Image by René Bittner from Pixabay
I had the opportunity courtesy of managing director Sheena Tu of Techequity-ai to attend a good portion of their one-day AI Summit on October 26, 2024 at the Computer History Museum in Mountain View, CA. What follows are a few of the highlights from that day.
An ingestible drone for endoscopic surgery
The highpoint of the day for me was Endiatx’s pill bot demo. We in the audience got to hear directly from humble co-founder and senior R&D engineer James Erd and junior software engineer Sean Talley about their development of PillBot™, which is in clinical trials now.
PillBot™Version 1 has three pumpjet motors, a transceiver, a camera, a lithium battery, and a space-saving, multi-layer flex circuit, according to the company. Talley said the motors are the same as in a Nokia phone.
Talley told us that the bot is designed to explore the stomach and other adjacent parts of the digestive tract, providing the endoscopic diagnostic capabilities of a remote-controlled, AI-assisted bot that’s been swallowed. Endiatx’s long-term plan is for the bot to deliver surgical capabilities as well.
Erd said he’d personally ingested four different prototypes of the bot for the company’s research purposes, and that another person at the company had actually swallowed 68 of them (if I heard Erd correctly) to date. Talley shared that one of the key challenges is getting the bot to swim in the stomach. He said he’d use a game controller to drive the bot, and that the bot was using AI to aid in navigation.
Erd and Talley then proceeded with the demo. Erd drank bottles of water beforehand, then swallowed the bot and drank more water. At least part of the reason for drinking the water was to make the fluid that the bot tries to swim in more voluminous and transparent, as opaque mucus in the stomach tends to cloud the picture.
Talley soon showed us a display of what he said the bot’s camera was seeing from inside Erd’s stomach. Much of it was bubbles and cloudiness in some areas. When there was a clear view, I saw what appeared to be blood vessels and the mucosa membrane that makes up the stomach’s inner lining.
I found it difficult to orient myself and interpret confidently what we were seeing. I had the impression the bot wasn’t moving much, if at all. But I appreciated the fact that these folks were willing to try to share the thinking behind their development process and give us a live demo of what they were up to.
Solving the gen AI data quality conundrum
Generative AI-enabled virtual assistants and knowledge graphs are both inside the bullseye in Gartner’s Impact Radar this year. As I’ve pointed out before, these two technologies are complementary. Virtual assistants (agents) provide a conversational user experience. Coupled with a purely probabilistic, neural net-trained large language model back end, these assistants offer a general question answering capability that’s sometimes surprisingly good, but other times clearly inaccurate or just questionable.
The challenge for enterprises, as Philip Rathle, CTO of graph database vendor Neo4j pointed out during his afternoon talk at the AI Summit, is that businesspeople can’t trust what the virtual assistant says well enough to base serious decisions on it.
Gartner drew a similar conclusion recently, saying that generative AI is quite useful for conversational user interfaces and content generation, but ranks low on the utility scale when it comes to forecasting and decision intelligence.
As many who read Data Science Central are aware, retrieval augmented generation (RAG) can boost the accuracy of generative AI. RAG to begin with relied on vector databases to improve accuracy scores. While Rathle agrees that vector-only RAGs improve concept accuracy, he and others at Neo4j recommend knowledge graph-based RAG as well. Why?
- A knowledge graph (KG) provides actionable, concrete answers that cite specifics. Each of these elements helps to engender trust.
- KGs are deterministic; they are graphs of facts, not suppositions.
- KGs support explainability, a must-have to be compliant with data protection laws and the EU’s AI Act.
- KGs allow fine-grained authentication, authorization and access control.
This is not to mention the ten-to-1 energy efficiency advantages evident when using database retrieval over LLM generation.
Rathle makes a helpful, well-founded argument. To just add one more point, I’d underscore that knowledge graphs make data reuse and deduplication possible. Much of the data wrangling that AI teams do has only a tactical focus, which implies rework and duplication of effort. (See https://www.datasciencecentral.com/prospecting-for-hidden-data-wealth-opportunities-2-of-2/ for more information.)
A scientist’s approach to data collection and curation
Emmanuel Turlay, founder and CEO of Airtrain AI, which offers a platform for unstructured data, started as a physicist at CERN, European Organization for Nuclear Research, before he moved to the private sector, working at companies such as Instacart.
Turlay, like Rathle, advocates the use of a high-quality knowledge base. Additionally, Turlay gave some advice for the best way to build datasets: Use real production data, he said. Start saving traffic from all your applications. Curate and browse your data–manually review it.
Turlay’s advice was music to my ears. Most enterprises need to relearn how to collect data and manage the data lifecycle in a more organic fashion from those who know how. Scientific researchers like Turley know how. (See https://www.datasciencecentral.com/a-12-step-fair-data-fabric-program-for-recovering-application-addicts/ for more information.)
After all the speaking sessions were done at the AI Summit, the organizers brought in a string quartet to end the event. We heard three cellos and a double bass play Erik Satie, Vivaldi, Bach…. Actual music to our ears.
