Frontier Models

The technical reality behind AI personality

When ChatGPT-5 rolled out, users immediately noticed something was off—the model felt different, less personable than ChatGPT-4o. Social media lit up with complaints about the "personality change," as if a familiar colleague had been replaced by someone more formal and distant. But here's the thing: LLMs don't actually have personalities. So what exactly is happening when an AI feels warm, quirky, or cold?

In a recent article, Ars Technica tackles the issue of personality in LLMs:

"...millions of daily users engage with AI chatbots as if they were talking to a consistent person—confiding secrets, seeking advice, and attributing fixed beliefs to what is actually a fluid idea-connection machine with no persistent self...LLMs are intelligence without agency—what we might call "vox sine persona": voice without person. Not the voice of someone, not even the collective voice of many someones, but a voice emanating from no one at all."

The article goes on to explain all the technical factors driving LLM personality:

• Pre-training

• Post-training

• System prompts

• Chat memory

• Context and RAG

• Manufactured spontaneity

For anyone working with AI, particularly those creating AI-powered experiences, it's essential to understand how each of these factors functions and the influence they can wield on AI personality.

The personhood trap: How AI fakes human personality
☕ Medium Read (11 minutes)

Designing with AI

ASCII wireframes: The surprisingly powerful way to prototype with AI

Guest author with the Design with AI newsletter, Nimisha Patil, presents two handy techniques for more effective vibe coding:

• ASCII wireframes to explore structure

• Figma-to-Code for pixel-perfect results

Instead of asking the AI to do everything at once, Nimisha's process separates layout structure from visual design details.

I especially like how she details the prompts and process around creating ASCII wireframes, and how this translates the traditional process of focusing on structure first (wireframing) into the age of AI. ASCII wireframes allow for faster iteration in the early stages of design and are resource-efficient (using fewer tokens/credits than visual generation).

A crucial part of her process is then asking the AI to review the ASCII wireframes with a panel of experts: "Review these wireframes with a panel of UX, marketplace design, and local market experts. What would they improve?" She has found that this step often reveals insights that she otherwise missed.

The article includes a real-world example of an ASCII wireframe, with associated prompts.

My 2 favourite vibe-coding techniques that help when prompts aren't enough
☕ Medium Read (8 minutes)

AI Product Development

Why AI features are eating your favorite app's profit margins

This week, the Wall Street Journal published an (unfortunately paywalled) article detailing why AI, at the same time, is getting both less expensive and more expensive:

"Despite [the] drop in cost per token, what’s driving up costs for many AI applications is so-called reasoning. Many new forms of AI re-run queries to double-check their answers, fan out to the web to gather extra intel, even write their own little programs to calculate things, all before returning with an answer that can be as short as a sentence. And AI agents will carry out a lengthy series of actions based on user prompts, potentially taking minutes or even hours."

Companies like Notion, which have AI in their products, purchase AI tokens from a frontier model provider (OpenAI, Anthropic, Google) to drive those features. This has shifted the economics of providing web-based services from one of near-zero marginal cost to incurring costs directly tied to users' usage of AI:

"Ivan Zhao, chief executive officer of productivity software company Notion, says that two years ago, his business had margins of around 90%, typical of cloud-based software companies. Now, around 10 percentage points of that profit go to the AI companies that underpin Notion’s latest offerings."

The costs are particularly acute for coding start-ups such as Cursor and Replit, whose entire service relies on the intensive use of frontier models. Both recently adjusted their pricing to account for increased costs, causing an outcry among some users. For product teams considering AI features, this serves as a reality check: every AI-powered interaction comes with a real cost that someone has to pay. The question isn't just "should we add AI?" but "which AI features deliver enough value to justify their ongoing expense?"

Cutting-edge AI was supposed to get cheaper. It's more expensive than ever. ($)
⚡ Quick Read (5 minutes)

Designing for AI

Chatbots and AI-driven search: Still falling short for everyday users

From the Nielsen Norman Group:

"AI chatbots and AI-driven search can help people find what they’re looking for — even if they don’t know what they’re looking for. In this way, generative AI tools can greatly alleviate a longstanding major pain point of a search-driven web. However, AI’s ability to help with this problem is still limited. In particular, many consumers are simply not aware of how powerful generative AI tools can be in information-seeking, and don’t know how to prompt to achieve better outcomes."

This article provides a great breakdown of traditional information-seeking behavior with search, including the concept of keyword foraging, a behavior users employ when they are unsure of the correct terms associated with the information they need (such as someone attempting a plumbing repair on their own). In theory, AI can eliminate the need for keyword foraging, but the reality is that the chatbot interface demands a level of understanding of AI that many everyday consumers lack.

How AI succeeds (and fails) to help people find information
☕ Medium Read (7 minutes)

Final quick thoughts…

From Ethan Mollick on X:

"The funny thing about the prediction that AI would be writing 90% of all code by now is that the prediction's failure distracts from the fact that AI adoption in code writing is actually extremely high, it was over 30% in December, 2024 according to one measure, with large impact. The speed at which AI has come to write a huge amount of code is pretty amazing, even if it is not as fast as predicted. Becoming a common theme in AI: boosters are more directionally correct than those who dismiss AI, but too early on timelines."

The Browser Company, maker of the AI-focused browser Dia, announced that they've been purchased by Atlassian. It's an interesting development in the world of AI-focused browsers (such as Perplexity's Comet). Atlassian plans to make Dia a go‑to browser for getting work done and supporting enterprise workflows, moving it away from being a consumer-oriented solution.

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