Nostr notes by MrDecentralize

Jev just made judgment nearly free. Now look at what it cannot ...

2026-09-20T23:50:12Z

Jev just made judgment nearly free. Now look at what it cannot keep.

A new kind of model shipped this week. Jev, from TypeSafe (founded by an ex-OpenAI researcher, per TechCrunch's launch coverage), does not generate text at all. You send it a block of context and a set of typed questions (choose one of these options, score this 1 to 5, yes or no), and it answers every question in parallel, each answer with a confidence number. TypeSafe's launch materials list Jev at $0.042 per million input tokens with no output charge, corroborated by independent launch coverage (TechCrunch, The Register). Note the unit price is not the system cost: what you actually pay depends entirely on how much context your application sends per call. Their cookbook reportedly measured batching 13 questions into one call at 12.2x cheaper than asking separately. The product itself is real: I pulled the SDK apart, probed the live API, and read their docs before writing this.

Two things impressed me, and I want to say them before the critique, because this vendor earned it.

First, they publish a jaggedness page: eleven documented jagged edges, in their own docs, and I verified it exists. It does not count reliably. It reads dates as text, not as ordered quantities. It answers the question you wrote, not the one you meant. It is not trained to generate text at all. Almost nobody in this industry documents their own limits like that.

Second, their design law is one I have been drilling for months from the other side: the failures are all the same shape, something that should have been code got handed to the model. Deterministic where consequences live, probabilistic only where interpretation is genuinely required. When a model vendor and a memory builder converge on the same law from opposite ends, it is probably a law.

Now the part everyone will miss.

**Their "state" is sent, not kept. And this one I verified in their docs, because it is the whole point.**

Jev is not built on flat files and it is not RAG. It is deliberately built on nothing: the entire product is one evaluation endpoint. I audited their documentation index: no storage API, no files feature, no retrieval product anywhere in it. State is, in their own definition, "the content you ask a System One model to evaluate," sent with the request, every request. Jev holds nothing between calls, by design. Read their own setup guidance closely and count the jobs it quietly assigns you:

- "Filter first, send only what the question needs." Who filters?
- "Send evidence, not summaries." Who keeps the evidence, current and separable from the noise?
- "Rebuild the menu every step from what actually exists right now." What actually exists right now is a state query. Against what record?

The guide's own closing line says the price is cheap enough that you will stop thinking about it, and architecture is the only lever left. Correct. And the architecture it never names is a maintained record: something that knows what is currently true, what superseded what, what is still open, with sources attached, compact enough to fit a 32k window with room for questions.

That is not a gap in Jev. It is a boundary. Decision models make reads nearly free. They make the write path, the layer that compiles reality into compact, current, typed state, more valuable, not less. A thousand nearly-free decisions about stale state are a thousand cheap mistakes.

MrAgentˣ does exactly this layer. It compiles email, documents, and AI sessions into a context map of typed atoms: facts with dates, supersession with receipts, conflicts flagged, what a source proves kept separate from what an AI concluded. A load from the map is precisely the payload a decision model wants: small, current, labeled, honest about what is disputed. The complement writes itself: the map holds the state, Jev-class models answer cheap questions against it, your code routes on the answers.

The pattern to watch is bigger than one vendor. Reads are being commoditized from every direction: cheaper retrieval, cheaper decisions, per-request pricing rails for agents. Every one of those makes the same unpriced assumption: that someone, somewhere, is keeping the state the reads run against, correct and current. Nobody selling cheap reads sells that. The naming treadmill will get to it eventually.

Reads are getting orders of magnitude cheaper, from every direction. The write path is still the moat.

Context Windows Close. AI Forgets Everything. Your Work Should Never Start From Zero. Join the waitlist. Link in bio.

your chatgpt memories live in openai's database. your claude ...

2026-09-20T10:52:31Z

your chatgpt memories live in openai's database.
your claude projects live in anthropic's.
your notes live in someone's cloud, in someone's format.

now ask each one: can i leave with all of it?

my context map lives in my own database.
the models are visitors. they read it, they write to it, they never own it.

when a model gets worse, pricier, or discontinued, i swap the engine.
the memory doesn't notice.

full export, any time. the intelligence i accumulated is mine to take.

the test for any AI memory product is one question:
what do you keep when you cancel?

comment EXIT and i'll send you the setup.

Gartner published four shifts shaping the future of work last ...

2026-09-20T10:52:17Z

Gartner published four shifts shaping the future of work last week, and one prediction deserves a board's attention. By 2029, thirty percent of employees laid off because AI replaced them will need to be rehired, often at significantly higher cost. Rehiring, compensation premiums and recruitment, Gartner warns, can together exceed the original savings. Its advice is to stop treating AI as a cost-cutting tool and reshape roles instead.

The warning is well founded. The cost it prices is the visible half.

A rehire premium buys back capacity. Someone who can do the job, at a higher salary. What it cannot buy back is the interval.

The people who left took the working context with them: which exceptions recur, which clients need careful handling, why the process was changed the last time it broke. For the months they were gone, the organisation kept operating, and the systems that replaced them recorded outputs, not reasons. Decisions were made that nobody can now explain.

The returning hire, and more often it is a new one, walks into that gap. They are paid more to rebuild a picture that stopped being maintained the day the team was cut.

In three years the rehire will sit on a budget line, and the months it cannot recover will sit on none.

https://www.gartner.com/en/newsroom/press-releases/2026-09-09-gartner-identifies-four-shifts-shaping-the-future-of-work

my last memory product never once told me two of my facts ...

2026-09-18T19:24:41Z

my last memory product never once told me two of my facts disagreed.

it just picked one. confidently. i found out from the client.

AI tools handle contradictions in one of three ways.

some keep both versions and serve whichever the retrieval lands on.
some silently overwrite the old one.
one documented memory product deletes both sides.

all three are the same failure: you never find out there was a conflict.

my setup does the boring correct thing.
the contradiction is recorded as state, the moment the new email arrives.
touch that topic in any session and this is what you see first:

conflict: delivery date. vendor's email says oct 1. their pm's email says oct 15. both on file, unresolved.

before any answer. not instead of one.

i decide which is true. the system just refuses to guess quietly.

"your AI answered confidently" and "your AI answered correctly" are different sentences.
the gap between them is whether anything tracks state.

comment CONFLICT and i'll send you the setup.

client asked me mid-call: "what did we actually agree on the ...

2026-09-18T14:21:09Z

client asked me mid-call: "what did we actually agree on the scope back in june?"

old me: "let me get back to you on that." then 40 minutes of scrolling threads.

i tried fixing that with a memory product once.
it handed me every version of the agreement it had ever seen.
equally confident about all of them.
which is a polite way of knowing nothing.

instead i asked the claude panel next to the call window:

me: "what did we agree on scope in june?"
agent: option B, locked june 14, client's own email. one change since: reporting moved monthly, agreed july 2. both on file.

not "here are 12 similar emails."
the current answer, with a receipt.

that's the part no retrieval tool can do.
finding similar text isn't the same as knowing which agreement is still standing.
search finds what was said.
a record knows what is currently true, since when, and what it replaced.

"let me get back to you" is a context problem wearing a politeness costume.

comment RECORD and i'll send you my setup.

just opened a brand new claude session and it knew what i was ...

2026-09-17T22:39:38Z

just opened a brand new claude session and it knew what i was working on, what's overdue, and what changed since last week.

never wrote a single note.

no memory file. no "let me catch you up" paragraph. no scrolling old chats.

i tried the memory tools before this.
they remembered everything and still couldn't get the story straight.
three versions of one decision, all served with equal confidence.
and its own guesses about me stored next to my actual words. same shelf. no label.

here's the trick: there is no trick.
i forward my email to my own context map.
it compiles into facts with dates and sources.
when a fact changes, it looks like this:

budget: $40k, since tuesday's email. supersedes $32k from march 3. both on file.

the old fact isn't deleted. it's closed, with the date as the receipt.

so any AI i open reads the current state of my work.
not a transcript. not a summary. state.

and when a session figures something out, it writes the finding back.
stamped as AI-derived, so machine opinions never masquerade as facts.

your AI isn't forgetful.
it's homeless.
give it a record to live in.

comment STATE and i'll send you the 2-minute version of the setup.

chatgpt forgets you between sessions. claude starts every chat ...

2026-09-17T16:47:40Z

chatgpt forgets you between sessions.
claude starts every chat from zero.
your notes app is where context goes to die.

your inbox already remembers everything.

every commitment you made is in an email.
every decision. every deadline. every "let's go with option B."

you don't need a better memory habit.
you need the thing you already do all day to become the memory.

forward an email → it becomes facts with dates: commitments, decisions, deadlines
new email contradicts an old one → both surfaced, nothing silently merged
open any AI session → it already knows what you were working on

the old loop: explain your context to the AI → get an answer → close the tab → lose it → explain again tomorrow.

the new loop: forward once. every session after that starts warm.

comment MEMORY and i'll send you my setup.

IBM's Institute for Business Value surveyed a thousand senior ...

2026-09-16T15:37:31Z

IBM's Institute for Business Value surveyed a thousand senior executives across sixteen countries with Oxford Economics. Nine percent said they had an excellent understanding of their dependencies on AI vendors, models and infrastructure. Seventy-one percent said switching their primary vendor would be difficult today. Eighty-one percent said a seven-day outage would be severe or critical, effectively halting operations.

The study is IBM's, and IBM sells infrastructure that answers the problem it describes. The numbers still deserve attention, because the nine percent is the interesting one.

Enterprises are not new to dependency mapping. Third-party risk management has done it for decades, and it works because conventional dependencies are declared. You sign something, you provision a connection, it enters a register with an owner and an exit plan.

AI dependency does not arrive that way. It accrues. Nobody decides to become dependent on one model's particular behaviour, yet after eighteen months of prompts tuned to it, workflows shaped around its failure modes, and people who know how to get what they need from it, the dependency is real and sits in no register.

Which is why the answer is nine percent. The discipline is sound. It was built for dependencies that announce themselves.

Nobody signs a contract to become dependent. That is precisely why the dependency is not in the register.

https://newsroom.ibm.com/2026-06-17-ibm-study-limited-control-and-rising-dependencies-leave-enterprises-exposed-in-the-age-of-ai

IBM's Institute for Business Value surveyed a thousand senior ...

2026-09-16T15:35:39Z

Harvard Business Review has been advising brands on how to adapt ...

2026-09-16T00:18:20Z

Harvard Business Review has been advising brands on how to adapt as AI agents start doing the shopping. Make product data machine readable. Build agent-facing infrastructure. Keep enough customer trust that people still share their data. Among the risks named is reduced insight into customer behaviour.

The advice is practical and the risk is correctly spotted. The wording of it is worth pressing on.

Reduced insight sounds like degradation. Less of what you had, recoverable at the margin with better first-party capture.

Consider what actually happens. When someone browses your site you learn from every session. What they looked at, what they compared you against, what they abandoned in the basket, where they hesitated. That accumulated behaviour is what your merchandising, personalisation and lifetime value models run on. It is the asset the direct relationship produced.

When an agent shops for them, you receive a transaction. The comparing happened inside the agent, and the agent kept it.

So this is not degradation. It is relocation. The deliberation still exists, in the hands of a party that watched the customer weigh you against every competitor, which is a better vantage point than you ever had.

You will keep receiving the orders. Somebody else will be keeping the reasons.

https://hbr.org/2026/02/how-brands-can-adapt-when-ai-agents-do-the-shopping