanding how these systems reason, what they’re capable of, and where they fall short.

New Data on Generative AI in the Workplace

Relatively few are using generative AI each day, but those who do are reaping real benefits.

This week, researchers published a study with new data on the proliferation and use of generative AI in the workplace. The research from scholars at Stanford, the World Bank, and Clemson surveyed over 4,000 U.S. workers using a new instrument designed to track adoption, usage patterns, and effects on productivity across industries and demographic groups.

Its headline finding is that 30.1% have used generative AI tools at work, with about 33% of those who have used it doing so each day. Below is a brief summary other key findings from the research by Claude:

This paper examines the labor market effects of generative AI through a new survey conducted in December 2024, finding that 30.1% of U.S. workers above age 18 have used generative AI tools at work since their public release. The research shows that generative AI use is most prevalent among younger workers, those with higher education levels, higher-income individuals, and workers in industries like information services and management. Among those who use generative AI at work, about one-third use it daily, though typically for limited hours per week. The study found significant productivity gains, with tasks that normally take 90 minutes being completed in 30 minutes when using generative AI. The authors also discovered interesting variations in adoption across demographics, with men reporting higher usage (38%) than women (27.8%), and notable differences across racial groups and industries. The paper contributes to understanding how generative AI is reshaping the labor market and has implications for policymakers, businesses, and researchers studying the economic impact of this technology.

Beyond the headline findings, a few things stand out. While we shouldn’t take this as gospel, the fact that those who use these tools estimate that generative AI helps them complete work three times more quickly demands attention. Speaking from our team’s experience, it feels like we’re able to get more done in less time than ever before. The gain isn’t theoretical — when work that took three hours now takes only one, there are two hours back in your day for other priorities. If you or individuals on your team aren’t working to leverage this technology on a regular basis, we believe you’re missing out on reclaiming significant time in your day.

Another notable finding concerns which tools respondents are using. According to the survey, ChatGPT and Gemini were the most popular generative AI tools, while Claude — our preferred tool for most tasks — ranked second to last, ahead of only Midjourney. Both ChatGPT and Gemini are capable tools, but for use cases related to writing, editing, summarizing, and qualitative analysis, we get more value out of Claude. We believe the 3.5 Sonnet, in particular, crosses an important threshold in writing quality as it can produce high-quality writing with straightforward prompting.

There are real barriers in using generative AI every day — there are certain tasks where its utility is limited, organizational security and IT policies often determine which tools people can and cannot use for work (and the tool makes a big difference), and there is a learning curve with the technology. But these barriers shouldn’t be the reason to retreat to the familiar. Our view is that we should be looking for ways to reduce these barriers so that we can take full advantage of the power these technologies afford.

---

We’ll leave you with something cool: Luma has released a new video model, Ray 2, and its image to video capabilities are impressive.

---

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.