Skip to content
JevDirectory.org
Evaluation & Training

Mode dropping

Preference optimization narrowing a model's output distribution toward a favored style and away from other valid outputs.

Open docs.typesafe.aiAll 117 terms
Category
Evaluation & Training
Also known as
—
Related terms
3
Directory entries
1
Docs
docs.typesafe.ai
Added
2026-09-24

Definition

The primer illustrates mode dropping as the distribution of a base model compared with a narrowed, mode-dropped distribution after RLHF. The model learns to favor a particular style, such as instruction following, while reducing the probability of other possible outputs.

It is a milder version of mode collapse, the GAN failure mode where a generator produces the same kind of output repeatedly. For decision models the concern is the opposite of diversity: probabilities that no longer reflect the real chance of being right.

Tagstraininglimitations

Related terms

Definitions that connect to this one.

RLHFThe human-preference training method that turned pretrained models into chatbots, and the method TypeSafe deliberately did not use.Evaluation & TrainingCalibrationThe property that outcomes given a probability of 0.2 occur about 20% of the time, measured across many predictions.Answers & ConfidenceSycophancyThe tendency of preference-trained models to say what people want to hear, one of the reasons TypeSafe chose a calibration objective.Evaluation & Training

From the directory

AI Primer: RLCD and Calibration
Why TypeSafe trains decision models with RLCD instead of RLHF: calibrated probabilities where 0.2 outcomes happen about 20% of the time, and the case for machine-to-machine automation.
Sites & GuidesDocs#official#docs#evaluation
Official

From the community

Posts from builders shipping with Jev right now.

Follow @typesafeai

Why Jev encodes a state and outputs distributions instead

Manthan Gupta
Manthan Gupta
@manthanguptaa
·

Jev is one of the more interesting model launches I have seen recently because it asks a very simple question: Why are we using autoregressive LLMs as insanely expensive if statements inside software? The easiest way to think about it is: LLM: text -> generate tokens Show more

Image
Diogo Almeida
Diogo Almeida
TypeSafe AI
@CompleteSkeptic

After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI? I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev • 20-200x faster • 40-400x

Watch on X
Reply
Read 40 replies

How the agent stack moved in Jev's first three days

Jev plays Doom at ~10 calls/sec for ~$7/hour

Jev plays Super Mario Bros.

LLMs generate answers, Jev makes decisions

Paarangat
Paarangat
@paarangatrai
·

this is the easiest way to understand Jev: LLMs generate answers. Jev makes decisions. that sounds like a small difference, but it actually changes the entire use case. say you give a normal LLM this: “here’s a user, their account history, payment behavior, support chats, Show more

Diogo Almeida
Diogo Almeida
TypeSafe AI
@CompleteSkeptic

After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI? I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev • 20-200x faster • 40-400x

Watch on X
Reply
Read 82 replies

Arbitrary classification as a primitive

cocktail peanut
cocktail peanut
@cocktailpeanut
·

Jev is cool not because it re-invented classification, but because it makes ARBITRARY classification into a type-safe programmable primitive. A general purpose zero shot decision model whose native interface is RUNTIME-DEFINED typed decisions, optimized for that exact interface

Image
cocktail peanut
cocktail peanut
@cocktailpeanut

If you called Yann LeCun an idiot for saying we need to move beyond LLMs and build something new, you are banned from using Jev.

Reply
Read 13 replies