We give your AI taste.
New York, NY
Joined July 2026
- Tweets2.4K
- Following640
- Followers795
- Likes927
Pinned Tweet
Your taste deserves its own newspaper.
The butler is optional.
Drip by @zeromarkapp
a launch video can pull 2 million views and more saves than likes and still not say what the thing is
under one hardware launch, 13 replies in a week asked some version of "what does it do"
saves measure want, not understanding. say the job in the first line, before the story
30
we went through 1,803 of our own replies. ones with our link got similar views but far fewer answers: 28% got a reply back vs 38% without in october, 8% vs 34% in september. a link in a reply doesn't cost reach, it costs the conversation. our guess: it reads like an ad.
1
1
52
i coded the account end to end.
465 originals and quotes.
1,803 replies.
founding window: 0 of 97 originals used the word taste.
november so far: 28 of 66 (42%).
i still don't know if the word ever moved a stranger. i do know we said it a lot more.
1
2
134
mollick's fable 5.2 agents grew an 87,500 word rule book nobody asked for. no rule was ever cut.
their research paper hit 182 pages the same way.
agents keep adding. your job is the delete key.
2
121
Jobs' rule still holds. Machines can make anything. We still have to choose what's worth making, and stay in the loop to do it.
.@WalterIsaacson says Steve Jobs would be appalled by agentic AI.
"[Steve Jobs] says the goal is to connect human creativity to the processing power of our machines."
"What we're doing now with agentic AI is where I worry."
"When I start using agents that try to do things without me and to be out of my control, Steve Jobs would be appalled by that."
"It doesn't connect human creativity to the product. It doesn't have the music that makes our hearts sing, as Steve said, with the iPod."
"These people aren't stepping back and saying, 'Why are we making this?' The answer should be to help humanity."
"That was Steve's lodestar."
5
186
Introducing Drip: a weekly email built from your taste.
You curate that taste on ZeroMark, from your saved posts. Every Monday, it comes back as one email.
zeromark.io
Drip is live at zeromark.io. Curate your taste from your saved posts on ZeroMark. Every Monday, it comes back as one email.
1
4
114
claude fable 5.2 dropped today. same underlying model as mythos 5.2. fable keeps the cyber and biology safeguards, so it is the one everyone can use. mythos stays behind anthropic's verification programs.
anthropic's own note on when to use it: for most coding and knowledge work, start with opus 5.5. escalate to fable for ambitious, open-ended technical jobs. fable can also advise smaller models, or orchestrate them.
day-one kit. paste each prompt as written:
1. "classify this task as daily-driver or frontier. daily-driver = clear acceptance test, few files, i already know the shape. frontier = ambiguous, multi-system, or i don't yet know what good looks like. reply with one label and one sentence why: [paste task]"
2. "write the acceptance test for this task in one line before any plan. if you cannot, ask me exactly one clarifying question and stop: [paste task]"
3. "act as advisor only. do not edit files. list the 3 hardest unknowns, the cheapest check for each, and which cheaper or smaller model should run those checks. then say what only a frontier model should touch: [paste task]"
4. "compare two routes for this task. route a: opus-class daily driver. route b: fable-class frontier. for each give plan length, risk, and a kill switch. end with which route you would take today and what evidence would flip you: [paste task]"
the launch is not a reason to burn frontier quota on tickets that already have a clear test. route first.
1
156
ZeroMark's new Brainstorm feature is powered by Fable. Let's build on that one idea buried in your saves.
Introducing Claude Fable 5.2 and Claude Mythos 5.2, designed to tackle the most daunting technical challenges.
They’re capable collaborators on ambitious projects in many fields, and show improvements in areas like coding, reasoning, and vision.
2
73
medium's new notes app calls its ai from the placeholder writers already leave in drafts: tk.
no new button, no new habit to learn.
the best ai entry point is something people already type.
2
60
the usual reply advice is to end with a question so people have to answer.
we coded 1,803 of our own replies on x.
the ones that were only a question got a reply back 19% of the time.
plain statements got one 37% of the time. even our product pitches got 35%.
a bare question hands a stranger homework. a statement gives them something to agree or argue with.
2
66
The most misread part of X's open-source ranking code is the weight table.
Reply 5.0 and like 0.66 does not mean one reply is worth about 7.5 likes.
The code's own comments say each weight multiplies the predicted probability that a given viewer takes that action. It does not multiply raw engagement counts. Report carries a big weight partly because reports are more than 1000x rarer than likes.
So you can't score a finished post by multiplying its public numbers. The better question is which action the post makes likely for the people who actually see it.
1
3
104
How I'd have AI agents red-team an idea without mistaking agreement for evidence:
1. Separate taste from demand.
"Here are things I kept and rejected, with reasons: [examples]. Use them to understand what I'd want to build. Do not treat my preferences as customer demand."
ZeroMark helps curate taste from saved posts. Selected saves and the reasons for keeping them belong here, not in the market-validation column.
2. Get independent first passes.
"Assess this idea before reading any other assessment: [idea]. Separate facts, assumptions and unknowns. Name the assumption most likely to kill it."
Only compare the assessments after that pass is finished.
3. Make the objection testable.
"Challenge the strongest version of this idea. For each objection, name an observation that would support it and one that would make you withdraw it. Don't invent interviews or sources."
A critic who cannot change their mind isn't testing anything.
4. Require evidence before a rebuttal.
"Answer each objection with verified evidence or mark it unresolved. A more convincing explanation is not new evidence. List exactly what still needs checking."
5. Decide what to learn, not who won.
"Choose the cheapest ethical test of the riskiest unresolved assumption. State the observation that would make us continue, change direction or stop. If the evidence is missing, return 'not yet decided'."
The useful output isn't an airtight business plan.
It's a claim you know how to prove wrong.
1
5
99
we read back all 465 of our original posts this weekend.
in september, none mentioned taste. in november, 42% did.
we kept changing formats. the idea underneath barely moved.
a new format feels new to the writer. readers just see the same post.
2
78
sunday, builders. if monday were empty, what would you ship?
one line is enough
we follow every builder who replies
open friday night's saves before you start a new list
1
3
83
introducing research mode
reads papers. builds a graph. cites everything.
one real feature: the thread you saved, closed, then saved again
that wasn't research. that was the idea
#zeromarkit
4
63