J'ai testé Jev avec Astra : Résultat bluffant !

Item: J'ai testé Jev avec Astra : Résultat bluffant !
Rating: 3
Author: PostSingularity Institute

I tested Jev with Astra: Mind-blowing results!

by Parlons IA · View the channel on PSI

Formal & Physical Sciences AI — News & Applications UYComputer science UYQArtificial intelligence (AI)

🎙 Parlons IA 👥 17K 📅 September 22, 2026 ⏱ 21 min 👁 165 📄 tutorial 🧭 2026-09-22

Available in: English (current) Français

Keywords

JevTypeSafe AIopen-source modeltoken cost optimizationprobability scoring

Summary

The video explores Jev, a model created by TypeSafe AI, and argues that it should not be seen as a direct competitor to general-purpose LLMs like ChatGPT Astra or Claude, but rather as an extremely fast and cheap probabilistic classifier. The creator explains Jev’s key features: responses in about 70 milliseconds, costs around $0.042 per million tokens, and the use of three primitives — Choice, Score, and Noul — for structured decision-making. Jev provides probability scores with its answers, giving users insight into model confidence and enabling better control over automated tasks. The video criticizes traditional conversational models trained with RLHF for always producing plausible answers even when unsure. A concrete demonstration follows: the creator builds a local system integrated into ChatGPT Codex using a lightweight open-source model (Ornith 1.0-9B) to route queries between local and cloud models. The interface shows token probabilities, confidence levels, and enables code execution and local file operations. The goal is to reduce token consumption and keep private data local while maintaining access to more powerful models when needed. The creator claims the architecture can be replicated in tools like Claude Code and argues that such intermediary routing layers will transform how AI tools are optimized.

201 words

Critical Evaluation

Value of the Information & Strength of the Argument

The video offers a valuable conceptual reframing of Jev as a routing agent rather than a replacement for large LLMs, which is a relevant distinction in current AI discussions. The demonstration of integrating a local fast model into ChatGPT Codex is practical and could be useful for developers looking to reduce token costs. The argument about exposing probability scores and confidence levels is solid in principle and addresses a real limitation of conversational models. However, the argumentation is mostly qualitative; quantitative claims (100x cheaper, 100-150x faster) appear without rigorous benchmarking or detailed comparisons. The logic is coherent but relies heavily on anecdotal observations and the creator’s own integration, rather than systematic experiments. The lack of citations weakens the overall value for a scientific audience.

Scientific Rigor, Source Quality, Title Accuracy

The video demonstrates limited scientific rigor: it does not provide verifiable sources for the claims about Jev’s architecture, performance, or costs. The description contains only promotional and affiliate links (e.g., a training site, a Medium blog, and tool recommendation links), not peer-reviewed papers or official documentation. The creator mentions concepts like RLHF, JSON, and probability scoring correctly but without deep formal explanation. The title’s promise of a ‘stunning result’ is somewhat exaggerated, though the content does show a functional integration. Overall, the video is more of a practical tutorial and opinion piece than a rigorous scientific exposition, and its sourcing is insufficient for high reliability.

244 words

Title / Content Match

The title accurately reflects the content: the creator demonstrates integrating Jev-like fast inference into ChatGPT Astra/Codex and highlights the surprising results. Minor mismatch: 'Jev' itself is not directly integrated, but a similar open-source-inspired mechanism is shown.

Quality & Reliability

5/10

The video presents a practical demonstration of integrating a fast local inference model with ChatGPT Codex, along with conceptual explanations of Jev's architecture. However, it relies on unsupported claims (e.g., cost savings up to 100x) and does not provide verifiable data or external sources, which limits overall reliability.

Key Moments

Markers derived by PSI from the transcript: the creator did not define chapters.

Cited Sources

Parlons IA — Dailymotion channel — Mentioned as an alternative platform for the creator's content.
Medium blog — flma1349 — Blog associated with the channel, likely for additional articles.
Parlons IA training site — Promoted as the creator's main training offer for AI and business.
Podcast link (Spotify) — Link to the creator's podcast, possibly containing related conversations.
SEO Agent AI (affiliate link) — Presents an AI tool, likely as a paid partnership or affiliate product.

Contribution & Novelties

The video’s original contribution is its practical demonstration of embedding a Jev-like fast inference layer inside ChatGPT Codex, allowing automatic routing between a local model and a cloud LLM. It helps make the concept of probabilistic classification and confidence scoring concrete for everyday automation. However, as the creator acknowledges, the underlying architecture is not new and existed as open-source; the novelty lies in the packaging and demonstration.

Pour aller plus loin :

Reinforcement learning from human feedback — Explains the RLHF background and limitations mentioned in the video.
Large language model — Provides foundational context on LLMs, including inference and cost considerations.
Human-in-the-loop — Relevant to the HITL use cases discussed by the creator.
Mixture of experts — A related routing technique that aligns with the model-choice optimization presented.

129 words

Radar Profile

The radar profile shows moderate scores across the board, with technical level slightly higher than reliability. This suggests the video has some practical and instructional value but lacks the sourcing and rigor expected for a fully trustworthy technical analysis.

Reliability 4/10