Career Opportunities

Facebook-f Twitter Linkedin-in Instagram

AI

Enabling the Adoption of Artificial Intelligence in the U.S. Government

Axiologic Solutions is leveraging the latest, powerful tools in the rapidly evolving area of artificial intelligence to create more powerful solutions for customers across the national security and federal civilian sectors.

Over the last twelve months, we have seen the emergence of multiple ML-generated Large Language Models (LLMs) that power text interpretation and image processing, leading to significant first-generation AI capabilities (OpenAI, Anthropic, Google, Stability.ai, LAION products, OpenAssistant) that can exceed human performance in over 50 tasks (Q&A, reasoning, summarization, translation, authoring, and painting). We have also seen image manipulation AI (StableDiffusion, MidJourney, and DALL-E-2) that can produce new high-quality images and video using natural language prompts.

The recent success in AI-powered tools (e.g., OpenAI ChatGPT) has led to an accelerated interest in the AI marketplace, which will drive even more powerful AI capabilities in the near term, covering an increasing set of data modalities (video, audio, biometric, physiological, psychological, emotional, etc.), and an ever-expanding set of human-centric tasks, and eventually non-human tasks (those performed by non-person entities).

We firmly believe we are now living through the fourth revolution, powered by AI, and the changes will come quickly and be significant.

There are currently three different flavors of AI:

  • AI as a collection of facts (data, information, and knowledge) in various modalities (text, image, audio, binary, etc.) from a variety of sources and has the ability to process those facts (e.g., answer a question; reason over the facts; use facts). The AI is limited to the facts it was trained on (or has access to). In some cases, the AI may hallucinate and make up facts without informing the user it has done so.
  • AI has the ability to create new objects of different modalities (text, imagery, music, video, etc.) – using training data as a “seed” for this creation. An AI must have some training baseline to extrapolate from.
  • AI has the ability to perform various tasks (it was trained to perform) often with better than human performance, typically against data that is submitted to it (e.g., text summarization, reasoning, improving, modifying). In a type of neuro-symbolic hybrid architecture, the AI can also use external tools (e.g., access a calculator; access a sandbox to execute software; access the Internet; access a private repository) to help it complete a task. Over time, we will see more of these traditionally symbolic tools becoming native to the AI in a neural form.

There is now a very wide gap between AI producers and ML model creators – and the gap will continue to grow. Reasons for this include:

  • It is extremely difficult for non-commercial (e.g., open source) entities to be competitive in the AI space due to the significant barriers of entry: expertise; GPU infrastructure; data investment; availability of expert human labelers, etc. The OpenAssistant project is attempting to duplicate ChatGPT using the open LLM LLaMA (Large Language Model Meta AI) and crowdsourcing techniques for model fine-tuning and alignment.
  • We have seen from ML model creators a proliferation of ML models that are basically doing the same thing trained against basically the same data. The availability of ML models will lead to significant model reuse/tuning – elimination of the focus on creating/training. ML foundational models from scratch. This is a type of ML-democratization.

There is also a re-emergence of symbolic AI (semantics, rules-based) and the idea of merging it with neuro-AI create new hybrid neuro-symbolic AI to address some weaknesses in neuro approaches (e.g., it is very difficult for an LLM to reason over large amounts of text logically), particularly in the areas of:

  • Knowledge Graphs in Data Fabric
  • Reasoning via Rules Engines
  • Planning and Optimization

We have seen that AI capabilities are very “easy” to consume naively but VERY difficult to consume in a sophisticated/secure fashion due to its inherent properties of non-determinism, uncertainty, error, and biases. This has led to a need to properly design systems to take into account these properties.

With these recent advancements in Artificial Intelligence (AI), Axiologic Solutions is now offering a series of services for US Government agencies to adopt proven, reliable and safe AI technologies and services quickly and efficiently into their mission.

ADDITIONAL RESOURCES

AI  |  Specializations