Jev Is Incredible  🔗 **Source:** [Jev Is Incredible](https://www.youtube.com/watch?v=F3YXg7AaKWE) > [!note]+ Document Notes > tags: [ai-models, structured-outputs, classification, llm-engineering, context-management] > author: Theo (t3.gg) > source: Theo - t3.gg (YouTube) > reference: https://www.youtube.com/watch?v=F3YXg7AaKWE > date created: 2026-09-21 > > > [!info] Generated analysis — not the author's words. Sections below are machine-written > > from the captured source. Only text inside `> [!quote]` blocks is verbatim from the author. > > Do not quote or attribute any other line to the named author. > > # 🏷️ Title: Jev is incredible > <!-- deep-read:sections start --> ## 🧭 Table of Contents - [📝 Summary](#-summary) · [📖 Main Content](#-main-content) · [🗒️ Notes](#️-notes) · [📚 References](#-references) · [🗣️ Paraphrase](#️-paraphrase) · [📰 Newsletter](#-newsletter) · [✅ Action Items](#-action-items) · [🧪 Hypotheses](#-hypotheses) · [📊 Methodology](#-methodology) · [❓ Questions](#-questions) · [🔮 Future Research](#-future-research) · [💡 Implications](#-implications) · [🧠 Learning Styles Assessment](#-learning-styles-assessment) · [📊 Accessibility Assessment](#-accessibility-assessment) · [🌍 DEIB Assessment](#-deib-assessment) · [✅ Logical Validity](#-logical-validity) · [🛡️ DARVO Analysis](#️-darvo-analysis) · [🎯 Conclusion](#-conclusion) · [🧱 Refactor Appendix](#-refactor-appendix) · [🧰 Template Library](#-template-library) ## 🔗 References (in-page) - Source video: https://www.youtube.com/watch?v=F3YXg7AaKWE — Theo (t3.gg), uploaded 2026-09-21, 30m29s - **Sponsored segment:** Depot (CI and build infrastructure), about [00:01:02]–[00:03:07]. - Subject: Jev, a "System 1" classification model from Typesafe AI. Its founder is described as a former OpenAI researcher who worked on the research behind ChatGPT and RLHF; the captions render his name "Dio," which is unverified. - Model names in the captions (Fable, Astra, Terra, Soul, Luna, Opus 5, Haiku, GPT-OSS-120B, DSV4 Flash) are transcribed as spoken; spellings may be off. - Key vendor figures (193.6x faster, 444.6x cheaper, 0% structured-output errors) come from Typesafe's own announcement; the benchmark set was not published. ## 📝 Summary **What happened.** Theo reviews Jev, a new model from Typesafe AI built for classification rather than text generation. It takes structured state in and returns typed JSON with calibrated confidence, fast and cheap enough to use like a function call or an "if statement" inside code. **What Jev is.** Named after the Jevons paradox (cheaper power increases total use) and framed with Kahneman's System 1/System 2 distinction, Jev handles fast, intuitive judgments. It gives up free-text generation, has a 32K-token context window, no vision yet, and — per the vendor — always returns the requested JSON shape. Theo stresses that the *format* is deterministic; the *content* is not. **The numbers.** - Price: about $0.04 per million input tokens versus about $10 per million on a frontier model; output tokens are free ("too cheap to meter"). - Speed: 70–500 ms per classification versus 3–300+ seconds for LLMs, which Theo calls "40 to 200x faster." - Typesafe's homepage claims: 193.6x faster and 444.6x cheaper, which Typesafe itself says is "the higher end of real world gains." - Structured-output error rate: 0% for Jev; Theo reads a 45.5% rate for Haiku off the vendor's chart. - Real use: a colleague classified 100 emails at about 200 ms average (p95 240 ms, 38/second); Theo classified 32,311 messages across 1,118 of his own chat threads. **Where it's good.** Moderation, email triage, sorting content, routing decisions in workflows, fast UI updates, navigating HTML, and games from state data — anything a person could answer in under about 10 seconds once they have the information. **Where it isn't.** Theo pushes back hard on two viral ideas. Using Jev as an LLM judge (suggested by the observability company Braintrust) makes little sense because it can't reason about complex outputs. Using it for context compaction is a bad idea: compaction is synthesis, not filtering; Jev's 32K window can't hold the context; labs don't expose reasoning traces; models are trained on their own compaction; and deleting earlier history invalidates prompt caches. **Why it matters.** A clear explanation of a new class of model — cheap, fast, typed classification — and an unusually direct warning against using it for the wrong jobs. **Risks.** The headline benchmarks are the vendor's and unpublished. The reference answers used an average of two frontier models, so "accuracy" means agreement with those models. The video opens with a sponsor read. And "can't hallucinate" is true only of the output format. **Bottom line.** Treat Jev like a library function for classification, not a chat model. Believe the speed and price claims more than the unpublished accuracy claims. ## 📖 Main Content Theo opens by saying this isn't the usual incremental model release. Jev is for data processing: taking data, organizing it, classifying it, and returning type-safe JSON. It won't replace general models, and if you're using expensive frontier models for this kind of work, "I have a lot of questions for you." After a sponsor segment for Depot — faster CI, Docker caching, GitHub Actions–compatible builds pitched as built "for agents today" — he gives the background. Jev comes from Typesafe AI, founded by a researcher who worked on instruction-following and RLHF at OpenAI. The core idea is System 1 models: fast, intuitive judgment like recognizing a shirt's color, leaving slow reasoning to other models. The Jevons-paradox name signals the pitch: not new capabilities, but making existing ones cheap and fast enough to use everywhere. Theo's own checkers demo shows the tradeoff: moves come back almost instantly, but the model plays badly because it isn't reasoning. Reading the announcement, he emphasizes that Jev is meant to be called from code, not by a human. You pass structured program state and get back typed decisions — "a frontier intelligence function call." He compares it to BAML, a language that defines typed interfaces between code and LLMs and repairs malformed JSON at runtime, which he found necessary because GPT-OSS-120B formatted JSON incorrectly about half the time. He anticipates confusion about determinism. Code that formats a name does the same thing every time; an LLM might not, and might return a float where you wanted an integer. Jev's output *shape* is deterministic; its *content* is probabilistic but, he says, reliable. It also runs many inputs in parallel. On cost, input is about $0.04 per million tokens and output is free. On speed, 70–500 ms versus 3–300 seconds. Calibration is a selling point: a model that's right 95% of the time but can't tell you when it's in the other 5% can't be used for automation, while Jev reports confidence with every answer. The vendor's comparison shows 27 questions answered in 0.114 seconds for a negligible cost, against about 9 seconds for a frontier model. Theo notes the vendor charted cost on a log scale when linear would have looked more dramatic, and that only two models ranked better on classification in their workflows. The benchmark wasn't published, though examples were: an alert-triage flow of the kind "is this unauthorized activity?" leading to incident-state classification and a recommended action. The headline 193.6x/444.6x figures come with the vendor's own caveat that they're the upper end. The reference answers averaged two frontier models, so those models' mistakes are baked in. On structured-output errors, Jev scores 0% and Haiku 45.5%, which prompts Theo to say Haiku "needs to stop being used for anything." Demos: Doom played from game state at 10 decisions a second for under $7 an hour, with visible flailing because each frame is a fresh decision with no memory; wiki racing; JSON Render from Vercel producing UI updates in one shot; and 1,500 exported emails classified at about 200 ms each. Theo imagines future vision support for checking video frames for personal data before release. Then the warnings. He's worried that excitement will produce bad ideas and calls out Braintrust for suggesting Jev as a judge for LLM outputs: prompting a general model as a judge takes seconds, and a non-reasoning classifier doesn't know enough to judge. He's harsher on using Jev for context compaction, listing five problems: compaction synthesizes rather than filters; 32K tokens can't hold the history it would be judging; labs no longer return reasoning traces, so key decisions are invisible; models are trained to compact their own contexts; and removing an early message invalidates the prompt cache for everything after it, where cache writes cost more than reads. He also notes that shallow classification can loop, and that people who benchmarked it found it performs poorly. What he does recommend for agentic development is analysis: his 1,118 chat threads, 32,311 messages, classified cheaply, showing nearly half his work was bug fixing and about 20% PR review. Because outputs are scored, he can adjust the threshold: at 80% confidence 22% of threads look like scope expansion; at 90%, 6.8%. When he asked it which threads would make good video content, it flagged about half — a task needing judgment, which it isn't built for. His closing rule of thumb: if a person could answer the question in under 10 seconds once they have the information, Jev is probably a good fit; if it needs thought, it isn't. Treat it as a library or a function, not a chat tool. ### 🗒️ Notes #### ⏱️ [00:00:00](https://www.youtube.com/watch?v=F3YXg7AaKWE&t=0) What Jev is - A classification model from Typesafe AI returning type-safe JSON; not for text or code generation. #### ⏱️ [00:01:02](https://www.youtube.com/watch?v=F3YXg7AaKWE&t=62) Sponsor: Depot - CI, Docker caching, GitHub Actions compatibility, pitched for agent workflows. Paid segment. #### ⏱️ [00:03:07](https://www.youtube.com/watch?v=F3YXg7AaKWE&t=187) Origin and framing - Founder previously worked on the research behind ChatGPT and RLHF. - System 1 vs System 2 (Kahneman); named for the Jevons paradox. #### ⏱️ [00:04:08](https://www.youtube.com/watch?v=F3YXg7AaKWE&t=248) Checkers demo - Near-instant responses; weak play — speed without reasoning. #### ⏱️ [00:06:13](https://www.youtube.com/watch?v=F3YXg7AaKWE&t=373) Call it from code - Structured state in, typed decisions out. BAML comparison; GPT-OSS-120B formats JSON wrong about half the time in Theo's experience. #### ⏱️ [00:09:21](https://www.youtube.com/watch?v=F3YXg7AaKWE&t=561) Determinism, cost, speed - Shape deterministic; content probabilistic. - About $0.04/M input tokens; outputs free; 70–500 ms. #### ⏱️ [00:11:26](https://www.youtube.com/watch?v=F3YXg7AaKWE&t=686) Calibration and use cases - Confidence on every output; "a smart if statement"; map-reduce over data; real-time UI; verification. #### ⏱️ [00:12:28](https://www.youtube.com/watch?v=F3YXg7AaKWE&t=748) What it can't do - Can't judge between LLM outputs; "roughly as intelligent as a switch statement"; 32K context. #### ⏱️ [00:13:30](https://www.youtube.com/watch?v=F3YXg7AaKWE&t=810) Benchmarks - 0.114 s for 27 questions; log-scale chart; benchmark unpublished; reference = average of two frontier models. - Vendor claims 193.6x faster, 444.6x cheaper, flagged by the vendor as the upper end. #### ⏱️ [00:16:41](https://www.youtube.com/watch?v=F3YXg7AaKWE&t=1001) Error rates - Jev 0% structured-output errors; Haiku 45.5%. #### ⏱️ [00:17:44](https://www.youtube.com/watch?v=F3YXg7AaKWE&t=1064) Demos - Doom from state at under $7/hour; wiki racing; JSON Render; emails at ~200 ms (p95 240 ms, 38/s). #### ⏱️ [00:21:57](https://www.youtube.com/watch?v=F3YXg7AaKWE&t=1317) Bad ideas: judge and compaction - Braintrust's judge suggestion criticized. - Five reasons compaction with Jev fails, including cache invalidation. #### ⏱️ [00:26:06](https://www.youtube.com/watch?v=F3YXg7AaKWE&t=1566) Good idea: analyzing history - 1,118 threads, 32,311 messages; about half bug fixing, ~20% PR work; threshold changes shift results (22% → 6.8%). #### ⏱️ [00:28:13](https://www.youtube.com/watch?v=F3YXg7AaKWE&t=1693) The 10-second rule - Under 10 seconds of thought: good fit. Needs reasoning: wrong tool. ## 📚 References - Primary: the video — https://www.youtube.com/watch?v=F3YXg7AaKWE - Typesafe AI's Jev announcement (read on screen; URL not captured). The benchmark set wasn't published. - BAML (BoundaryML) — the typed LLM interface language Theo compares against. - Braintrust — the observability company whose judge suggestion Theo criticizes. - Vercel JSON Render — the UI demo mentioned. - Daniel Kahneman, *Thinking, Fast and Slow* (2011) — the System 1/System 2 framing. - The Jevons paradox — William Stanley Jevons, *The Coal Question* (1865). - Sponsor: Depot (depot.dev). ## 🗣️ Paraphrase 1. There's a new kind of model out, built for sorting and classifying data rather than writing text. 2. It returns data in exactly the shape you ask for, very fast and very cheaply. 3. It's meant to be called from your code like a function, not chatted with. 4. The format is guaranteed; the answer itself is still a probabilistic judgment, but it comes with a confidence score. 5. It's dozens to hundreds of times faster and cheaper than general models for this kind of work. 6. It's good at quick decisions — sorting email, moderation, routing — and bad at anything that needs thinking. 7. Don't use it to judge other models' output, and definitely don't use it to compress an agent's memory. 8. It's great for analyzing large amounts of your own data, where you can set confidence thresholds. 9. Rule of thumb: if a person could answer it in under 10 seconds, it's a good fit. ## 📰 Newsletter **Subject: A model that's as smart as a switch statement (on purpose)** Theo's [breakdown of Jev](https://www.youtube.com/watch?v=F3YXg7AaKWE) is worth a watch mostly for what he tells you *not* to do with it. Jev, from Typesafe AI, is a classification model: you hand it structured data and a JSON schema, and it hands back typed decisions with confidence scores in 70–500 ms, at roughly $0.04 per million input tokens with free output. Theo's framing is the right one — it's "a smart if statement," something you call from code, not something you chat with. Where it shines: email triage (a colleague ran 1,500 emails at about 200 ms each), moderation, sorting a big pile of content, routing decisions inside a workflow. Where it doesn't: judging other models' output, and — Theo is emphatic — compressing an agent's context. Compaction is synthesis, not filtering; a 32K window can't see what it's deciding about; and deleting earlier history blows your prompt cache. Two grains of salt. The headline "193.6x faster, 444.6x cheaper" comes from Typesafe's own unpublished benchmark, scored against an average of two frontier models, and Typesafe itself calls it the upper end. And the video opens with a sponsor read. WordPress angle: this is exactly the tool for spam and comment moderation, classifying support threads, or tagging a content backlog — anywhere a site runs lots of small, cheap decisions. ## ✅ Action Items **DevRel** - A strong teaching example for "right model for the job": System 1 classification versus System 2 reasoning. - The 10-second rule is a memorable heuristic for developer audiences. **OSPO** - Structured-output reliability varies widely by model (the 45.5% Haiku figure, GPT-OSS-120B's JSON problems). Projects depending on structured output should test models before relying on them. **Community** - Not applicable to this source directly; no community content. **WordPress subteams** - *Core AI* — the Abilities API depends on structured input and output; reliable typed JSON is exactly what makes ability calls safe. Worth evaluating classification models for low-risk routing decisions, and noting error rates when choosing defaults. - *Plugins* — comment moderation, spam detection, and content tagging plugins are natural fits for cheap classification models. - *Support* — classifying forum threads by topic and urgency at scale. - *Meta* — triage for Trac tickets or plugin review queues. - *Core Performance* — the cache-invalidation point applies to any AI feature that edits conversation history. - *Core*, *Core Program Team*, *Mobile*, *Playground*, *Community*, *Training*, *Marketing*, *Design*, *Documentation*, *Photos*, *TV*, *Polyglots*, *Accessibility*, *CLI*, *Hosting*, *Tide*, *Openverse*, *Themes* — no direct action. **The WPOCC** - Not applicable to this source. **FAIR** - Classification models could help triage package metadata or flag suspicious plugin updates cheaply, as a first-pass filter before human review. ### → Things 3 - [ ] Consider a cheap classifier for triaging the Readwise/vault backlog (tagging, not synthesis) ## 🧪 Hypotheses 1. **Typed classification beats prompted LLMs on format errors.** Across matched schemas, a dedicated classifier will produce fewer structured-output errors than general models. Testable on public schemas. 2. **Calibration enables automation.** Workflows using confidence thresholds will route fewer wrong cases to automation than workflows without them. Testable in a triage pipeline. 3. **Classifier compaction degrades agent performance.** Agents whose context is compacted by a non-reasoning classifier will complete fewer tasks than agents using native compaction. Testable on an agent benchmark. 4. **Speed changes use.** Per the Jevons framing, teams adopting a cheap classifier will increase total classification calls substantially. Testable from usage data. 5. **The 10-second rule predicts fit.** Tasks humans answer in under 10 seconds will show higher classifier accuracy than slower tasks. Testable by timing humans on task sets. ## 📊 Methodology **Design:** a sponsored creator review mixing the vendor's announcement, the reviewer's own quick demos, and community demos from social media. **Evidence:** - Vendor benchmarks — unpublished benchmark set; reference answers are an average of two frontier models; the vendor flags its own numbers as the upper end. - Theo's demos — checkers (qualitative), his own chat-history classification (real use, no accuracy check). - A colleague's email test — latency measured; accuracy not reported. - Error-rate charts — the vendor's. **Strengths:** Theo separates format determinism from content accuracy, notes the log-scale charting, and flags the unpublished benchmark. **Weaknesses:** no independent accuracy test; a sponsor read at the start (for a different company); social-media demos are selected for virality. ## ❓ Questions - How accurate is Jev on classification tasks with ground-truth labels rather than agreement with frontier models? - How well calibrated are its confidence scores in practice? - What happens when inputs exceed 32K tokens — truncation, error, or silent loss? - Is "can't hallucinate" meaningful when the content of a well-formed answer can still be wrong? - How does it compare with small open models fine-tuned for classification? - What are the privacy terms for data sent to it (emails, chat histories)? ## 🔮 Future Research - Independent benchmarks with labeled datasets and calibration curves. - Agent-compaction experiments comparing classifier-based and native compaction. - Cost studies of the Jevons effect after adoption. - Evaluation of vision support once released, especially for PII detection in video frames. - Comparison against fine-tuned small models on cost, speed, and accuracy. ## 💡 Implications **Practice.** Use classification models for high-volume, low-judgment decisions with confidence thresholds, and keep reasoning models for judgment. Don't put a classifier in the middle of agent memory management. **Policy.** Moderation and triage at this price will spread quickly; platforms need policies on appeal and human review for automated classifications. **Security.** Theo lists jailbreak detection and guardrails as use cases. A fast, cheap first-pass filter is useful, but a non-reasoning classifier can be fooled by subtle attacks, so it should back up, not replace, deeper checks. Sending private data (emails, chat histories) to a new vendor carries data-handling risk. **Education.** System 1/System 2 is a useful way to teach model selection. **Funding.** Cheap classification lowers the cost of moderation for community-run sites and open-source forums. ## 🧠 Learning Styles Assessment - **Visual** — moderate: on-screen demos and charts (not visible in the transcript). - **Aural** — strong: energetic narration that works as audio for the conceptual parts. - **Verbal** — strong: clear analogies (if statement, switch statement, knee reflex). - **Kinesthetic** — moderate: demos invite trying it; access is invite-only. - **Logical** — strong: a clear argument for fit and misfit, with a heuristic. - **Social** — moderate: community demos and responses to chat. - **Solitary** — moderate. - **Naturalistic** — strong in theme: the whole subject is classification. **Reading level:** spoken, about grade 11–12, with dense AI-engineering jargon (RLHF, structured outputs, p95, prompt caching, compaction). ## 📊 Accessibility Assessment Mapped to WCAG 2.2 AA, evaluated as published on YouTube. 1. **Captions (1.2.2)** — auto-generated; model and company names are garbled ("Chaz GBT," "anthrowing," "Enthropic"), profanity is bleeped as "[ __ ]." *Fail for accuracy.* 2. **Audio description (1.2.5)** — the demos (checkers, Doom, charts) carry information shown on screen without full narration; the log-scale point depends on the chart. *Partial.* 3. **Keyboard operation (2.1.1)** — inherited from YouTube. *Pass.* 4. **Navigation (2.4.x)** — chapters not visible in the capture for a 30-minute video. *Partial.* 5. **Readable language (3.1.x)** — heavy jargon. *Partial.* 6. **Sponsor identification** — the sponsor segment is announced verbally. *Pass.* 7. **Timing (2.2.x)** — not applicable. **Top fix:** corrected captions for technical terms, and verbal description of chart values. ## 🌍 DEIB Assessment **Cultural breadth.** Anglophone tech industry; one mention of a comment from China. **Inclusive language.** Some condescension toward commenters ("I'm increasingly tired of comment sections that fundamentally don't understand") and viewers who "don't code anymore sadly." **Neurodiversity.** Fast pacing and frequent asides make it demanding; the closing heuristic helps. **Representation.** Male-led; people named are mostly men. **Psychological safety.** Mixed. The warnings are useful, but mockery ("you just fundamentally don't understand") could discourage beginners from asking questions. ## ✅ Logical Validity ### Strong reasoning - Distinguishing format determinism from content accuracy. - Flagging the vendor's log-scale chart, unpublished benchmark, and self-described upper-end numbers. - A detailed technical argument against classifier-based compaction. - A clear, testable heuristic (10 seconds). ### Fallacies 1. **Repeating a vendor overclaim.** *Where:* [00:07:16] "it can't hallucinate." *Why it fails:* the format can't be hallucinated; the content can still be wrong, which Theo himself explains later. *Fix:* "it can't break the format." 2. **Assertion as verification.** *Where:* [00:10:23] "There's no doubt on that claim. That is real and true." *Why it fails:* the speed range comes from the vendor; Theo's own tests support it qualitatively but aren't measured. *Fix:* cite his measurement. 3. **Circular accuracy metric (noted by Theo).** *Where:* reference answers averaged from two frontier models. *Why it matters:* measures agreement, not truth. 4. **Hasty generalization.** *Where:* "Haiku… needs to stop being used for anything," from one structured-output chart. *Fix:* limit the claim to structured output. 5. **Mockery as argument.** *Where:* the compaction section ("never ever stray from the defaults"). *Why it matters:* the technical case is strong on its own; the ridicule doesn't add evidence. ### Tone Energetic, opinionated, occasionally dismissive of audience confusion, technically careful where it matters. ## 🛡️ DARVO Analysis **Does the source perform DARVO?** No. There's no accused party defending itself. **Is DARVO described?** No. Theo criticizes Braintrust's suggestion, but that's disagreement, not a DARVO sequence. **Assessment:** not applicable to this source. ## 🎯 Conclusion **Bottom line.** A useful, mostly careful review of a new class of model: fast, cheap, typed classification meant to be called from code. Theo's best contributions are the distinction between format and content determinism, the 10-second rule, and a detailed case against using Jev for context compaction or LLM judging. The headline performance claims come from the vendor's unpublished benchmark, the video is sponsored (by a different company), and "can't hallucinate" should be read as "can't break the format." **Next steps.** 1. Use the 10-second rule to pick tasks for classification models. 2. Wait for independent accuracy and calibration benchmarks before relying on Jev for decisions with consequences. 3. Keep classifiers out of agent memory management. ## 🧱 Refactor Appendix ### Literature Split Pad **L1 — System 1 models.** Fast, typed classification; structured state in, JSON with confidence out. Source: Theo, 2026-09-21. **L2 — Format vs content determinism.** The shape is guaranteed; the answer is probabilistic. **L3 — Cost and speed.** About $0.04/M input tokens, free output, 70–500 ms. **L4 — Misuse: compaction.** Synthesis not filtering; 32K window; hidden reasoning; trained compaction; cache invalidation. **L5 — The 10-second rule.** #### Atomic Split Pad **A1 — A guaranteed format isn't a guaranteed answer.** Type safety constrains shape, not truth. Links: [[Structured outputs]], [[Hallucination]]. **A2 — Calibration is what makes automation possible.** A model that can't tell you when it's wrong can't run unattended. Links: [[Confidence thresholds]]. **A3 — If a person answers in under 10 seconds, a classifier may do.** Links: [[Model selection]]. **A4 — Editing early history breaks the prompt cache.** Everything after the change must be rewritten. Links: [[Prompt caching]], [[Context management]]. **A5 — Cheaper intelligence means more of it.** The Jevons paradox for inference. Links: [[Jevons paradox]]. #### Fleeting Queue - Verify the founder's name and Typesafe AI's announcement URL. - Check model-name spellings from the captions. - Does Jev publish calibration data? <!-- deep-read:sections end --> 📄 **Full text:** [[Full Document Contents/Articles/Jev Is Incredible.md|Jev Is Incredible by Theo (t3.gg)]] ## Highlights - First thing, compaction isn't a filter. We're not just going through your history and selectively deleting lines from it. We are synthesizing a summary based on everything that's happened so far. ([View Highlight](https://read.readwise.io/read/01m31zf7ge8bkw6wnv1cwye1rv)) - Note: L4 — why a classifier shouldn't compact agent context: compaction is synthesis, not filtering, and Jev's 32K window can't see what it's deciding about. Related: editing early history invalidates the prompt cache (Atom A4). - Tags: context-management, atom/literature/h1 - If the answer is under 10 seconds, this model's probably good for it. If it's over 10 seconds, this model's likely less good for it. ([View Highlight](https://read.readwise.io/read/01m31zfaegxy3t4szb5nsacm2v)) - Note: Atom A3 — the 10-second rule for model selection. If a person could answer in under 10 seconds once they have the information, a System 1 classifier is probably a fit. - Tags: classification, ai-models, atom/atomic/h1 - this model is incredible. It's a system one model. ([View Highlight](https://read.readwise.io/read/01m31zer15hr38wsy737psezq1)) - Note: L1 — System 1 models: fast, typed classification, not reasoning. Theo's closing statement of the model's purpose. - Tags: atom/literature/h2, classification