Introducing Orca, Microsoft’s Breakthrough in Teaching Machine Learning Models How to Teach Themselves


Microsoft AI Introduces Orca

Microsoft just introduced Orca, a massive 13-billion parameter model. Orca’s a whale of a project that aims to take what those big-brained Large Foundation Models have been doing and kick it up a notch.

These large models are already pretty smart on their own, learning from zero to hero with very little help. But Microsoft’s asking the big question: Can these models teach themselves or, better yet, other models? That’s where Orca comes in.

Orca’s job is to understand how the GPT-4 thinks and then pass that knowledge onto others. It does this by dissecting the conversation-like exchanges between questions and answers from GPT-4, giving its students a clear path of how to get from A to B. This fine-tuning process has seen significant success in overcoming challenges with task diversity, query complexity, and data scaling. Quite the teacher, our Orca is.

Microsoft researchers don’t stop at just explanation, though. They’re spicing up Orca’s training diet with a cocktail of complex tasks from the Flan 2022 Collection. This diverse menu of brain teasers ensures Orca’s ready for whatever the world can throw at it.


Meta reportedly making LLaMA commercially available, despite lawmaker inquiries

Meta is making a run for it, set to take their open-source LLaMA out of the research-only barn and make it available for commercial use. This is even after a couple of U.S. senators put their heads together and sent a stern letter to Meta’s sheriff, Mark Zuckerberg, about a leak of the LLaMA to 4chan. Oh, the drama.

Meta’s Fundamental AI Research Team, or FAIR has been sharing LLaMA’s model weights for educational and research purposes. But, much like a juicy neighborhood secret, these weights got leaked, letting developers worldwide get a taste of a GPT-level LLaMA.

Zuckerberg announced plans to weave generative AI into all its products, clinging tightly to their commitment to an open, science-based approach to AI research. In a chat with Lex Fridman, he also declared that LLaMA is basically going to be the engine that powers access to AI for small businesses and content creators that use Facebook’s apps.


Workato partners with OpenAI to ease business automation

Workato, a company specializing in office automation, has joined forces with OpenAI to make automating office tasks even easier. They’re incorporating AI models from OpenAI into their own platform. Imagine a coworker who’s super smart and really good at making things efficient – that’s the kind of benefit they’re aiming for.

Their platform is getting some snazzy new features, like Workato Copilots, which lets folks create automation and connect applications using plain ol’ English. They’ve also got something called WorkbotGPT, which is like having a watercooler chat with your enterprise apps and data via popular chat apps.

This AI-powered coworker, the Copilot, is like a 24/7 office guru, offering onboarding support, learning new capabilities, offering recommendations, and troubleshooting issues. It’s not quite like your office buddy Jim, but it’s getting there.

READ THE ARTICLE ON VENTUREBEAT looks to innovation to chip away at Google’s search dominance

Richard Socher, the big-brained wizard behind, has his eyes set on the prize: Google’s search crown. He ain’t no rookie to AI; in fact, his 2014 doctoral thesis was pretty much a red carpet entrance for the AI show we’re all tuning into today.

Socher, after sprinkling some AI magic at Salesforce with an AI layer called Einstein, started in 2020. He’s not backing down from this uphill battle. After all, he’s got time and innovation playing in his team.

Even when Neeva, another search engine kid on the block, got picked up by Snowflake after failing to make a significant dent, Socher isn’t flinching. His site, he claims, has more users than Neeva ever managed to hook.

Back in 2014, Socher and his Stanford pals turned the key to some significant AI breakthroughs. Now, as he sees his research getting all jazzed up in real-world applications, he’s living a scientist’s dream. “Being ahead of your time makes you a visionary,” he says.


Grammys outline new rules for AI use

The Grammys just laid down the law for AI-made music. The Recording Academy, aka the Grammy, decided a song cooked up by an AI could still get a nod, but only if a real live person helped stir the pot significantly.

“Look, if we hear an AI crooning or strumming, we’re all ears,” says Harvey Mason Jr., the Grammy bossman. “But for a songwriting award, we need a human to have penned most of it.”

Apparently, Mason sees AI not just as the future, but as the present of the music industry. His attitude isn’t “out with the old,” but more “how can we get this new stuff to play nice with our standards?” Mason’s asking the hard questions on how to make room at the table for AI, without letting it eat everyone’s lunch.


AI Could Predict Pancreatic Cancer Early In Some Cases

AI may just have gotten a step closer to beating pancreatic cancer at its own sneaky game. Now, this cancer’s a real bad apple, known for springing up on folks too late for doctors to do much. But, researchers publishing in Nature Medicine reckon they’ve got an AI tool that could help catch the snake in the grass early.

These scientists used an AI tool to analyze medical records. They trained the machine to predict who might get a visit from this cancer, based on the medical codes in these records and symptoms, some as odd as gallstones and belly aches.

Sounds weird, but apparently, if you were pegged as high risk by this tool, you’re 320 in a 1,000 to actually get diagnosed with the disease. No small beans when early detection can swing you from being a goner to standing a fighting chance.

Chris Sander, a biology bigwig from Harvard and co-author of the study, said this tool could sharpen up the decisions doctors make. Not only that, it could add a few extra innings to folks’ lives and give treatments a better shot at working.