Deep Currents 03.03.25

Deep Currents is a new format I’m experimenting with to surface timely and exciting developments in generative AI. My goal is to provide a curated list of recent breakthroughs, product updates, and helpful articles that have resonated with me as a design director trying to stay on top of this rapidly evolving field. Please let me know if you find it useful or if you have thoughts about how it could be better!

Okay, let's dive into this week's currents…

AI Tool Spotlight

• Midjourney finally improved the image organization features in the web app, and now allows you to group folders and easily move images into, and more significantly out of folders (a previous limitation that drove me nuts). This update follows a recent enhancement to allow mixing of multiple moodboards and style reference codes, furthering Midjourney’s lead as the most flexible and creative image generation tool.
• OpenAI has rolled out its Operator agent tool to more countries including Australia, Brazil, Canada, India, Japan, and the U.K. All you need now is a $200 USD/month Pro plan and you can get ChatGPT to buy your groceries for you.
• OpenAI just made their Deep Research report-generating feature available to all paid users. Previously Deep Research was only available to users with Pro plans. Subscribers on the Plus and Teams plans can now generate up to 10 reports per month, while Pro plan users have a much higher limit of 120 reports. Generating each report uses significant computing resources, and given that it can take up 30 minutes to generate a complex report, you’ll want to use those requests wisely.
• Perplexity launched their own Deep Research tool last month and made it available to all users, not just to Pro subscribers. It’s not as good at generating reports as the ChatGPT version, but hey it’s free!
• If you’re looking for help with your coding, Google just launched a free version of Gemini Code Assist for individuals. Code Assist is similar to GitHub Copilot but offers higher usage limits.

Learning Corner

• What Can Language Models Actually Do? by Dan Shipper is a really helpful article to understand why LLMs are really good at summarizing.
• LLMs Turn Every Question into an Answer by Dan Shipper. This is a followup from the previous article. In part 2 Dan explains how and why LLMs are great at elaborating.

Technical Breakthroughs

• Last week Anthropic finally released a new frontier model called Claude Sonnet 3.7. It’s the first hybrid reasoning model to appear on the market, offering users the option to turn on an “extended thinking” mode for prompts that require more complex analysis, such as coding applications or generating research reports. Claude is apparently holding onto its crown as the top coding model with this release.
• Just a few days later OpenAI released ChatGPT 4.5. Initially only available to Pro and API users, they’ve promised to make it available to Plus and Team subscribers this week. This is not a reasoning model so improvements are not expected for coding or statistical analysis (that’s expected in an upcoming GPT-5 hybrid model). However overall performance is said to be noticeably better with less hallucinations and more nuanced vibes.
• Earlier this month, Perplexity released an open source “de-censored” version of the DeepSeek R1 model called R1 1776. The number 1776 is a reference to the United States’ Declaration of Independence, symbolizing a commitment to freedom of information, as Perplexity employed an interesting post-training technique to undo CCP brainwashing of the original R1 model. This means you can ask it about the Tiananmen Square protests and Uyghur reeducation camps and it will provide more well-rounded answers.

That’s all for now! Of course a lot more has happened lately in the world of generative AI. What has resonated with you lately? Let me know in the comments, or send me an email.