The conversation regarding the evolution of artificial intelligence is hitting a crescendo with experts highlighting that the technology may have already peaked after ingesting the entirety of the internet. However, Microsoft’s new Orca AgentInstruct framework is the company’s workaround that presents added information to AI to help it to continue to learn and scale.
Orca-AgentInstruct framework was recently released by Microsoft to feed the models that power AI solutions with synthetic data. Microsoft’s framework takes advantage of Generative Teaching processes that allow models to learn new skills based on synthetic data that includes datasets created from raw inputs such as text files and pre-written code.
Ideally, models would ‘practice’ on tasks using those randomly generated text files and code to then produce their own versions or responses to the data served.
While the process may bring up the concepts of cloning degradation like in the movie Multiplicity, Microsoft says defends its framework in a new write up on its Microsoft Research Blog titled, Orca-AgentInstruct: Agentic flows can be effective synthetic-data generators.
The worry about data scarcity when evolving AI has been on Microsoft’s radar for a bit and back in July of 2024, the company produced a video to help explain how it would tackle the problem going forward.
Over the past few months Microsoft’s AgentInstruct framework has generated 25-million-pair datasets based on synthetic data, and those results helped power a Mistral 7-billion parameter base model. But it is not only AgentInstruct framework’s ability to recreate data based on synthetic data, but it was able to do so while also boasting a 40 percent performance boost on the human-centric benchmarking tool AGIEV and 54 percent on benchmarked data sets tool GSM8K.
Microsoft achieved its feat of training AI on AI-generated data with agentic flow that allows the framework to help refine output by leveraging a processes called reflection as its own version of checks and balances to prevent model collapse or degenerative outcome that lead to that dreaded Multiplicity outcome of dumb or nonsensical results.
Microsoft’s partner OpenAI has implemented its own version of AgentInstruct called Orion which also attempts to head off AI model optimizations and hindered scaling due to data constraints. NVIDIA is another company invested in AI that is using a similar synthetic data patchwork called Nemotron-4 340B to offer other companies running into model refinements, a way to keep scaling and growing their AI solutions.
