This Might Be the Best AI Release of 2026 - CourtneyR.dev's 2nd Brain

This Might Be the Best AI Release of 2026 ![rw-book-cover](https://i.ytimg.com/vi/BHPDsGVciDk/sddefault.jpg) 🔗 **Source:** [This Might Be the Best AI Release of 2026](https://youtube.com/watch?v=BHPDsGVciDk&is=wSbl9CG471-76h3x) > [!note]+ Document Notes > tags: [system-one-models, decision-classifiers, inference-cost, real-time-moderation, computer-use-testing, jevons-paradox, vendor-benchmarks] > author: ThePrimeagen (The PrimeTime) > source: The PrimeTime (YouTube) > reference: https://www.youtube.com/watch?v=BHPDsGVciDk > date created: 2026-09-26 16:25 > > > [!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 Do not quote or attribute any other line to the named author.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: This Might Be the Best AI Release of 2026 >  ## 🧭 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--this-might-be-the-best-ai-release-of-2026) ## 📝 Summary - **What happened.** ThePrimeagen's clips channel, The PrimeTime, uploaded a 12-minute reaction on 2026-09-25 to Jev, the "System 1" decision model from Typesafe AI (typesafe.ai). He reads the Typesafe manifesto page on screen, runs his own prompts in the Typesafe playground, shows Jev playing Balatro live, and shows a Jev-driven computer-use session. The auto-transcript renders the product as "Jev," "Jeff," and "Jevon," the game as "Bellotro," "Batro," and "Bellatro," and the vendor's numbers as "190* 3 faster and 444 times cheaper." - **The claim (source's).** Typesafe's site says Jev is roughly 190x faster and 444x cheaper than frontier models (Typesafe's announcement gives 193.6x and 444.6x; the transcript garbles the first figure). Prime reads this aloud at 00:02:50 and immediately says nobody knows which model is the baseline. - **The claim (Prime's).** Jev is not a chatbot and cannot write code; it takes a state plus a set of questions with fixed answers and returns choices with confidence numbers, all questions running in parallel. He argues this shape fits real-time work a generative agent can't do: chat moderation at a 2,000-person Smash Bros event, pruning agent context, form-filling in computer use, and playing Doom or Balatro frame by frame. - **The mechanism.** Three question types appear in his demo: choice (pick one option, returned with 95% and 85% confidence in two runs), score (relative position across options), and a 0-to-1 value the captions render as "new," which he ties to the Bernoulli distribution. Billing is on input tokens only; his dashboard reads 345 input tokens for one Smash Bros question and he computes about three million asks for $44, quoting $42 per billion tokens at 00:09:58. - **Why it matters.** If a closed-set decision costs 100 milliseconds and a fraction of a cent, the decision can sit inside program control flow rather than inside a prompt. That changes where the expensive generative model runs: only where prose is the deliverable. - **Key stakeholders.** Typesafe AI (vendor), ThePrimeagen (creator, unpaid by Typesafe as far as the transcript shows, paid by WorkOS for the 00:01:25 sponsor read), developers paying frontier prices for classification, and moderators of live events. - **Risks and tradeoffs.** The speed and cost multipliers come from the vendor and carry no baseline. Prime's "a thousand times better than almost any other solution" at 00:09:58 is his own unmeasured figure. The Balatro agent wins about 30% of the time by his count and he calls its play "not helpful" twice; the demo shows the shape of the tool more than its accuracy. Nothing is said about calibration, out-of-set answers, data retention, or availability. - **Bottom line.** The architectural point is the same one Firecrawl and Theo made about Jev the week before: a closed answer set wants a classifier, not a generator. What Prime adds is the real-time case (per-message moderation as messages arrive, game control at frame rate) and a concrete billing readout. The multipliers stay unverified. ## 📖 Main content **Who is reacting to what.** The PrimeTime is ThePrimeagen's secondary channel, where his live-stream segments are cut into standalone videos. The source on screen is Typesafe AI's manifesto and product pages for Jev (he says "if you go to their manifesto" at 00:02:08), then the Typesafe playground where he types his own state and questions, then his own Balatro-playing script ("jev play" at 00:05:42), then an Omarchy QA session fed by screenshots from a remote QEMU virtual machine (the captions give "Omari uh QA" and "QMU" at 00:07:08). Two separate people's claims run through the video: Typesafe's (the multipliers, the manifesto framing, the Jevons name) and Prime's (everything about what he built and how fast it felt). This analysis keeps them apart. **The framing Prime supplies.** He opens with a System 1 analogy: a ball is flying at your face and you duck without deliberating (00:00:42). That is his gloss, not the vendor's copy, though Typesafe uses the Kahneman "System 1" label too. He is emphatic about what Jev is not: not a chatbot, not conversational, not able to write quicksort (00:00:00). It is, in his words, a decision bot. **The framing Typesafe supplies.** He reads the manifesto tagline, which the captions render as "composable AI build prod not god" (00:02:08), and the horse-to-car analogy: early cars kept whip holders because designers carried assumptions from the previous technology. The manifesto's application is that people use LLMs to make decisions the way early cars kept whip holders: an inherited shape, not a fit. Then the multipliers, which the captions give as "190* 3 faster and 444 times cheaper" (00:02:50). Prime's reaction is the right one: "We don't know what model they're talking about." He jokes that the baseline might be "the fabled fable" or a fast Grok tier. The transcript's "Grock 45" is an unverifiable model name. **The interface, as demonstrated.** State plus questions plus a fixed answer set per question, all evaluated in parallel (00:03:33). His first demo is deliberately crude (a weightlifting-competition accident) and returns "go clean up" at 95% confidence in one run and 85% in a scored run. The second is the one he cares about: state describes a Super Smash Bros match where Donkey Kong is receiving chat messages; two questions ask whether the message is appropriate and whether it helps Donkey Kong win. Changing the message from an insult to "smash all the opponents" flips the helpfulness score. He notes he mislabeled "is appropriate" when he meant "not appropriate," and that adding more state would sharpen the answer (00:04:59). The third question type, a 0-to-1 value, is transcribed as "new" and attributed to the Bernoulli distribution; the exact type name needs the Typesafe docs. **The real-time argument.** This is Prime's original contribution to the Jev coverage. He ran a live event where about 2,000 people drove four Smash Bros characters through chat. Moderation ran through a generative agent that took about a minute and a half per batch, which meant a minute and a half of stage time to fill (00:05:42). With a parallel per-message classifier, each message is judged as it arrives and the agent's only remaining job is turning approved messages into game inputs. His estimate is minutes of dead air saved per round. That is a latency argument, not a cost argument, and it is the strongest use case in the video because the human audience is synchronously waiting. **The game demos.** Balatro plays live during the middle third. The agent picks a strategy (pair scaling), buys jokers to match, and wins about 30% of runs by his count (00:06:25). He disagrees with its play at least twice and says outright "It's not going to win" before it clears a blind at 00:07:51. Doom, at 00:11:25, is explained as several parallel yes/no questions per frame: dodge or stand, fire or hold. The point is that per-frame decisions at 100 milliseconds each are feasible on a hosted model, which no token-generating model does. **Context pruning and computer use.** Two further uses get one line each. First, ask Jev "is this thing relevant to the job at hand?" for each chunk of agent context and drop what it rejects (00:07:51). Second, drive computer-use testing: he shows Jev clicking through a UI and argues that forms are classification problems (00:11:25), so a decision model is the right tool for "go do all these activities and make sure it actually works" as a replacement for hand-written Playwright tests (00:10:43). Neither claim is measured in the video. **The billing readout.** He shows a usage dashboard with 8 million tokens consumed during the Balatro session, refreshes and sees another million, and says the price has barely moved (00:07:51). At 00:08:34 he reads 345 input tokens for the Smash Bros question and says input is the only billed dimension, then computes about three million asks for $44. At 00:09:58 he quotes $42 per billion tokens. The arithmetic is internally consistent: 345 tokens times three million is 1.035 billion tokens, times $42 per billion is about $43.50. It is also consistent with the $0.04 per million input tokens Theo reported from Typesafe's page on 2026-09-20. **Contested or unverifiable in this source.** The 190x and 444x figures carry no baseline, no task set, and no methodology, and Prime says so. His "a thousand times better than almost any other solution" (00:09:58) and "a hund times faster" for computer use (00:07:08) are estimates from feel. The 30% Balatro win rate is his eyeball count over an unstated number of runs. "Soul 6" and "Opus 55" at 00:09:16 are model releases he name-drops as having shipped the same day; the second is almost certainly Opus 5.5, and the first is unverifiable from captions. Steve Faulkner (captions: "Steve Falner"), who proposed the `if email feels urgent` syntax he shows at 00:09:58, is most likely the Cloudflare Workers engineering director of that name; the transcript does not say. ## 🗒️ Notes #### ⏱️ [00:00:00](https://www.youtube.com/watch?v=BHPDsGVciDk&t=0) Cold open and what Jev is not - On-screen: an AI drawing boxes and arrows in real time, then Mario, then Doom. He can't say what the diagram is, only that it's fast. - Definition by negation: not a chatbot, no text in and text out, can't be asked to write quicksort. - Reframes Jev as a "decision bot" and introduces System 1 thinking with the ball-to-the-face soccer example. > [!quote] [00:00:00] > It's not a chatbot, meaning that you cannot give it text and it'll give you text back. > [!quote] [00:00:42] > But that's called system one thinking. It is fast, quick decisionmaking. #### ⏱️ [00:01:25](https://www.youtube.com/watch?v=BHPDsGVciDk&t=85) Sponsor read: WorkOS - Sponsor is WorkOS, introduced with the joke "a quick thank you from the sponsor, Vibe Coders. Silence." Enterprise auth pitch: SSO, passkeys, roles, recovery. Named customers: OpenAI, Anthropic, Cursor. - Free for the first million users; the caption "9999,999 free seats" is the garbled 999,999. - Disclosure is explicit and separated from the Jev content. No Typesafe relationship is disclosed or implied. #### ⏱️ [00:02:08](https://www.youtube.com/watch?v=BHPDsGVciDk&t=128) Reading the Typesafe manifesto - Tagline as captioned: "composable AI build prod not god." The whip-holder analogy: early cars inherited horse-era fittings; LLMs used for decisions are the inherited fitting. - Vendor multipliers read aloud, then immediately questioned for a missing baseline. > [!quote] [00:02:08] > It says composable AI build prod not god. > [!quote] [00:02:50] > It's 190* 3 faster and 444 times cheaper. We don't know what model they're talking about. #### ⏱️ [00:03:33](https://www.youtube.com/watch?v=BHPDsGVciDk&t=213) Playground demo: state, questions, three types - Crude first demo returns "go clean up" at 95% confidence; the scored variant gives 85% to option one with a small residual on "keep working out." - Smash Bros demo: state is a match where Donkey Kong receives chat messages; two questions test appropriateness and helpfulness. He mislabeled "is appropriate" and corrects himself on air. - Third type introduced: a 0-to-1 value, captioned "new," tied to the Bernoulli distribution. Changing the message to "smash all the opponents" makes it "fairly helpful." > [!quote] [00:04:16] > By the way, this is the third type, a new. The name new comes from Berni distribution as far as I can tell. #### ⏱️ [00:05:42](https://www.youtube.com/watch?v=BHPDsGVciDk&t=342) The live-event moderation problem - About 2,000 attendees drove four Smash Bros characters via chat. Moderation ran through a slow generative agent, about 90 seconds per batch, then a second pass turned approved messages into code. - Proposed fix: judge each message in parallel as it arrives, leaving the agent one job. > [!quote] [00:05:42] > Well, that takes like a minute and a half. So, that means I have to be funny on stage for a long time. #### ⏱️ [00:06:25](https://www.youtube.com/watch?v=BHPDsGVciDk&t=385) Balatro live and the Omarchy QA feed - "jev play" starts a Balatro run: small blind, discards, full house, pairs. Strategy is pair scaling; he dislikes it but reports multiple wins and about a 30% win rate. - Second window: an Omarchy QA session receiving screenshots from a remote QEMU VM, with a claim that Jev plus "a really dumb model" could run it about a hundred times faster. Not measured. > [!quote] [00:06:25] > I do not like how it plays, but it has won multiple times cuz it sets out a strategy. #### ⏱️ [00:07:51](https://www.youtube.com/watch?v=BHPDsGVciDk&t=471) Context pruning, token dashboard, Jevons - Use case: ask "is this relevant to the job at hand?" per context chunk and drop rejects before the agent call. - Dashboard: 8 million tokens spent on Balatro, one million more after a refresh, price barely moving. Balatro clears a blind he predicted it would lose. - 345 input tokens per Smash Bros question, input-only billing, about three million asks for $44. Name explained via Jevons paradox. > [!quote] [00:07:08] > Though you can use it say to kind of make your context you give to your agent smaller. > [!quote] [00:08:34] > That Smash Brothers question that I had, 345 input tokens, and you're only charged for input tokens. > [!quote] [00:08:34] > That's why it's also called Jev, named after Jevon's paradox. #### ⏱️ [00:09:16](https://www.youtube.com/watch?v=BHPDsGVciDk&t=556) The same-day releases he ignores, and `feels urgent` - Two other models shipped the same day (captioned "Soul 6" and "Opus 55"); he says he hasn't tried either and doesn't care. A chat member interrupts and gets a one-hour timeout. - Steve Faulkner's suggestion: `if email feels urgent` as a language construct backed by Jev. Prime says you don't need a new language, `if await feels urgent` in any language does the same. - His own figures: 100 milliseconds per call, "a thousand times better," $42 per billion tokens. > [!quote] [00:09:16] > Hey, guess who hasn't even tried him because I honestly don't give a rat's ass. This guy. > [!quote] [00:09:58] > It takes a 100 milliseconds to execute. This is a thousand times better than almost any other solution. #### ⏱️ [00:10:43](https://www.youtube.com/watch?v=BHPDsGVciDk&t=643) Computer use, forms as classification, how Doom works - Pitch to the unconvinced: replace hand-written Playwright tests with "go do all these activities and make sure it actually works," driven by Jev. - Forms are classification problems, so a decision model fits. - Doom explained: several parallel yes/no questions per frame (dodge or not, fire or hold), visible as the agent holds fire until the enemy is closer, then dodges backward. Video ends on this demo with no outro. > [!quote] [00:10:43] > I just want you to imagine all your crappy playwright tests you have ever written in your lifetime. > [!quote] [00:11:25] > This this is just a classification problem. That's right. Forms, they're just classification problems. ## 📚 References The transcript itself contains no links, and the video description was not captured with it; the following come from the transcript's named entities and are primary sources for them. - [This Might Be the Best AI Release of 2026, The PrimeTime](https://www.youtube.com/watch?v=BHPDsGVciDk): the source video, 12 minutes, uploaded 2026-09-25. - [Typesafe AI](https://typesafe.ai/): vendor of Jev; the manifesto he reads at 00:02:08 lives here. - [Introducing System One Models and Jev, Typesafe blog](https://typesafe.ai/blog/introducing-system-one-models-and-jev): the announcement carrying the 193.6x and 444.6x figures the captions garble, and the Doom demo he references at 00:11:25. - [Typesafe docs](https://docs.typesafe.ai/introduction): where the real name of the third question type (captioned "new") should be confirmed. - [WorkOS](https://workos.com): the 00:01:25 sponsor. - [Jev Is Incredible, Theo (t3.gg), 2026-09-20](https://www.youtube.com/watch?v=F3YXg7AaKWE): earlier coverage; reports $0.04 per million input tokens, consistent with Prime's $42 per billion. - [Jev Is Not an LLM (And That's Why It Wins), Firecrawl, 2026-09-21](https://www.youtube.com/watch?v=PNbbMi2h9VA): earlier coverage with the same 200x/400x-class claims and no benchmark. - *Thinking, Fast and Slow*, Daniel Kahneman: origin of the System 1 / System 2 framing both Prime and Typesafe use. Not named in the transcript. - Jevons paradox (William Stanley Jevons, *The Coal Question*, 1865): the naming source he cites at 00:08:34. - [Balatro](https://www.playbalatro.com/): the card game the agent plays; captions render it "Bellotro," "Batro," and "Bellatro." - [Omarchy](https://omarchy.org/): the Arch-based Linux setup whose QA session appears at 00:07:08 (captions: "Omari uh QA"). - [QEMU](https://www.qemu.org/): the VM supplying screenshots to that session (captions: "QMU"). ## 🗣️ Paraphrase Prime's argument runs like this. Everything you've used so far under the label "AI" is a talker: text in, text out. Jev doesn't talk. You describe the world, list the questions you need answered, and fix the set of answers each question can have. It hands back the pick and how sure it is, for every question at once. That's the reflex that moves your head out of the way of a ball, not the deliberation that decides which house to buy. The reason he cares is a night he spent on stage. Two thousand people were steering four Smash Bros characters through chat, and every batch of messages went through a generative agent that took a minute and a half to say which ones were out of line. That's a minute and a half of him improvising for a crowd. Per-message, parallel, hundred-millisecond judgments would have collapsed that wait to nothing and left the agent with one job. From there he generalizes: prune agent context chunk by chunk, fill forms in a browser, play Doom by asking "dodge?" and "fire?" every frame. It's billed on input only, a few hundred tokens a question, so his three-million-question thought experiment costs about $44. What he doesn't do is verify the vendor's speed and cost multipliers. He reads them, says nobody knows the baseline, and moves on to his own demos. That's the honest version of a reaction video: the mechanism is shown, the marketing numbers are quoted with a shrug. ## 📰 Newsletter **A model that only answers multiple-choice questions, and why that's the interesting release this week.** ThePrimeagen spent 12 minutes on Jev, the "System 1" decision model from Typesafe AI, and his best material isn't the demos of it playing Balatro and Doom. It's the story of running chat moderation for a 2,000-person live event through a generative agent that took 90 seconds per batch, and realizing that a per-message classifier returning a yes/no with a confidence number at 100 milliseconds would have erased that wait entirely. Billing is input-only; he reads 345 tokens per question off his dashboard and works out roughly three million questions for $44. Where this lands for WordPress people: plugin and theme review queues, support-forum triage, Openverse moderation, and "which handbook page answers this" lookups are all fixed-answer questions we currently pay generation prices to answer. The probability that comes back with the label is the useful part, because it gives you a threshold for when a human steps in. Prime's framing, that forms are classification problems, also points at a different shape of end-to-end test than the Playwright scripts most plugin repos carry. Caveat you should carry with the link: the 190x-faster and 444x-cheaper figures are Typesafe's, and Prime says on air that nobody knows the baseline model. Watch for the mechanism, not the multipliers. Links: [the video](https://www.youtube.com/watch?v=BHPDsGVciDk), [Typesafe's announcement](https://typesafe.ai/blog/introducing-system-one-models-and-jev), [Theo's earlier walkthrough](https://www.youtube.com/watch?v=F3YXg7AaKWE). ## ✅ Action items #### DevRel - Developer advocacy: build the two-model pattern as a demo (classifier gates each incoming item, generative model runs only on what passes) and publish the per-request token bill from the provider dashboard, the way Prime shows his at 00:08:34. - Education: use the Smash Bros moderation story (00:05:42) as the opening case in a session on latency-bound versus cost-bound AI decisions; it's the clearest human-waiting example in recent coverage. - Developer experience: add a "decision or generation?" checklist to agent quickstarts, with the context-pruning use case from 00:07:51 as the worked example. - Developer marketing: quote the mechanism (input-only billing, no output tokens) and never the vendor multipliers without a benchmark citation. Prime himself flags the missing baseline at 00:02:50. - Community management: the one-hour chat timeout at 00:09:16 is a live moderation moment; pair it with the moderation use case when discussing how classifiers change moderator workload. - Internal/external: confirm the real name of the third question type (captioned "new") and the pricing page figures with Typesafe before anyone writes a tutorial. #### OSPO - Policy and governance: no direct action from this source; nearest relevance is deciding whether a hosted, closed-weights classifier is acceptable inside contributor tooling before a team ships one. - License compliance: check Typesafe's terms for input retention and training-on-input before any community data (forum posts, chat logs) is sent as state. - Security management: chat messages and scraped UI text entering model state are injectable input. Add the per-message moderation loop to the threat model as a distinct surface. - Software inventory: a decision model in a moderation path is a network dependency with an availability SLO; record it as such, not as a library. - Developer support: publish guidance on confidence thresholds; 95% and 85% at 00:03:33 are single-run readouts, not calibration. - Community engagement: no direct action; nearest relevance is tracking whether a second System 1 vendor appears, since a single-vendor category is a lock-in risk. - Strategic alignment: set a KPI of inference spend per resolved decision so a classifier swap is measurable. #### Community - Maintainers: pilot per-item classification on one project's issue or review queue for 30 days and measure agreement with human labels before trusting it. - Contributors: document the escalation rule (what confidence routes to a person) so contributors know when the machine decided and when a human did. - Sponsors: if a vendor funds a moderation or triage pilot, require published before/after latency and cost figures. - Ecosystem: run the same classifier against the same labeled corpus across two projects so results are comparable rather than anecdotal. #### WordPress - Core: no direct action from this source; nearest relevance is none, this is tooling around the project, not code in it. - Core Performance: no direct action; the 100-millisecond latency figure is a hosted-API number, not a PHP request budget. - Core AI: add closed-set classification as a distinct pattern in any responsible-AI guidance, with its own evaluation criteria (accuracy, calibration) separate from generation quality. - **Core Program Team:** flag that "System 1 model" is a vendor coinage borrowed from Kahneman, not an established category, before it enters cross-team notes. - Mobile: no direct action from this source. - Playground: a browser demo running the same input through a classifier and a generator side by side would teach the distinction in one page. - Community: the live-event moderation story at 00:05:42 is directly relevant to WordCamp and meetup livestream chat; evaluate per-message classification for those channels. - Training: draft a short lesson on choosing classification versus generation, using the Smash Bros message flip at 00:04:59 as the example. - Marketing: no direct action from this source. - Support: evaluate per-thread routing with a classifier, measured against current human routing accuracy on a labeled sample. - Design: no direct action from this source. - Documentation: if System 1 / System 2 language enters WordPress AI docs, add a glossary entry crediting Kahneman and noting it is an analogy for an interface, not an architecture claim. - Photos: closed-set moderation tagging is a candidate; test against a human-labeled sample first. - TV: caption-quality scoring maps onto the score question type; this video's own captions garble seven proper nouns and would be a fair test set. - Polyglots: no direct action from this source. - Accessibility: no direct action from this source. - Meta: no direct action from this source. - CLI: no direct action from this source. - Hosting: no direct action from this source. - Tide: score-type classification could rank scan findings by severity; run against known-labeled results before adoption. - Openverse: moderation-queue classification with a human-review threshold is the strongest fit named here. - Themes: review-queue prioritization is a closed-set ranking problem; worth a measured 30-day test. - Plugins: same as Themes, with a mandatory human gate on anything security-adjacent. #### The WPOCC - Programs: no direct action from this source; nearest relevance is that The WP Open Community Collective could track inference cost per decision as a line item in any tooling stipend. - Projects: no direct action from this source. - Funding experiments: if a microgrant covers AI-assisted moderation or triage, require published latency and cost numbers before and after as a condition. ### FAIR - Technical independence: Jev is hosted and network-only. Any FAIR-adjacent tooling that depends on it inherits a single-vendor availability dependency; document that before adoption. - Policies and covenants: if a classifier enters metadata or provenance checks, log the decision and its confidence, not just the outcome, so the call is auditable. - Productized governance artifacts: add a "hosted classifier dependency" row to operator scorecards alongside hosted LLMs. ## 🧪 Hypotheses 1. On a live chat stream of 500 or more messages, per-message parallel classification with Jev reduces median moderation decision latency from the 90-second batch figure Prime reports (00:05:42) to under 2 seconds, with label agreement against human moderators within 5 percentage points of the batch-agent baseline. 2. Using a closed-set classifier to prune agent context (00:07:51) reduces total input tokens to the generative model by at least 30% on long-running coding tasks without lowering task completion rate by more than 2 percentage points. 3. Jev's per-call wall-clock latency from a US data center, measured p50 over 1,000 calls, is under 150 milliseconds; his 100-millisecond figure (00:09:58) holds at p50 but not p95. 4. For UI form-filling tasks, a Jev-driven computer-use loop completes at least 80% of a 50-form suite that a hand-written Playwright suite covers, at lower per-run cost but higher variance across runs. 5. The 30% Balatro win rate he reports (00:06:25) is not reproducible above 20% across 100 seeded runs, because the strategy selection is his script's, not the model's, and the model's per-hand choices are the weak link. 6. Vendor multipliers of 190x and 444x fail to reproduce within a factor of 5 against a named frontier model on a published closed-set benchmark; this is the same hypothesis the Firecrawl and Theo coverage left open. ## 📊 Methodology - **Design for H1.** Replay a recorded chat log from a real live event (2,000 participants, per his event) through three arms: batch generative agent, per-message Jev, per-message frontier LLM with a boolean schema. Record decision latency per message and label agreement against a human-moderated ground truth. Report p50 and p95 latency and macro-F1. - **Design for H2.** Instrument one agent harness twice on the same 50-task suite: with and without a Jev relevance filter over context chunks. Count input tokens billed to the generative model and task completions. Divide spend by completions, not by requests. - **Design for H3.** Fire 1,000 identical 345-token requests from three network locations, time from request issue to parsed response, and separate transit from model time using server-reported timing headers if available. - **Design for H4.** Take 50 forms from real WordPress admin and plugin settings screens. Run the existing Playwright suite and a Jev-driven computer-use loop five times each. Report pass rate, per-run cost from the provider invoice, and run-to-run variance. - **Design for H5.** Seed 100 Balatro runs with fixed decks, hold the strategy script constant, and count wins. Then swap the per-hand decision from Jev to a rule table and count again; the difference isolates the model's contribution. - **Design for H6.** Pull Typesafe's published benchmark if one exists. If none does, that absence is the finding. Any replication must name the baseline model, version, context length, and date. - **Controls throughout.** Fix state text and question wording across arms. Pre-register the agreement threshold. Report confidence calibration with a Brier score, since the video shows single-run confidence numbers only. ## ❓ Questions - Which frontier model, at what context length, is the baseline for 190x and 444x? Prime asks this at 00:02:50 and nobody in the coverage so far has an answer. - What is the real name of the third question type the captions render as "new," and is it a Bernoulli-style probability or a bounded score? - Is the 100-millisecond figure a p50 from his own runs or a number from the vendor page? He states it without a source at 00:09:58. - When a chat message matches none of the offered answers, does Jev return a low-confidence pick or signal out-of-set? The moderation use case depends on this. - Does Typesafe retain or train on input state? Chat logs from a 2,000-person event are personal data. - How many Balatro runs sit behind the 30% win rate, and how much of the strategy lives in his script rather than the model? - What does the computer-use loop at 00:10:43 actually use Jev for: element selection, action choice, or done/not-done? The demo doesn't say. - Why does he treat the same-day Opus release as irrelevant at 00:09:16? Is the argument that decision models and generative models are on different curves, or just fatigue with release cycles? ## 🔮 Future research - **Latency-bound AI decisions as a category.** The moderation story separates two reasons to want a classifier: cost and human-waiting latency. The coverage so far (Firecrawl, Theo) argued cost. Nobody has published p50/p95 latency for hosted decision models under real-time load. - **Classifier-driven context pruning.** "Is this chunk relevant?" as a pre-filter before the generative call is proposed at 00:07:51 with no measurement. The tradeoff between tokens saved and task failures from over-pruning is unstudied. - **Forms as classification.** If UI form-filling decomposes into closed-set decisions, computer-use agents could split into a cheap decision layer and an expensive planning layer. Whether the split holds on messy real-world forms is an open question. - **Injection through moderated content.** Chat messages become model state in his loop. A message written to look like an instruction is an attack on a classifier that the generative-model literature does not yet cover. - **Calibration under adversarial input.** Live chat is adversarial by nature. Whether reported confidence stays meaningful when users learn to game the classifier is untested. - **Reaction-video epistemics.** Prime quotes the vendor's numbers, disclaims them, then runs his own unmeasured demos. What a viewer takes away from that mix, versus from a vendor demo alone, is a media-literacy question worth a small study. ## 💡 Implications - **Practice.** Sort your AI calls by whether a human is waiting on the result. Where one is, a 100-millisecond decision model changes what's feasible even if it saved no money. Where none is, the cost argument from the earlier Jev coverage applies instead. - **Policy.** Per-message automated moderation needs a written escalation threshold and an appeal path. His demo shows 95% and 85% single-run confidences; a policy that acts on 51% without a human gate is a governance failure. - **Security.** Any text a classifier reads is a control surface. Chat, form labels, and scraped UI text all enter state in his demos. Treat them as untrusted. - **Education.** The ball-to-the-face System 1 analogy at 00:00:42 and the Smash Bros message flip at 00:04:59 are two clean teaching devices. Credit Kahneman and name the framing as an interface analogy. - **Funding models.** Input-only billing at about $42 per billion tokens moves spend to volume of questions asked. Teams budgeting by token count need a per-decision meter, and a Jevons-style demand increase is the vendor's stated expectation. - **Vendor literacy.** A creator who reads the vendor's multipliers, says on air that the baseline is unknown, and then shows his own dashboard is a better model of how to cover a release than a vendor demo. It is still not a benchmark. ## 🧠 Learning styles assessment - **Reading level estimate.** Flesch-Kincaid Grade Level on the transcript body: approximately grade 5 to 6, Flesch Reading Ease near 80. Sentences are short and conversational; domain terms (System 1, Bernoulli distribution, Jevons paradox, Playwright, QEMU, context window) push effective comprehension to a working developer. #### 👀 Visual learners - Present: live playground UI with state and questions visible, a token-usage dashboard, Balatro and Doom gameplay, a screenshot feed from a VM, and a code snippet (`if email feels urgent`). - Improve: no diagram of state, questions, and typed answers, which is the one idea that most needs one. The two-question Smash Bros example would fit a single slide. Zoom into the dashboard when reading numbers. #### 🔊 Aural learners - Present: continuous narration in a high-energy register; the argument is fully spoken. - Improve: a 90-second audio summary would serve podcast listeners; the ending cuts mid-demo at 00:11:25 with no spoken recap. #### 🗣️ Verbal learners - Present: definition by negation (not a chatbot, can't write quicksort), the ball-to-the-face analogy, the whip-holder analogy from the manifesto, Jevons paradox explained in one sentence. - Improve: Bernoulli distribution, QEMU, Omarchy, and Playwright go undefined. One clause each would cost 10 seconds. The garbled type name needs the real word spoken clearly. #### 🧩 Kinesthetic learners - Present: he types state and questions live and reruns with a changed message, which is a reproducible experiment a viewer can copy into the Typesafe playground. - Improve: no repository link for the Balatro or Smash Bros scripts, no quickstart. The viewer can imitate the playground steps but not the game loops. #### 🧠 Logical learners - Present: the moderation argument has a clear structure (90-second batch wait, per-message parallel alternative, dead air saved), and the billing arithmetic at 00:08:34 is checkable. - Improve: "a thousand times better" at 00:09:58 has no derivation. Show the two latencies being compared and the logical learner can extend it. #### 🧑‍🤝‍🧑 Social learners - Present: live chat participation, including a viewer suggestion (Steve Faulkner's syntax) adopted on air, and a moderation event that the video is itself about. - Improve: no discussion prompt or community link in the transcript. "Which of your current LLM calls has a human waiting on it?" would convert the argument into a conversation. #### 🧘‍♀️ Solitary learners - Present: 12 minutes, self-contained, rewatchable; the playground demo can be reproduced alone. - Improve: no reflection prompt or self-check. Listing three closed-set decisions in your own codebase would make the video an exercise. #### 🌿 Naturalistic learners - Present: Jevons paradox is itself an energy-economics idea (cheaper coal, more coal burned), named at 00:08:34. - Improve: he uses Jevons to mean "more demand" without the resource-consumption half. Naming the compute-per-decision and energy angle would complete the analogy and is a claim the vendor could substantiate. ## 📊 Accessibility assessment #### Screen reader semantics - Present: YouTube supplies page landmarks and player controls; the auto-transcript is machine-readable. - Improve: the playground UI, the token dashboard, the manifesto text, and both games carry information the narration only partly speaks. WCAG 2.2 AA 1.1.1 (Non-text Content) and 1.2.5 (Audio Description, Prerecorded) are not met for the dashboard and the manifesto page. #### Keyboard navigation - Present: the YouTube player is keyboard operable with visible focus. - Improve: nothing in the video adds barriers. 2.1.1 and 2.4.7 pass via the platform, not the content. #### Contrast and readability - Present: unverifiable from a transcript. The playground and dashboard are likely dark-theme web UI; Balatro and Doom are high-motion game footage. - Improve: small UI text at default player size risks failing 1.4.3 (Contrast Minimum) on mobile. Zoom into the 345-token readout and the pricing line when citing them. #### Text resizing and reflow - Present: the player scales; captions resize independently. - Improve: burned-in UI text and code (`if email feels urgent`) don't resize. 1.4.4 (Resize Text) and 1.4.10 (Reflow) fail for text baked into the frame. #### Media and captions - Present: an auto-generated transcript exists, which is how this analysis was produced. No chapter list is present in the transcript metadata available here. - Improve: the captions garble the product (Jev, Jeff, Jevon), the game (Bellotro, Batro, Bellatro), the vendor's figure (190* 3), the third question type ("new"), Omarchy, QEMU, WorkOS ("work OS"), "auth" ("OOTH"), and a contributor's name. That is a 1.2.2 (Captions, Prerecorded) failure in substance: captions exist and are wrong on nearly every proper noun. Human-corrected captions for 12 minutes are cheap. Doom and Balatro footage should be checked against 2.3.1 (Three Flashes or Below Threshold). #### Forms and interactions - Present: none in the video itself. The sponsor URL is spoken and shown. - Improve: put the sponsor and Typesafe URLs in the description as selectable text. 3.3.2 (Labels or Instructions) applies to the landing pages, not to the video. #### Standards reference - Primary gaps against WCAG 2.2 AA: **1.2.2** (inaccurate captions on proper nouns), **1.2.5** (no audio description for the dashboard, manifesto page, and game state), **1.1.1** (numbers conveyed visually only), **1.4.4 / 1.4.10** (burned-in text), and **2.3.1** to verify for game footage. The captions failure is the highest impact and the easiest fix. ## 🌍 DEIB assessment #### Cultural breadth - Present: soccer named with a nod to "football" elsewhere (00:00:00). Otherwise the references are US developer-culture staples: Smash Bros, Doom, Balatro, Playwright. - Improve: one non-gaming, non-US example of a closed-set decision would widen who sees themselves in the use case. Multilingual moderation would also test the language-understanding claim. #### Inclusive language - Present: second person, direct, no exclusionary terms aimed at groups. - Improve: the demo state at 00:03:33 is a scatological joke, "brother" at 00:03:33 and "bros" at 00:09:16 default to a male audience, "a really dumb model" at 00:07:08 uses ability language for a capability tier, and profanity at 00:09:16 and 00:10:43 is casual. Low severity individually; together they set the room for a narrow audience. #### Neurodiversity support - Present: each segment does one thing and the structure is predictable: what it isn't, how it works, where he'd use it, what it costs. - Improve: two games play on screen while he talks, and a chat interruption derails the argument at 00:09:16. Chapter markers and a slower cut of the demo segments would reduce the processing load. The ending stops mid-demo with no signal that the video is over. #### Representation - Present: one on-camera voice. One chat participant is credited by name for the `feels urgent` idea. The vendor is unnamed as people. - Improve: no skeptical outside voice, and the same-day competing releases are waved off rather than compared. One counter-perspective would improve both credibility and representation. #### Psychological safety - Present: the vendor's numbers are questioned rather than repeated as fact, and his own agent's mistakes are shown, not hidden. The sponsor is clearly separated. - Improve: a viewer is timed out for one hour on air for interrupting (00:09:16), framed as a joke. The moderation demo and the live moderation action sit side by side; a line about how that channel's rules work would make the joke land safely. "If you still haven't dipped your little toesies in the old AIS" at 00:10:43 teases non-adopters. ## ✅ Logical validity #### Strong reasoning - **Causal, 00:05:42.** A batch moderation agent that takes 90 seconds produces 90 seconds of stage dead time; a per-message parallel classifier that judges each message on arrival removes the batch wait. The mechanism is stated and the arithmetic is simple. Clarify by giving the per-message latency he observed, not the vendor's. - **Deductive, 00:08:34.** If billing is input-only and one question costs 345 input tokens, then three million questions cost about 1.035 billion tokens, which at $42 per billion is about $43.50. His "$44" checks out. This is the one number in the video a viewer can verify from what's on screen. - **Anticipatory skepticism, 00:02:50.** He reads the vendor multipliers and immediately says the baseline is unknown. That is the correct move and it separates his coverage from a vendor demo. - **Controlled single-variable change, 00:04:59.** Changing only the chat message from an insult to "smash all the opponents" flips the helpfulness answer while the appropriateness answer stays. Narrow evidence for the narrow claim (state drives the choice). - **Decomposition, 00:11:25.** Explaining Doom as several parallel yes/no questions per frame (dodge or not, fire or hold) shows the mechanism rather than asserting a result, and it matches what appears on screen. #### Fallacies - **Bare assertion (unsupported quantification), 00:09:58.** "This is a thousand times better than almost any other solution." No two latencies are named. *Fix:* state what the "other solution" is and its measured latency, or say "faster than the generative agent I used, which took 90 seconds." - **Hasty generalization, 00:07:08.** "I bet you I could go through this like a hund times faster." A wager about an unrun experiment presented alongside measured demos. *Fix:* run the QA loop both ways and report the time. - **Anecdotal evidence, 00:06:25.** "It will win about like 30% of the time" is an eyeball count over an unstated number of runs, and the strategy is his script's, not the model's. *Fix:* count runs and separate script from model. - **Appeal to novelty and dismissal, 00:09:16.** Same-day releases from other vendors are waved off ("I honestly don't give a rat's ass") without comparison. This isn't a claim about Jev, so it doesn't damage the argument, but it substitutes attitude for evaluation. *Fix:* one sentence on why a decision model and a generative model are on different curves. - **Equivocation on "cheap," 00:07:51 to 00:09:58.** "Dirt cheap" mixes per-token price (a fact from the dashboard) with total cost (which scales with the millions of questions he's proposing). Jevons paradox, which he cites, is precisely the warning that cheap per-unit raises total spend. *Fix:* give the per-decision price and the projected volume together. - **Suppressed baseline (inherited), 00:02:50.** The vendor's 190x and 444x figures have no baseline. He flags this, so it's not his fallacy, but he repeats the numbers, and viewers remember numbers over caveats. *Fix:* don't read the multipliers aloud until a benchmark exists. #### Tone of voice - Loud, fast, profane, and enthusiastic, in the live-stream register his audience expects. Second person and contractions throughout ("you don't program with this thing," "you're only charged for input tokens"). The tone shifts three times: to skeptical at 00:02:50 when reading vendor numbers, to exasperated at 00:07:51 when the Balatro agent won't buy a joker, and to combative at 00:09:16 when a chat member interrupts. The most careful reasoning (moderation latency, billing arithmetic) sits in the middle third, and the loosest claims ("a thousand times better") arrive near the end as he winds up. The 00:01:25 sponsor read is delivered in a deliberately flatter voice and is unmistakable as an ad, which is a disclosure done well. ## 🛡️ DARVO analysis Not applicable to this source. DARVO (Deny, Attack, Reverse Victim and Offender) describes a response pattern to accusation or criticism, and this is a reaction video with no accuser and no defensive posture toward the product. For completeness across the three prongs: **Deny** — nothing is denied; he volunteers that the Balatro agent plays badly (00:06:25), that he mislabeled a question (00:04:16), and that the vendor baseline is unknown (00:02:50). **Attack** — the nearest thing is the dismissal of same-day competing releases at 00:09:16 and a one-hour timeout for a chat member who interrupted; these are dismissals of bystanders, not attacks on a critic, and no critic of Jev appears. **Reverse victim/offender** — no such move. The only self-as-victim moment is a joke about gray hairs and a missing razor at 00:01:25. ## 🎯 Conclusion The durable idea in this video is not the one in the title. Whether Jev is "the best AI release of 2026" rests on vendor multipliers Prime himself declines to endorse at 00:02:50. What survives is narrower and more useful: when a human is waiting on each decision, a closed-set classifier that answers in about 100 milliseconds changes what a system can do, and cost is a second benefit rather than the first. His Smash Bros moderation story at 00:05:42 is the cleanest statement of that anyone has made about this product, because the 90-second batch wait was his own stage time. The numbers you can trust are the ones on his dashboard: 345 input tokens for a two-question moderation check, input-only billing, and a $42-per-billion price that matches Theo's $0.04-per-million reading from five days earlier. The numbers you can't are 190x, 444x, "a thousand times better," "a hundred times faster," and a 30% win rate from an uncounted set of runs. **Next steps for readers:** list the AI calls in your own tooling where a person is synchronously waiting and the answer comes from a known set. That list is where a decision model pays off first, and you can make it today without adopting anything. Then confirm the third question type's real name and the pricing from Typesafe's docs, because the captions on this video get nearly every proper noun wrong. ## 🔄 Refactor Appendix — This Might Be the Best AI Release of 2026 > Density: 2–5 Literature atoms (H3 pad), 0–3 Permanent candidates (H4 pad). > Emitted: 6 Literature, 2 Permanent. Resonance filter applied — the Balatro win rate, > the "hundred times faster" QA estimate, and the same-day-release dismissal were dropped > as unmeasured or off-topic. > > Konik claim-header rules: present tense, declarative, ≤10 words, searchable, unique. > > Atomicity: one declarative sentence per atom. "and"/"also" → split. > > Tag pool: #bmf · #addendum · #xref · #articleseed · #storystem · #share · #aphorisms ### 📚 Literature Split Pad #### [H1] A decision model returns choices not text > [!quote] [00:00:00] > It's not a chatbot, meaning that you cannot give it text and it'll give you text back. **Context & use**: - 00:00:00, the definition by negation that opens the video before any demo. - The cleanest one-line boundary between a classifier endpoint and a generative model; pairs with the Firecrawl and Theo coverage of the same product. **Tags**: #xref  #### [H2] Vendor multipliers arrive without a named baseline > [!quote] [00:02:50] > It's 190* 3 faster and 444 times cheaper. We don't know what model they're talking about. **Context & use**: - 00:02:50, reading Typesafe's manifesto page; the caption garbles 193.6x to "190* 3." - A creator flagging the missing baseline on air is the right move and still leaves the numbers in the viewer's memory. Teaching specimen for reading AI marketing. **Tags**: #xref #share  #### [H3] Batch moderation latency becomes stage dead time > [!quote] [00:05:42] > Well, that takes like a minute and a half. So, that means I have to be funny on stage for a long time. **Context & use**: - 00:05:42, recounting a 2,000-person live Smash Bros event moderated through a generative agent. - The human-waiting case for a fast classifier, stated in cost-to-the-human terms rather than dollars. This is the video's original contribution. **Tags**: #xref #articleseed  #### [H4] A classifier prunes agent context before the LLM call > [!quote] [00:07:08] > Though you can use it say to kind of make your context you give to your agent smaller. **Context & use**: - 00:07:08, one sentence after saying you don't program with Jev; the proposed loop is "is this relevant to the job at hand?" per chunk. - Unmeasured in the video. Worth an experiment because it turns context management into a cheap pre-filter rather than prompt engineering. **Tags**: #xref  #### [H5] Input-only billing makes each question a few hundred tokens > [!quote] [00:08:34] > That Smash Brothers question that I had, 345 input tokens, and you're only charged for input tokens. **Context & use**: - 00:08:34, reading his own usage dashboard; he extends it to about three million questions for $44 and $42 per billion tokens at 00:09:58. - The one verifiable number in the video; matches Theo's $0.04 per million input tokens from 2026-09-20. **Tags**: #xref  #### [H6] Forms are classification problems > [!quote] [00:11:25] > This this is just a classification problem. That's right. Forms, they're just classification problems. **Context & use**: - 00:11:25, explaining why a decision model can drive computer-use testing as a replacement for hand-written Playwright suites. - A reframing of end-to-end UI testing as a sequence of closed-set choices. Asserted, not demonstrated on a real form suite. **Tags**: #xref #articleseed  ### 🧠 Atomic Split Pad #### [P1] A human waiting on each decision is the first case for a fast classifier **Claim** (one sentence, Courtney's voice): When a person is synchronously waiting on each AI decision, a closed-set classifier answering in about 100 milliseconds changes what the system can do, and its lower cost is a second benefit rather than the reason to adopt it. **Evidence pointer**: [[#H3]] · [[#H5]] **Scope & conditions**: Holds when decisions arrive one at a time with a human or a real-time loop blocked on each (live chat moderation, game control, interactive form filling). Breaks for offline batch work, where throughput and cost dominate and a slower generative model is fine. Also breaks if the answer set is open or the classifier has no out-of-set signal. **Links**: - [[A Known Answer Set Calls for a Classifier Not a Generative Model]] — because extends the cost argument with the latency case - [[Agentic Does Not Mean Latency-Sensitive Human-in-the-Loop Does]] — because applies-in the human-waiting condition that this note makes the adoption trigger - [[Route Judgment to the Expensive Model and Bulk Execution to the Cheap One]] — because extends the routing rule to per-item decisions inside a live loop  #### [P2] A cheap per-decision price predicts more decisions not lower spend **Claim** (one sentence, Courtney's voice): When a model decision drops to a fraction of a cent, the budget question shifts from price per call to how many new places the code will now ask, because the vendor's own naming after Jevons paradox predicts demand growth. **Evidence pointer**: [[#H5]] · [[#H4]] **Scope & conditions**: Holds when the decision call is cheap enough to embed in control flow (the `if email feels urgent` pattern at 00:09:58). Breaks when volume is fixed by an external rate, such as one decision per support ticket. The tell that demand will grow: the team starts asking the model questions it previously answered with a hard-coded rule. **Links**: - [[A Cheap Enough Model Turns a Gated Feature Into a Free One]] — because contrasts on direction: that note has cheapness removing a vendor gate, this one has cheapness multiplying a buyer's calls - [[Cheap Generation Raises the Value of Deterministic Verification]] — because extends the "cheap raises the bill elsewhere" pattern from generation to decisions - [[A Vendors Capability Claim Needs a Paired Comparison Not a Demo]] — because evidence-for the rule that the 190x and 444x figures stay claims until a paired run exists  ### ✏️ Fleeting Queue - [F1] Confirm the real name of Jev's third question type (captions: "new," tied to the Bernoulli distribution) and the current input-token price at docs.typesafe.ai before citing either — https://www.youtube.com/watch?v=BHPDsGVciDk #fleeting #inbox - [F2] Article seed: "Latency-bound versus cost-bound AI decisions," using the 90-second live-moderation wait at 00:05:42 as the opening case and Theo's and Firecrawl's cost framing as the contrast #fleeting #inbox #articleseed - [F3] Open question: does per-message classification of live chat create a prompt-injection path where a message shaped like an instruction flips the "appropriate" answer? The Smash Bros loop at 00:04:59 is the concrete instance #fleeting #inbox - [F4] Identify the two same-day releases captioned "Soul 6" and "Opus 55" at 00:09:16 (the second is likely Opus 5.5, already in the vault via Firecrawl's 2026-09-25 video) and note whether either ships a classification tier #fleeting #inbox ### → Things 3 (max 2 actions) - [ ] List the AI calls in my own tooling where a person waits synchronously on a closed-set answer, and log the count — AI cost review, 20 min #things - [ ] Verify Jev's third question-type name and input pricing against docs.typesafe.ai before reusing any figure from this video — research, 10 min #things  📄 **Full text:** [[Full Document Contents/Articles/This Might Be the Best AI Release of 2026.md|This Might Be the Best AI Release of 2026 by ThePrimeagen (The PrimeTime)]] ## Highlights - It's not a chatbot, meaning that you cannot give it text and it'll give you text back. ([View Highlight](https://read.readwise.io/read/01m3fppnjvn52h0y2h2vre7btp)) - Note: [H1] A decision model returns choices not text - Tags: atom/literature/h1 - cheaper. We don't know what ([View Highlight](https://read.readwise.io/read/01m3fpw8atdcb3m0rkpjyy5rn1)) - Note: [H2] Vendor multipliers arrive without a named baseline - Tags: atom/literature/h2 - Well, that takes like a minute and a half. So, that means I have to be funny on stage for a long time. ([View Highlight](https://read.readwise.io/read/01m3fpq0tj6k90syq6ch7cpjby)) - Note: [H3] Batch moderation latency becomes stage dead time - Tags: atom/literature/h3 - Though, you can use it, say, to kind of make your context; you give to your agent smaller ([View Highlight](https://read.readwise.io/read/01m3fpq74s1sceyk2r0y3tw6ed)) - Note: [H4] A classifier prunes agent context before the LLM call - Tags: atom/literature/h3 - That Smash Brothers question that I had, 345 input tokens, and you're only charged for input tokens. ([View Highlight](https://read.readwise.io/read/01m3fpqbex0w3g573g976hr2ym)) - Note: [H5] Input-only billing makes each question a few hundred tokens - Tags: atom/literature/h3