CHAPTERS
- 0:00 – 1:31
Jev week kickoff: why this “decision model” beats frontier LLMs for daily work
Claire frames a week of rapid model launches (Opus 5.5, GPT-6 variants, etc.) and explains why Jev has become her most-used model despite the hype elsewhere. She positions Jev as a fast, cheap, high-agency decision engine that unlocks a surprising number of productivity and product workflows.
- •Jev stands out amid a flood of new model releases
- •Framed as a fast/cheap “System 1 decision model” from TypeSafe AI
- •Claim: explodes practical use cases (personal productivity, code, product)
- •Cost pitch: meaningful work done with sub-$10 in tokens
- •Preview: what Jev is, what it returns, and real use cases
- 1:31 – 2:32
Sponsor break: OpenArt Arena and evaluating models by task
A brief sponsor segment introduces OpenArt Arena as a creative-model leaderboard judged via blind professional evaluations. The emphasis is choosing the best model for specific creative jobs rather than a single overall winner.
- •OpenArt Arena ranks models across real creative tasks
- •Blind evaluations by professionals reduce hype bias
- •Covers both image and video creative workflows
- •Positioning: pick the right model per job to save time/cost
- 2:32 – 3:32
What makes Jev different: text-in, type-safe values-out
Claire explains Jev’s core distinction versus standard LLMs: outputs aren’t free-form text but predefined, type-safe values selected from a constrained set. She also notes Jev is text-only (no images), though image descriptions can be used.
- •Standard LLMs: text in → generated strings out
- •Jev: text in → predefined type-safe values out
- •No native image input; can use text descriptions of images
- •Constraining outputs enables reliable routing/decisions
- 3:32 – 5:34
Why Jev is so cheap and fast: pricing + real-time feasibility
Claire highlights Jev’s economics: it charges only for input tokens because outputs are minimal, making it dramatically cheaper than text-generating models. She argues this speed/cost profile makes Jev viable in real-time loops where other models feel too slow or expensive.
- •Jev costs ~$0.04 per million input tokens (no output charge)
- •Traditional LLMs often rack up expensive output tokens
- •Best fit: decisions, routing, classification at scale
- •Latency advantage enables real-time product loops
- 5:34 – 7:35
Jev primitives: Choice, Score, and Null (yes/no probability)
Claire breaks down Jev’s three fundamental output types and how to think about them in software terms. The simplicity (pick from options, score severity, estimate yes/no likelihood) is positioned as exactly what many production systems need.
- •Choice: select one item from a list of options
- •Score: rank/severity on a numeric or ordered scale
- •Null: boolean-like probability of “yes” for a question
- •Maps neatly to common engineering patterns: routing, triage, filtering
- 7:35 – 9:07
Use case 1: PR categorization via pairwise relatedness + clustering
Claire demos using Jev to analyze pull requests by comparing PR pairs for thematic relatedness, then clustering them. A second cheap model (Gemini Flash-Lite) labels the clusters, producing initiative-level breakdowns that leaders often need.
- •Pull thousands of PRs from GitHub and categorize them
- •Jev excels at pairwise “related/not related” decisions for clustering
- •Use a second model to label clusters after Jev groups them
- •Example: marketing site (~112 PRs) cost ~1.1 cents
- 9:07 – 11:08
Scaling the PR analysis: initiative reporting for a larger repo (pennies)
Claire scales the same approach to a larger product repo (ChatPRD), analyzing ~1,700 PRs and ~17,000 PR pairs. The output gives a credible distribution of engineering effort across initiatives like infra/security, reliability, integrations, editing, and prototyping—at a cost of only a few cents.
- •ChatPRD repo analysis: ~1,700 PRs and ~17,000 pairs
- •Total cost cited: ~9 cents and ~2 minutes runtime
- •Generates initiative breakdowns useful for exec/board reporting
- •Findings: significant investment in platform/security/infra and reliability
- 11:08 – 12:39
Use case 2: analyzing local Claude Code/Codex sessions for work patterns
Claire shows how to mine locally stored agent sessions to quantify what you spend time on over months. She compares distributions by turns and by session days, showing a shift from mostly engineering to more agent work and media/publishing.
- •Claude Code/Codex sessions are stored locally and analyzable
- •Categorize sessions to see effort mix over time
- •Observed trend: engineering-heavy early; later more agent + publishing work
- •Outputs can be grouped by turns or by session-days
- 12:39 – 14:11
Use case 3: Gmail triage—Jev scoring as a fast filter + ‘LLM buddy’ follow-up
Claire describes a private workflow where Jev scores emails (using subject/snippet) for deletability and applies clean tags. She emphasizes a pattern: use Jev to cluster/filter first, then send only the relevant subsets to a more capable generative model/agent for deeper action.
- •Score/categorize emails quickly from subject + snippet
- •Create delete/ignore vs. needs-attention groupings
- •Jev alone is helpful; Jev + a generative model is “super powerful”
- •General pattern: tag/cluster/filter → apply precise actions per cluster
- 14:11 – 16:14
Use case 4: ChatPRD product insights graph—where Jev fits in a multi-model pipeline
Claire explains how Jev helped “crack” a difficult product insights graph problem by separating classification/clustering from deeper reasoning and generation. Jev handles labeling and assignment; a stronger reasoning model (Astra) synthesizes insights; generative models (Sol/Luna) handle output where needed.
- •Prior approach with frontier models was too nuanced/expensive to brute-force
- •Jev inserted specifically for classification, clustering, mapping steps
- •Astra used for cross-cluster analysis/strategy/insight extraction
- •Sol/Luna used for generation tasks; Jev orchestrates efficient structure
- 16:14 – 18:48
Scale + economics of the insights graph: 1,100 signals and 200k+ pairwise/groupings
Claire quantifies the system: ~1,100 raw signals (PRs, support tickets, conversations, Linear tickets) and ~200,000 classification/pairwise operations. Jev’s portion costs about $4, with higher spend on the deeper analysis model, yielding a more margin-accretive product feature.
- •Data sources: PRs, support, conversations, Linear, and more
- •Goal: measure gaps between customer asks vs. team execution over time
- •~200k classification/pairwise groupings; Jev cost ~4 dollars
- •Result: richer trends, categorization, and ‘auto-wiki’ style context
- 18:48 – 22:21
Demo 1: How I AI audience signal dashboard—comment sentiment + episode ideas + live search
Claire demos a dashboard built by pulling ~4,500 YouTube comments via the YouTube v3 API and using Jev for sentiment and idea detection. She also shows fast live search across comments, noting batching/scoring architecture to keep latency low.
- •Fetch comments with YouTube v3 API and categorize sentiment
- •Detect comments containing episode ideas; summarize requests
- •Dashboard shows sentiment distribution and per-episode sentiment
- •Live search across thousands of comments using batching + scoring
- 22:21 – 25:02
Demo 2: real-time voice-to-color emotion mapper + quote selection (latency in practice)
Claire demonstrates a real-time app that transcribes voice, uses Jev to pick an emotion color, then selects a matching quote via an external Quotes API. She explains the mechanics: predefined hex color choices and quote-category filters, with observable latency largely coming from external calls.
- •Uses OpenAI Realtime Voice API for live input
- •Jev selects from predefined hex color values based on sentiment
- •Quotes API adds a bit of latency; pipeline shows real constraints
- •Architecture: rank colors → pick top → filter/rank quotes → render
- 25:02 – 26:24
Wrap-up: Jev 101 recap and what’s coming in episode 2
Claire summarizes Jev’s role: cheap, fast decision-making for filtering/classifying, best paired with stronger models for deep reasoning and generation. She teases a follow-up Jev episode with a popular guest and closes with standard subscribe/comment calls to action.
- •Recap: Jev for decisions/classification; pair with ‘brainy’ models for depth
- •Encourages viewers to ask Jev questions for the next episode
- •Teases midweek episode featuring a popular guest
- •Closing CTAs: like, subscribe, comment, and listen on podcast platforms
