Funding Activity

Blattner Technologies acq. Superwise

Blattner Technologies, a digital transformation and predictive analytics provider, acquired Superwise.ai to add to their portfolio of AI-related offerings and services.

Read PR

Data for AI

Good data is more important than ever.

Within the Data for AI space, the big theme (intertwined of course with generative AI) is data sourcing and labeling. Not only are LLMs data-hungry, they also perform proportionally with data quality: for organizations building their own LLMs, data quantity is often better than data quantity because they allow for smaller, more lightweight models that still perform exceptionally well. In addition, representative data that has been cleared for copyright purposes and purged of toxicity is what is required to build models that can surpass OpenAI's in terms of usefulness.

We've learned a lot from OpenAI's less than ideal methods - from paying Kenyan workers $2/hour for toxicity labeling to copyright infringement, the GPT series is a reiteration of some of the worst aspects of machine learning and artificial intelligence. The importance of responsibility in these instances cannot be overstated.

In the sourcing space, labelers like Snorkel AI are accompanied by more innovative approaches such as StageZero (assembling global language datasets through mobile games for more representative and performant audio/voice algorithms) and GCX (a source of music and audio content that has been cleared for use in algorithms). The value of data itself has skyrocketed because the modeling problem has been solved.

MLOps and ModelOps

MLOps companies take advantage of confusion around Large Language Models (LLMs).

Much of this report covers the generative AI hype and the response across the ecosystem. The MLOps and ModelOps companies in EAIDB have largely reacted positively and have adapted or built aspects of LLMOps to capture some of the use cases in language modeling. There is a great opportunity here simply because the number of new tools, models, papers, and ideas around the topic has skyrocketed in such a way that contrasting various approaches is near impossible.

However, we expect that these solutions will not dominate the market going forward. Just as traditional machine learning turned from an art to a science with low-code platforms, AutoML, and the entire “model and platform builders” space, so too will language modeling become a science. In fact, this has already begun to take place with new research and platforms like MosaicML (acq. by Databricks). There is also some new research on "AutoGPT" (think of a self-training, self-finetuning GPT variant). Read the paper here.

In addition, Fiddler was recently funded by Dentsu Ventures to scale their business in Japan and Southeast Asia. This is one of the first instances of international recognition of a responsible AI startup through a funding round, and Fiddler certainly leads the pack in momentum.

AI GRC

AI GRC startups face threats from all over the ML value chain, but the leading products are clearly differentiated.

AI GRC products cover everything from organization-wide transparency to holistic legal and regulatory compliance to model value and risk assessment. As the sort of "catch-all" at the end of a machine learning pipeline, AI GRC products face consistent competition from literally all sides. One of the most prevalent trends here are the sheer number of companies from other primary business lines revealing new GRC products to build on their existing customer base and provide their users with a more all-encompassing view of their model lifecycles.

Ultimately, time will tell who the winners in this space are, though some companies like Monitaur and Credo have pulled ahead due to their positioning and holistic reach over the machine learning lifecycle.

Another interesting trend is the absence of vertical AI GRC products. Of the 32 AI GRC companies in EAIDB, only two are more vertical-aligned. There are a few companies that focus on the full array of use cases for a particular vertical, such as FairNow for HR and Monitaur for insurance, but it seems relatively rare to encounter companies like these. This may be because global demand is still low and in order to capture a wider client base, most AI GRC companies prefer to be horizontal. There is certainly space for vertical solutions that are specific to the nuances and use cases present in highly regulated industries.

Some companies are choosing to remain horizontal but are focusing more on the language modeling and generative AI side - Preamble, Deepbrainz AI, and Titaniam are examples.

Model and Platform Builders

HR model-builders are dying out as other horizontal-oriented companies take over. Finance and healthcare are rapidly scaling.

Within the model and platform builders, we have startups that expressly develop and sell their algorithms or proprietary technology to clients. Examples include toxic content moderation services (e.g., Modulate, Trust Lab) and HR algorithms (e.g., Kanarys, Pave).

Human Resources (HR)

In this report, we focus on two verticals: HR and finance. The decline of new HR-specific startups and algorithms has been repeatedly noted in EAIDB's previous reports and continues into the first half of 2023. Most notably, the rate of new startups from 2015-2020 was about five per year.There has not been a single such company on EAIDB's radar targeting the same space since 2020.

This may simply be because HR teams are implementing their own ML - as barriers to this technology continue to decrease, the build vs. buy decision leans increasingly towards “build.” Newer EAIDB startups in the HR space target GRC or other high-level, post-implementation use cases.

Lastly, as the more mature vertical in EAIDB, this space has also seen its fair share of exits thus far. Four acquisitions (Happyr, Praice, Headstart, Talenya) and one liquidation (Ceretai) have given us a glimpse at what consolidation might look like for other verticals.

Financial Services

Finance is an almost opposite story. There has not been much momentum in the finance vertical for responsible AI enablement just yet, but there are certainly relevant catalysts that are pushing the industry towards adopting better AI. From the Consumer Financial Protection Bureau's (CFPB) report to the US Congress on fair lending in 2022, there were 174 institutions cited for fair lending-related violations coupled with an increase of 146% in fair lending examinations and targeted reviews from 2020-2022. The CFPB has also historically referred to “digital marketing fairness” and redlining as other adjacent concerns. This market is wide open with only a few strong players such as FairPlay AI, Evispot, and Stratyfy. Building this technology in a financial context requires an exceptional technical discipline as well as the ability to demonstrate value (a recurring theme within the responsible AI ecosystem).

Alternative Machine Learning

Causal and federated AI are growing at faster rates than GenAI.

On the Alternative ML side, there are really three frameworks that stand out against the backdrop of centralized ML: causal AI, federated learning, and neurosymbolic AI. We won't cover neurosymbolic AI in this report because UMNAI is really the only startup spearheading this effort.

Causal AI

It is important to establish that, according to some estimates, causal AI is growing at a much faster rate than generative AI. This is potentially because of its essentially unlimited use cases in healthcare and other experiment-driven fields where generative AI is just not as useful. There are discussions, however, of uniting the two - why not give generative AI models the ability to make causal inference? This may not even be so far away. A paper released earlier this year covers exactly this topic.

Within the causal AI market, EAIDB has recognized CausaLens as one of the only fully horizontal causal AI platforms focused on delivering high-fidelity and explainable causal AI capabilities to businesses.

Federated Learning

Still a growing market, federated learning solves problems associated with high data transfer costs and sensitive data usage by allowing machine learning algorithms to be constructed in a distributed fashion such that data never leaves its home.

Products like Integrate AI and Flower Labs are bringing federated learning to businesses with their platforms. However, there is also additional interest in Large Language Models (LLM) trained in a federated environment. Companies like DynamoFL and FedML attempt to provide this functionality to alleviate some of the privacy concerns with language modeling. This is an early and very specific market that depends on both technical details (such as increased computational complexity) and federated learning adoption trends.

AI Security

AI Security is fast becoming a VC-favorite product type.

The newest category addition to EAIDB, AI Security consists of “security for AI,” not “AI for security.” Startups in this area address some of the more dangerous aspects of models in production. This category also comprises the greatest amount of current investor attention because it is the most approachable and understandable from a business value perspective.

Companies in this space offer model testing solutions (like adversarial testing, inversion testing, penetration testing, prompt injection, etc.) and sometimes couple them with dashboards or model inventories (akin to some products in the GRC space). Those that provide a more holistic view of an organization's model attack surface (Protect AI, Cranium AI, etc.) target CISOs or other C-level executives. Others have strong proprietary technology meant to thoroughly test and secure AI pipelines (Adversa AI, Robust Intelligence, etc.).

NightDragon released an insightful post recently that broke down their perspective of this budding market. Though their list of displayed companies and funding information is somewhat incomplete, it does provide a holistic look at what they describe as "a new frontier of AI innovation." Astoundingly, every one of the AI Security companies in EAIDB has raised funding already.

Firms like Unusual Ventures have recently voiced their interest in the space, while others like Paladin Capital Ventures, Flying Fish Partners, and more have made active investments already.

Open Source

The rise of freemium business models and generative AI moves open source forward.

Continuing our theme from the end of 2022, we find that there are still large gaps in the open source domain for responsible enablement. The two most abundant types of libraries in this space revolve around MLOps and explainability. There are a few libraries for fairness, bias mitigation, etc., but these are undermaintained and understaffed relative to some of the more operational repositories.

Recently, however, there has been a large influx of open source tools released (both by companies in the MLOps space and by independent organizations) for LLM maintenance, development, and validation. Nearly every major responsible MLOps company in EAIDB (Arize Phoenix, Fiddler Auditor, TruEra TruLens, etc.) released packages to assist with LLMOps. In addition, several entities have created open source libraries for regulating LLM behavior and security - Microsoft Guidance, Microsoft Counterfit, Rebuff AI, Guardrails AI are all examples.

One notable event from this year is Mozilla VC's recent investment into Fiddler AI. Mozilla's commitment to open source principles and vision of supplying powerful tools to developers at low costs certainly motivated their entry into the space: their funding round preceded Fiddler's launch of their open-source Fiddler Auditor.

Outside of LLMs and generative AI, there is still a limited set of tools for processes like compliance and governance. For small-to-medium enterprises (SMEs), the price tag on larger solutions is steep. Some companies, like KomplyAI, aim to provide these capabilities at a lower cost.

Another interesting trend is the rise of freemium models within EAIDB companies. This type of business model is similar to HuggingFace's - provide an open source library for free but charge for scale and additional features/tools. HuggingFace democratizes large models and provides tooling to implement them, but charges for scale via inference endpoints, CPUs/GPUs, AutoTrain, etc. Examples of this are Flower Labs, Giskard AI, Themis AI, and Deepchecks. All of these companies were founded post-2018 with the majority founded post-2020.