This week in AI research: Enterprise firms warming up to autonomous agents, Sora debuts

This week, Meta’s unveiled of a “concept-predicting” language model and AI-powered Metaverse avatars. Meanwhile, industry leaders are forecasting autonomous AI agents by 2025. Deloitte has projected that by 2025, 25% of enterprises using generative AI will deploy such agents, expanding to 50% by 2027. Also this week, OpenAI finally opened up its video creation tool “Sora” after teasing the platform in spring. In other news, we’re also seeing a move away from simply scaling up models to focus on algorithmic efficiency, a humanoid robot inspiring STEM interest in Zimbabwe, and AI models tackling complex problems with “long thinking.”

Meta unveils “Meta Motivo” AI model to enhance Metaverse avatars and large context models

Source: Internal announcements, Meta

Meta has launched “Meta Motivo,” a behavioral foundation model designed to control virtual embodied humanoid agents’ movements through unsupervised reinforcement learning, potentially enabling more lifelike NPCs and character animations in the Metaverse. The company also released its Large Concept Model (LCM), which takes a novel approach to language modeling by decoupling reasoning from language representation and predicting concepts rather than tokens. Additionally, Meta launched Video Seal, an open-source framework for embedding imperceptible watermarks into videos, building upon their previous Audio Seal technology and demonstrating resilience against common editing modifications.

Executives predict emergence of autonomous AI agents in 2025 and profitability gains in AI sector

Source: Reuters NEXT conference

At the Reuters NEXT event, industry leaders projected that autonomous agents capable of independent action—like making purchases and scheduling meetings—will become viable by 2025 due to advancements in step-by-step reasoning. Sarah Friar, CFO of OpenAI, anticipates achieving artificial general intelligence (AGI) within a few years. Venture capital sentiment is shifting from prioritizing revenue growth to focusing on profitability, signaling a maturation of the AI market. Concrete example from portfolio companies shows 75% labor cost reduction through AI sales team deployment.

AWS Re:Invent 2024 unveils Nova foundation models and healthcare AI initiatives

Source: Re:Invent analysis by Himanshu Jain

Amgen’s Technology Strategy and Innovation Leader Himanshu Jain has a rundown on Amazon Web Services’ platform evolution at Re:Invent 2024, where the cloud services pioneer unveiled its Nova foundation model family with four distinct text processing variants, alongside Canvas and Reel modules for image and video generation respectively. Healthcare sector implementations are maturing with Merck, Amgen, Geisinger, Gilead, and Natera pioneers in adopting genAI in their respective businesses.

Broadcom ascends to trillion-dollar valuation club thanks to AI chip alliances

Source: Reuters

Broadcom has crossed the $1 trillion valuation threshold thanks to its expanding artificial intelligence semiconductor strategy. The company’s strategic positioning includes development partnerships with major cloud and technology providers, notably OpenAI’s inference chip development program targeting 2026 deployment through TSMC’s manufacturing capabilities. CEO Hock Tan projects the company’s AI chip division could generate revenues approaching $90 billion by 2027.

Wayve’s end-to-end autonomous driving system learns directly from video

Source: Wayve
London-based startup Wayve introduced an AI-driven autonomous driving system that forgoes traditional high-definition mapping in favor of learning from raw, unlabeled driving footage. This approach allows the model to mimic human driving behavior, with near-term plans to offer Level 3 driver assistance and eventual progression toward Levels 4 and 5 full autonomy. Tesla’s autonomous mode, including Full Self-Driving (FSD), also relies heavily on video data captured by its onboard cameras.

Biotech sector navigates transformative 2024 amid AI integration and funding shifts

Source:  Industry analysis from Andrii Buvailo, Ph.D. via LinkedIn
The biotech sector experienced significant structural shifts in 2024 with both strategic consolidation and technological integration. Notable developments include Xaira Therapeutics’ $1 billion funding haul, contrasting with an overall dip in M&A activity. Clinical advances centered on gene-editing technologies, building upon Casgevy’s 2023 approval, while FDA approvals expanded across multiple therapeutic areas including oncology (Tecelra, Bizengri, KRAZATI), neurology (Vyalev), and rare diseases (Yorvipath). The integration of AI technologies accelerated through partnerships with NVIDIA and Google, while “Techbio” ventures demonstrated early clinical validation. Yet layoffs were another common theme with workforce reductions affecting companies like Ginkgo Bioworks and 23andMe. Meanwhile geopolitical tensions especially U.S.-China relations, prompted regulatory responses through the BIOSECURE Act. .

NVIDIA CEO highlights “long thinking” AI models for improved reasoning

Source: Jensen Huang at industry conference presentations, WSJ
NVIDIA CEO Jensen Huang’s emphasis on “long thinking” AI models marks a pivotal shift in computational methodology, with systems designed for extended reasoning cycles. Analysis of NVIDIA’s quarterly earnings reveals this evolution could exponentially scale inference computing demands, particularly for complex scientific and mathematical challenges. These systems, capable of 100-day computational processes, represent a departure from rapid-response frameworks, with OpenAI’s o-series models already demonstrating value in Harvard Medical School’s genetic research. Mirroring Kahneman’s System 2 cognitive framework, this deliberative AI processing approach signals fundamental changes in infrastructure requirements while potentially reducing error rates. This development aligns with NVIDIA’s record $35.1 billion quarterly revenue, reflecting strong market validation of advanced AI architectural demands.

Exxon enters AI data center power fray with natural gas CCS facility

Source: TechCrunch

TechCrunch reported taht Exxon Mobil announced plans to construct a 1.5+ gigawatt natural gas power plant specifically targeting AI data center operations, marking its first external power generation project. The facility will incorporate carbon capture and storage (CCS) technology targeting 90% emissions reduction. Operating independently from grid infrastructure, the plant addresses critical power supply constraints projected to affect nearly 50% of new AI data centers by 2027.

Alphabet’s stock surges and quantum computing chip announcements

Source: Alphabet investor updates, Google
Alphabet’s stock hit record highs following the unveiling of its Gemini 2.0 model (see below for more context), which offers superior performance over previous versions. Upcoming projects like Project Mariner and the Jules coding agent, plus the Willow quantum computing chip capable of outperforming today’s supercomputers, signal Alphabet’s robust roadmap. Market excitement reflects confidence in Alphabet’s leadership in advanced AI and quantum technologies.

Google DeepMind debuts Gemini 2.0 and OpenAI introduces Sora

Source: Philipp Schmid (LinkedIn), Google, Sora official site
Google DeepMind released Gemini 2.0, operating twice as fast as Gemini 1.5 Pro with high-accuracy scores. This marks a significant leap forward in the search giant’s pursuit of agentic AI, moving beyond simply organizing information to creating AI that can actively assist users. Available now for testing is Gemini 2.0 Flash, an experimental model designed for low latency and enhanced performance. It not only surpasses its predecessor in speed but also introduces several new capabilities including the ability to process multimodal inputs like images, video, and audio, and now also generate multimodal outputs such as natively created images and steerable text-to-speech audio. In addition, Gemini 2.0 Flash natively supports tool usage and integrate with Google Search, code execution, and user-defined functions.

Meanwhile, OpenAI launched Sora, a model that supports highly accurate text-to-video generation. The video below provides a sample of Sora’s capabilities. The website produced the five-second video in about one minute with a simple prompt related to AI agents.

Converge Bio and Remilk collaborate on LLM-based protein generation

Source: Dov Gertz (LinkedIn)
Converge Bio announced a partnership with Remilk to employ an LLM-based GenAI platform for developing protein alternatives to traditional dairy. This collaboration highlights AI-driven biotechnology’s role in sustainable food production, enabling more efficient identification and synthesis of novel protein structures.

“Tech and Drugs” podcast series to feature Kiin AI

Source: Thibault GEOUI (LinkedIn)
The inaugural episode of the “Tech and Drugs” podcast launches on December 17th at 11 AM CET, featuring guests from Kiin AI, including Filippo Abbondanza and Mark Davies. Hosted by Thibault Geoui, the series focuses on the intersection of technology and pharmaceuticals, providing industry insights on AI-driven drug development. Geoui is the heat of science analytics at Charles Riber Laboratories in Frankfurt.

Insilico Medicine’s AI-driven drug discovery progress: ISM5411 featured

Source: Andrii Buvailo, Ph.D. (LinkedIn)
Insilico Medicine has achieved a milestone with ISM5411, an AI-designed PHD inhibitor targeting Inflammatory Bowel Disease (IBD), now documented in Nature Biotechnology. The compound, developed using Insilico’s Pharma.AI and Chemistry42 generative chemistry platforms, reflects the accelerating impact of AI in pharmaceutical development. It progressed from target selection to preclinical candidate in 12 months.

Acknowledgements

Special thanks to Andrii Buvailo, Ph.D., Science & Technology Communicator and founder of BioPharmaTrend.com and Frederic Célerse, Ph.D., Research Scientist in AI for Chemistry at EPFL for their work in identifying quality content related to AI in biopharma and beyond.