Baita Services

We have a brand new feature for Baita: Real-time learning for our intermediate language model (ILM). We have a system called a resolver that catches when the ILM is not confident about a match, and determines how to handle it with a statistical model. The new feature works out what training data would have made the ILM more accurate given the resolver's output, and adds that back in to the original training dataset (only for trusted users). And it can do this all on the fly! Right now this feature is limited to mismatches caught by the resolver (an automated system), but we are working to engage real time learning directly from user feedback in the conversation itself. Once done, this will be a huge step forward in AI development. See a demo here: https://lnkd.in/gZkT99YK #artificialintelligence #ai #languagemodel

ChatGPT had quite a hiccup yesterday* with spouting unexpected and irregular outputs. This is a big problem with black box models, particularly LLMs: without explainable AI there is no good way to predict or detect these events. Effectively they are SLA failure events that cannot be easily monitored with automated tooling. And unfortunately as a lion's share of AI company's are built upon OpenAI's models, these errors propagate downstream and impact end users. Something I put a lot of effort into Baita are developer dashboards, logging, and inspection tooling. Every model, base model, and submodel comes with state tracking tools and all state data is open for developers to inspect. Plus runtime debugging tools, it's one of the benefits of running AI models on a virtual machine! #artificialintelligence #gpt #explainableai *Article: https://lnkd.in/gW8x9rQB

Here is a follow up to my last AGI theory video. This one goes into more detail about how we design algorithms that recursively break down problems without prior training or coding by a developer. Fair warning, it does go quite a bit into math and AI abstractions. #artificialintelligence #artificialgeneralintelligence #agi https://lnkd.in/gxJbiu9M

Hah, talk about crazy timing. Last week I demoed Baita's ability to easily store and recall long term information across multiple integrated apps (see https://lnkd.in/gZdyY65f), and guess what OpenAI just published today today: their experiments with long term memory (https://lnkd.in/gWZ2cnpd). Question is, will they be able to keep up with our tech? If you have long-term data needs with AI, we are ready to help you today. #artificialintelligence #ai #gpt

Our Baita AI can now be trained to interface with multiple applications and submodels at the same time, each tuned to solve a specific problem. We also avoid the accuracy limitations of context windows as seen in traditional LLM-backed AI. In the following demo, the submodels are previously trained on checkers, tic tac toe, and a separate video generation app. All running locally snappily on laptop hardware (although we can use gpt models if needed). If you want to know how to cut your AI costs by an order of magnitude while improving reliability, ping me. For the latest demo: https://lnkd.in/gZdyY65f

Here's the latest demo of Baita, now with a chat engine: https://lnkd.in/gt9JAsMN I briefly mention automated self-optimization for improving objectives like maximizing conversion rates, or reducing user turn over. For some more technical context, if you watched the last demo about dynamically generated actions spaces in checkers, this is the real motivation. As Baita partners set up a chat instance, our AI can procedurally identify machine learning problems on the fly (action space, constraints, objective), and deploy its own dedicated models behind the scenes. I'm still finishing up development of data management and serialization for this feature, but it will be awesome. Both for revenue maximization, as well as a advancement in pure computer science!

For analysts and data scientists, two of the biggest challenges are getting the desired stakeholders to act upon your findings, and getting attribution for the fruits of your labor. This is a problem I’ve been tackling in some form or another for years, and is one of the reasons why I created Baita Services. There should be a seamless pipeline of analysis to a business action, without the consumer of that analysis needing to invest any extra time or work to gain a benefit. If you are curious how we solved this problem while preserving attribution, feel free to ping me.