Databricks’ new open-source AI model could offer enterprises a leaner alternative to OpenAI’s GPT-3.5 

Welcome to AI Decoded, Fast Company’s weekly newsletter that breaks down the most important news in the world of AI. Sign up to receive this newsletter every week via email here.

Databricks unveils “mixture of experts” AI model

The San Francisco-based company Databricks announced on Wednesday a new natural language model called DBRX, which it says performs better than a number of popular and comparably sized LLMs, including Meta’s Llama 2-70B and OpenAI’s GPT-3.5.

Databricks rose to prominence for its ability to host, protect, and enable analytics on the proprietary data of large enterprises and organizations in its cloud. With DBRX, it can offer its some 12,000 customers a secure cloud where they can also expose their data to advanced AI models, which it argues helps avoid the security risks of sending data to an outside foundation model.

DBRX, which is available as open source, isn’t as capable as such state-of-the-art models as Google’s Gemini or OpenAI’s GPT-4 but, as Databricks CEO Ali Ghodsi pointed out at a press event this week, many enterprises don’t require gigantic models for the kinds of applications they’re looking to carry out. A financial institution, for example, he said  might be able to use DBRX to look for signs of fraud among its databases of numbers, and a healthcare organization could use the AI to look for patterns of disease across thousands of electronic patient records.

Many of today’s LLMs  expend too much energy to tackle simple problems, which both uses up compute power and slows delivery of an answer to a user. DBRX addresses this issue by using a “mixture-of-experts” design that divides up the model’s brain into 16 specialized “experts.” When a specific type of calculation is requested, a “router” inside the model knows which “expert” to call on. The whole DBRX model contains 132 billion parameters, but because of that division of labor, it uses only 36 billion parameters at any given time, Ghodsi explained. For businesses that want to use AI for day-to-day operations, this style of LLM architecture could lower the barrier to entry.

When is “personal AI” too personal?  

What happens when personal AI gets so personal that it involves a significant other? Of course, AI girlfriends are a thing, but the girlfriends typically aren’t based on real people. The Talk to Your Ex app, which was featured on Product Hunt last summer but apparently still hasn’t launched, lets the user upload past text messages with an ex-partner in order to simulate conversation long after the glow of love has faded. (It’s possible to do something similar using a GPT if you have a ChatGPT Plus account.) But does this violate the ex’s rights? And is it good for the user to cling to a past partner’s words instead of moving on? 

Future personal AIs will offer to read our emails, listen to our phone calls, access our health data, and track our emotional state. They might tap sensors or cameras to catch signals from the environment we move in. Those same sensors and cameras might capture and analyze the images of others. AI companies will want to offer this level of personalization because consumers will want it. Like most tech things, such personal agents will have beneficial and healthy applications (an elderly person reminiscing with a bot when no one else is around, perhaps), but likely some completely unforeseen consequences too. Consumers need to get used to this situation; ultimately it’ll be on them to put clear boundaries on what their personal AI can know.

ChatGPT is moving quickly toward the mainstream

ChatGPT is rapidly carving out a place for itself in Western culture. A new Pew survey, conducted in February, shows that the number of Americans who have used the AI tool is rising. Just 18% had used it last July; that number rose to 23% in February.  

Not surprisingly, the uptake is driven mainly by younger generations. Pew found that 43% of the under-30 crowd had used ChatGPT. That’s up 10 points since last July. Pew says that among people aged 30 to 64, usage of ChatGPT is up only “slightly.” 

More Americans are also using ChatGPT at work. The share of adults who have used ChatGPT for work has risen by double digits in the past year, from 8% in March 2023, to 20% in February 2024. Pew found that most people wanted to use the chatbot either for educational or entertainment purposes (17% in both cases).

Not that people trust what comes out of the chatbot, especially when it comes to important stuff such as election information: 38% of the people Pew spoke with said “they have not too much trust or no trust at all” for AI-generated information about the 2024 election. That may not bode well for OpenAI, but it may be a good sign for democracy: It strongly suggests that Americans have gotten the message they’re at high risk of exposure to AI-generated disinformation in this election year.

More AI coverage from Fast Company: 

• Microsoft’s Inflection AI grab likely cost more than $1 billion, says an insider
• 4 ways to make AI a force multiplier for workers with disabilities
• Grammarly’s AI can now offer suggestions to make your work messages clearer
• One year in, Khan Academy’s AI has 65,000 students, and is still learning new skills

Want exclusive reporting and trend analysis on technology, business innovation, future of work, and design? Sign up for Fast Company Premium.

https://www.fastcompany.com/91070566/databricks-ai-model-enterprise?partner=rss&utm_source=rss&utm_medium=feed&utm_campaign=rss+fastcompany&utm_content=rss