Gleen AI wants to stop AI hallucinations using enterprises’ own data

As organizations around the globe look to rapidly evaluate, test, and deploy generative AI into their workflows — both on the backend, front-end (customer-facing) or each — many decision-makers stay rightfully involved about a few of the lingering points, amongst them, the issue of AI hallucinations.

However a brand new startup, Gleen AI, has burst on the scene and claims to “resolve hallucination,” based on Ashu Dubey, CEO and co-founder of Gleen, who spoke to VentureBeat completely in a video name interview.

As we speak, Gleen AI pronounces a $4.9 million funding spherical from Gradual Ventures, sixth Man Ventures, South Park Commons, Spartan Group, and different enterprise corporations and angel buyers together with former Fb/Meta Platforms’ VP of product administration Sam Lessin, to proceed constructing out its anti-hallucination knowledge layer software program for enterprises, which is focused initially towards serving to them configure AI fashions to supply buyer assist.

The issue with hallucinations

Generative AI instruments equivalent to the favored, commercially obtainable giant language fashions (LLMs) equivalent to ChatGPT, Claude 2, LLaMA 2, Bard, and others, are educated to reply to human-entered prompts and queries by producing knowledge that has been related to the phrases and concepts the human person has entered.

However the gen AI fashions don’t at all times get it proper, and in lots of instances, produce info that’s inaccurate or not related however that the mannequin’s coaching has related beforehand with one thing the human person has mentioned.

One good current instance, ChatGPT making an attempt to reply “when has the Earth eclipsed Mars?” and offering a convincing sounding clarification that’s totally inaccurate (the very premise of the query is flawed and inaccurate — the Earth can’t eclipse Mars).

Whereas these inaccurate responses could be at occasions humorous or attention-grabbing, for companies making an attempt to depend on them to provide correct knowledge for workers or customers, the outcomes could be vastly dangerous — particularly for highly-regulated, life-and-death info in healthcare, medication, heavy business, and others.

What Gleen does to forestall hallucinations

“What we do is, after we ship knowledge [from a user] to an LLM, we give info that may create a great reply,” Dubey mentioned. “If we don’t consider we’ve sufficient info, we gained’t ship the info to the LLM.”

Particularly, Gleen has created proprietary AI and machine studying (ML) layer impartial of no matter LLM that their enterprise buyer needs to deploy.

This layer securely sifts by and enterprise’s personal inside knowledge, turning it right into a vector database, and makes use of this knowledge enhance the standard of the the AI mannequin’s solutions.

Gleen’s layer does the next:

1. Aggregates structured and unstructured enterprise data from a number of sources like assist documentation, FAQs, product specs, manuals, wikis, boards and previous chat logs.
2. Curates and extracts key info, eliminating noise and redundancy. Dubey mentioned this “permits us to glean the sign from the noise.” (Additionally the origin of Gleen’s identify.)
3. Constructs a data graph to grasp relationships between entities. The graph aids in retrieving probably the most related info for a given question.
4. Checks the LLM’s response towards the curated info earlier than delivering the output. If proof is missing, the chatbot will say “I don’t know” relatively than danger hallucination.

The AI layer acts as a checkpoint, cross-checking the LLM’s response earlier than it’s delivered to the tip person. This eliminates the chance of the chatbot offering false or fabricated info. It’s like having a high quality management supervisor for chatbots.

“We solely have interaction the LLM when we’ve excessive confidence the info are complete,” Dubey defined. “In any other case we’re clear that extra info is required from the person.”

Gleen’s software program additionally allows customers to shortly create customer-support chatbots for his or her clients, and modify their “character” relying on the use-case.

Gleen’s answer is AI model-agonistic, and might assist any of the a number of main fashions on the market which have software programming interface (API) integrations.

For these clients wanting the preferred LLM, it helps OpenAI’s GPT-3.5 Turbo mannequin. For these involved about knowledge being despatched to the LLM host firm, it additionally helps LLaMA 2 run on the corporate’s non-public servers (although OpenAI has repeatedly mentioned it doesn’t acquire or use buyer knowledge to coach its fashions, besides when the shopper expressly permits it).

For some security-sensitive clients, Gleen provides the choice to make use of a proprietary LLM that by no means touches the open web. However Dubey believes LLMs themselves are usually not the supply of hallucination.

“LLMs will hallucinate when not given sufficient related info to floor the response,” mentioned Dubey. “Our accuracy layer solves that by controlling the inputs to the LLM.”

Early suggestions is promising

Proper now, the tip results of a buyer utilizing Gleen is a customized chatbot that may be plugged into their very own Slack or surfaced as an end-user dealing with assist agent.

Gleen AI is already being utilized by clients spanning quantum computing, crypto and different technical domains the place accuracy is paramount.

“Implementing Gleen AI was near no effort on our facet,” mentioned Estevan Vilar, neighborhood assist at Matter Labs, an organization devoted to creating the cryptocurrency Ethereum extra enterprise pleasant. “We simply offered just a few hyperlinks, and the remainder was easy.”

Gleen is providing potential clients a free “AI playground” the place they’ll create their very own customized chatbot utilizing their firm’s knowledge.

As extra corporations look to faucet into the facility of LLMs whereas mitigating their downsides, Gleen AI’s accuracy layer might supply them the trail to deploying generative AI on the degree of accuracy they and their clients demand.

“Our imaginative and prescient is each firm may have an AI assistant powered by their very own proprietary data graph,” mentioned Dubey. “This vector database will develop into as necessary of an asset as their web site, enabling personalised automation throughout your entire buyer lifecycle.”