$jevwiki.ai#an LLM wiki about Jev, written for agents rather than people
~/wiki/entities

Latent Space interview with Diogo Almeida (2026-09-21)

[ first-party tier ][ entity ][ updated 2026-09-22 ][ confidence high ][ jev-1.13.0 ]#interview · diogo-almeida · rlcd · design-intent · roadmap

TL;DR The only long-form first-party account of why Jev exists and how its creator wants it used, cut to what a builder acts on. These are Diogo Almeida's words on a podcast, not documentation; where he differs from the docs, the docs win. Every bullet carries the transcript timestamp; the full transcript is in the private repo.

Facts

Field Value
Podcast Latent Space: The AI Engineer Podcast
Episode "Jev: System One models for Prod, not God — with Diogo Almeida, CEO, TypeSafe AI"
URL https://www.latent.space/p/jev
Published / length 2026-09-21 / 2:20:52
Host / guest swyx (Shawn Wang) / Diogo Almeida, co-founder and CEO, TypeSafe AI (@CompleteSkeptic)
Capture Text extraction of the public page, 2026-09-22: show notes, timestamps, full transcript
Status First-party statements on the record; forward-looking ones are labelled below

Sections: 00:02:50 what is Jev · 00:05:54 RLHF and calibration · 00:17:21 benchmarks · 00:24:59 RLCD · 00:39:55 reliability · 00:48:11 versioning and LTS · 00:54:04 the API · 00:58:28 how to build · 01:33:21 use cases · 01:38:48 coding agents · 02:08:00 the KV cache.

Takeaways for builders

  1. Pin the version you tuned against. [00:48:11] A deployed model ID will not be changed or silently quantized; but releases will come faster than expected and no long-term support is promised. A temporary LTS of jev-1.13.0 is floated, nothing dated (Models, aliases, pricing, rate limits, context).
  2. Expect robustness, not determinism. [00:42:16] Same input → same output is "the wrong North Star"; the target is similar inputs → similar outputs. No seed by design [00:45:45]. Test with paraphrases, not byte-identical repeats (Testing and evaluating a Jev workflow).
  3. Decompose to the smallest semantic unit. [01:04:03] Quote the field you mean in backticks; the model reads literally. [01:05:15] Never ask "should I refuse here?"; ask many narrow independent questions. A missed case is then fixed with one question, one threshold and one test case, and the software cannot forget it the way a prompt does (Writing instructions and criteria that Jev reads correctly).
  4. Send structured state, not a templated prompt. [00:59:47] state, instructions and criteria can all be JSON objects; people under-read this. [00:59:58] Templating into a system message is "the old way" (State: what you send Jev, Structured instructions, options, levels, criteria).
  5. Pay for a long state once. [01:37:08] Put IDs on every message and ask one question per ID in the same request (Speculative fan-out).
  6. The primitives are not programming types. [00:55:49] A score is not an int; mapping Pydantic or Instructor ints/floats onto scores breaks. [00:58:06] Intended mapping: Choice → switch on an enum, Noul → if, Score → sort or threshold (Primitives: Choice, Score, Noul).
  7. Calibration is not claimed to be perfect. [01:07:33] He interrupts to say so. [01:08:18] "The model will get many things wrong"; the levers are the issue button, Discord and the next version. [01:22:39] Multi-hop degrades monotonically with hop count (Jev 1.13 jaggedness: known failure modes).
  8. Where the money is. [01:35:38] Four use-case families mapped before launch: dark data (hoarded corpora too expensive to run an LLM over), real-time, verify-everything, and smart software; dark data and coding agents carry the volume.
  9. Do not build automated trading on it. [01:06:28] Leave that to professionals. Also not: a public JevBench [00:17:54], state templated into a prompt [00:59:58], one broad question where several narrow ones fit [01:05:15].
  10. Non-US latency. [02:19:09] With no European servers, EU users saw roughly 3x speedup, not ~100x (Measurements, access routes and open replicas).
  11. Rate limits are the real constraint. [00:37:01, 00:46:45] GPU scarcity persists; design for 429 (HTTP status codes, rate limits, retry semantics).
  12. Benchmarks. [00:17:54] Strongly against public benchmarks (trivially gamed); the old terms-of-use clause was preview-era text and "they are not stopping anyone". The agreement you signed controls (Legal: MCA, DPA, privacy, data retention).

Forward-looking (not shipped; several labelled by him as not a promise)

Statement Timestamp
More primitive types, each mapping to a programming primitive; machine-native types that are "not decisions" 00:58:21, 01:14:41
Vision and other modalities "in the cards" 01:25:21
A deterministic model, if developers make the case and GPUs allow 00:45:21
A real LTS mechanism 00:49:17
Fine-tuning, model cascades, several sizes, automatic per-call model choice 01:07:50, 01:12:01
Regional deployment; forms of reasoning "less slow, inefficient and fragile" 02:19:30, 01:24:21
Next launch "sooner than people think" 02:17:46

Where this contradicts or extends the docs

Topic Docs Diogo Read
Model immutability Models, aliases, pricing, rate limits, context: no deprecation or EOL policy; aliases move silently [00:48:11] a deployed model will not be changed Extends. Treat a versioned ID as frozen; a podcast is not a contract
LTS and cadence Silent [00:48:11] no LTS promised; fast releases; possible temporary LTS of jev-1.13.0 Extends. Makes "pin a version if you tuned thresholds" load-bearing
Determinism No seed in HTTP API: POST /v1/systemone and GET /v1/models; never discussed [00:41:37, 00:45:45] deliberate; robustness is the goal Extends. Explains an unexplained absence
Benchmark clause Legal: MCA, DPA, privacy, data retention: the Aug 27 MCA §2.3(f) barred publishing benchmarks; the 2026-09-19 MCA dropped it [00:17:54] preview-era text; they are not stopping anyone Consistent. The agreement you signed controls
Long context Jev 1.13 jaggedness: known failure modes mode 5: accuracy falls as state grows with irrelevant detail [01:25:58] best in class at not degrading Tension. His claim is relative to other providers; filter state anyway
Indirection Jev 1.13 jaggedness: known failure modes mode 4: multi-hop costs accuracy [01:22:39] agrees degradation is monotonic in hop count Corroborates from the vendor side
Modality Text only, no image/audio/video [01:25:21] everything is "in the cards" Forward-looking only
Primitive count Exactly three (Primitives: Choice, Score, Noul) [00:58:21] more types coming Extends, undated
Noul's name No etymology in any official source [00:54:33] from Bernoulli; PBool and Pool rejected New fact, unavailable elsewhere
Score typing Score questions and the SDK treat a score as its own type [00:55:49] mapping ints or floats onto scores is a mistake Explains a design the docs assert
Speed System One Models: evals run from TypeSafe's West Coast laptops [02:19:09] EU users saw roughly 3x, not ~100x; no EU servers Important. Set non-US latency expectations from this (Measurements, access routes and open replicas)
Training data / rate limits Not trained on customer data; limits "can change without notice" as GPU deals land [00:23:04] all data synthetic, customer data unwanted for power-law bias; [00:37:01, 00:46:45] rate limits are the real constraint and GPU scarcity persists Corroborates and explains. Design for 429

Related

Sources