An ex-OpenAI researcher just deleted language from the LLM...

Tech concept / product being discussed: Jev (a “system one model”)

• Jev is introduced as a new frontier AI model from Diogo Almeida (ex-OpenAI) and his company TypeSafe AI.
• The central claim: Jev “deleted language” from a large language model, turning it into a fast, type-safe classifier instead of a chatty text generator.
• Jev is positioned as a cheaper/faster option for applications that need quick decisions—a “gut instinct”-style model.

Core feature: Type-safe, schema-constrained outputs

Unlike normal LLMs that can output arbitrary text, Jev requires the prompt to specify a strongly typed return shape.

Allowed output shapes

Jev’s outputs are constrained to one of three shapes:

1. Choice
2. Score
3. Null
   • Described as a yes/no style response.

Key emphasis

• The video highlights guaranteed schema matching
• Claims that type errors are mathematically impossible

“System 1” vs “System 2” framing

• Jev is called a “system one model”: fast and instinctive.
• It’s contrasted with slower “system two” reasoning models, described as burning huge token budgets for deliberation.
• The goal is to be practical for app developers.

Performance / cost / hallucination claims

The video claims Jev offers:

• ~200x faster
• ~400x cheaper
• Free output tokens
• Zero hallucinations (as stated in the subtitles)

It’s also claimed to be suitable for real-time usage.

Developer use cases highlighted

AI moderation / safety-like gating example

• Example prompt: “Is this a horse?”
• If Jev returns null for invalid accounts, the app auto-instaban users that don’t match the expected type (e.g., “horse” accounts only).

Game NPC behavior

• Developers use Jev to drive real-time agent/NPC decisions.

Real-time AI calculator

• A developer claims to have built a “world’s first real-time AI calculator” using Jev.

Confidence calibration (how quality is represented)

• Jev is described as returning a calibrated confidence number (e.g., “60%”).
• Training technique mentioned: RLCD (reinforcement learning for calibrated decisions)
  • Aim: confidence values correspond to actual correctness frequency.

How Jev works (and uncertainty)

• The architecture is described as kept private
• The CEO is quoted/characterized as keeping details “close to the chest,” with a potential paper later.

Criticism / alternative claims

Potential similarity to older approaches

• Doubters argue Jev may be similar to older zero-shot classification techniques.
• The video mentions disputes including:
  • The company giving no credit to earlier pioneers of zero-shot classification
  • A claim that prior work (from about a year earlier) is “the exact same thing” as Jev

“OpenJeb” reproduction

A developer reportedly built OpenJeb with the following characteristics:

• Reproduces the interface by reading option probabilities from a frozen Qwen-4B
• Claimed to run in a single forward pass
• Reportedly requires no new training
• Reported to run on a 3090
• Includes a web GPU demo runnable in-browser

Tutorial/review style structure of the video (what the host plans to cover)

The video explicitly states it will:

1. Take a look at Jev’s code
2. Discuss “trust-me-bro benchmarks”
3. Cover a person who claims they built an open-source Jev over a year ago

Sponsor mention (unrelated to Jev)

Mux is promoted for video infrastructure, including:

• Transcripts, thumbnails, storyboards, clips
• “Mux robots” for AI workflows (e.g., translation, moderation)
• Pay-per-job pricing and a free plan

Main speakers / sources (as named in subtitles)

• Jev (speaker/identity): “My name is Jev.”
• Diogo Almeida (ex-OpenAI researcher; creator, described as associated with TypeSafe AI)
• Daniel Kahneman (referenced for “Thinking Fast and Slow” system 1 vs system 2)
• Jin Yang (referenced as an earlier zero-shot classifier pioneer)
• Qwen-4B (referenced in the OpenJeb reproduction approach)
• The host: the narrator (unnamed in the subtitles) who introduces The Code Report
• OpenJeb builder (“this guy” / developer mentioned as reproducing Jev)