(Jonathan-Karwacki/Shutterstock)
Databricks today launched a bricks agent, a new offer aimed at helping AI Agent Systems customers quickly in operation and starting costs, security and efficiency.
Many companies invest large sources in AI agent. While the potential payout of automation is large, they find that the actual process of building and deploying agents is so difficult.
“You can give everyone the tools they need to build agents,” says databricks vice president Joel Minnick. “But the real process of getting agents into production could be much easier for many customers.”
There is no problem with the AI models themselves. Rather, ISSU is with the quality and cost of AI models will meet expectations, says Minnick.
Databricks saw three main problems. The first was the lack of a good way to evaluate AI applications on the type of workload in the real world that customers want to run.
“Model providers boast the latest and largest version of this model can answer questions about the Mathematical Olympics, or it’s really great in vibration coding,” tents Minnick Bigdatawire. “But it doesn’t reflect the problem of real life.”
For example, if the company has tried to create an agent product recommendation, the agent could recommend a product that does not exist, refuse to confirm the real product it reserves, or even recommends that the customer buy a competitive product. “So these are the types of real evaluation problems in which customers are located,” says Minnick.
Another is the availability of data. While companies may have quite large amants of data, they could miss enough data to train on the agent to perform a specific task. It can also take a large number of chips to practice models on their data and the cost of these tokens can be added.
Finally, Databricks saw that optimization was an unsenoid. The solution for quality and efficient and efficient requires barnsing of competing requirements, and this may take some degree of effect and sophistication, partly when the basic models change from providers per month.
“The task of Hercules is to keep up with the latest and greatest research of optimizing technical technical technology,” says Minnick. “It’s hard, even for society like databricks. So for an average business to keep up, he often only becomes just what he feels fine, but I often feel like I don’t think I have found the best script so far.”
Agencies are attempting to solve these challenges of AI-resistance, data availability and optimization-by the context of specific cases of use, including structured extraction of information, reliable knowledge, transformation of your text and organized systems with multiple agents.
The offer uses research techniques developed by Mosaic AI Research to create an AI rating system to help users adapt their models and agents to the task and data. Agent bricks will automatically generate a set of judges that will be used to test and evaluate AI of AI customers.
While the product can control the process evaluation, customers can have full control over specific configurations and criteria, MINNICK says. “You have full control over the adjustment, adding to them, taking things, and make sure they feel exactly what you want to evaluate these Judes against the agent,” he says.
If the customer does not have enough data for the agent training, agents can generate synthetic data to be for training. On the front of the optimization, the software can attribute different techniques to find the right balance between quality and cost.
“We give you scorecards … (it) says I used the Lambda furnace more techniques, I got 95% quality in the various checks we agreed on. And I did it for this price.
Databricks CEO Ali Ghodsi delivers a key comment on data + AI Summit 2024
But 3x lower costs of operating this model, “says Minnick.” So we give customers a lot of choice, for this case of use, where I want to fall on this quality versus cost curve and be able to be a ceiling to get. ”
Databricks is not new to evaluate the machine learning model. The company supports classic working flows ML surrounding things like clustering and classification. What is different in the world of generative AI and agency AI is that the processing process is much more organic and requires a more flexible approach, says Minnick.
“As you evaluate them, it was more improved to understand what quality it has,” he says. “That’s why these judges are so important to understand, well, what exactly I think this agent will have to encourage the real world?… What do I think a good look?
Agent bricks are now in beta beta. One of the early tests was Astrazneneca, which lasted about 60 minutes for the knowledge agents to be built, which is able to extract news from 400,000 research documents, says Minnick.
“For the first time, businesses can move from Idea to production AI to their own speed and confidence, control of compromise with quality and cost,” said Ali Ghsti, CEO and co -founder of Databricks. “No manual tuning, no guessing and all databricks security and management has been offered.
Databricks made a notification on its Summit Data AI +, which is too much placed in San Francisco this week.
Related items:
Databricks NABS Neon to solve the narrow section of the AI database
Databricks to increase 5, $ for 55 B valuations: Message
It’s Snowflake vs. Databricks in a duel of large data conferences
(Tagstotranslate) agent brick