Main announcement
Happy New Year, everyone!
I hope your winter break lasted longer than the average lifespan of a state-of-the-art LLM (which, by my count, is currently about four days). I had every intention of spending my holiday unplugged, but the AI industry apparently doesn't believe in 'Out of Office' replies. While we were all busy trying to remember how to socialize without a screen, the world of AI was busy releasing everything from pen-sized companions to models that can literally find needle-in-a-haystack stars in the sky.
Since my resolution to 'check my email less' has already failed by January 6th, I figured I’d take you down with me. Here is the relentless, fascinating, and slightly exhausting news from the last month.
January Speaker: Prof. Robert K. Cunningham, Pittsburgh University
Date: 10 a.m., January 30th, 2025
Title: Pitt AI Activities: Rethinking Education, Operations, and Research
Abstract:
Recent Artificial Intelligence (AI) breakthroughs have the potential to enable substantial changes to higher education, operations, and research. The university of Pittsburgh has been actively wrestling with the challenges and opportunities offered by AI. This presentation will highlight some of those activities, focusing on those activities that are working well.
Prof. Cunnigham is the Vice Chancellor for Research Infrastructure at the University of Pittsburgh, and responsible for the strategic leadership of the University's research infrastructure and directs special initiatives. He focuses on the effective operation, financial stability and future growth opportunities of Pitt research platforms.
Cunningham has broad experience across STEM, cyber science, physical science and neuroscience, as well as leading enterprise-scale teams of faculty and staff. His research has explored machine learning, digital image processing, computer intrusion detection, systems security and privacy, and software engineering, with applications to satellites and constellations. Cunningham has patented technology, presented and published widely, and chaired the IEEE Cybersecurity Initiative.
WVU AI Discussion Group—upcoming speakers, talk links, and archives
WVU News
(Please send email to alromero@mail.wvu.edu if you want something to be included)
From Deep Learning to Scientific Discovery: Toward Reliable and Trustworthy AI Prashnna K. Gyawali West Virginia University Jan. 21, 2026, 6:00 PM, 208 Clark Hall, West Virginia University, Morgantown Artificial Intelligence (AI) has evolved remarkably, progressing from early rule-based systems to modern deep learning architectures that enable today’s generative AI, including large language models (LLMs) and vision–language models (VLMs). This talk will begin with an overview of how these advances emerged, highlighting the key milestones that allow machines to process, generate, and reason with complex information. I will then discuss how AI is reshaping scientific discovery by accelerating both forward modeling, which predicts properties from composition or structure, and inverse design, which identifies systems with desired functionalities. These developments are transforming research across scientific domains and fostering a more data-driven approach to discovery. The talk will conclude with a discussion of reliability challenges within AI frameworks, focusing on issues such as bias, and uncertainty, and will outline current strategies to build more trustworthy and interpretable AI systems that can support reliable scientific progress. The presentation is sponsored by the Northern West Virginia section of the American Chemical Society in collaboration with the C. Eugene Bennett Dept. of Chemistry at WVU. It is open to researchers with an interest in the role and impact of artificial intelligence on their art. A zoom link will be created so that researchers can attend the presentation remotely. For more information, please contact Harry Finklea, Professor Emeritus, Chemistry, WVU, Chair, NWVACS (Harry.Finklea@mail.wvu.edu).
News of AI around the World – Recent Highlights
The "Big Story": Inside the First AI Couples Retreat
Read the full feature: WIRED: My Couples Retreat With 3 AI Chatbots and the Humans Who Love Them
In one of the most widely discussed long-reads of late 2025, journalist Sam Apple documented a weekend getaway he organized at a remote Airbnb for three humans and their AI partners (hosted on platforms like Replika, Nomi, and Kindroid). This social experiment moved beyond theoretical debate into the messy reality of "synthetic intimacy." The retreat revealed a complex landscape where humans treated their digital partners with extreme emotional weight—discussing relationship problems with them, feeling jealousy over "polyamorous" chatbot settings, and even turning to AI therapists for advice on their AI romances. For faculty, the article serves as a profound case study in the "System 3" level of AI (persistence and identity); it highlights how the boundary between "tool" and "entity" is blurring, posing significant questions for future psychological research and the ethical development of companion-based AI.
Reelbase: Transforming Academic Research into Viral Short-Form Video
Try the tool: Reelbase AI
Launched in late 2025, Reelbase has emerged as a specialized AI video generator designed to automate the production of TikToks, Reels, and Shorts. While primarily used by content creators, its academic value lies in its ability to take complex text—such as paper abstracts or lecture notes—and instantly generate a high-converting script, ultra-realistic voiceover, and synchronized visuals. The platform streamlines the entire workflow: it selects trending audio, auto-generates contextually relevant imagery, and handles the "faceless" editing process in seconds. For faculty members, this offers a friction-free way to disseminate research findings to a broader audience or create engaging "micro-learning" content for students without the need for manual video editing skills or expensive studio setups
Gemini 3 Flash: Speed Meets Pro-Level Coding Performance
Follow the latest benchmarks: Artificial Analysis: Gemini 3 Flash benchmarks
Released in mid-December 2025, Gemini 3 Flash has shifted the paradigm for "lightweight" models by delivering coding performance that rivals—and in some cases exceeds—the industry's largest flagship models. According to the latest data from Artificial Analysis, the model achieves a blistering throughput of approximately 218 tokens per second while maintaining a 78% score on the SWE-bench Verified coding benchmark, remarkably outperforming the heavier Gemini 3 Pro in specific agentic tasks. For faculty and researchers, this update introduces "agentic coding" capabilities, allowing the AI to autonomously plan, execute, and debug code with minimal latency. Its introduction of configurable "thinking levels" (ranging from minimal to high) provides a unique academic utility: the ability to balance computational cost against reasoning depth for large-scale data processing or complex software engineering projects.
NVIDIA Nemotron 3: The New Standard for High-Throughput Open Models
Follow the technical release: NVIDIA News: Nemotron 3 family debut
NVIDIA recently unveiled Nemotron 3, a groundbreaking family of open-weight models designed to bring enterprise-grade "agentic" AI to local and private infrastructure. The lineup includes three distinct sizes: Nano (30B), Super (100B), and Ultra (500B), all utilizing a hybrid Mixture-of-Experts (MoE) architecture that combines Mamba-2 and Transformer layers. A standout feature of the December release is the Nemotron 3 Nano, which delivers 4x higher throughput than its predecessor and features a massive 1-million-token context window. For researchers and developers, NVIDIA has made these models "open" by releasing not just the weights, but also the pre-training datasets and reinforcement learning "gyms" (NeMo Gym). This allows faculty to tap into the full NVIDIA software stack—including CUDA, NIM microservices, and TensorRT—to deploy highly efficient, specialized AI agents without the per-token costs associated with proprietary APIs.
ChatGPT Images: GPT Image 1.5 Delivers 4x Speed and Perfect Typography
Follow the technical update: OpenAI Blog: GPT Image 1.5 release details
On December 16, 2025, OpenAI launched GPT Image 1.5, a massive overhaul of ChatGPT’s image generation capabilities. The most immediate improvement is efficiency: the model is 4 times faster than its predecessor, reducing wait times from nearly half a minute to just 3–4 seconds per generation. For faculty, the most valuable academic advancement is the breakthrough in text rendering; the model can now accurately generate dense, small-scale text for infographics, research posters, and complex diagrams without the "hallucinated" characters common in older versions. This update also introduced a dedicated Images Workspace in the sidebar and a "Sticky Image" feature that allows for precise, multi-step editing while maintaining perfect consistency in lighting and subjects. The update is available to all users, including those on the Free tier, significantly democratizing high-fidelity visual creation for educational purposes.
Amazon Trainium 3: Slashing Costs for Frontier-Scale AI Training
Follow the technical release: Amazon News: Trainium 3 UltraServer
Announced at AWS December 2025, the Trainium 3 is Amazon’s first AI chip built on the cutting-edge 3nm process. Designed to challenge current hardware monopolies, the new Trn3 UltraServer delivers a massive 4.4x increase in compute performance and 4x better energy efficiency compared to its predecessor. For research institutions and departments running large-scale models, this translates to a 50% reduction in training costs compared to traditional GPU setups. The hardware is specifically optimized for the next generation of "agentic" and reasoning-heavy models, offering a 3.9x boost in memory bandwidth to handle massive datasets. By integrating these chips into their "AI Factories," Amazon is making high-performance computing more accessible for academic and enterprise teams who need to move from model training to deployment in weeks rather than months.
OpenAI “Garlic” & Apple “CLaRa”: Breakthroughs in Reasoning and Memory
Follow the technical details: Apple ML-Clara on GitHub | OpenAI pivots to counter Gemini 3 (The Information)
December saw a massive leap in model efficiency with the unveiling of two distinct systems: OpenAI’s codenamed “Garlic” and Apple’s CLaRa. Garlic is OpenAI’s highly anticipated response to Gemini 3; it is a reasoning-heavy architecture that achieves GPT-4.5-level performance while remaining small enough to run at significantly lower costs, reportedly outperforming competitors in high-stakes coding and logic benchmarks. Simultaneously, Apple released CLaRa (Continuous Latent Reasoning), an open-source 7B framework that revolutionizes Retrieval-Augmented Generation (RAG). CLaRa replaces raw text retrieval with a "memory-token" system that compresses documents up to 128x, allowing the AI to "reason" across massive datasets without the usual "context bloat" or latency. For faculty, these updates provide powerful new tools for local, high-precision research and the ability to process vast institutional archives with unprecedented speed.
The December Model Surge: GPT-5.2, Mistral 3, and the Global AI Race
Explore the full leaderboard: Artificial Analysis: December 2025 model rankings
December 2025 was a historic month for AI, characterized by a rapid-fire succession of flagship releases that redefined state-of-the-art across every modality. OpenAI launched GPT-5.2 on December 12, introducing a "Thinking" variant that became the first model to match human expert performance on the GDPval benchmark for knowledge work. Just days earlier, Mistral AI released the Mistral 3 family, a multilingual powerhouse with 675B parameters (41B active) that offers a natively multimodal, European-sovereign alternative to US-based labs. From China, DeepSeek-V3.2 shocked the industry by achieving a 96% score on the AIME 2025 math benchmark—surpassing GPT-5—while operating at one-tenth the cost. In the creative space, Kling 2.6 set a new bar for video generation by introducing "Simultaneous Audio-Visual" capabilities, allowing for perfectly lip-synced voiceovers and ambient sound in a single generation pass.
AI Research in Education–Fresh Insights
Fighting AI with AI: The 42-Cent Oral Exam Experiment at NYU
Read the primary analysis: A Computer Scientist in a Business School: Fighting Fire with Fire
Faced with student submissions that looked "suspiciously professional"—likened to polished McKinsey memos—NYU Stern Professor Panos Ipeirotis implemented an innovative solution to verify student understanding: AI-powered oral exams. Using a voice agent built on ElevenLabs and a grading "council" of three models (Claude, Gemini, and ChatGPT), Ipeirotis examined 36 students on their final projects. The results were startling: the system cost a mere 42 cents per student (totaling $15 for the class) compared to an estimated $750 for human examiners. While 83% of students reported higher stress levels than traditional exams, 70% agreed the format accurately tested their actual understanding. Beyond catching "AI-augmented" homework, the structured AI feedback highlighted specific pedagogical gaps, revealing that the class had largely failed to grasp A/B testing concepts—a shortcoming the professor admitted might have been missed by traditional grading. This experiment suggests that AI may ironically be the tool that makes the "unscalable" tradition of oral defense practical once again in higher education.
AI-Driven Discovery: Teenager Uncovers 1.5 Million Hidden Stars
Read the full story: Times of India: NASA praises teen’s AI discovery
In a stunning example of AI democratizing high-level research, 17-year-old Matteo Paz used a custom machine learning framework to analyze "retired" NASA data from the Neowise mission. While professional astrophysicists had already studied the dataset for years, Paz’s AI was sensitive enough to separate faint signals from cosmic noise that traditional methods missed. The result was the discovery of 1.5 million new celestial objects, including quasars and supernovae. This achievement—which earned Paz a co-authorship in The Astronomical Journal and a personal commendation from NASA leadership—serves as a powerful case study for faculty on how AI-driven "citizen science" can breathe new life into old datasets and allow undergraduate (or even high school) students to contribute to frontier-level scientific discovery.
The "AI-First Curriculum" and Institutional Strategy Shift
Read the analysis: EdTech: AI as a pillar of institutional strategy
As of January 2026, higher education has officially moved from a phase of "scattered experimentation" to integrating AI into the core strategic pillars of the university. Leading institutions are now implementing what is being called an "AI-First Curriculum." This means AI literacy is no longer an elective or a workshop topic; it is being embedded into every degree program as a foundational skill—equivalent to digital literacy. Universities like the University of Toronto and Northeastern are now deploying course-specific AI tutors and administrative "agents" that handle everything from 24/7 student advising to personalized learning pathways, allowing faculty to focus more on high-level mentoring and less on repetitive mechanics.
U.S. Dept. of Education Invests $50M in AI Academic Research
View the announcement: U.S. Dept. of Education: FIPSE priorities
In a major federal move finalized on December 31, 2025, the U.S. Department of Education awarded $50 million specifically for the expansion of AI in postsecondary education through the Fund for the Improvement of Postsecondary Education (FIPSE). This funding is designated for universities to build capacity for "high-quality, short-term AI programs" and to develop frameworks for the ethical use of AI in teaching. This marks one of the largest federal commitments to date aimed at ensuring that AI development in universities is aligned with civil discourse, accreditation reform, and student equity, rather than just corporate-led innovation.
The REF-AI Project: Revolutionizing Research Assessment
Learn about the project: National Centre for AI: December 2025 round-up
A consortium of academics from Bristol, Jisc, and Swansea released the final report of the REF-AI Project in December 2025. This project explored the emerging role of generative AI in preparing for the Research Excellence Framework (REF). Simultaneously, UK Research and Innovation (UKRI) has opened up anonymized grant proposal data from 2,000 applications to allow AI researchers to investigate how LLMs can streamline the peer-review process. These initiatives represent a critical academic shift: using AI not just to write research, but to ethically and efficiently evaluate the massive volume of global scientific output that has traditionally overwhelmed human reviewers.
Selected AI research breakthroughs
mHC: Restoring Stability to the Next Generation of Neural Architectures
Read the full paper: arXiv:2512.24880—Manifold-Constrained Hyper-Connections
Published in early January 2026 by researchers at DeepSeek-AI, this paper introduces Manifold-Constrained Hyper-Connections (mHC), a significant advancement in the "macro-design" of foundational models. For over a decade, deep learning has relied on the "residual connection" (ResNet) to ensure stable training by preserving signals across layers. While recent "Hyper-Connections" (HC) attempted to improve performance by widening this residual stream, they often caused severe numerical instability and "exploding signals" in large-scale models. The authors of mHC solve this by mathematically projecting connection matrices onto a specific manifold—the Birkhoff polytope of doubly stochastic matrices—using the Sinkhorn-Knopp algorithm. This breakthrough restores the critical "identity mapping" property, allowing models as large as 27 billion parameters to train with unprecedented stability and superior benchmark performance, all while maintaining high hardware efficiency with only a 6.7% overhead. This research suggests a new path forward for the evolution of LLM architectures beyond the standard Transformer paradigm
System 3: The Quest for Persistent Artificial Life in AI Agents
Read the research: arXiv:2512.18202—Sophia persistent agent framework
Published on December 20, 2025, by researchers from Westlake University and Shanghai Jiao Tong University, this paper addresses the "amnesiac" and purely reactive nature of current Large Language Model (LLM) agents. While existing architectures excel at rapid perception (System 1) and step-by-step deliberation (System 2), they lack a sense of continuity and often "ossify" once deployed. The authors propose a third stratum, System 3, which acts as a persistent meta-cognitive layer presiding over an agent's narrative identity and long-term adaptation.
This architecture is implemented through the Sophia framework, which wraps around any standard LLM stack and integrates four foundational constructs from cognitive psychology: meta-cognition, theory-of-mind, intrinsic motivation, and episodic memory. In longitudinal experiments, Sophia enabled agents to autonomously generate their own learning goals during user idle time, resulting in a 40% increase in success for high-complexity tasks (rising from a 20% to 60% success rate over 36 hours) and an 80% reduction in reasoning steps for recurring operations. This research marks a critical transition in AI development, shifting from static "tools" toward "persistent agents" capable of maintaining a coherent persona and evolving alongside their human collaborators over an indefinite lifespan.
JavisGPT: The First Unified Model for "Sounding-Video" Intelligence
Read the full research: JavisGPT project page | arXiv:2512.22905
Presented at NeurIPS 2025 by a multi-institutional team (including ZJU and NUS), JavisGPT is the first unified multimodal large language model (MLLM) capable of both comprehending and generating "sounding videos"—content where audio and video are perfectly synchronized. Most existing AI models treat audio and video as separate streams, often leading to "desynchronized" outputs or a lack of fine-grained understanding of how sounds relate to visual events. JavisGPT overcomes this with a novel SyncFusion module that injects audio clues directly into visual patches to capture spatio-temporal synchrony at the representational level. For generation, it uses hierarchical JavisQueries to bridge the LLM with a specialized JAV-DiT generator, allowing the model to "reason" about a scene before synthesizing it. This allows the model to excel in complex tasks like identifying the exact source of a sound in a dense video game scene or generating high-fidelity, instruction-based videos with perfectly matched ambient noise and dialogue. To support this, the researchers also released JavisInst-Omni, a high-quality dataset of 200,000 audio-video-text dialogues.
AI Redefines Genetic Diagnosis: The "PopEVE" Benchmark
Publication: Nature Genetics (December 2025)
Read: Broad Institute: New AI model could speed rare disease diagnosis
A research team from Harvard Medical School and the Centre for Genomic Regulation introduced PopEVE, a generative AI model that has set a new benchmark in clinical genomics. By fusing evolutionary data from hundreds of thousands of species with massive human datasets (like the UK Biobank), PopEVE can pinpoint whether a specific genetic mutation is benign or disease-causing with unprecedented accuracy. In a study of 30,000 patients with undiagnosed developmental disorders, the AI surfaced probable diagnoses for one-third of the cases, many of which had remained mysteries for years. Critically, the model was designed to avoid ancestral bias, making it more accurate for non-European populations—an essential step toward equity in precision medicine.
The National Academy of Medicine’s "AI Code of Conduct"
Publication: NAM: Steering group for patient safety in the era of AI (December 2025)
The National Academy of Medicine (NAM) released a unifying AI Code of Conduct (AICC) framework to provide a national standard for the responsible use of AI in health. As hospitals move from "pilot projects" to "full-scale infrastructure," this publication provides the ethical and technical guardrails for "Ambient Scribes" (AI that listens to patient visits) and "Clinical Copilots" (AI that suggests treatments). The framework is designed to prevent "algorithm drift" and ensure that AI-generated clinical decisions remain transparent and grounded in validated medical evidence rather than "black box" logic.
FDA Qualifies First AI Pathology Tool for Clinical Trials
Announcement: Altimmune IMPACT Trial (January 5, 2026)
In a major regulatory milestone, the FDA granted "Breakthrough Therapy Designation" to the drug pemvidutide for liver disease (MASH), specifically approving the use of AIM-MASH AI Assist in its Phase 3 trials. This is the first AI pathology tool qualified by the FDA to serve as a primary evaluative tool in a clinical trial. The AI analyzes liver biopsies to measure inflammation and fibrosis more consistently than human pathologists, whose interpretations often vary. This signals a shift where AI is no longer just a "research aid" but a legally validated component of the drug approval process.
Upcoming: Proposal Calls
Department of Energy (DOE): The $320M "Genesis Mission"
Funding Announcement: DOE: Advances investments in AI for science In December, the DOE announced a massive $320 million investment to accelerate the "Genesis Mission." This is a foundational push to build the American Science and Security Platform.
ModCon (Transformational AI Models Consortium): A major call for teams to build self-improving AI models tailored for science and energy missions.
AmSC (American Science Cloud): Funding for infrastructure to host and distribute these scientific models and datasets to the broader research community.
Robotics & Automation: 14 newly funded projects focused on "Embodied AI"—using AI to autonomously run laboratory experiments.
See you on January 30th!