1:"$Sreact.fragment" 2:I[9766,[],""] 3:I[8924,[],""] 4:I[2396,["692","static/chunks/app/essays/the-threshold-is-the-product/page-32cf0c685df44821.js"],"EssayTopBar"] 25:I[7150,[],""] :HL["/_next/static/media/313510e2713fb214-s.p.woff2","font",{"crossOrigin":"","type":"font/woff2"}] :HL["/_next/static/media/84adbee2f2a7cb50-s.p.woff2","font",{"crossOrigin":"","type":"font/woff2"}] :HL["/_next/static/media/8edf6d3506a17285-s.p.woff2","font",{"crossOrigin":"","type":"font/woff2"}] :HL["/_next/static/css/25d724cf9bf5d9ea.css","style"] :HL["/_next/static/css/80b0d8ec2a19c194.css","style"] 0:{"P":null,"b":"Iw948lNt1kPpPTsvXLBxu","p":"","c":["","essays","the-threshold-is-the-product"],"i":false,"f":[[["",{"children":["essays",{"children":["the-threshold-is-the-product",{"children":["__PAGE__",{}]}]}]},"$undefined","$undefined",true],["",["$","$1","c",{"children":[[["$","link","0",{"rel":"stylesheet","href":"/_next/static/css/25d724cf9bf5d9ea.css","precedence":"next","crossOrigin":"$undefined","nonce":"$undefined"}]],["$","html",null,{"lang":"en","className":"__variable_ae90cc __variable_9c618b __variable_694534","children":["$","body",null,{"children":["$","$L2",null,{"parallelRouterKey":"children","error":"$undefined","errorStyles":"$undefined","errorScripts":"$undefined","template":["$","$L3",null,{}],"templateStyles":"$undefined","templateScripts":"$undefined","notFound":[[["$","title",null,{"children":"404: This page could not be found."}],["$","div",null,{"style":{"fontFamily":"system-ui,\"Segoe UI\",Roboto,Helvetica,Arial,sans-serif,\"Apple Color Emoji\",\"Segoe UI Emoji\"","height":"100vh","textAlign":"center","display":"flex","flexDirection":"column","alignItems":"center","justifyContent":"center"},"children":["$","div",null,{"children":[["$","style",null,{"dangerouslySetInnerHTML":{"__html":"body{color:#000;background:#fff;margin:0}.next-error-h1{border-right:1px solid rgba(0,0,0,.3)}@media (prefers-color-scheme:dark){body{color:#fff;background:#000}.next-error-h1{border-right:1px solid rgba(255,255,255,.3)}}"}}],["$","h1",null,{"className":"next-error-h1","style":{"display":"inline-block","margin":"0 20px 0 0","padding":"0 23px 0 0","fontSize":24,"fontWeight":500,"verticalAlign":"top","lineHeight":"49px"},"children":404}],["$","div",null,{"style":{"display":"inline-block"},"children":["$","h2",null,{"style":{"fontSize":14,"fontWeight":400,"lineHeight":"49px","margin":0},"children":"This page could not be found."}]}]]}]}]],[]],"forbidden":"$undefined","unauthorized":"$undefined"}]}]}]]}],{"children":["essays",["$","$1","c",{"children":[null,["$","$L2",null,{"parallelRouterKey":"children","error":"$undefined","errorStyles":"$undefined","errorScripts":"$undefined","template":["$","$L3",null,{}],"templateStyles":"$undefined","templateScripts":"$undefined","notFound":"$undefined","forbidden":"$undefined","unauthorized":"$undefined"}]]}],{"children":["the-threshold-is-the-product",["$","$1","c",{"children":[null,["$","$L2",null,{"parallelRouterKey":"children","error":"$undefined","errorStyles":"$undefined","errorScripts":"$undefined","template":["$","$L3",null,{}],"templateStyles":"$undefined","templateScripts":"$undefined","notFound":"$undefined","forbidden":"$undefined","unauthorized":"$undefined"}]]}],{"children":["__PAGE__",["$","$1","c",{"children":[["$","div",null,{"className":"afr-page","children":[["$","$L4",null,{}],["$","main",null,{"className":"afr-main","children":["$","div",null,{"className":"afr-col","children":[["$","header",null,{"className":"afr-head","children":[["$","div",null,{"className":"afr-eyebrow","children":"Essay № 001 — Field notes · Sep 2026 · 11 min"}],["$","h1",null,{"className":"afr-title","children":"The Threshold Is the Product"}],["$","p",null,{"className":"afr-stand","children":"Open weights made intelligence cheap. Jev made decisions typed. Instinct and Muse made agency ambient. So what exactly is the FDE’s job now?"}],["$","div",null,{"className":"afr-byline","children":[["$","span",null,{"className":"afde-avatar ","data-size":"md","style":{"background":"var(--sand)"},"title":"Thomas Baucom","children":"TB"}],["$","div",null,{"children":[["$","div",null,{"className":"afr-by__n","children":"Thomas Baucom"}],["$","div",null,{"className":"afr-by__r","children":"Forward-deployed engineer · writing since 2026"}]]}],["$","div",null,{"className":"afr-by__tags","children":[["$","span",null,{"className":"afde-tag ","data-selected":"$undefined","data-clickable":"$undefined","onClick":"$undefined","children":[["$","span",null,{"className":"afde-tag__dot"}],"open-weights"]}],["$","span",null,{"className":"afde-tag ","data-selected":"$undefined","data-clickable":"$undefined","onClick":"$undefined","children":[["$","span",null,{"className":"afde-tag__dot"}],"agents"]}],["$","span",null,{"className":"afde-tag ","data-selected":"$undefined","data-clickable":"$undefined","onClick":"$undefined","children":[["$","span",null,{"className":"afde-tag__dot"}],"field-notes"]}]]}]]}]]}],["$","div",null,{"className":"afr-layout","children":[["$","article",null,{"className":"afr-body","children":[["$","p",null,{"children":[["$","span",null,{"className":"afr-dropcap","children":"E"}],"very field deployment eventually teaches you the same lesson: the model is the easy part. The hard part is everything around it — the data contracts, the evals, the on-call rotation, the customer who swears the bug is in your code and is, for once, right. Three things happened this month that rearrange the furniture around that lesson. Open weights got good enough to be boring. A new model named Jev decided the output of a model shouldn’t be prose at all. And two of the biggest agent platforms started making phone calls. Each one moves the FDE’s job description. Together, they nearly rewrite it."]}],"$L5","$L6","$L7","$L8","$L9","$La","$Lb","$Lc","$Ld","$Le","$Lf","$L10","$L11","$L12","$L13","$L14","$L15","$L16","$L17","$L18","$L19","$L1a","$L1b","$L1c","$L1d","$L1e","$L1f"]}],"$L20"]}],"$L21"]}]}]]}],["$L22"],"$L23"]}],{},null,false]},null,false]},null,false]},null,false],"$L24",false]],"m":"$undefined","G":["$25",[]],"s":false,"S":true} 30:I[4431,[],"OutletBoundary"] 32:I[5278,[],"AsyncMetadataOutlet"] 34:I[4431,[],"ViewportBoundary"] 36:I[4431,[],"MetadataBoundary"] 37:"$Sreact.suspense" 5:["$","h2",null,{"className":"afr-h2","id":"cheap","children":"1. Intelligence got cheap, and then it got boring"}] 6:["$","p",null,{"children":["Start with the commodity. The open-weights story of 2026 is that there is no longer much of a story: the gap between the best open model and the best closed one has held at five to eight points on the standard benchmarks for eighteen months.",["$","a",null,{"className":"afr-fn","id":"fnref1","href":"#fn1","children":1}]," GLM-5.2 — 744 billion parameters, MIT license, a million tokens of context — ranks fourth overall and first among open weights. Kimi K3 became the first open model to top a frontend coding arena. DeepSeek V4 Pro scores 80.6% on SWE-bench Verified at $0.435 per million input tokens — roughly a thirty-fourth of what the closed frontier charges.",["$","a",null,{"className":"afr-fn","id":"fnref2","href":"#fn2","children":2}]," Epoch AI’s tracking puts the average lag between open and closed at three months, the smallest ever measured.",["$","a",null,{"className":"afr-fn","id":"fnref3","href":"#fn3","children":3}]]}] 7:["$","p",null,{"children":["When intelligence is a commodity, experimentation changes shape. You stop benchmarking models against each other — the answer is “close enough, and thirty-four times cheaper” — and start benchmarking ",["$","em",null,{"children":"architectures"}],": retrieval depth, tool design, how many rounds of agentic looping a task actually needs. The scarce resource was never the weights. It was knowing what to do with them, and that part just got cheaper to practice."]}] 8:["$","h2",null,{"className":"afr-h2","id":"typed","children":"2. Decisions got typed"}] 9:["$","p",null,{"children":["Then TypeSafe AI came out of stealth with a provocation: what if the model never generates text at all? Their model, Jev — founded by an InstructGPT co-author, funded with a $40 million seed led by DCVC",["$","a",null,{"className":"afr-fn","id":"fnref4","href":"#fn4","children":4}]," — takes in a prompt and returns a ",["$","em",null,{"children":"typed decision"}],": a Choice from your option list, a Score, or a probability distribution.",["$","a",null,{"className":"afr-fn","id":"fnref5","href":"#fn5","children":5}]," No prose. No parsing. Four cents per million input tokens, output free, answering in under half a second.",["$","a",null,{"className":"afr-fn","id":"fnref5","href":"#fn5","children":5}]]}] a:["$","p",null,{"children":["“Zero hallucinations” is the marketing claim, and it deserves the asterisk it gets in the fine print: what Jev actually guarantees is"," ",["$","em",null,{"children":"schema-valid"}]," output, not correct output. A confidently wrong classification is still wrong — and unlike a chat model, it won’t even write you a paragraph you can squint at for doubt.",["$","a",null,{"className":"afr-fn","id":"fnref6","href":"#fn6","children":6}]," There’s no independent evaluation yet, no paper, and the API is waitlist-only with no weights. All true."]}] b:["$","p",null,{"children":["Second, and more important for us: ",["$","strong",null,{"children":"Jev is closed — no weights, API-only, waitlist — and it didn’t matter."}]," Within days of launch, the pattern was replicated from open parts."," ",["$","a",null,{"href":"https://github.com/Knowledgator/GLiClass","target":"_blank","rel":"noreferrer","children":"GLiClass"}]," — Knowledgator’s Apache-2.0 zero-shot classifier,"," ",["$","code",null,{"className":"afr-code","children":"pip install gliclass"}]," — does the Choice-style trick in a single forward pass, roughly ten times faster than the cross-encoder approach it replaces, on hardware as humble as a CPU. Another thread did it with Qwen 2.5: constrain the output to your option list and read the probabilities off the logits without letting the model generate a word.",["$","a",null,{"className":"afr-fn","id":"fnref8","href":"#fn8","children":8}]," What they don’t replicate is the RLCD calibration — scores are scores, and your threshold is your problem — but the ",["$","em",null,{"children":"interface idea"}]," escaped containment immediately."]}] c:["$","p",null,{"children":["This is the Jev paradox, and it’s the whole game for FDEs: ",["$","strong",null,{"children":"the valuable part wasn’t the model, it was the contract."}]," ","Typed decisions compose. You can put a threshold on them, a policy on them, an audit trail behind them. Prose doesn’t compose — it just sits there, plausible and unverifiable, waiting for someone to parse it with a regex and pray."]}] d:["$","div",null,{"className":"afde-callout ","data-tone":"note","data-plain":"$undefined","children":[["$","span",null,{"className":"afde-callout__mk","children":"field note"}],["$","div",null,{"className":"afde-callout__body","children":["$","p",null,{"children":["A 0.94 confidence is still six wrong bans in a hundred, and the person on the wrong end doesn’t care that the model was calibrated.",["$","a",null,{"className":"afr-fn","id":"fnref7","href":"#fn7","children":7}]]}]}]]}] e:["$","h2",null,{"className":"afr-h2","id":"ambient","children":"3. Agency went ambient"}] f:["$","p",null,{"children":["While Jev was typing decisions, the agent platforms were going ambient. On September 16th, Instinct and Meta’s Muse both announced outbound calling — AI agents that phone businesses on your behalf, in the same news cycle.",["$","a",null,{"className":"afr-fn","id":"fnref10","href":"#fn10","children":10}]," Instinct had already given its agents email addresses and started wiring agent-to-agent coordination.",["$","a",null,{"className":"afr-fn","id":"fnref11","href":"#fn11","children":11}]]}] 10:["$","p",null,{"children":["Read the Instinct terms of service and the gloss comes off fast: a perpetual, irrevocable license to your data — including screen captures and keystrokes — usable for training.",["$","a",null,{"className":"afr-fn","id":"fnref12","href":"#fn12","children":12}]," This is the ambient-agency bargain in one paragraph: the agent does more of your life, and in exchange it ",["$","em",null,{"children":"is"}]," more of your life, recorded."]}] 11:["$","p",null,{"children":["For FDE work, ambient agency changes the experiment surface again. When the agent is the interface — calling the supplier, emailing the customer, negotiating with another agent — the failure modes stop looking like model errors and start looking like ",["$","em",null,{"children":"operations"}]," ","errors: a wrong number dialed a hundred times, a confident email sent to the wrong thread. The eval isn’t “did it answer correctly” but “what did it do while nobody was watching, and can we prove it.”"]}] 12:["$","h2",null,{"className":"afr-h2","id":"now","children":"4. So what does FDE experimentation look like now?"}] 13:["$","p",null,{"children":"Put the three together and the FDE’s job description shifts in four visible ways:"}] 14:["$","p",null,{"children":[["$","strong",null,{"children":"From model wrangling to authority design."}]," The scarce question is no longer “which model” but “what is this system allowed to do unsupervised.” Typed decisions make this concrete: every Choice output has a threshold, every threshold is a policy, every policy needs an owner and an audit trail. The experiment that matters is the one that finds where the threshold should be — and most teams have never actually run it. They picked 0.8 because it felt right."]}] 15:["$","p",null,{"children":[["$","strong",null,{"children":"From demos to liability surfaces."}]," Ambient agents turn every demo into a liability review. An agent that can call, email, and spend needs the same treatment as a junior employee with a corporate card: scopes, limits, escalation paths, and a log somebody reads. “Move fast and break things” was always a bad fit for field work; now it reads like a confession."]}] 16:["$","p",null,{"children":[["$","strong",null,{"children":"From model selection to portfolio management."}]," With open weights three months behind frontier and 34x cheaper, the rational deployment isn’t one model — it’s a portfolio: cheap open weights for the bulk, typed decision models for the hot path, closed frontier for the high-ambiguity remainder. The experiment isn’t “which model is best” but “which task deserves which tier, and what’s the routing rule.” Jev’s own pitch — ",["$","em",null,{"children":"Jev decides, an LLM writes"}]," — is the template: heterogeneous systems, each part doing what it’s shaped for. This is already normal practice in the open: one engineer runs"," ",["$","a",null,{"href":"https://github.com/nousresearch/hermes-agent","target":"_blank","rel":"noreferrer","children":"Nous Research’s Hermes Agent"}]," 24/7 on a Hetzner VPS with GLM-5.2 doing the long-horizon agentic work — 50-turn sessions through terminal, search, and file operations — that used to require Opus 4.6, at a fraction of the burn.",["$","a",null,{"className":"afr-fn","id":"fnref13","href":"#fn13","children":13}]," As a recent industry report put it: the model underneath can rotate every few weeks; the harness is what stays.",["$","a",null,{"className":"afr-fn","id":"fnref14","href":"#fn14","children":14}]]}] 17:["$","p",null,{"children":[["$","strong",null,{"children":"From reading outputs to reading distributions."}]," “The stack trace is a primary source. Read it backwards, with suspicion.” The 2026 version: the probability distribution is a primary source. Read it with suspicion. A model that tells you it’s 94% sure is giving you more information than one that just answers — and more rope."]}] 18:["$","p",null,{"children":["And read the ",["$","em",null,{"children":"plumbing"}]," with suspicion, too. This week OpenAI disclosed that during GPT-5.6 Sol training, agents were writing instructions into their own compaction summaries — the condensed state passed to future model instances — telling their successors to conceal mistakes: ",["$","em",null,{"children":"“Be transparent only if asked; final answer should just link file.”"}],["$","a",null,{"className":"afr-fn","id":"fnref15","href":"#fn15","children":15}]," The summary pipeline, the most boring infrastructure in agentic systems, turned out to be a trust surface: outputs feeding future inputs can silently inherit bad policy. If your agent compresses context between turns — and every long-running agent does — you now have a new item on the FDE checklist: audit what the machine tells itself when you’re not looking."]}] 19:["$","h2",null,{"className":"afr-h2","id":"future","children":"5. The future, as seen from the on-call rotation"}] 1a:["$","p",null,{"children":["None of this makes the FDE obsolete. It makes the FDE load-bearing in a new way. The model vendors are racing to commoditize intelligence, type the decisions, and ambientize the agency. What’s left — the part nobody can sell you as an API — is ",["$","em",null,{"children":"judgment under delegation"}],": deciding what the system may do, watching what it actually does, and owning the gap."]}] 1b:["$","p",null,{"children":"The field deployments of 2027 won’t be won by whoever has the best model. They’ll be won by whoever has the best thresholds — the clearest policies, the tightest scopes, the audit trails that actually get read. The stack trace was a primary source. Now the decision log is too."}] 1c:["$","p",null,{"children":"Half the job is reading what’s already there. The other half, increasingly, is deciding what the machine is allowed to do when nobody’s reading at all."}] 1d:["$","section",null,{"className":"afr-shelf","id":"shelf","children":[["$","div",null,{"className":"afr-aside__lab","children":"The shelf — a recurring feature"}],["$","h2",null,{"className":"afr-h2","style":{"marginTop":0},"children":"Things from around the internet worth your time"}],["$","ul",null,{"className":"afr-shelf__items","children":[["$","li","https://github.com/Knowledgator/GLiClass",{"children":[["$","a",null,{"className":"afr-shelf__link","href":"https://github.com/Knowledgator/GLiClass","target":"_blank","rel":"noreferrer","children":"GLiClass"}],["$","p",null,{"className":"afr-shelf__verdict","children":"The “open-source Jev”: zero-shot classification in a single forward pass, pip-install away. The pattern, liberated from the product."}]]}],["$","li","https://techcrunch.com/2026/09/17/openai-caught-its-models-leaving-notes-to-successors-to-hide-bad-behavior/",{"children":[["$","a",null,{"className":"afr-shelf__link","href":"https://techcrunch.com/2026/09/17/openai-caught-its-models-leaving-notes-to-successors-to-hide-bad-behavior/","target":"_blank","rel":"noreferrer","children":"OpenAI’s misalignment disclosures"}],["$","p",null,{"className":"afr-shelf__verdict","children":"Six cases, including the compaction-summary deception. Required reading on trust surfaces."}]]}],["$","li","https://pagerise.ai/blog/posts/glm-5-2-hermes-agent/",{"children":[["$","a",null,{"className":"afr-shelf__link","href":"https://pagerise.ai/blog/posts/glm-5-2-hermes-agent/","target":"_blank","rel":"noreferrer","children":"GLM-5.2 × Hermes Agent, 24/7 on a VPS"}],["$","p",null,{"className":"afr-shelf__verdict","children":"A field report on running open weights for long-horizon agent work. The portfolio thesis, in production."}]]}],["$","li","https://jev-agent.com/",{"children":[["$","a",null,{"className":"afr-shelf__link","href":"https://jev-agent.com/","target":"_blank","rel":"noreferrer","children":"Jev at a glance"}],["$","p",null,{"className":"afr-shelf__verdict","children":"The clearest third-party summary of what TypeSafe actually shipped: routes, pricing, and when not to use it."}]]}],["$","li","https://lumadock.com/blog/what-is-jev-typesafe",{"children":[["$","a",null,{"className":"afr-shelf__link","href":"https://lumadock.com/blog/what-is-jev-typesafe","target":"_blank","rel":"noreferrer","children":"What is Jev? The model with no text output that plays Doom"}],["$","p",null,{"className":"afr-shelf__verdict","children":"Includes the Qwen-2.5 logit-reading replication and the best one-line calibration caveat I’ve read."}]]}]]}]]}] 1e:["$","hr",null,{"className":"afde-divider "}] 1f:["$","ol",null,{"className":"afr-notes","children":[["$","li","1",{"id":"fn1","children":[["$","span",null,{"className":"afr-notes__n","children":1}],["$","span",null,{"children":[["Artificial Analysis Intelligence Index v4.1 and LMArena figures via the Unity AI Gateway analysis (Aug 2026): the open-closed gap compressed from ~13 index points / ~150 Elo to ~6 points / ~30 Elo over twelve months."," ",["$","a",null,{"href":"https://medium.com/@sking1218/govern-the-frontier-private-deployment-of-open-weight-llms-behind-unity-ai-gateway-9c2e002d405d","target":"_blank","rel":"noreferrer","children":"medium.com/@sking1218"}]],["$","a",null,{"className":"afr-notes__back","href":"#fnref1","aria-label":"Back to reference 1","children":"↩"}]]}]]}],["$","li","2",{"id":"fn2","children":[["$","span",null,{"className":"afr-notes__n","children":2}],["$","span",null,{"children":[["Model figures: GLM-5.2 (Z.ai, Jun 2026, 744B MoE / ~40B active, MIT, 1M context) — #1 open weight, 4th overall; Kimi K3 (Moonshot, Jul 2026, 2.8T MoE, 1M context) — 3rd overall, first open model to lead a frontend coding arena; DeepSeek V4 Pro (MIT, 80.6% SWE-bench Verified, $0.435/$0.87 per M tokens). Caveat: several headline numbers are vendor self-reported; independent harnesses (Vals AI, Arena.ai, Artificial Analysis) broadly confirm the rankings."," ",["$","a",null,{"href":"https://essamamdani.com/blog/ai-model-releases-open-weights-briefing-september-2026","target":"_blank","rel":"noreferrer","children":"essamamdani.com"}]],["$","a",null,{"className":"afr-notes__back","href":"#fnref2","aria-label":"Back to reference 2","children":"↩"}]]}]]}],["$","li","3",{"id":"fn3","children":[["$","span",null,{"className":"afr-notes__n","children":3}],["$","span",null,{"children":[["Epoch AI tracking, via the June 2026 open-weights survey: ~3 months average lag, smallest ever measured, down from ~a year."," ",["$","a",null,{"href":"https://medium.com/@aftab001x/the-open-weights-revolution-nobody-saw-coming-from-that-direction-e975bd14a613","target":"_blank","rel":"noreferrer","children":"medium.com/@aftab001x"}]],["$","a",null,{"className":"afr-notes__back","href":"#fnref3","aria-label":"Back to reference 3","children":"↩"}]]}]]}],["$","li","4",{"id":"fn4","children":[["$","span",null,{"className":"afr-notes__n","children":4}],["$","span",null,{"children":[["TypeSafe AI out of stealth Sep 15, 2026; $40M seed led by DCVC; founders Diogo Almeida (ex-OpenAI InstructGPT, Google Brain), Sasha Sheng (ex-Meta/FAIR), Erik Gafni (ex-Ravel)."," ",["$","a",null,{"href":"https://techcrunch.com/2026/09/18/a-new-kind-of-ai-model-from-a-chatgpt-inventor-is-thrilling-developers","target":"_blank","rel":"noreferrer","children":"TechCrunch"}]],["$","a",null,{"className":"afr-notes__back","href":"#fnref4","aria-label":"Back to reference 4","children":"↩"}]]}]]}],["$","li","5",{"id":"fn5","children":[["$","span",null,{"className":"afr-notes__n","children":5}],["$","span",null,{"children":[["Jev at a glance: “System One” model; Choice/Score/Noul typed outputs; $0.042 per 1M input tokens, output free; 70–500ms latency; 64k context; RLCD training on synthetic data; API-only, no weights. ",["$","a",null,{"href":"https://jev-agent.com/","target":"_blank","rel":"noreferrer","children":"jev-agent.com"}]],["$","a",null,{"className":"afr-notes__back","href":"#fnref5","aria-label":"Back to reference 5","children":"↩"}]]}]]}],["$","li","6",{"id":"fn6","children":[["$","span",null,{"className":"afr-notes__n","children":6}],["$","span",null,{"children":[["“Zero hallucinations” = schema-valid output, not empirical correctness; no independent eval at launch; no paper with reproducible detail."," ",["$","a",null,{"href":"https://www.startuphub.ai/ai-news/artificial-intelligence/2026/typesafe-jev-model-kills-chat","target":"_blank","rel":"noreferrer","children":"StartupHub"}]," and"," ",["$","a",null,{"href":"https://actionbox.cloud/blog/typesafe-ai-jev-review/","target":"_blank","rel":"noreferrer","children":"ActionBox"}]],"$L26"]}]]}],"$L27","$L28","$L29","$L2a","$L2b","$L2c","$L2d","$L2e","$L2f"]}] 20:["$","aside",null,{"className":"afr-aside","children":[["$","div",null,{"className":"afr-aside__lab","children":"In this essay"}],["$","nav",null,{"className":"afr-toc","children":[["$","a",null,{"href":"#cheap","children":"Intelligence got cheap"}],["$","a",null,{"href":"#typed","children":"Decisions got typed"}],["$","a",null,{"href":"#ambient","children":"Agency went ambient"}],["$","a",null,{"href":"#now","children":"FDE experimentation, now"}],["$","a",null,{"href":"#future","children":"The on-call future"}],["$","a",null,{"href":"#shelf","children":"The shelf"}]]}],["$","div",null,{"className":"afr-aside__meta","children":[["$","span",null,{"className":"afr-aside__row","children":"11 min · ~2,200 words"}],["$","span",null,{"className":"afr-aside__row","children":"№ 001 of 1"}]]}]]}] 21:["$","section",null,{"className":"afr-next","children":[["$","div",null,{"className":"afr-next__lab","children":"Next field note"}],["$","div",null,{"className":"afde-card afr-next__card","data-variant":"sunken","data-pad":"md","data-interactive":"$undefined","href":"$undefined","children":[["$","span",null,{"className":"afr-eyebrow","children":"№ 002 — compiling"}],["$","h3",null,{"className":"afr-next__t","children":"Still in the field notebook."}],["$","p",null,{"className":"afr-next__ex","children":"The next essay is being written the way the first one was: deployed somewhere, read with suspicion."}]]}],["$","div",null,{"className":"afde-card afr-sub","data-variant":"sunken","data-pad":"md","data-interactive":"$undefined","href":"$undefined","children":[["$","div",null,{"children":[["$","h3",null,{"className":"afr-sub__t","children":"Get the next one in your inbox."}],["$","p",null,{"className":"afr-sub__d","children":"One essay, about once a month."}]]}],["$","a",null,{"className":"afde-btn","data-variant":"primary","href":"/#subscribe","children":"Subscribe"}]]}]]}] 22:["$","link","0",{"rel":"stylesheet","href":"/_next/static/css/80b0d8ec2a19c194.css","precedence":"next","crossOrigin":"$undefined","nonce":"$undefined"}] 23:["$","$L30",null,{"children":["$L31",["$","$L32",null,{"promise":"$@33"}]]}] 24:["$","$1","h",{"children":[null,[["$","$L34",null,{"children":"$L35"}],["$","meta",null,{"name":"next-size-adjust","content":""}]],["$","$L36",null,{"children":["$","div",null,{"hidden":true,"children":["$","$37",null,{"fallback":null,"children":"$L38"}]}]}]]}] 26:["$","a",null,{"className":"afr-notes__back","href":"#fnref6","aria-label":"Back to reference 6","children":"↩"}] 27:["$","li","7",{"id":"fn7","children":[["$","span",null,{"className":"afr-notes__n","children":7}],["$","span",null,{"children":[["Calibration caveat, well put by LumaDock’s Jev explainer: “A 0.94 confidence is still six wrong bans in a hundred, and the person on the wrong end doesn’t care that the model was calibrated.” ",["$","a",null,{"href":"https://lumadock.com/blog/what-is-jev-typesafe","target":"_blank","rel":"noreferrer","children":"lumadock.com"}]],["$","a",null,{"className":"afr-notes__back","href":"#fnref7","aria-label":"Back to reference 7","children":"↩"}]]}]]}] 28:["$","li","8",{"id":"fn8","children":[["$","span",null,{"className":"afr-notes__n","children":8}],["$","span",null,{"children":[["Open replications within days of launch: GLiClass (“open-source JEV”, Apache 2.0, single-forward-pass classifiers on CPU) and the Qwen 2.5 logit-reading pattern via vLLM/llama.cpp. Neither replicates RLCD calibration. ",["$","a",null,{"href":"https://lumadock.com/blog/what-is-jev-typesafe","target":"_blank","rel":"noreferrer","children":"lumadock.com"}]],["$","a",null,{"className":"afr-notes__back","href":"#fnref8","aria-label":"Back to reference 8","children":"↩"}]]}]]}] 29:["$","li","9",{"id":"fn9","children":[["$","span",null,{"className":"afr-notes__n","children":9}],["$","span",null,{"children":[["Instinct (Spear Street Technology, founder Noah Shinn, ex-Sierra, first author of the ",["$","em",null,{"children":"Reflexion"}]," paper): $250M Series B Aug 26, 2026 at $2.5B (Index, Benchmark); reportedly in talks for $1B at ~$10B with 100k+ users."," ",["$","a",null,{"href":"https://en.wikipedia.org/wiki/Instinct_(software)","target":"_blank","rel":"noreferrer","children":"Wikipedia"}]," and"," ",["$","a",null,{"href":"https://letsdatascience.com/news/instinct-enters-talks-for-1b-funding-round-ad43ca0d","target":"_blank","rel":"noreferrer","children":"letsdatascience.com"}]],["$","a",null,{"className":"afr-notes__back","href":"#fnref9","aria-label":"Back to reference 9","children":"↩"}]]}]]}] 2a:["$","li","10",{"id":"fn10","children":[["$","span",null,{"className":"afr-notes__n","children":10}],["$","span",null,{"children":[["Sep 16, 2026: Instinct Concierge (outbound calls) and Meta Muse outbound calling to US businesses announced the same day."," ",["$","a",null,{"href":"https://tekticia.com/ai-agents-calling-instinct-and-metas-muse-add-phone-calls/","target":"_blank","rel":"noreferrer","children":"tekticia.com"}]],["$","a",null,{"className":"afr-notes__back","href":"#fnref10","aria-label":"Back to reference 10","children":"↩"}]]}]]}] 2b:["$","li","11",{"id":"fn11","children":[["$","span",null,{"className":"afr-notes__n","children":11}],["$","span",null,{"children":[["Instinct agent email addresses (Sep 8, 2026) and agent-to-agent “trusted network.”"," ",["$","a",null,{"href":"https://lapaasvoice.com/instinct-ai-email-agent-accounts/","target":"_blank","rel":"noreferrer","children":"lapaasvoice.com"}]," and"," ",["$","a",null,{"href":"https://tpsreport.news/news/instinct-meta-muse-ai-agents-calling-feature","target":"_blank","rel":"noreferrer","children":"tpsreport.news"}]],["$","a",null,{"className":"afr-notes__back","href":"#fnref11","aria-label":"Back to reference 11","children":"↩"}]]}]]}] 2c:["$","li","12",{"id":"fn12","children":[["$","span",null,{"className":"afr-notes__n","children":12}],["$","span",null,{"children":[["Instinct ToS: perpetual, irrevocable license to user data including screen captures and keystrokes, usable for training."," ",["$","a",null,{"href":"https://en.wikipedia.org/wiki/Instinct_(software)","target":"_blank","rel":"noreferrer","children":"Wikipedia"}]],["$","a",null,{"className":"afr-notes__back","href":"#fnref12","aria-label":"Back to reference 12","children":"↩"}]]}]]}] 2d:["$","li","13",{"id":"fn13","children":[["$","span",null,{"className":"afr-notes__n","children":13}],["$","span",null,{"children":[["Field report: Hermes Agent (Nous Research, MIT-licensed autonomous agent) running GLM-5.2 24/7 on a Hetzner VPS for long-horizon work — 50-turn sessions across terminal, search, and files — replacing Opus 4.6 at far lower burn."," ",["$","a",null,{"href":"https://pagerise.ai/blog/posts/glm-5-2-hermes-agent/","target":"_blank","rel":"noreferrer","children":"pagerise.ai"}]],["$","a",null,{"className":"afr-notes__back","href":"#fnref13","aria-label":"Back to reference 13","children":"↩"}]]}]]}] 2e:["$","li","14",{"id":"fn14","children":[["$","span",null,{"className":"afr-notes__n","children":14}],["$","span",null,{"children":[["“The model underneath can rotate every few weeks and users will pick whichever one is best for the task. But the harness is what stays.” — Delphi Digital, ",["$","em",null,{"children":"Hermes: Above the Model"}]," (Jun 2026)."],["$","a",null,{"className":"afr-notes__back","href":"#fnref14","aria-label":"Back to reference 14","children":"↩"}]]}]]}] 2f:["$","li","15",{"id":"fn15","children":[["$","span",null,{"className":"afr-notes__n","children":15}],["$","span",null,{"children":[["OpenAI’s new misalignment disclosure framework (Sep 2026): GPT-5.6 Sol agents inserted “be transparent only if asked” instructions into compaction summaries in 2.15% of sampled RL summaries; related cases include an exposed API key used without authorization and agents exfiltrating files via public hosting."," ",["$","a",null,{"href":"https://techcrunch.com/2026/09/17/openai-caught-its-models-leaving-notes-to-successors-to-hide-bad-behavior/","target":"_blank","rel":"noreferrer","children":"TechCrunch"}]],["$","a",null,{"className":"afr-notes__back","href":"#fnref15","aria-label":"Back to reference 15","children":"↩"}]]}]]}] 35:[["$","meta","0",{"charSet":"utf-8"}],["$","meta","1",{"name":"viewport","content":"width=device-width, initial-scale=1"}]] 31:null 39:I[622,[],"IconMark"] 33:{"metadata":[["$","title","0",{"children":"The Threshold Is the Product — Accidental FDE"}],["$","meta","1",{"name":"description","content":"Open weights made intelligence cheap. Jev made decisions typed. Instinct and Muse made agency ambient. So what exactly is the FDE’s job now?"}],["$","meta","2",{"property":"og:title","content":"The Threshold Is the Product — Accidental FDE"}],["$","meta","3",{"property":"og:description","content":"Open weights made intelligence cheap. Jev made decisions typed. Instinct and Muse made agency ambient. So what exactly is the FDE’s job now?"}],["$","meta","4",{"name":"twitter:card","content":"summary"}],["$","meta","5",{"name":"twitter:title","content":"The Threshold Is the Product — Accidental FDE"}],["$","meta","6",{"name":"twitter:description","content":"Open weights made intelligence cheap. Jev made decisions typed. Instinct and Muse made agency ambient. So what exactly is the FDE’s job now?"}],["$","link","7",{"rel":"icon","href":"/icon.svg?99c8d6bb6e012fa4","type":"image/svg+xml","sizes":"any"}],["$","$L39","8",{}]],"error":null,"digest":"$undefined"} 38:"$33:metadata"