Frontier Models

OpenAI hits 'code red' as rivals close in

This past week's Hard Fork episode provides an excellent recap and analysis of the AI landscape over the past few weeks. OpenAI reportedly declared a "code red" and is pulling engineers off other projects to focus on improving ChatGPT, responding to competitive pressure from Google's Gemini 3 and Anthropic's Claude Opus 4.5—both of which have released state-of-the-art models that challenge OpenAI's position. For years, OpenAI maintained a lead through superior models, but that advantage appears to be narrowing. Gemini 3 has impressed users with its speed and capability, while Claude Opus 4.5 has drawn praise for what users describe as more natural, empathetic interactions.

The overall takeaway: the models are getting noticeably better—faster, more capable, and increasingly useful for everyday work tasks like research, fact-checking, and writing. But there's no clear "best" choice anymore. Each model has strengths, and power users are finding they need to experiment across platforms rather than committing to just one.

Hard Fork | Here's why OpenAI is spooked
Watch Time: 56 minutes

AI and Creative Work

Study: AI-created ads outperformed human-made ones

New research is surfacing a picture of AI in creative work. In one study, AI-generated visual ads achieved 20% higher click-through rates than ads created by human experts, but only when viewers didn't know they were AI-made. Once that label was revealed, performance flipped dramatically, dropping to 31% below human-created ads. A key element: the winning AI ads weren't used straight out of the model. Human experts selected them from a larger pool of AI-generated options, suggesting a workflow where the value comes from curation, not just generation.

The discussion on LinkedIn about this finding points to a broader shift many are observing across industries. As one commenter put it, the role of the expert is moving from "Creator" to "Editor," with AI expanding the range of options that can be explored quickly and cheaply, while human judgment determines what actually ships. It's a reminder that the gains from AI often depend less on the technology itself and more on how it's integrated with human expertise.

LinkedIn thread | The impact of visual generative AI on advertising effectiveness

Designing for AI

How AI is changing how people think

AI isn't just speeding up how we work: it's reshaping how we think. That's the argument from Caitlin McCurrie, whose team has been studying the subtle behavioral shifts emerging as people increasingly collaborate with AI.

Rather than navigating software through clicks and menus, people now externalize their intent in natural language—transforming their mental model from "I operate the system" to "I collaborate with an intelligence." They ask AI to challenge their assumptions, explore ideas they wouldn't raise in meetings, or serve as a sounding board before committing to an opinion.

Caitlin sees the following implications for product teams:

1. Prompts are the new UI.

   What people type reveals friction points far earlier than click maps ever could.

2. Personalization isn’t optional anymore.

   LLMs flatten experience levels, so tools must adapt to match wildly different mental models.

3. Research needs to focus on relationships, not just interactions.

   Human-AI collaboration is becoming relational. People describe their models with trust, affection, annoyance. They’ll use pronouns and nicknames like, “Let's ask my Chat, he’ll know”. It’s closer to ethnography than usability testing.

4. The bar for clarity is higher than ever.

   As AI generates more content, teams must provide stronger narrative scaffolding so meaning doesn’t get lost in a swirl of text.

Caitlin McCurrie | Why AI is changing how people think, not just how they work
⚡ Quick Read (3 minutes)

AI Agents

Finding the right work for AI agents

Vercel shares what they've learned from months of building internal AI agents. Their key insight: The projects most likely to succeed today target work that's low cognitive load but highly repetitive. These are tasks too dynamic for traditional automation but predictable enough for AI to handle reliably. Think lead qualification, abuse report triage, and data entry.

By asking their teams a simple question ("What do you wish you never had to do again?"), they identified the mundane, repetitive work that humans don't love anyway. One example: a lead-qualification agent now lets one person handle work that previously required 10, freeing the other nine employees for higher-value sales work.

This post contains some great, practical suggestions for getting started with agents in your organization.

Vercel Blog | What we learned building agents at Vercel
⚡ Quick Read (4 minutes)

A final thought…

Zara Zhang posting on X:

1. Open a voice memo app.

2. Brain dump your thoughts on a topic; record it.

3. Transcribe it.

4. Import transcript into NotebookLM.

5. Get it to generate a slide deck using Nano Banana Pro.

6. See your rambling thoughts visualized into structured & beautiful slides.

❤️

That’s it for this week.

Thanks for reading, and see you next Wednesday with more curated AI/UX news and insights. 👋

All the best, Heidi