Amazon is unveiling its multimodal “Olympus” AI model featuring advanced video analysis capabilities, with development versions ranging from 400 billion to 2 trillion parameters. In other news, Mount Sinai Health System opened a 65,000-square-foot AI research center housing 40 Principal Investigators and 250 technical staff, while Recursion Pharmaceuticals completed its merger with Exscientia, reporting a combined pipeline of 10 potential drug candidates and capital runway through 2027. In technical developments, IDEA Research’s DINO-X achieved 56.0 AP on COCO benchmarks with a 5.8 AP improvement on rare class detection, while Anthropic open-sourced the Model Context Protocol for standardizing AI system integration. To learn more about what’s going on in the AI research landscape, read on.

Amazon develops multimodal “Olympus” AI model with advanced video analysis capabilities

Source: The Information

Amazon has developed a new multimodal AI system, code-named Olympus, that extends beyond traditional image-text processing to include sophisticated video analysis capabilities. The model represents a strategic step forward in Amazon’s AI portfolio, though technical analysis indicates its general language capabilities may not match current industry leaders like Claude or GPT-4. Key technical specifications include video scene comprehension exceeding traditional recognition systems and granular motion analysis capabilities (e.g., precise trajectory tracking). It is built on larger parameter models, with development versions reportedly ranging from 400 billion to 2 trillion parameters. It also supports integration of smaller vision-based models with larger text processing architectures. The system’s development aligns with specific industrial applications, including sports analytics requiring frame-precise motion tracking, underwater equipment inspection for oil and gas operations, and media archive search and content discovery optimization. This development marks Amazon’s strategic foray into the video AI market, currently dominated by specialized providers like Twelve Labs and Google’s Gemini. The system is expected to make a formal debut at AWS re:Invent.

Study demonstrates 50% parameter reduction in Llama with minimal performance loss

Source: LinkedIn / Philipp Schmid

A new study from Neural Magic demonstrates successful pruning of Llama 3.1 8B, maintaining 98.4% accuracy while removing half of its parameters. The optimized model achieves 30% higher throughput and 1.8x lower latency, with improvements up to 5.0x when combined with quantization. The pruning process, completed in 26 hours using 32 H100 GPUs and 13 billion training tokens, preserves performance across fine-tuning tasks including GSM8K and Ultrachat-200K. The research, as Hugging Face’s Technical Lead Philipp Schmid shared, represents a significant advance in LLM optimization techniques, especially for NVIDIA Ampere and newer GPUs.

Anthropic releases Model Context Protocol (MCP) for standardized AI system integration

Source: Fast Company

Anthropic has open-sourced the Model Context Protocol (MCP), introducing a standardized interface protocol for AI system integration that addresses architectural challenges in connecting AI models with external data sources and tools. The protocol’s technical implementation enables bidirectional communication between AI assistants and various data stores, including knowledge bases, business intelligence graphs, and development environments.

Novel machine learning pipeline accelerates computational chemistry modeling in complex solvents

Source: LinkedIn/Journal of Chemical Theory and Computation

Research scientist in AI for chemistry Frédéric Célerse, Ph.D. announced on LinkedIn the launch of a novel computational pipeline that combines active and transfer learning methodologies to enhance neural network potentials (NNPs) for chemical solute-solvent reactions. A paper on the research was published in the Journal of Chemical Theory and Computation. Joining Célerse in authoring the document were Veronika Juraskova, Shubhajit Das, Matthew Wodrich, and lead author Clémence Corminboeuf, The technical implementation integrates several unique approaches: well-tempered metadynamics with neural architectures, a dual-learning system combining active and transfer methodologies, and ab initio level dynamic simulations. This architecture enables precise modeling of solute-solvent interactions, particularly demonstrating effectiveness in analyzing methanol and trifluoroethanol (TFE) systems.

Insilico Medicine achieves 10th AI-driven IND clearance for cancer drug

Source: LinkedIn

Insilico Medicine has received FDA IND clearance for ISM5939, marking its 10th AI-enabled drug candidate. The compound, developed using their PharmaAI platform, is an oral small molecule ENPP1 inhibitor targeting solid tumors. The development process only took a few months from initiation to lead optimization. “Insilico’s AI platform, PharmaAI developed this novel compound in just three months—yes, three months—from project initiation to lead optimization,” wrote longevity and biotech VC Garri Zmudze on LinkedIn. “That’s not just speed; it’s a new level of efficiency in drug discovery in my opinion. A hard one to match.”

Since 2021, Insilico has nominated 20 preclinical candidates and secured 10 IND clearances using its proprietary platforms, including Chemistry42. Its flagship molecule ISM001_055 recently showed positive Phase IIa clinical trial results.

AI startup talent migration accelerates as technical founders shift to major platforms

Source: The Information

More AI technical leadership continues as to gravitate from startups to major technology platforms. Two notable moves represent the trend: Fixie.ai CTO Justin Uberti is joining OpenAI’s real-time AI division and Reka AI’s chief scientist Yi Tay is moving to Google DeepMind. Uberti’s move, following OpenAI’s summer acquisition overtures to Fixie (which had raised $17 million), particularly highlights the industry’s focus on real-time AI capabilities, building on his prior work developing WebRTC at Google. This pattern of technical leadership migration emerges against a backdrop of increasing AI infrastructure costs and cooling startup valuations.

Recursion and Exscientia wrap up AI-driven drug discovery merger

Source: LinkedIn, STAT

Recursion Pharmaceuticals has completed its strategic merger with Exscientia, while simultaneously implementing workforce reductions following the deal closure. The consolidation unites two leading clinical-stage TechBio companies, integrating their AI-driven drug discovery platforms and development pipelines. The combined company, which currently has 10 potential drugs in development, reports sufficient capital runway through 2027. This strategic realignment comes as part of post-merger integration efforts, though the firm did not disclose specific figures regarding the workforce reduction. In a LinkedIn post, Recursion CEO Chris Gibson noted on LinkedIn that there was more to the story than a simple headcount reduction. Referring to a story mentioning the job cuts, Gibson noted, “A more accurate headline might be: “Recursion grows from ~550 to ~800 employees with close of Exscientia merger.” While acknowledging the cuts, Gibson stressed that the “merger is about more than integrating two teams.” He continued: “it’s about creating something larger than the sum of its parts to bring life-changing medicines to patients faster.”

The company’s therapeutic pipeline spans oncology and rare disease with an other category including candidates for c. diff and pulmonary fibrosis. “We have more than 10 active programs with $450 million in upfront and milestone payments already received and which could yield over $20 billion in additional milestones before royalties,” noted the company’s CSO David Hallett over LinkedIn.

Bluesky faces first major data scraping incident with AI training dataset collection

Source: Mashable via Yahoo Tech

The decentralized microblogging social network Bluesky has had roughly one million of its public posts systematically aggregated for AI training purposes and uploaded to Hugging Face’s platform. The dataset, created by machine learning librarian Daniel van Strien, contained comprehensive user data including decentralized identifiers (DIDs) and post metadata, using  Bluesky’s firehose API architecture. The incident highlights the technical tensions between decentralized social protocols and AI training data collection: while Bluesky’s architecture explicitly enables public data access through its firehose API — providing “aggregated, chronological stream of all public data updates” — the platform lacks granular user controls for AI training consent. Following disclosure, the dataset was removed from Hugging Face.

Mount Sinai launches AI research center

Source: Mount Sinai Health System

Mount Sinai Health System has inaugurated the Hamilton and Amabel James Center for Artificial Intelligence and Human Health, marking a strategic expansion of its AI research capabilities within the academic medical sector. The 65,000-square-foot facility, housed in a repurposed 12-story building at 3 East 101st Street, represents a significant infrastructure investment in healthcare AI development. Technical specifications include dedicated research space across eight floors, housing approximately 40 Principal Investigators and 250 supporting staff members, including graduate students, postdoctoral fellows, and computer scientists. The center’s computational infrastructure builds upon Mount Sinai’s established high-performance computing foundation, which includes the Minerva supercomputer platform operational since 2013.

Analytics engineering role combines software and data analysis skills

Source: LinkedIn

Analytics engineering is emerging as a high-demand hybrid role in data science, explains data professional Redha Cherif. These specialists build data pipelines, bridging software engineering and data analysis. The position requires a 50-50 split between technical capabilities (Git, CI/CD, Airflow) and data modeling expertise (Kimball, Inmon methodologies). Core tools include DBT for data processing, with practitioners needing both domain knowledge and software development best practices. The role represents a strategic middle ground between traditional data engineering and analysis positions.

Mistral AI expands to US while OpenAI sets up shop in Paris

Source: LinkedIn

According to a French-language LinkedIn post by Riskintel Media CEO Yasmine Douadi (pictured right), French AI startup Mistral AI is establishing operations in Palo Alto, California, while OpenAI expands to Paris. The move aims to recruit engineers and scientists while growing U.S. commercial operations, with co-founder Guillaume Lample potentially relocating from Paris. Mistral AI, valued at €6 billion after a €600 million funding round, joins a concentrated hub of tech giants in Silicon Valley including Google, Meta, and Anthropic.

AstraZeneca forges strategic alliance with Lunit Oncology

Source: LinkedIn

Hanna Foster, Scientific Product Sales Manager at Quadratech Diagnostics Ltd, shares on LinkedIn that AstraZeneca entered a strategic collaboration with Lunit Oncology.  This partnership represents a strategic move to integrate advanced AI diagnostic and therapeutic development tools into AstraZeneca’s oncology pipeline. The collaboration aims to tap Lunit’s AI expertise in medical image analysis and biomarker detection, potentially enhancing clinical decision-making and treatment optimization in oncology.

DINO-X unifies open-world object detection capabilities

Source: Papers with Code / Arxiv

IDEA Research launched DINO-X, a unified object-centric vision model achieving state-of-the-art performance in open-world object detection. Built on the transformer-based encoder-decoder architecture of Grounding DINO 1.5, DINO-X enhances object-level representation through flexible prompt inputs, including text, visual, and custom prompts, enabling prompt-free detection of any object in an image. It uses a large-scale dataset, Grounding-100M, containing more than100 million high-quality grounding samples to improve open-vocabulary detection and support various perception tasks such as detection, segmentation, pose estimation, and object-based QA. Experimental results highlight its superior performance, with the DINO-X Pro model achieving 56.0 AP on COCO and significant on LVIS zero-shot benchmarks, including a 5.8 AP boost on rare class detection.

“OminiControl” framework achieves minimal parameter overhead for diffusion models

Source: Papers with Code

Researchers introduced OminiControl, demonstrating significant efficiency gains in Diffusion Transformer model adaptation. The framework achieves image condition integration using only 0.1% additional parameters, marking a substantial reduction compared to conventional approaches requiring separate encoder modules. The system leverages parameter reuse mechanisms within existing DiT architectures, enabling multi-modal attention processing without architectural complexity. The research includes the release of Subjects200K, comprising over 200,000 identity-consistent images for advancing subject-consistent generation research.

OneDiffusion introduces unified approach to generative and predictive tasks

Source: Papers with Code/Arxiv

A new large-scale diffusion model demonstrates unified capability across bidirectional image synthesis and understanding tasks. OneDiffusion handles multiple input conditions including text, depth, pose, and layout while supporting reverse processes such as depth estimation and segmentation. The system implements a frame sequence approach with variable noise scaling, enabling conditional generation and analytical tasks within a single architecture. Notable technical achievements include resolution-independent processing and competitive performance across multiple domains despite utilizing a relatively constrained training dataset.

Marco-o1 advances open-ended reasoning capabilities in AI systems

Source: Papers with Code/Arxiv

Marco-o1 expands the capabilities of large reasoning models (LRMs) beyond conventional bounded-solution domains. The system integrates Chain-of-Thought fine-tuning with Monte Carlo Tree Search, specifically targeting scenarios where clear evaluation metrics are absent. This methodological approach represents a shift from traditional reinforcement learning in standardized domains toward generalized reasoning in open-ended problem spaces. The framework emphasizes adaptability to complex real-world problem-solving tasks where rewards are challenging to quantify.

South Korea launches strategic AI Safety Institute with cross-sector consortium

Source: Korea Herald, Newswise

South Korea has officially inaugurated its AI Safety Institute (AISI), establishing a dedicated research hub for artificial intelligence safety and governance. Located at the Pangyo Global R&D Center, the institute marks a strategic implementation of commitments made during the May 2024 AI Seoul Summit, where ten nations agreed to strengthen international AI safety collaboration. Based in Pangyo near Seoul, the institute will evaluate the risks related to AI, including the misuse and loss of control. It will also serve as an AI safety research hub.

Enveda Therapeutics raises $130M Series C for AI-driven drug discovery

Source: Bloomberg/LinkedIn

Enveda Therapeutics has secured $130 million in Series C funding led by Kinnevik and FPV Ventures, with participation from Baillie Gifford, Lux Capital and the Nature Conservancy. The Boulder-based company, which combines AI and natural product chemistry to identify therapeutic compounds, is launching a Phase I clinical trial for an oral eczema treatment. The funding continues the trend of significant investment in AI-driven pharmaceutical development platforms despite broader market challenges in the sector.

Report examines current state and limitations of AI in drug discovery

Source: DeepMirror

A new analysis from DeepMirror explores AI’s current role in drug discovery, finding that despite significant advances, the technology faces key limitations in clinical translation. The report highlights two major constraints: data relevance (how well training data applies to specific problems) and dataset size limitations compared to other AI applications. While modern AI can process larger datasets than traditional QSAR models, DeepMirror’s research shows that newer architectures like graph neural networks often don’t significantly outperform classical methods in real-world drug discovery applications. The analysis suggests AI’s current utility lies mainly in iterating on structurally related molecules rather than generating novel clinical candidates.

Isomorphic Labs CEO outlines AI’s impact on accelerating research in TED talk

Source: LinkedIn / TED Talk

In a recent TED talk, Isomorphic Labs’ Chief AI Officer Max Jaderberg discussed how AI is revolutionizing scientific research through “AI analogs” of biological systems. Highlighting AlphaFold 3’s capabilities, Jaderberg explained how the technology can simulate molecular interactions and protein folding at unprecedented speeds. The system enables thousands of AI agents to work simultaneously on therapeutic design under human guidance, potentially reversing Eroom’s law (the observation that drug discovery becomes slower and more expensive over time in contrast to Moore’s law, which predicted semiconductor progress for decades). Jaderberg emphasized applications in cancer treatment, where AI could help design proteins to selectively target cancer cells while preserving healthy tissue.

SmolVLM debuts as compact, high-performance vision language model

Source: LinkedIn / HuggingFace Blog

Hugging Face has released SmolVLM, a 2B-parameter vision language model optimized for on-device inference. The model demonstrates superior performance in its size class, generating tokens 7.5-16x faster than Qwen2-VL while requiring only 5 GB of GPU RAM. Features include the ability to run at 17 tokens/second on MacBooks, support for fine-tuning on Google Colab, and unexpected proficiency in video benchmarks despite no video training. The model achieves competitive scores on key benchmarks like MMMU (38.8), MathVista (44.6), and DocVQA (81.6), making it particularly suitable for edge devices and resource-constrained environments.

Commonwealth Bank of Australia’s AI implementation improves customer service

Source: Bloomberg

Commonwealth Bank of Australia (CBA) is investing in AI to enhance operations and customer service. To date, its deployment of AI has cut call center wait times by 40% and halved scam losses. The bank is developing 50 additional AI use cases, aiming to improve automation and productivity. The Bank expects the initiatives to increase profitability and reduce operational costs over time.

The bulk of AI-generated long-form content is now AI-generated

Source: WIRED

A new analysis from Originality AI revealing that 54% of English-language posts exceeding 100 words are now AI-generated. The research, examining 8,795 public posts from 2018 to 2024, identified a 189% surge in AI-authored content following ChatGPT’s release in early 2023. This transformation coincides with LinkedIn’s strategic integration of AI writing tools for Premium subscribers, enabling automated content creation for posts, profiles, and messages. While the platform maintains robust defenses against duplicate and low-quality content, the distinction between AI-generated and human-authored professional communications has become increasingly blurred.

LinkedIn has “robust defenses in place to proactively identify low-quality, and exact or near-exact duplicate content,” Adam Walkiewicz, LinkedIn’s head of “feed relevance,” told Wired. “We see AI as a tool that can help with review of a draft or to beat the blank page problem, but the original thoughts and ideas that our members share are what matter.”

Elon Musk’s xAI scales infrastructure in race against OpenAI

Source: Wall Street Journal

Elon Musk’s artificial intelligence venture xAI is rapidly scaling its infrastructure to compete with OpenAI, achieving a $50 billion valuation despite its late market entry. The company has constructed a massive data center in Memphis housing 100,000 NVIDIA GPUs in just 122 days, while securing $11 billion in funding. Though xAI’s current revenue of $100 million annually trails significantly behind OpenAI’s projected $4 billion, the company exploits exclusive data from X and Tesla to differentiate its technology. With plans to double its GPU capacity and launch a standalone consumer app, xAI’s aggressive infrastructure buildout demonstrates Musk’s determination to establish dominance in the AI sector. Musk was a co-founder of OpenAI, but has criticized its current trajectory.

Research assistance: Frédéric Célerse, Ph.D., Research Scientist in AI for Chemistry, Ecole Polytechnique Fédérale de Lausanne.