TypeSafe AI releases Jev, a 'System One' decision model that returns typed probabilities instead of text

Description by Gemini, which watched the video:

Summary
Claire Vo, host of How I AI, introduces and demonstrates Jev, a fast, low-cost "System 1" decision model developed by TypeSafe AI. She contrasts its structured, type-safe output paradigm with standard generative LLMs and demonstrates how she integrates Jev into multi-model workflows, local developer data analysis, product intelligence, and real-time interactive apps.

---

What is shown

• [01:42] Sponsor segment: Overview of OpenArt Arena, showcasing creative model rankings across video and image generation tasks.
• [02:50] Architecture & documentation walk-through: TypeSafe AI documentation comparing standard LLMs with System 1 models, detailing Jev's primitives: Choice [05:42], Score [06:16], and Noul (calibrated probability/Boolean) [06:38].
• [07:48] GitHub PR analysis in Codex: Using Jev for pairwise comparisons and Gemini 3.5 Flash-Lite for theme labeling across pull requests:
  • First run: 112 PRs (6,216 pairwise comparisons) clustered into 39 groups across 6 themes for $0.011 [07:48].
  • Second run: 1,745 PRs (approx. 17,000 pairwise evaluations) analyzed in under two minutes for $0.09 [09:28].
• [11:20] Local session log analytics: Meta-analysis running across local Claude Code and Codex session logs from January to September 2026, plotting shifts in engineering versus agent-directed work [11:37].
• [15:16] ChatPRD architecture overview: Multi-model pipeline diagram pairing Jev for high-throughput classification and clustering with Astra and Sol/Luna for deeper reasoning and text synthesis.
• [19:10] Comment Lab dashboard & live search: Analysis of 4,483 audience comments, categorized into sentiment tones, 58 episode ideas, and 465 quality praise tags [20:11], followed by live search filtering queries like "comments about screenshare" [21:20] and "slop" [21:29].
• [22:52] Real-time voice-to-quote app: A live browser application pairing OpenAI's Realtime voice API with Jev to detect emotional sentiment, dynamically change background hex colors, and query matching quotes as Vo speaks [23:20–24:10].

---

Claims & numbers

• Vo states that models released in the preceding five days include Opus 5.5, GPT-6 Sol, and GPT-6 Luna [00:14].
• Jev is described as an unstructured-text-input, type-safe output decision model with response latencies between 70 ms and 500 ms [02:50].
• Vo notes that Jev costs $0.042 per million input tokens (or $42 per billion tokens), while output tokens are free because outputs are structured classifications rather than generated strings [02:50, 04:12].
• Vo claims running Jev on 1,745 PRs with roughly 17,000 pairwise comparisons cost 9 cents and completed in approximately two minutes [09:40].
• Vo notes her local developer activity shifted from nearly 100% manual product engineering in January 2026 to under 40% in September 2026, with agentic and tooling workflows expanding [12:04].
• Vo states her ChatPRD product intelligence pipeline ingested 1,100 raw signals, ran over 200,000 classifications and pairwise groupings via Jev, and cost approximately $4 in Jev compute [17:42].

---

Notable quotes

• [03:20]: "With Jev, you are getting text in, type-safe values out."
• [04:12]: "It is four cents per million input tokens. It is like dirt freaking cheap."
• [13:26]: "Jev alone is okay. Jev with an LLM buddy is super powerful."

---

Assessment
A hands-on technical review and practical demonstration by a creator/founder. The showcased workflows in Codex, ChatPRD, and custom web applications reflect working developer implementations, with live performance, cost breakdowns, and API response latencies shown directly on screen.

Described by gemini-3.8-flash on 2026-09-29 from the video's audio and frames.

Description by Gemini, which watched the video:

Summary
Mark Kashef demonstrates how to build a local, open-source multimodal classifier pipeline inspired by Jev using Claude Opus 5.5 and open-source models. He details an end-to-end workflow to fine-tune an encoder model (such as ModernBERT) to evaluate travel terms, verify photo evidence, and match client requirements locally.

What is shown

• [00:00 - 00:35] Demo of "Away Together," a travel agency app matching 12 customer profiles against hotel packages and cancellation terms.
• [01:02 - 02:08] Breakdown of classification queries (cancellation refund, late arrival, pool access, wheelchair accessibility) and the 4-step framework.
• [02:52 - 04:15] Whiteboard explanation of encoder-only vs. decoder-only architectures and context priming.
• [04:16 - 04:48] Open-source model alternatives shown on Hugging Face and GitHub, including ModernBERT-base-zeroshot-v2.0 and Diffusion Gemma.
• [05:34 - 07:04] Prompts and instructions provided to Claude to configure local training, evaluation benchmarks, and image recognition.
• [07:05 - 09:54] The 8-part prompt structure (Job, Computer, Data, Baseline, Training, Final test, App + Images, Delivery) for Claude.
• [09:55 - 10:55] Visual diagram explaining overfitting risk and separating test/validation sets.
• [11:04 - 11:49] JSON data format structure with classification criteria (meets, violates, insufficient_evidence).
• [12:08 - 12:43] Accuracy comparison charts: first model (60.28%), V2 model (95.28%), and closed Jev model (98.61%).
• [12:44 - 13:26] Image verification flow overriding text classification (e.g., detecting steps or identifying a pond instead of a pool).
• [13:27 - 14:22] Querying SuperGrok to locate recent open-source Jev derivatives on GitHub and generating an automated training system prompt for Claude Opus 5.5.

Claims & numbers

• The presenter claims the system runs entirely locally on consumer hardware for free without ongoing API token costs.
• Training on a local computer without a dedicated GPU takes between 3 to 6 hours per retraining cycle, according to the presenter [07:38].
• Benchmark figures shown: the initial travel model scored 60.28% accuracy, the V2 fine-tuned model achieved 95.28%, compared to Jev's 98.61% on 360 test scenarios (1,440 text decisions) [12:08].
• Another test graphic displays a baseline accuracy improvement from 74.75% before travel training to 93.63% after training across 500 decisions [04:49].

Notable quotes

• "So I took the idea behind Jev and made a version that runs entirely on my computer, completely for free." [00:00]
• "Jev is what's called pretty much a classifier model, specifically it's called an encoder-only model." [02:58]
• "So I wasn't able to quite beat Jev, but I got close enough on a local model running on this computer..." [12:33]

Assessment
This is a technical tutorial and hands-on workflow demonstration. While the web interface, architecture concepts, and prompt engineering methods are shown clearly, long training runs and complete model code execution are abbreviated for presentation purposes.

Described by gemini-3.8-flash on 2026-09-29 from the video's audio and frames.