The more perceptive of you would know that I had spoken at Hasgeek Fifth Elephant last month. Some of you have asked me about the video, and here goes.
For a month after the event, the video was exclusive to Hasgeek’s members (among the many carrots they have to spur membership). Now that a month has passed, they have removed the paywall, and here it is.
All I remember is that I had slept really badly the previous night (checked my logs - some 4 hours, with a 4 hour gap in between). I had got to the event late morning, and had been pleasantly surprised by the number of people I knew who were there. I was reconnecting with some of them after a really long time.
In any case, this talk reveals “the knowledge” as it stood then (July 13th). We’ve built a lot more since then, but the basic principle remains the same - get LLMs to do creative stuff, and then translate it in a way that a more deterministic system can execute. The more deterministic system (uncorrelated with the LLM) can then “verify” the output of the LLM and provide feedback. This way we’ll get the upside of the LLM while hedging for its downside.
At the time of the talk, the only “module” of our product that we had built out was for cleaning data (which is why I focus so much on that there). Now, we’ve built out plenty more, and are offering Babbage as a service. If you or someone you know is interested, do give us a shout!