Looking to the Near Future

The most exciting part of working with large multimodal models is that the ground keeps shifting beneath us. In the coming seasons I’ll keep returning to a single question in many guises: How well can an artificial system notice, interpret, and respond to the subtle patterns of human thought and feeling? What follows is a sketch of where I hope to travel, not a schedule of guaranteed milestones.

Models in Motion

GPT-4o already behaves very differently from its predecessor, and its successors will not wait politely for me to finish one experiment before the next one lands. My ambition is to treat each silent model update as data in itself: to log emergent behaviours, compare them against earlier baselines, and share concise reflections wherever a shift looks meaningful. Think of these research pages as a lab diary rather than as polished papers.

Personas with Personality

The small family of psychometrically profiled “AI personas” introduced on this page has proved surprisingly revealing. Instead of racing to add dozens more characters, I want to deepen the existing set—running longer dialogues, trying prompts from more cultures, and seeing whether stable moral or emotional signatures really survive contact with new training data. Whether those signatures persist or dissolve, the outcome will sharpen what we mean by “consistency” in a language model.

Theory-of-Mind, Revisited

Pilot runs suggest GPT-4o can pass classical false-belief tests in several languages, but the results wobble with minor prompt tweaks. My plan is to tighten the methodology, invite a handful of colleagues to replicate the protocol, and publish a brief methods note if the findings converge. If they don’t, the divergence will be documented just as openly. The goal is to map why the model sometimes succeeds, not merely to prove that it can.

Generative Item Crafting

Automated question writing has long tempted test developers. I am experimenting with an item-generator that drafts both personality and cognitive-ability questions and then vets them against human pilot data. If the signal-to-noise ratio improves, I’ll release a petite public sample—enough to demonstrate the workflow without suggesting that quality control is solved. The wider ambition is to show that generative tools can compress, not replace, painstaking psychometric labour.

Narrative Feedback for Test-Takers

Raw scores rarely tell a respondent anything they can use. I’m exploring whether a large language model can draft plain-language narrative reports that translate psychometric results into clear, actionable stories: what a pattern of answers might mean, how confident we can be, and where the limits of interpretation lie. The focus is on readability, fairness, and tone—testing whether the same core findings can be framed in ways that are both accurate and genuinely helpful. If the early prototypes feel promising, I’ll share anonymised examples to invite critique on language, cultural nuance, and potential bias.

Multimodal Ethics Conversations

Voice-based Socratic dialogues with two prototype agents—AI Athenus and AI Skeptos—are already generating rich transcripts. Next I hope to analyse how spoken exchanges differ from text-only ones: Do participants feel more engaged? Less defensive? If the gains look real, a lightweight browser demo may follow; the allure of a VR version remains on the horizon, but only if simpler forms prove their worth.

Staying Responsive

Rumours swirl about models with million-token context windows and broader multimodal reach. Rather than chasing every headline, I plan to anchor the programme to a few core questions—consistency of persona traits, depth of social reasoning, quality of generative items, and usefulness of narrative feedback—then revisit the same questions whenever a major model iteration arrives. This rhythm should keep the work cumulative while leaving room for serendipity.

In short, consider this roadmap a statement of ambitions, not promises: directions I intend to explore, subject to evidence, time, and the inevitable surprises that make research worthwhile.