Generative AI Digest: Focus shifts from risk to investment

Generative AI market developments can almost be read as a call and response. In many ways, OpenAI LLC framed the early debate after the popularity of ChatGPT, with every announcement leading to a chase by foundation model providers and requiring answers from regulators, technologists and other startups. OpenAI is no longer the sole initiator of the "call"; in some cases, it is now responding to the calls of others. Perplexity AI Inc. introduced Deep Research as a counter to OpenAI's research tool, and X.AI Corp.'s newly announced Grok 3 has rolled out DeepSearch. OpenAI was also called upon to address DeepSeek's low token output costs, which likely influenced the pricing of the newly released o3-mini. Perhaps the most intriguing question for OpenAI now relates to AI safety. With rhetoric in the USA, the United Kingdom and France signaling a more relaxed regulatory stance, it seems that the responsibility for self-governance will increasingly fall on generative AI leaders.

Many of our earliest AI digests outlined the heated debates between the AI accelerators and decelerators. Notable figures in the AI industry — as well as some politicians, investors and academics — were engaged in discussions related to AI safety and perceived existential threats. The Artificial Intelligence Action Summit held in Paris underlined that this debate is largely over. Showcased by the summit's name change — it was AI Safety Summit when it was inaugurated at Bletchley Park in the UK in 2023 — the French, US and UK administrations all evidenced that the AI accelerationists are in the driving seat. It is tempting to point to the new US administration as a major force for change in this regard, but it could also be felt in France's championing of the French AI industry, and the UK government joining the US in refusing to sign an AI Action Statement. It may be that a significant safety incident will realign the scales, but the prevailing winds seem to suggest a more liberal regulatory environment than many had projected, with countries vying to build a competitive domestic AI industry.

Product releases and updates

Adobe Inc.'s Firefly Video Model has been made available in public beta. Adobe presents the release as the first "commercially safe" AI video-generation model, supporting features like "generate video" and "generative extend" in Adobe Premiere Pro. Other generative AI features in the new Firefly application are "scene to image" — where users can build 3D shapes on a canvas and use those shapes to generate images — and a spoken dialogue translation tool. Adobe also introduced new subscription plans — Firefly Standard and Firefly Pro — with Firefly Premium coming soon for those needing more extensive video and audio access.

Glean Technologies Inc., as part of its pivot from enterprise search to "work AI," has announced Glean Agents. The platform allows for the creation of AI agents using natural language and provides access to a library of prebuilt agents. The platform's agentic reasoning engine is pitched as enabling complex task execution, while "universal knowledge" provides access to structured data from databases and applications and has online search capabilities.

Social media company Snap Inc. announced a text-to-image model for mobile devices. The on-device capability can generate images quickly, with Snap suggesting that it will integrate the model into Snapchat's AI features. It appears to have been built on the foundations of a 2023 whitepaper entitled "SnapFusion: Text-to-Image Diffusion Model on Mobile Devices," released in partnership with Northeastern University.

OpenAI has launched "deep research," a new agentic capability in ChatGPT that conducts multi-step research on the internet for complex tasks, available to Pro users. This tool leverages the upcoming OpenAI o3 model optimized for web browsing and data analysis, providing comprehensive reports with clear citations. In January OpenAI also announced the o3-mini model, which it suggests is optimized for cost-effective reasoning in STEM fields, offering features like function calling, structured outputs and developer messages. This model is available to ChatGPT Plus, Team and Pro users.

Perplexity Deep Research, a tool also designed to conduct thorough research and analysis, was announced shortly afterward. Perplexity's capability is available for free — unlike OpenAI's offering — although with usage limits. Pro subscribers can get unlimited queries. As with other deep search outputs, the tool performs multiple searches, reviews numerous sources and aims to compile the information into a detailed report.

DeepSearch was highlighted as part of x.AI's Grok 3 model family announcement. The launch video suggested that in addition to the Grok 3 flagship model, a set of Grok 3 mini, Grok 3 Reasoning and Grok 3 mini Reasoning models would be made available. The video indicated that Grok 3 outperforms other frontier models across several benchmarks related to math, science and computer science, and displayed strong performance on the crowdsourced chatbot arena tables. X's Premium+ tier users will get access to Grok 3 initially. However, it appears some usage will be gated behind a separate paywall known as SuperGrok, including DeepSearch and reasoning models. A focus of the announcement was the high volumes of GPUs x.AI has access to, in order to train its models — citing 200,000 GPUs as part of its Colossus project to build the largest fully connected cluster.

Mistral also announced new models and AI assistant features. Mistral Small 3 is a 24-billion-parameter model released under the Apache 2.0 license. It is designed to deliver faster millisecond-per-token speeds than models of comparable performance. The company also announced Mistral Saba, a model "custom trained" to meet "specialized regional" needs in the Middle East and South Asia, and that seems specifically geared toward Middle Eastern audiences. Performing well in Arabic, Mistral suggests the model can grasp local idioms and cultural references. AI assistant Le Chat saw an upgrade and is now able to provide a rapid response setting and a code interpreter, and has been integrated with a Black Forest Labs Flux Ultra image-generation model.

Zyphra Technologies has launched Zonos AI, an open-source text-to-speech model, in beta. It is an open-weight text-to-speech model trained on more than 200,000 hours of varied multilingual speech, and can be used for generating voices from text or voice cloning if given a reference clip.

Anthropic PBC announced Claude 3.7 Sonnet alongside an agentic AI coding tool called Claude Code. While many were waiting on a disruptive Claude 4.0 release, this updated version offers significant enhancements. As a hybrid reasoning model, 3.7 Sonnet can use chain-of-thought reasoning if required, and using the API can configure the depth to which the model "thinks."

Alphabet Inc.'s Google announced the general availability of Gemini 2.0 Flash, a model optimized for high-volume tasks, and introduced Gemini 2.0 Pro, an experimental model designed for coding and complex prompts. Additionally, Google released Gemini 2.0 Flash-Lite, a highly cost-efficient model, in public preview. All models feature multimodal input with text output, with both Gemini 2.0 Pro and 2.0 Flash soon to receive image and audio output capabilities.

Bytedance Inc. announced a family of joint image-and-video-generation models, known as Goku AI. A whitepaper outlines a novel "rectified flow formulation" approach designed to handle images and videos in a unified framework. The outcome appears to be high-quality videos that can be over 20 seconds in length. The company also announced AI video-generation tool OmniHuman-1. The tool can generate realistic videos from an image and a short audio or video clip. One showcased capability is lip-syncing. A challenge that Bytedance cites is that the model is very resource-intensive, requiring significant computational power.

Funding and M&A

The OpenAI board rebuffed a $97 billion bid from a consortium of investors led by Elon Musk to buy the nonprofit that controls the company. This appears less to have been a serious bid, and more of an attempt to make the process of OpenAI's shift to a for-profit venture more challenging, as this involves compensating the nonprofit. By making the bid, Musk has potentially elevated the value of the nonprofit.

Safe Superintelligence, co-founded by former OpenAI chief scientist Ilya Sutskever, announced that it will raise more than $1 billion in a round of funding led by new investor Green Oaks Capital Partners. The startup has yet to deliver a product, and suggests it will avoid short-term commercial pressures. However, the profiles of Sutskever and other founders — Daniel Levy (formerly at OpenAI) and Daniel Gross (formerly at Apple Inc.) — have driven interest in the company.

AI inferencing provider Groq Inc. secured a $1.5 billion commitment from the Kingdom of Saudi Arabia. The commitment was announced at the LEAP 2025 event, alongside updates that Groq would be delivering its AI inferencing capabilities partly through a new datacenter in Dammam, Saudi Arabia, which it claims is the largest inferencing cluster in the region.

Tana Labs Inc. emerged from stealth with $14.5 million in funding from Lightspeed Ventures, Tola Capital, Firstminute Capital, Northzone Ventures UK and Alliance VC. Tana is developing an AI-native workspace powered by a knowledge graph. The workspace is designed to help users access the information they need quickly, generate content using generative AI, and use their voice to trigger actions.

MongoDB Inc. has acquired Voyage AI to enhance its capabilities in embedding and re-ranking models for information retrieval. This integration aims to improve the accuracy and relevance of data retrieval in AI applications, addressing issues like AI hallucinations. Voyage AI's technology will remain available through various platforms, with further integrations into MongoDB expected later this year.

HP Inc. is reportedly set to acquire assets from Humane and bring in the majority of the AI device startup's employees. The startup was known for its AI pin, a poorly received wearable device designed to screenlessly interface with AI.

Red Hat Inc. announced in January that it had completed its acquisition of Neuralmagic Inc., a startup focused on accelerated inference serving. Red Hat announced its intent to acquire the company in November 2024.

Politics and regulations

In February France hosted the Artificial Intelligence Action Summit, gathering at the Grand Palais. Many of the speeches at the event focused on investment rather than risk. US Vice President JD Vance stated that "I'm not here this morning to talk about AI safety, which was the title of the conference a couple of years ago. I'm here to talk about AI opportunity." French President Emmanuel Macron suggested France was "back in the AI race," citing major investments in France. There were multiple warnings issued around overregulation.

The UK and US refused to sign the joint international declaration "Statement on Inclusive and Sustainable Artificial Intelligence for People and the Planet," which resulted from the summit. The statement reaffirmed priorities to "ensure AI is open, inclusive, transparent, ethical, safe, secure and trustworthy," "avoid market concentration," and "encourage AI deployment that positively shapes the future of work," among others. The UK government suggested it did not sign the declaration due to a lack of "practical clarity on global governance."

In a further indication of the UK's increasingly acceleratory AI attitudes, the UK's AI Safety Institute has been renamed the AI Security Institute. Additionally, the Department for Science, Innovation and Technology, in a notice detailing a memorandum of understanding, outlined a partnership with Anthropic around AI opportunities. Light on detail, the notice suggested the collaboration would initially focus on improving how UK citizens can access government information and services online using AI.

DeepSeek has encountered regulatory and political resistance in Europe and the US. Italy's Data Protection Authority instructed DeepSeek to cease processing data of Italian citizens immediately and initiated an investigation into the company. Additionally, Ireland and Italy requested further information regarding DeepSeek's data processing practices. Additionally, two US congressmen introduced a bill to ban DeepSeek from being installed on government devices.

Gain access to our full news & research coverage and the industry-specific data that informs our insights(opens in a new tab)

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.