Top 5 AI Advances You May Have Missed in 2025

Beyond headlines shouting "AGI", "DATA CENTERS" and "JOB REPLACEMENT", 2025 produced quieter breakthroughs aimed at making AI more interpretable, dependable and safe in the real world. Here are five that I believe deserve a mention.

1) Baby Dragon Hatchling (BDH): building reasoning from local “neurons”Baby Dragon Hatchling swaps the usual Large Language Model (LLM) layout for a network structure of locally interacting neuron-like units (attempting to more closely mimic the human brain). The authors report GPT level results at similar sizes, while being able to continually learn.

Paper: https://lnkd.in/gGTxyNEK

‍Code: https://lnkd.in/gEvY22ZB

‍2) Hallbayes: a calculator for hallucination riskHallbayes takes a prompt and outputs a risk score that the model will make something up, then recommends “answer” or “refuse” according to the service level target. It works without retraining, using an OpenAI-compatible chat API and transparent maths you can log for audit.

Paper: None

Code: https://lnkd.in/ga7t-vc2

‍3) RAIS: deepfake detection that keeps up with new attacksAudio deepfake detectors often degrade as attackers change tactics. RAIS uses continual learning with a small replay buffer, but selects that buffer using auxiliary labels generated by a separate network, to preserve diversity and reduce forgetting. It reports an average equal error rate of 1.953% across five stages.

Paper: https://lnkd.in/g__nAkjR

‍Code: https://lnkd.in/g67ZbeKX‍

4) DETree: detecting human and AI collaboration, not a simple binary Workplace text is frequently co-authored by people and models. DETree treats different collaboration routes as a hierarchy and trains representations to match it, supported by a hybrid-text benchmark called RealBench. The goal is robustness when the writing style or model changes.

Paper: https://lnkd.in/gJRQrfVe

‍Code: https://lnkd.in/g4R7Cjqx

‍5) PD-SGD: privacy via “plausible deniability”Models can leak whether an individual was in the training data. PD-SGD uses rejection sampling to sometimes skip parameter updates when a mini-batch cannot be plausibly denied, aiming for a better privacy and performance balance than standard approaches.

Paper: https://lnkd.in/g_mzcjVJ

‍Code: Private repository

Although these tech headlines were dwarfed by the hype. The next wave of useful innovation in AI is coming in the form of trust infrastructure.

For business leaders, that means selecting vendors that go beyond a stylish chat interface, and looking to work with those that are taming the AI dragons beginning to hatch.

‍

22 December 2025