Generative AI Digest: Google, AWS announce agent offerings

December brought a long-awaited flurry of video generator announcements. The availability of OpenAI LLC's Sora to ChatGPT Plus and Pro users and Google LLC's announcement of Veo 2 garnered significant attention. The related area of interactive world generation saw a boost with demonstrations of World Labs' capabilities and Google's Genie 2. These were just a part of a larger array of end-of-year announcements, with AWS re:Invent taking place, new open-source models from Meta Platforms Inc. and Alibaba Group Holding Ltd., and notable funding rounds. To end the year, o3, the latest edition of OpenAI's reasoning model series, was announced alongside o3 mini. While o3 demonstrates major improvements in addressing complex logic, coding and math tasks, running the model is set to be very expensive, with o3 processing significantly more tokens per task.

With agentic AI projected to be one of the biggest themes in tech in the coming year, many of tech's biggest names are framing their agent offerings to be able to hit the ground running. Google and AWS both announced their own variations of agent suites and studios, following a similar set of announcements from Microsoft Corp., as well as enhancements to their models that improve agent-specific capabilities and search. Salesforce Inc. announced Agentforce 2.0, claiming improved information retrieval, new "skill" templates and enhanced integration opportunities. Perplexity AI Inc.'s Carbon acquisition showcases a growing fixation on search and retrieval fundamentals as core enabling infrastructure to support this shift. Although too costly for widespread use in many agent tasks, the advancements in reasoning demonstrated by o3 may establish a basis for more adaptable agents suitable for cognitively demanding tasks. Agents were not the only nascent technology on show in December, with the emergence of interactive world-generation foundation models.

Product releases and updates

Alibaba's Qwen with Questions "reasoning" model was released. The model, available in preview at 32 billion parameters, is designed for challenging problems, which may require multiple reasoning steps. It represents an open-source alternative to OpenAI's o1 model, which showcased the potential of chain-of-thought reasoning, and more compute and time dedicated to inference. This release followed the 100-plus open-source models the Chinese technology giant released in September.

Luma Labs announced several upgrades to its Dream Machine creative generative AI platform. The well-capitalized startup, best known for its video-generation capabilities, announced text-to-image model Photon, alongside a cheaper and faster Photon Flash. As part of its transition to v1.0, the company revamped the interface for Dream Machine and added new tools for modification and support for reference images.

Meta has announced Llama 3.3, a new 70-billion-parameter large language model. This marks a substantial reduction in size compared with the 405 billion parameters of Llama 3.1, while reportedly attaining higher performance levels. The announcement emphasizes the improvements in cost and efficiency and the implications this efficiency improvement has for broadening the use of LLMs. These advancements may have significant implications for organizations seeking local deployments.

Audio generator Eleven Labs Inc. announced a conversational AI, which brings together its text-to-speech capabilities with an LLM of a user's choice. Suggested use cases include customer support, scheduling, outbound sales and gaming.

At its re:Invent conference, Amazon Web Services Inc. made a sequence of announcements, including the release of a new Amazon Nova family of multi-modal foundation models (Micro, Lite, Pro, Premier, Nova Canvas, Nova Reel) that can process text, images and video prompts through Amazon Bedrock. Additionally, AWS unveiled new EC2 instances powered by Trainium2 chips, specifically designed for high-performance training of GenAI models, with the ability to scale workloads significantly. Further enhancements were made to Amazon Bedrock, expanding its capabilities for building and deploying GenAI applications, including a broader selection of managed models and features for managing prompts and knowledge bases. The event also showcased improvements to Amazon Q Business and Q Developer, AI-powered assistants designed to enhance business insight and streamline software development processes.

A blog authored by the technical team at World Labs publicly showcased the startup's advanced interactive world-generation capabilities for the first time. The company asserts that its models can create persistent worlds that can be navigated in a manner consistent with real-world geometry.

In November, Hitachi Vantara Corp. introduced "Hitachi iQ with NVIDIA HGX," a technology stack designed to provide organizations with AI-ready infrastructure. This platform brings together an AI software stack, networking, and Hitachi's high-performance parallel file system and object storage, among other components.

Couchbase Inc. has launched AI services within its Capella platform to assist developers in creating and managing AI agents. Services include an unstructured data service, a vectorization capability and an agent catalog. The business is best known for its NoSQL database, and it sees an opportunity to extend into agentic applications.

Salesforce announced Agentforce 2.0, with an enhanced library of pre-built "skills" that span Salesforce's product portfolio, and customers can leverage pre-built connectors to extend this reach across other applications. Of particular note was the integration with Slack, where users can now call on agents as "co-workers" and engage in enterprise search. Salesforce's acquisition of MuleSoft Inc. in 2018 appears to have been a key enabler to this connectivity drive, with the business announcing that new MuleSoft capabilities would become available in the first quarter of 2025. Improvements to the Atlas Reasoning Engine, with reportedly enhanced data retrieval and models with enhanced reasoning capabilities, were also cited as part of the release. A new agent builder assistant was also mentioned to help users create agents more easily.

OpenAI launched its "Shipmas" campaign, a 12-day initiative featuring product releases, demos and new features leading up to the holidays. The campaign kicked off with the full release of its updated reasoning model, o1, including the ability for users to fine-tune o1 on their datasets, and the introduction of a new tier called ChatGPT Pro, priced at $200 per month. It also announced the long-awaited public launch of its AI video generator Sora, which can create videos up to 20 seconds long at 1080p resolution from text, image or other video inputs. Other highlights include integrations with Apple Inc. devices, video capabilities in ChatGPT's Advanced Voice Mode and the introduction of a Projects tool for organizing conversations. OpenAI has also rolled out its search platform SearchGPT to all logged-in users, and is introducing access to ChatGPT via phone calls and WhatsApp. To end Shipmas, OpenAI announced o3 — the successor to the o1 model that focuses on tasks requiring step-by-step logical reasoning. The company had to skip a stage of its naming convention (o2) to avoid trademark issues with UK mobile carrier O2.

Google announced updates to both models and agent strategy, including releases of Veo and Imagen 3 on Vertex AI. Veo, now in private preview, allows companies to generate high-quality videos from text or image, while Imagen 3 offers improved image generation capabilities with photorealistic detail and customization options. Additionally, the company introduced an experimental version of Gemini 2.0 Flash, a low-latency model optimized for agentic workflows. Finally, Trillium, the giant's sixth-generation tensor processing unit, became generally available, reportedly delivering a fourfold increase in training performance and a threefold increase in inference throughput compared with its predecessor.

Funding and M&A

Perplexity snapped up Seattle-based startup Carbon, which specializes in connecting AI systems to external data sources. Carbon's expertise in retrieval-augmented generation should enable LLMs to access external databases for more informed responses and enhance Perplexity's capabilities to search files and messages across platforms. This move positions Perplexity in the competitive enterprise AI search market and follows industry reports that the company recently raised another $500 million at a $9 billion valuation.

OpenAI is facilitating a tender offer allowing employees to sell approximately $1.5 billion in shares to SoftBank Corp., priced according to the startup's most recent funding round, which closed at a valuation of $156 billion. The investment, sourced from Vision Fund 2, follows SoftBank's previous investment of $500 million in OpenAI as SoftBank seeks to increase its stake in the generative AI company.

X.AI Corp. initially announced that it would receive $5 billion in a round of funding on Nov. 20, raised at a post-money valuation of $50 billion. Two days later, on Nov. 22, the company raised the figure to over $6 billion. The transaction included participation from new investor Qatar Investment Authority, as well as returning investors Sequoia Capital, Andreessen Horowitz and Valor Management. The company suggests part of the proceeds will go to acquiring additional NVIDIA GPUs.

AI chip developer Tenstorrent Inc. has announced a series D funding round that includes $593 million in equity and a $100 million convertible note from 2023 that has been converted to equity. The round was co-led by returning investors Samsung Securities and AFW Partners and was raised at a pre-money valuation of $2 billion. This brings the startup's total funding to $951 million.

Datacenter operator Vultr has raised $333 million in capital as it positions itself as an alternative to major hyperscalers like Microsoft and Google. Vultr plans to utilize these funds to expand its datacenter offerings and acquire GPUs, amid a surge in demand for AI datacenters. This funding round, led by LuminArx and including chipmaker Advanced Micro Devices Inc., values Vultr at $3.5 billion; the company has not previously sought external equity.

Cartesia AI Inc has announced $22 million in new funding, led by Index Ventures, bringing its total capital raised to $27 million. This investment will accelerate the development of Cartesia's state-space models (SSMs), which enable highly efficient AI applications across various sectors. Unlike traditional transformers, Cartesia's SSM architecture scales linearly with sequence length, allowing for faster inference and better long-term memory.

Politics and regulations

Brazilian Data Protection Authority (ANPD) has included planned actions on AI as part of its 16-point agenda for 2025–2026. In November, the ANPD announced a public consultation on AI and data protection, with several questions related to the relationship between AI and the country's 2018 Data Protection Act.

The British government has initiated a consultation to regulate the use of copyrighted material in training AI models, aiming to clarify how intellectual property laws apply to AI. The consultation suggests creating an exception to copyright law for AI training while proposing new transparency requirements for AI developers regarding the content used for training. These efforts come amid ongoing legal disputes over copyright issues in AI development, highlighting the need for clear guidelines in the industry.

Texas Attorney General Ken Paxton has initiated an investigation into Character.AI and 14 other tech platforms, (including Reddit, Instagram and Discord) over concerns related to child privacy and safety to determine whether these platforms comply with Texas law, including the Securing Children Online Through Parental Empowerment Act and the Texas Data Privacy and Security Act. The investigation follows multiple lawsuits against Character.AI, where parents allege that AI chatbots made inappropriate comments to children, including instances of romantic involvement and harmful suggestions. In response, Character.AI has announced new safety features aimed at protecting teens and is working to develop separate models for adult and minor users.

The Bipartisan House Task Force on AI has released its findings in a 253-page report, with over 80 recommendations. The document is important as it is likely to shape US policymaking over the next few years. Many of the recommendations are vague, although some include applying AI to reduce government administrative burden, that IP laws need further clarification, and that highly consequential decision-making should be made with a human in the loop.

Gain access to our full news & research coverage and the industry-specific data that informs our insights

This article was published by S&P Global Market Intelligence and not by S&P Global Ratings, which is a separately managed division of S&P Global.
451 Research is a technology research group within S&P Global Market Intelligence. For more about the group, please refer to the 451 Research overview and contact page.