This $40M Startup Was Replaced After Just 3 Days!
Read the Portuguese translation →
Author: Better Stack · Watch on YouTube
Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.
Original language: English
Original video: This $40M Startup Was Replaced After Just 3 Days! · Channel: Better Stack · 01:34 · Watch on YouTube
Topics: AI agents, AI evaluation and benchmarks, AI models, Open source and software ecosystems
Summary
Section titled “Summary”Type-safe AI’s Jev model was introduced for fast, low-cost classification and structured decisions in software. Three days later, a free open-source alternative called Llama was released, with better performance on the benchmark described in the video. Both models return structured probabilities, but that structure does not guarantee correct answers. The video says Llama achieved 0.766 accuracy versus Jev’s 0.727 and highlights the potential for inexpensive, fast AI decisions.
Key points
Section titled “Key points”- Jev was designed for classification and strict decisions inside software.
- The video says Jev runs up to 200 times faster and 400 times cheaper than using a traditional LLM for the same work.
- A free open-source alternative called Llama appeared three days after Jev’s release.
- Both models return a probability in a defined structure, such as whether an email is spam.
- Structured output does not ensure that a model’s answer is correct.
- Jev scored 0.727 accuracy on a benchmark of typed decisions.
- Llama scored 0.766 on the same benchmark.
- The video suggests these models can enable fast, low-cost decisions and processing large datasets.
Chapters
Section titled “Chapters”| Start | Chapter | Summary |
|---|---|---|
| 00:00 | Jev and its open-source alternative | The video introduces Jev’s classification use case and compares it with the free Llama alternative released three days later. |
| 00:30 | Structured output is not accuracy | An email spam example explains how the models return probabilities while still potentially making incorrect decisions. |
| 01:03 | Benchmark results and use cases | Llama outperforms Jev on the cited benchmark, and the video points to its potential for fast, inexpensive AI decisions. |
Quotes
Section titled “Quotes”This 40-million AI startup just got cloned in 3 days. — 00:00
And then it returns a probability that the email is spam or not. — 00:30
But that doesn’t mean that the answer is correct. — 00:30
Llama, however, scraped past and beat Jev on this — 01:03
the cost and speed is so impressive. — 01:34
Concepts and entities
Section titled “Concepts and entities”TypeSafe AI, Jev, Llama, Hugging Face, Better Stack, decisões tipadas
Full transcript (original language)
Section titled “Full transcript (original language)”00:00 This 40-million AI startup just got cloned in 3 days. Type-safe AI went viral this week after releasing Jev, a brand new model designed for classification and the strict decisions inside your software. It runs up to 200 times faster and 400 times cheaper compared to doing the same work with a traditional LLM. But even this wasn’t enough because only 3 days later, a free open-source alternative called Llama was released and it has even better performance. Both Jev and Llama work in the same way and they’re effectively
00:30 if statement. You just give it a question and it returns a probability. For example, I could just ask the question, is this email spam? And then I give it the body of the email and then it returns a probability that the email is spam or not. And then you could use that information to choose to delete the email. But the problem isn’t with the structured output. Both Jev and Llama are guaranteed to provide the response in the structure that you define. But that doesn’t mean that the answer is correct. It could confidently mark an email as spam that wasn’t and vice versa. And in fact, Jev is wrong about 1/4 of the time with an accurate with an
01:03 accuracy of 0.727 on a benchmark of typed decisions. Llama, however, scraped past and beat Jev on this, coming in at 0.766 accuracy on the exact same benchmark. And remember, this tool is completely free. So it seems in 2026, the lifespan of software has shrunk to just a few days. You can check out Llama over on Hugging Face. It’s fantastic for processing large data sets or making cheap, quick decisions inside your software. This unlocks an entire new class of use cases for AI because the
01:34 cost and speed is so impressive. And if you want to stay up-to-date with tech and AI news, subscribe to Better Stack.