Generative AI Digest: Established AI players shift beyond model development

Even by the standards set by preceding months, September was a big month for generative AI news. In part as a reflection of peak conference season, we have seen new models, new offerings and a flurry of major fundraising rounds. In notable news for market direction, this month saw pivots from former generative AI high flyers that are extending beyond model development. We also saw two mega-rounds from Ilya Sutskever's safety-focused startup and "large world model" startup World Alliance Laboratories Ltd., better known as WorldLabs, and two acquisitions from generative AI startup Typeface Inc.

The new o1 model that OpenAI LLC released garnered significant press attention, and the past few weeks brought a number of new cutting-edge models. That said, it is hard not to feel that the competitive environment is shifting, with growing emphasis on the tooling offered around the models and less on the models themselves. This is perhaps most visible with Aleph Alpha transitioning from a model provider to releasing model-agnostic developer tools; also with Midjourney Inc. and OpenAI's exploration of hardware devices, Anthropic PBC's "artifacts" and the more holistic agent builders emerging from Salesforce Inc. This transition is driven partly by how challenging smaller foundation model providers are finding competing with the latest flagship models directly, and partly by a drive to improve the performance of open-source models.

Product releases and updates

Open AI's new model o1 is a text-based model set designed for more complex reasoning tasks. The model is not yet augmented with the file parsing and image processing capabilities of flagship GPT 4o and is currently rate-limited to 50 queries a week. What stands out from the release is the model's ability to walk through multiple reasoning steps — a seemingly automated alternative to disaggregating an ask to provide a model with multiple step-by-step instructions. The response can take a lot longer than prompting GPT 4o but is seen as well suited to challenging coding tasks or those requiring math. In other Open AI news, there appears to be confirmation that the company is working on a new device, with a project team helmed by Jony Ive.

Apple Inc.'s "It's Glowtime" event in early September showcased a number of new AI features. Most notably, the company revealed that the release date for Apple Intelligence, its new AI assistance system, would be as early as October in the US. Apple also announced new iPhone models, suggesting that chip upgrades would allow the series to better support Apple Intelligence capabilities.

Chinese startup Shanghai Xiyu Jizhi Technology Co. Ltd., better known as MiniMax, released its first video model at the end of August. A text-to-video model, with a suggestion the release will later be joined by image-to-video capabilities; the outputs are currently limited to six seconds. The high-resolution outputs, and virtual camera control MiniMax is able to support, has generated significant interest. The company has raised $850 million in total funding, which includes a $600 million round led by Alibaba Group Holding Ltd. that was announced in March.

AI Agents were the focus of Salesforce's annual Dreamforce conference. These "Agents" refer to autonomous applications that can execute tasks, with pre-built frameworks for sales development representatives and personal shopping agents, among others. Facets include Agentforce Studio — a low-code builder to configure and test agents — and an associated partnership network, with Google Cloud, Amazon Web Services Inc., International Business Machines Corp. and NVIDIA Corp. among the cited partners.

Runway AI Inc. announced a video-to-video capability as part of its Gen-3 upgrades. This allows a user to change the style of video content using text prompts, with the company showcasing the ability for a user to control the scale of change that takes place. A user may configure settings to have most of the original video retained or to make the transformation more substantial. Transformed clips are currently limited to 10 seconds. Accompanying this announcement was the launch of a new API service, designed to support the integration of its Gen-3 Alpha Turbo model into other applications.

Midjourney appears to intend to establish its own hardware, creating a new team based in San Francisco. Speculation, fueled by the background of Midjourney's head of hardware and the company's social media posts, is that it might be looking to enable a virtual environment that a user could "enter." The image generator has already denied that the device will be a wearable.

Ideogram 2.0 was released in late August, alongside an Ideogram iOS app. The text-to-image model is optimized around depicting text and graphic design but can also be used to generate general-purpose images. The announcement comes after the company completed $80 million in series A funding in February, the proceeds of which it said would be used to generate more capable generative media models.

Anthropic announced an enterprise-focused subscription plan, Claude for Enterprise. The foundation model provider already offers a Team plan, but Claude for Enterprise adds additional security and administrative controls. The upgrade also comes with an expanded context window (500k) and additional integrations, starting with a GitHub integration. A major area of focus for Anthropic has been artifacts — reusable pieces of content that can be iterated upon or referenced; the company cites diagrams, code snippets and plain-text documents — and they feature heavily in the Claude for Enterprise positioning.

German generative AI startup Aleph Alpha GmbH announced a new product, PhariaAI, a strategic pivot. The company was best known for its foundation models, but PhariaAI is presented as a toolset to "unify and streamline the development of AI applications in enterprises and public institutions." The focus for the business appears to remain "sovereignty" and control — its developer tools are designed to support hybrid execution where sensitive information is processed only in secure environments — but is model-agnostic, designed to also support third-party commercial and open-source models. The company did release two new 7bn parameter foundation models, but it is clear that foundation model development is no longer the core mission of the business.

Chinese startup DeepSeek AI announced DeepSeek-V2.5, an open-source model specializing in math, code and reasoning. At 236bn parameters, and a 128k context window, unsurprisingly the model outperforms other well-known frontier models in Chinese. The company cites benchmarks suggesting it is competitive in coding and arithmetic relative to leading Open AI and Anthropic models. The model is a combination of earlier models DeepSeekV2-Chat and DeepSeek-Coder-V2-Instruct.

Matt Shumer, CEO of OthersideAI Inc., a company best known for its HyperWrite product, announced an open-source, reflection-tuned mode (an optimization technique where models learn to improve their decision-making by "reflecting" on past actions or predictions), Reflection 70bn. Claimed to be a variant of Llama 3.1 70B Instruct, the architecture is supposedly designed with separate planning and generation steps with the ability to catch errors in its reasoning. Shumer cited benchmarks that suggested Reflection 70bn was the highest-performing open-source model of its size upon its release. Significant skepticism has been directed at the claimed model lineage of Reflection 70bn, as well as its ability to reproduce benchmark scores.

In other open-source news, the Allen Institute for AI released a mixture of expert models, OLMoE-1B-7B. The Allen Institute for AI is exposing the training data and source code behind the model, rather than merely the model weights, with the intent to make a mixture of expert models more accessible for researchers.

Funding and M&A

Generative AI technology provider Typeface acquired personalized photography startup Treat Technologies Inc and AI content platform Narrato Inc. Typeface suggested that these acquisitions would be used to improve its platform's "multimodal capabilities." Typeface disclosed $165 million in total funding since its 2022 incorporation, with Treat receiving $15 million and Narrato $1 million, according to S&P Capital IQ Pro.

OpenAI is finalizing a new fundraising round, with the reported intent to raise over $6 billion. Reportedly, investors include Thrive Capital Management, Apple, NVIDIA and returning investor Microsoft Corp. OpenAI will receive $1 billion from venture capital firm Thrive Capital Management, with an indication that the round will value the generative AI company at $150 billion.

AI research and code generation startup Magic AI Inc. announced a $320 million venture round, bringing the company's total funding to $465 million. The round was led by individual investor Eric Schmidt, formerly CEO of Google LLC, who was joined by returning investor CapitalG Management Company and new investors Atlassian Corp., Jane Street Group, Sequoia Capital Operations, Elad Gil, Nat Friedman and Daniel Gross. The company claims its objective is to "build frontier-scale code models to build a coworker," and will invest the proceeds in bolstering its AI infrastructure.

OpenAI alumni Ilya Sutskever's new AI startup Safe Superintelligence Inc. (SSI), focused on safe AI development and has raised $1 billion in funding to grow its core team and computing assets. A post-money valuation was not disclosed but industry sources claim it is close to $5 billion. The mega-round shows that investors will shell out for top AI talent even as overall investment in generative AI has slowed; Sutskever founded the company only in June. Investors include Andreessen Horowitz, Sequoia Capital, DST Global and SV Angel.

There is a $30 billion AI infrastructure fund in the works in a partnership among BlackRock, MGX and Microsoft. NVIDIA will serve in an advisory capacity to support the Global AI Infrastructure Investment Partnership, which is reported to be set to raise up to $30 billion in equity investments and up to $70 billion in additional debt financing.

WorldLabs, a startup led by Stanford professor Fei-Fei Li, emerged from stealth with the news that it has raised a $230 million venture round at over a $1 billion valuation to build large world models, a type of model that is spatially aware and capable of reasoning and interacting within a 3D world. The startup has an extensive cap table; this round was led by Andreessen Horowitz, NEA and Radical Ventures, and includes many other individual investors (Marc Benioff, Jim Breyer and Andrej Karpathy, to name a few) as well as strategic investors Adobe Ventures, AMD Ventures, Databricks Ventures, NVentures, the venture capital arm of NVIDIA and Shinrai Investments LLC.

Generative-rich media company WOMBO Studios Inc, best known for its lip-syncing and image-generation capabilities, announced a C$12.2 million (about US$9 million) round co-led by NVIDIA and returning investor Digital Asset Fund. Other investors included CoreWeave Inc., State Bank of India and Web3.com Ventures. The round brought the Toronto-headquartered startup's total funding to C$20.3 million (about US$15 million).

Enterprise search company Glean Technologies Inc. has raised another $260 million in a series E funding round at a $4.6 billion valuation, doubling the latter in the six months since its February series D round. Simultaneously, it repositions itself within a much broader category of Work AI and claims to have more than $550 million in cash on hand to tackle the market opportunity. New investors include the round's co-leaders Altimeter and DST Global, Craft Ventures, Sapphire Ventures and SoftBank Vision Fund 2.

Other rounds of interest are in the table below.

Politics and regulations

The EU, US and UK signed the Council of Europe Framework Convention on AI, a pact related to the regulation of AI systems. The strategy is to continue to onboard jurisdictions. States are instructed to carry out AI risk and impact assessments related to human rights, democracy and the rule of law and to support establishing procedural rights and safeguards. One notable ask is that all interactions with an AI should include notice that it is an AI, to ensure the model cannot be confused with a human being.

The government of Singapore is updating its Smart Nation strategy, with AI presented as a critical technology framing the initiative. The Smart Nation initiative was first announced in 2014 as a framework to encourage digitalization of government services, as well as drive digital advances in Singapore more generally. Early analysis suggests the update will encourage AI investments in the public sector and may include items on semiconductors.

Voluntary AI guidelines have been announced in Australia, accompanied by a consultation on whether they should be formally instituted for high-risk systems. The consultation will take place from Sept. 4 to Oct. 4. The guidelines include human oversight in all high-risk AI application areas, maintaining records at various stages of the AI model life cycle and establishing processes where those that have been negatively affected can escalate their challenge.

Adobe Inc., OpenAI and Microsoft have all announced their backing for a watermarking bill in California. The bill will require AI companies to label AI content and prohibit companies from selling software designed to remove provenance data. The bill has been revised after an initial backlash from some tech industry bodies, so the backing of the bill by large AI players is seen as a sign of progress.

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.
 

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.