The "Interface-Free" Future Envisioned by an OpenAI Co-Founder

Greg Brockman, co-founder of OpenAI, recently shared his vision for the future of AI. He candidly admitted that ChatGPT's plugin feature—an extension heavily promoted back in 2023—was a failure. The reason, he says, is that "the model wasn't ready yet."

What Brockman envisions instead is not app extensions, but agents that invisibly understand context and take action automatically. He says he is aiming for a world where users never need to learn how to use software—a "future with almost no interface." That said, even OpenAI's own coding agent "Codex" is reportedly still far from this vision.

While industry-wide interest in realizing agentic AI continues to grow, it's clear that significant technical hurdles remain.

Survey of 26,000 Students: The Learning Cost of AI Use Surfaces Two Years Later

A large-scale survey of more than 26,000 students in China has revealed the impact of AI use on learning.

Students who used AI finished their homework faster and achieved higher short-term scores, but their exam performance dropped by as much as 24%. Even more notably, it took roughly two years for the negative effects on entrance exam results to fully materialize.

What looks like improved learning efficiency in the short term may erode foundational skills in the long run—the study suggests that more careful evaluation is needed before rushing to adopt AI education tools.

Academic Society Detects AI-Generated Peer Reviews with "Hidden Prompts," Drawing Criticism from Researchers

It has come to light that an academic society had been embedding "hidden prompts" in manuscripts to detect AI-generated peer review reports during the scientific paper review process.

The technique is intended to detect whether reviewers are using AI, but criticism is mounting within the scientific community, with concerns that it "violates reviewers' privacy" and "undermines trust." As AI-generated paper reviews become more widespread, the question of how to preserve academic integrity is being raised once again.

Google Research Releases TabFM, a Foundation Model for Tabular Data

Google Research has released "TabFM 1.0.0," a foundation model capable of zero-shot classification and regression on tabular data.

Its key feature is the ability to make predictions on structured data mixing numerical and categorical values without any fine-tuning or hyperparameter tuning. It works by passing training examples as context and returning predictions in a single inference. The PyTorch weights are publicly available on HuggingFace.

From the perspective of democratizing tabular data analysis, practical adoption is anticipated.

Distilled Model "Qwythos-9B" with 1M-Token Support Runs on 4GB VRAM

"Qwythos-9B," released by Empero AI, is generating buzz. It is a reasoning model that can handle an ultra-long context of 1 million tokens, and reportedly runs on just 4GB of VRAM.

By distilling into a 9B-parameter model, ultra-long-context inference becomes feasible even in local environments. In the Japanese developer community, practical insights on coding with local LLMs—such as splitting tasks, avoiding MoE models, and the importance of session management—are being actively discussed, showing that local AI adoption is steadily expanding.

---