Stephen Council/BI
- An ex-OpenAI researcher built Jev, a popular new AI model that takes a different approach than LLMs.
- The model is made for code, makes decisions in milliseconds, and includes probabilities in answers.
- Developers at a Saturday hackathon told Business Insider what they're building with the new tool.
When a new twist on AI called Jev arrived this month, San Francisco's tech scene responded as it does: lots of hype, and then a hackathon.
Jev is a new and viral AI model built by the startup TypeSafe AI, and developers are flocking to try it out. Jev's launch video from mid-September has racked up almost 40 million views on X, throwing founder Diogo Almeida into the spotlight.
Notably, Jev marks a break from large language models (LLMs) at OpenAI and Anthropic. Rather than spitting out text, Jev makes decisions. It's also dirt cheap and incredibly fast. An ex-OpenAI researcher, Almeida has said that the focus on chatbots doesn't do enough to automate rote work.
I attended a Jev hackathon on Saturday at the startup CodeRabbit's office to find out what the hype is all about. The event became something of a coming-out party for Jev. Some developers say they have already woven Jev into their products and feel that TypeSafe has delivered a true breakthrough. Sourabh Mane, a design engineer at CodeRabbit, called it a "breath of fresh air."
Venture capitalists are buzzing around Jev as well. The Information reports that some VCs have offered to invest in TypeSafe at a $10 billion valuation.
How Jev differs from chatbots like ChatGPT
Jev takes in questions and rapidly responds with structured answers that include probability estimates of accuracy. It works best when engineers break down their questions into the most specific possible pieces, Allie Laabs, TypeSafe's head of developer relations, told the crowd at the start of the hackathon.
Laabs told me that, with Jev, her startup hopes to advance a new school of thought in an industry dominated by LLMs.
"What if this path that we've been so horse-blinders on, the LLM path, what if that's not the only path?" she said.
Laabs gave the developers some other pointers, too. Jev works well for classifying — it could mow through customer support queries and slot them into buckets, or moderate a fast-moving online chat feed. She hopes that Jev's speed and low cost will help some developers build tools they'd previously thought were too ambitious.
Developers sprint to inject Jev into their code
Fueled on croissants and coffee, the developers were thrilled to get going.
Cofounder Jalaja Kurubarahalli said that if Jev can sort questions more quickly, it could reduce the lag in the voice tools that her startup, Quintess AI, builds for maintenance workers. Hitesh Viswasam, a robotics engineer, suggested that a model like Jev could let robots respond to heat or danger at human-like speeds.
Some attendees had already improved their code with Jev. Developer Lawrence Liang showed me the personal AI agent he'd built for himself. He said this was his second Jev hackathon of the week and that he'd built the model into 80% of his agent's code infrastructure. Jev figures out much more quickly than an LLM which tools are needed for any given query.
"It almost blows your mind," Liang said.
Engineer Avram Cheaney said that adding Jev to his AI search product for eBay, called JunkDrawer.ai, had made it "drastically cheaper and drastically faster" than using just LLMs. Those few seconds of speedup are particularly important for consumer products, where people have become used to instantaneous loading.
"This is what's been missing," Cheaney said.
TypeSafe peddles a new way to think about AI
It's been a hectic week at TypeSafe. Jev's launch catapulted the company from anonymity to a closely watched AI player. Its Discord server now has more than 100,000 members.
"It is definitely completely nuts," Laabs said, adding that the early success is something "very few people ever get to experience in their lives."
Laabs said she's been delighted to see that people have quickly begun to understand what Jev can do.
Before, she was worried because Jev doesn't align precisely with how AI is widely understood, where a text box spits out code or detailed answers to questions.
"A software engineer has to put it in something; it isn't anything until that," Laabs said. She's found that the explanation helps: "Then people go, 'Ohhh!'"
Have a tip? Contact this reporter via email at scouncil@businessinsider.com, or over text, Signal, Telegram, or WhatsApp at 415-757-8198. Use a personal email address, a nonwork WiFi network, and a nonwork device; here's our guide to sharing information securely.
Read the original article on Business Insider