Specializes in judgment and classification instead of long texts... Attention among developers
Quick Look
- TypeSafe AI, founded by an Open AI graduate, introduced the judgment model Zev.
- Zev is a specialized AI that only presents yes/no or choices within 0.1 seconds instead of generating long text.
- It is attracting attention among developers due to its ultra-low cost of about 56 won per million input tokens and no formatting errors.
AI-generated summary
Why It Matters
Existing large-scale language models have strengths in generating long texts, but it has been raised in the development field that they are inefficient in terms of cost and speed for real-time judgment and classification tasks.
Specializes in judgment and classification instead of long texts... Attention among developers
1 million tokens at an ultra-low price of 56 won... Aiming at corporate SW ‘traffic control’
(Seoul = Yonhap News) Reporter Oh Ji-eun = Artificial intelligence (AI) that only presents 'yes/no' or options in 0.1 seconds, instead of writing long articles fluently like ChatGPT, is causing a sensation among developers.
According to the information and communication technology (ICT) industry on the 5th, TypeSafe AI, founded by Diogo Almeida and others from Open AI, attracted investments worth $40 million (about KRW 54 billion) and introduced the judgment model Jev.
Zev is not a large language model (LLM) for writing articles or writing long lines of code. I can't explain the context at length, and I can't do complex arithmetic operations or date calculations.
Instead, Zev looks at the given data and professionally only makes judgments and classifications based on clear criteria, such as “Is this post a malicious comment?”, “Where should I send this customer’s question to the accounting team or technical support team?”, and “To what extent does it comply with regulations?”
Unlike existing LLMs that take several seconds to complete a sentence line by line, Zev completes his judgment in about 0.1 seconds as soon as he is asked a question.
The cost is also $0.042 (about 56 won) per million input tokens, which is tens to hundreds of times cheaper than existing large models, and there is no output cost for the results at all.
The part that receives the greatest response from developers is that 'formatting errors' do not occur structurally.
In actual development sites, AI judgments are needed thousands of times per second for customer service classification or at each crossroads of the automation system, but models such as existing ChatGPT often get in trouble by giving incorrect answers outside of the established format or sending errors that the computer cannot read.
On the other hand, Zev was designed to prevent the creation of text other than the initially specified options or values between 0 and 1, dramatically increasing system stability.
The industry believes that AI is evolving beyond the stage of writing and imagining into component-type AI that makes quick and accurate decisions within actual company software.
An industry insider analyzed, "Even if complex report writing is left to smart models like ChatGPT or Claude, the division of labor has begun to leave simple traffic control, which occurs thousands of times per second, to light and fast models like Zev."
What to Watch
AI outlook — possibilities, not facts
Companies will increasingly adopt AI models specialized for judgment and classification tasks.
Likely · Within months
Open Questions
- What is the level of accuracy of the Zev model compared to the existing LLM?
- How will TypeSafe AI’s future business model be structured?
- What technical support will companies need to integrate Zev into their existing systems?
