GPT-5.5 : A Sign of the Future

GPT-5.5 : A Sign of the Future
3 min read

After gaining early access to GPT-5.5, academic Ethan Mollick believes this model is “a big deal.” He sees it as proof that AI progress continues rapidly, while also noting that the capabilities of these systems remain highly uneven.

To illustrate this progress, the author submitted the same programming challenge to several models: create a procedurally generated 3D simulation showing the evolution of a port city from 3000 BC to 3000 AD, aesthetically pleasing and allowing user control. The responses range from o3, OpenAI’s first reasoning model, to what he considers the current best open-weight model, Kimi K2.6, up to GPT-5.5 Pro. Only the latter actually simulates a city that evolves, rather than simply replacing buildings over time. It is also faster than its predecessor: GPT-5.4 Pro took 33 minutes to complete the task, compared to 20 minutes for GPT-5.5 Pro. The responses are published in a gallery that readers can view.

Models, Applications, and Tools Progress Together

The author suggests distinguishing three dimensions: the models, like Opus 4.7, Gemini 3.1, or GPT-5.5; the applications, meaning the products that allow interaction with a model and task assignment; and the harnesses, which are the tools the AI can use and how they are connected to the models. These tools enable the AI to control a computer, write code, conduct research, or create images. In his view, OpenAI has advanced on all three fronts. The GPT-5.5 family is powerful, with GPT-5.5 Pro, accessible only on the web, being the most capable. The Codex application, for its part, is approaching Claude Code and now offers an accessible desktop application.

The new image generation model, sometimes called GPT-imagegen-2, can render good-quality text and create almost any described image. The author tested it with his “Otter Test, ” which asks an AI to depict an otter on a plane using wifi. The model produced images of this scene, including an art gallery in the style of several painters.

Ethan Mollick also tasked Codex, powered by GPT-5.5, with sorting hundreds of anonymized data files from his crowdfunding research, in STATA, CSV, XLS, and Word formats. In four queries, the tool produced an academic paper including a literature review, statistics, and a hypothesis. According to the author, the references are real and the statistics accurate, but the hypothesis seems uninteresting to him and the question of causality remains open. He also asked Codex to design a tabletop role-playing game, with a world and rules, and then simulate sessions. The result is a 101-page PDF, illustrated by the image generator.

Persistent Limitations

Upon closely reading the role-playing game, the author notes that the frontier of capabilities remains jagged. He cites several recurring weaknesses in AI-written fiction: a taste for the bizarre, overly complex ideas that don’t resolve, awkward metaphors like “weather and architecture are the same argument at different speeds, ” overly ornate sentences, and characters who all speak in the same tone. He adds that the first name “Mara” appears frequently.

For the author, GPT-5.5 shows that models are gaining intelligence, applications are becoming more capable, and tools are improving. However, he specifies that “the fiction remains flat” and hypotheses are sometimes uninteresting, even when the statistics are solid. He finally describes a pace of a new model every few months, with increasingly large leaps. “The frontier remains jagged, but it is much farther out than before.”

S’abonner
Notification pour
guest

0 Commentaires
Le plus ancien
Le plus récent Le plus populaire
0
Nous aimerions avoir votre avis, veuillez laisser un commentaire.x