Медиа про AI для креаторов, предпринимателей, дизайнеров и творческих единиц.
Исследуем, что отдавать людям, а что машинам. Обсуждаем с экспертами. Тренды и классные кейсы.
By Анфиса Савченко и Артем Семьянов
Post #1034
30
Forwarded from Varga Startup [Live Updates]
Today, ex-openAI Diogo Almeida introduced Jev — a new class of ML models.
He claims it’s the jump from LLMs, same as LLMs did to RNNs (pre 2017)
The idea is to predict next category probability, instead of predicting next token.
Example:
Question: “I paid twice, which email I should use?”
Variants: billing / technical / returns.
Result:
billing → 0.97
technical → 0.01
returns → 0.02
Why it is fast?
All probabilities can be calculated in parallel, it is much faster then sequential token-by-token in LLMs.
How fast? ~75x faster response time VS GPT-5.6 Terra
How cheap? ~170x cheaper vs GPT-5.6 Terra
You can watch a 4 min video, it’s wow!
What are my thoughts about it:
It’s impressive: for gaming, robotics, data scrapping, any environment where number of posible actions is limited that is huge: for example a real-intelligence video game will be only $7/hour in tokens (they messured it on Doom)
What is a TL;DR about it? I would make a joke and say it’s a Categorized Markov Chain trained on large corpus beating best LLMs on cost and latency for structured decision tasks
Really-really impressive idea and implementation!
He claims it’s the jump from LLMs, same as LLMs did to RNNs (pre 2017)
The idea is to predict next category probability, instead of predicting next token.
Example:
Question: “I paid twice, which email I should use?”
Variants: billing / technical / returns.
Result:
billing → 0.97
technical → 0.01
returns → 0.02
Why it is fast?
All probabilities can be calculated in parallel, it is much faster then sequential token-by-token in LLMs.
How fast? ~75x faster response time VS GPT-5.6 Terra
How cheap? ~170x cheaper vs GPT-5.6 Terra
You can watch a 4 min video, it’s wow!
What are my thoughts about it:
It’s impressive: for gaming, robotics, data scrapping, any environment where number of posible actions is limited that is huge: for example a real-intelligence video game will be only $7/hour in tokens (they messured it on Doom)
What is a TL;DR about it? I would make a joke and say it’s a Categorized Markov Chain trained on large corpus beating best LLMs on cost and latency for structured decision tasks
Really-really impressive idea and implementation!
