CrewAI AI industry newsletter

View in browser Hi there, Last month, TypeSafe AI released a model called Jev. TypeSafe is a San Francisco lab founded by Diogo Almeida, who co-invented the training method behind ChatGPT. What's special about this model Jev is what it can't do. It can't chat. It can't write, code, or draw. You hand it a situation and a question. It hands back a decision. That's the whole product. It answers in three ways: Choice. Pick one from a fixed menu: order, return, or clarify. Score. How strongly does this fit, on a scale. Yes or no. Should we refund this? TypeSafe calls this one "Noul." Every answer comes with a number that says how sure it is. Developers moved fast. Vercel says Jev reached nearly 13% of paid teams on its AI gateway within 24 hours of launch. Splitting the Work Think of a good AI setup like a well-run office. It has three kinds of workers: The checklist: Some jobs are the same every time, like filing a form or updating a record. You don't need AI for these. A simple rule does it for free. The receptionist: Some jobs are quick calls. Is this a refund or a delivery question? Is this customer urgent? Are we sure enough to act? You need judgment here, but only a second's worth. The expert: A few jobs are genuinely hard, like researching a new client or writing a thoughtful reply. That's where the big, expensive AI earns its keep. The problem is, most of a company's daily work is checklist and receptionist work. Yet most companies send all of it to the expert. Since implementing jev.ai, an analysis of 12,759 launch posts found developers reporting a median 7x speedup, and about 30x in cost savings. Bryo AI's CTO found Gemini slightly more accurate at sorting emails, but 10 to 20 times pricier. TypeSafe's own docs warn that Jev is unreliable at counting, arithmetic and comparing dates. Keep the math in your code. And Jev never explains itself. If you need a reason on record, pair it with a chat model or a person. And that’s completely fine. A decider that tells you when it's unsure always beats the one that guesses with a straight face. The CTO question: How many of your AI's decisions are you paying a genius to make? What we're doing at Crew At CrewAI, we're testing Jev inside Conversational Flows, the part of our product that decides where each customer message should go. So far, it's helping in two ways: 1. On some tasks, those routing decisions are getting faster and more accurate. 2. And on voice agents, we're seeing them start talking sooner after a customer finishes speaking. Engineers call that "time to first speech." It's still early, but for anyone building an AI that talks to customers, those fractions of a second are the difference between a conversation and an awkward pause. That’s all from my end. Lorenze Hernandez Lead OSS Enginerer Follow me on LinkedIn or X This Week's Treats One great long-form read - Fix your life in 1 Day Two tweets worth checking out: Exploding Body Cow vs Gold prices Three songs you should add to your work playlist Cler Achel - Tinariwen White Gloves - Khruangbin Albatross - Fleetwood Mac Signal vs Slop Every week we plot what happened in AI on two axes. Slop ↔ Signal: was there actual substance, or was it noise? Now ↔ Later: is this something you should act on this quarter, or a signal about where things are headed in 1-3 years? Let's see what happened this week. NOW / SIGNAL 1. Humans have started to talk like chatbots The WSJ found people catching themselves using prompt-style phrasing, explicit goal framing and over-polite instructions in meetings, texts and everyday conversation. 2. Robinhood just handed 29 million customers an AI that trades for them Over 150,000 customers already run AI agents on their accounts, and those agents now use Robinhood's tools almost 30 million times a day. Customers carry the full risk of their agents' trades, and Robinhood says it hasn't measured how those trades perform. NOW / SLOP 3. 661 engi