Frontier Models

OpenAI's GPT-5: Ambitious promise but a rocky rollout

OpenAI launched the much-anticipated GPT-5 this past Thursday, making it available to all ChatGPT users, including free tiers. The new flagship model promises to be smarter and faster, with 45% fewer factual errors than GPT-4o. The announcement highlighted overall better performance in real-world tasks, including coding, writing, and healthcare-related queries. Most notably, GPT-5 introduces automatic model selection that chooses between reasoning and quick response models based on a query's complexity. This, in theory, removes the guesswork of choosing the right model for a given task.

But the rollout has been, in CEO Sam Altman's own words, "bumpy." The routing system that was meant to select the correct model automatically wasn't working correctly when it rolled out Thursday, making "GPT-5 seem way dumber," according to Altman. More controversially, OpenAI retired access to popular older models like GPT-4o, o3, and o4-mini. The backlash was swift and severe—with one user describing the new model as “polished, clipped, and weirdly impersonal … like they replaced your favorite coffee shop with a vending machine.” Within 24 hours, Altman reversed course, announcing users with (paid) Plus accounts could continue using GPT-4o while the company monitors usage patterns.

Nate Jones, in his paid newsletter, provides some succinct analysis:

"Architecturally, GPT-5 is a leap forward: it can reason more deeply, integrate tools more flexibly, and work with larger contexts than GPT-4o. But scale has introduced new failure modes that undermine that capability in practice. Misroutes can push prompts to sub-optimal models. Hardware variance and routing issues can produce sudden dips in quality. And changes in tone or creativity hit hardest for users who built habits and workflows around the 'feel' of GPT-4o."

It will take some time for the dust to settle on GPT-5. Amidst the controversy, there have been some impressive results reported, including experiments done by Nate Jones and Ethan Mollick (link immediately below).

GPT-5: It just does stuff
☕ Medium Read (8 minutes)

The GPT-5 rollout has been a big mess
⚡ Quick Read (5 minutes)

Frontier Models

OpenAI's new open-weight models give organizations more options to explore AI

On August 5, OpenAI released two open-weight models—gpt-oss-120b (120B) and gpt-oss-20b (20B)—its first open-weight models since GPT-2 in 2019.

Both are high-performing models, which are freely downloadable and customizable. The larger 120B model reaches near-parity with OpenAI's o4-mini, and the smaller 20B model roughly matches o3-mini. The 20B model can run on a laptop with just 16GB of memory; the 120B version requires only a single 80GB GPU, dramatically lowering the infrastructure barrier that has kept many teams from experimenting with AI integration. It’s important to note that these are not knowledge models; they are optimized for API access to information, “instead of trying to memorize the entire internet.”

With these models, teams can now prototype AI features locally without API costs, maintain complete control over sensitive data (crucial for industries with strict compliance requirements), and customize models for specific use cases without vendor lock-in.

OpenAI just released its first open-weight models
since GPT-2
☕ Medium Read (8 minutes)

AI for Research

Digital twins: representing real-life users with AI

"A digital twin is a genAI-based model of a particular individual that can be used to predict both individual and population-level preferences and behaviors."

In a recent article, the Nielsen Norman Group takes on the topic of creating and using digital twins in user research. They characterize a digital twin as an "artificial digital clone: a system that can complete surveys, predict choices, or interact in real time on behalf of a real individual."

Digital twins are different from synthetic users in that they represent specific individuals, whereas synthetic users represent segments or characteristics. However, as the article points out, these two approaches, in reality, exist on a continuum:

• At one end, a synthetic user is built from attributes that many real people share.

• At the other end, a digital twin is based on information gathered about very few people — often, just one.

Use cases for digital twins include:

• Filling in skipped survey items using the twin's predictions

• Asking respondents a subset of questions and inferring the rest using the twin (resulting in shorter surveys)

• Using data from one-time interviews to build twins that stand in for groups that are expensive or impractical to follow over time

The article goes on to describe options for building digital twins and the ethical questions that should not be overlooked.

Digital twins: simulating humans with Generative AI
⚡ Quick Read (5 minutes)

AI in the Enterprise

How AI-forward companies drive AI adoption

Many product leaders are looking to accelerate AI adoption in their organizations. But how to accomplish that? Lenny Rachitsky and Peter Yang interviewed executives leading that effort in AI-forward organizations. They identified five key steps to success:

1. Explain the how.

2. Track and reward adoption.

3. Cut the red tape.

4. Turn enthusiasts into teachers.

5. Prioritize the high-impact tasks.

The post on X goes on to specify 25 associated best practices. These are a great read to consider which practices might be a fit for your organization.

@lennysan post on X
⚡ Quick Read (4 minutes)

A final quick thought:

• Al doesn't do it end-to-end.

• It does it middle-to-middle.

• The new bottlenecks are prompting and verifying.

@balajis post on X

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