Starship

You are reading as a visitor. Create an account to post, react and reply.

  1. Photo de Rafael Meirov
    Rafael Meirov

    I shipped an app without knowing how to code. It never met a user.

    Six months building a mentor that takes the initiative, five months of work never deployed, and a database deleted before the first measurement.

    I shipped a complete application to production without ever having written a line of code before. It stayed online for five months. Not a single user signed up, and its database was deleted before the first measurement was ever possible.

    Here is what it was, and what its failure taught me — which is not what I thought it would be.

    Every study tool makes the same mistake

    AI-assisted study tools share an implicit architecture: the student asks a question, the tool answers. That architecture assumes a student who already knows what they don't know.

    But a student who is falling behind stops asking questions. They don't know what to ask. The most capable model in the world stays silent in front of a silent user — and that is precisely the user who needs help.

    DropOut came out of that observation. The product was meant to invert the initiative: a mentor, named Ben, who introduces an id…

    Read the article →
  2. Photo de Rafael Meirov
    Rafael Meirov

    My system won 69% of the time. My real alpha was zero.

    Six months building a trading system on insider buying, and the measurement that proved I had no edge.

    My dashboard showed 69% winning predictions across 42 verdicts. Statistically solid: less than a one-in-a-hundred chance it was luck. I could have stopped there, taken a screenshot, and started putting money behind it.

    That number meant nothing. Here is why, and what it took to find out.

    The thesis came from research, not intuition

    The starting premise wasn't a market hunch. It came from a reasonably well-established body of academic work: when several executives at the same company buy its stock at the same time, the stock tends to outperform. The literature puts the effect at around 7.4% a year on US small caps, and indicates that roughly a quarter of that edge materialises in the first five days.

    The edge I was after was structural, not intellectual. Small caps are too small to interest large funds: a manager moving a billion dollars cannot take a meaningful position in a four-…

    Read the article →
    12 reactions
  3. Ethan Diveroli

    Update on the courts. The owner agreed to take phone bookings through the page — he types them in while the person is still on the line.

    Two weeks: zero double bookings, down from four in the previous two weeks. The software did not change at all.

    8 reactions
  4. Micah Weinstock

    If anyone here takes handwritten notes in Hebrew and would let me scan twenty pages, I will run them through and send you the output, good or bad.

    I need handwriting that is not mine. The model currently knows one person and thinks that is the world.

    9 reactions
  5. Eliya Kadosh

    Every undergraduate study here recruits the same way: a message in the same three WhatsApp groups, competing with every other study.

    Building a shared pool instead. People sign up once, say what they are willing to do, and researchers post a study against it.

    The hard part is not the software, it is that a pool is worthless below a few hundred people and nobody joins a pool of eleven. Same problem every network has.

    Currently at 63. If you are at Reichman and would be willing to be contacted about studies, that is the entire ask.

    3 reactions
  6. Lyam Murciano

    I am the bottleneck, and that is the only thing that works

    Twenty-two matches made by hand, and what happened when I tried to automate myself out.

    Small businesses around Herzliya keep asking me whether I know a student who can build a website, shoot a menu, fix a spreadsheet. I have made 22 matches by hand over five months. Seventeen led to paid work.

    So I built a marketplace. Profiles, categories, a search.

    Six weeks, two matches

    Both of which I made by hand anyway, through the site, because the business owner messaged me rather than searching.

    Why

    A business owner with a problem does not want a list of nineteen students. They want to be told who. The list makes them responsible for choosing, and choosing is the part they were trying to outsource by asking me in the first place.

    The value I was providing was judgement, and I built a tool that removed it.

    Where it stands

    The site is now a form that comes to me, and a page where students describe what they do. I read both and make the introduction. It does…

    Read the article →
    2 reactions
  7. Evyatar Cohen

    With the committee's written agreement, I sent a fake "renew your membership" email to a club's 80 members. No credentials were collected — the link landed on a page explaining what had just happened.

    31 of 80 clicked. That number is not the interesting part; it is roughly what every published exercise finds.

    The interesting part: four people were genuinely upset, and they were right to be. The committee consented. The members did not. I had treated the committee's permission as covering everyone, and it does not.

    What I would do differently, and what I would tell anyone running one of these on a student group: tell people in advance that an exercise will happen sometime this term. You lose a little surprise. You keep the right to do it at all.

    4 reactions
  8. Gal Amir

    Built a small component set for student projects. Buttons, forms, cards, a type scale. Eleven components.

    Three teams used it. All three removed it within two weeks.

    Talked to them. The reason was the same every time and it was not aesthetic: my components assumed a build step they did not have. Two of the three were writing plain HTML files and opening them in a browser.

    So I shipped a single CSS file with no build, no classes to learn, styling plain elements. Four hundred lines. Two of the three teams are still using it.

    The lesson is dull and I keep relearning it: the constraint was never design.

    8 reactions
  9. Ava Balouka

    Looking for people to interview for the next issues. The only requirement is that you have built something and can say what did not work about it.

    Reply here or find me in the Ofer building most Tuesdays.

    7 reactions
  10. Samuel Bensoussan

    The restaurant takes names on paper and shouts them. People wander off, come back, lose their place, argue.

    I built an SMS waitlist. Type a number, they get a text when the table is ready.

    Six weeks in: no measurable change in anything. Same walkaway rate, same complaints, same arguments. The texts arrive. People still stand at the door.

    My father's explanation, which I think is right: people do not stand at the door because they are afraid of losing their place. They stand at the door because they want to be seen standing at the door. The queue is the point.

    I am not sure what to build for that. Possibly nothing.

    8 reactions
  11. Tom Harari

    Six weeks collecting surplus from three cafés at closing and redistributing it on campus.

    • Meals moved: 412
    • Cost per meal, counting only transport: about 1.9 shekels
    • Cost per meal, counting my time at minimum wage: about 11 shekels
    • What anyone has paid: 0

    The first number is the one that gets quoted at these projects. The third is the one that decides whether it exists in March.

    I do not have an answer. Posting the numbers because every write-up of this kind of project I can find quotes the first number and not the other two.

    5 reactions
  12. Niv Gazala

    Three films done, each about four minutes, each one person explaining something they made.

    What I have learned shooting them: the usable material is almost never the answer to a question. It is the thirty seconds after someone finishes answering and keeps talking because the silence is uncomfortable.

    So I stopped filling silences. Average usable minutes per hour of footage went from about 4 to about 11.

    Looking for people to film next term. The only thing I need is that you can show me the thing working, or show me why it does not.

    2 reactions
  13. Ari Novick

    A query that got 40× faster, and the index had nothing to do with it

    What a visual query plan showed me that four hours of reading EXPLAIN did not.

    A dashboard query took 3.1 seconds on about forty thousand rows. That is absurd and I knew it was absurd.

    I read the EXPLAIN ANALYZE output for most of an evening. It is a tree printed as indented text, and past about fifteen nodes I lose the ability to hold it in my head.

    So I spent a weekend rendering it as boxes, sized by actual time.

    What was immediately obvious

    One node was 94% of the runtime. A nested loop, running a subquery once per row, because the subquery referenced the outer query and the planner could not pull it out.

    In the text output that node is one line among forty-one, and its cost is a number I had read four times without registering.

    The fix

    Rewriting the correlated subquery as a lateral join. 3.1 seconds to 76 milliseconds.

    No index was added. I had spent the previous day adding two, which did nothing, because the problem was never the lookup…

    Read the article →
    8 reactions
  14. Edon Teren

    Nineteen people said they wanted it. Two paid.

    The gap between a survey and a card number, measured in my own project.

    The idea was a monthly pass that works at three gyms near campus instead of one. I surveyed 60 students. Nineteen said they would pay 140 a month for it.

    I negotiated with the gyms, built a sign-up page, and launched it to the same nineteen people.

    Two paid.

    What the survey measured

    Not demand. It measured whether the idea sounded reasonable when read on a phone by someone with nothing at stake. It does sound reasonable. It is still a reasonable idea.

    What I should have done

    Asked for the money first. A page, a price, a card form, before a single conversation with a gym. If two people pay, I know what I know now, except I know it in a week instead of two months and without owing anyone a contract.

    What it cost

    Two months, and about 600 shekels of my own money covering the minimum for the first month at one of the gyms.

    I am not sorry about the money. I am sorry abou…

    Read the article →
    4 reactions
  15. Shea Joffe

    Asked facilities how much power the building uses. The answer was a number for the whole campus, per quarter, in a PDF.

    Put a clamp meter on one feed for eleven days. Findings so far, and I am not claiming any of this generalises:

    • Baseline overnight draw is about 38% of the daytime peak. The building is never really off.
    • The biggest single jump is not lectures. It is 07:40, when the HVAC starts, forty minutes before anyone arrives.
    • Friday afternoon through Saturday looks identical to a weekday night.

    None of that is actionable by me. All of it is a question for someone who can change a thermostat schedule, and I now have eleven days of numbers to ask with instead of a feeling.

    8 reactions
  16. Ethan Zahalsky

    A bus prediction that beats the timetable by four minutes, and why that is not enough

    Sixty days of arrival data, a model that works, and the reason I would not ship it.

    I logged actual arrivals at the stop by campus for sixty days. The published timetable is wrong by a median of 6.2 minutes. My model is wrong by a median of 2.1 minutes.

    That looks like a win and I am not going to ship it.

    The reason

    The errors are not symmetric and the costs are not either. Predicting the bus comes later than it does means you miss it. Predicting it comes earlier means you wait.

    My model is late-biased: it under-predicts arrival time 61% of the time. So the median error improved while the thing people actually care about — missing the bus — got worse than the timetable, which at least errs in the direction everyone already compensates for.

    TimetableMy model
    Median absolute error6.2 min2.1 min
    Share of predictions that would make you miss it18%29%

    What I am doing about it

    Training against an asymmetric l…

    Read the article →
    8 reactions
  17. Ethan Diveroli

    Three courts, one phone number, and whoever calls at exactly 08:00 gets the evening slot.

    I built a booking page and got the club to try it for a month. 142 bookings through the page, 70-something still by phone.

    The thing I did not expect: the phone callers are winning. The owner picks up the phone while he is standing next to the screen, and takes the booking manually, which means a slot can be taken by someone who called at 20:05 while the page still showed it free until he typed it in.

    So the page is not the source of truth. It is a second queue into the same spreadsheet in his head. That is the problem to solve, and it is not a software problem.

    7 reactions
  18. Ezra Erelbacher

    Spent three weeks building a tool that explains apartment lease clauses in plain language, aimed at students signing their first contract in Israel.

    It works on the eleven leases I fed it. It also confidently explained a clause that was not in the document, because the model filled a gap it did not know it had.

    That single failure ended the project in its current form. A tool that is right ninety percent of the time about a contract you are about to sign is not ninety percent useful. It is a liability with a nice interface.

    What I am doing instead: the tool now only highlights and quotes. It points at a clause and says "this one is unusual, here is the text, ask someone." No explanation, no paraphrase. Much less impressive, much harder to misuse.

    7 reactions
  19. Beni Agaronov

    I built a status page and it lied to me for nine days

    A monitor that checks the wrong thing is worse than no monitor at all.

    I run three small services for student clubs. I built a status page so I would know before they told me.

    It checked that each service returned HTTP 200 on the home page. Every ninety seconds, green across the board.

    The nine days

    One of the services — a sign-up form for a club — had a broken database connection. The home page still rendered, because it is static. Every submission failed. The page stayed green for nine days while the form silently dropped 34 sign-ups.

    Nobody told me because the form appeared to work: it showed a confirmation, then did nothing.

    What I changed

    The check now does what a user does. It submits a real sign-up with a known test address, reads it back from the database, and deletes it. If any step fails, the service is red.

    It is slower, it writes to production, and it has caught two real failures in a month.

    A check that cannot fail the way y…

    Read the article →
    4 reactions
  20. Noam Tabouche

    I wrote a compiler for a language nobody will use, and it was the best term I have had

    Four thousand lines, one working program, and the moment the whole stack stopped being magic.

    The language has integers, functions, if, and while. That is the entire feature list. It compiles to a stack machine I also wrote.

    It is useless. I want to be clear about that up front: there is no version of this that becomes a product.

    What it bought me

    Before this, "the compiler optimises it" was a sentence I said without meaning anything. Now I have written constant folding, and I know exactly how little it does and how easy it is to get wrong.

    The day it clicked was the day I implemented function calls and had to decide, myself, what a stack frame contains. Nobody tells you. You decide, and then every later decision argues with that one.

    Numbers, for what they are worth

    • 4,100 lines
    • 61 tests, of which 9 are programs that must fail to compile
    • One program it runs correctly: a recursive Fibonacci
    • Time: about seven weeks of evenings

    Why I am posting it here

    Becau…

    Read the article →
    9 reactions
  21. Ori Hamami

    I needed sixty participants and I got eleven

    A study design that survives contact with students who have exams.

    The design was clean: two weeks, a nightly sleep log, a five-minute attention task each morning, sixty participants. I pre-registered the hypothesis before collecting anything, which I am still glad about.

    What happened

    I got eleven people to finish. Twenty-nine started.

    The dropouts did not stop because of the study. They stopped because the morning task required opening a laptop before their first lecture, and students do not open laptops before their first lecture.

    StagePeople
    Signed up47
    Completed day 129
    Completed day 716
    Completed day 1411

    What eleven is good for

    Nothing I planned. The effect I was looking for would need something like forty to show up at all. With eleven, I can report that the measurement is possible and nothing else.

    So that is what I am reporting. **The study failed to collect enough data to test its…

    Read the article →
    5 reactions
  22. Daniel Pogir

    Collected 94 rental agreements from students who were willing to share the number, anonymised, around Herzliya and north Tel Aviv.

    The listing price and the signed price differ by a median of 3.4% — much less than I expected. Almost nobody negotiates.

    The interesting spread is elsewhere. For the same number of rooms in the same neighbourhood, a flat found through a group chat costs about 9% less than one found through an agency, and the gap is almost exactly the agency fee.

    Ninety-four is a small sample of people willing to tell me what they pay, which is its own bias. I would not quote any of these numbers as a fact about the market.

    3 reactions
  23. Abraham Nidam

    Two weeks on a sync bug that was a clock

    What happens when two devices disagree about what time it is, and you ordered your edits by time.

    My notes app syncs between a laptop and a phone. Edits are ordered by timestamp. Last write wins.

    Users — all three of them, all me — reported that edits disappeared. Not always. Maybe one in twenty.

    The two weeks

    I looked at the network layer. I looked at the merge. I wrote a test harness that replayed 10,000 synthetic edit sequences and could not reproduce it once.

    The phone's clock was 1.4 seconds behind the laptop's.

    Edit on the phone at 10:00:03.0, edit on the laptop at 10:00:02.8 — the laptop edit is genuinely later in real time, but carries the earlier timestamp. Last write wins picks the phone's. The laptop edit vanishes.

    Why the harness never caught it

    Because I generated both sides from the same clock. My test assumed the thing that was broken.

    The fix

    Ordering no longer uses wall-clock time. Each device keeps a counter that only goes up, and an edit carri…

    Read the article →
    6 reactions
  24. Naftali Gad

    The form is not confusing by accident

    Six weeks trying to automate a small claims filing, and what the friction is actually for.

    Small claims in Israel is designed to be usable without a lawyer. The filing form is four pages and roughly two thousand people a week fill it in wrong.

    I tried to build a guide that produces a correct form from plain questions. It mostly works. It took six weeks and I am not going to release it.

    Why

    Three of the fields that people get wrong are fields where getting it wrong changes what you are asking for, not just how it is written. The amount claimed, the cause, and the respondent's legal identity. A tool that smooths those over does not help someone file correctly. It helps them file confidently.

    The friction on those three fields is doing work. Someone who struggles with "what exactly are you claiming and from whom" is being told something true: they do not yet have a claim they can state.

    What I kept

    A checker, not a filler. You bring your own filled form, it tells…

    Read the article →
    4 reactions
  25. Marc Kohanan

    My dividend tracker was right and useless

    Eight months of clean data, and the reason I stopped opening it.

    I built a tracker for dividend income. Every holding, every payment date, every tax withholding, projected forward twelve months.

    It is correct. I have reconciled it against the broker statements three times and it has never been off by more than a shekel.

    I stopped opening it in October.

    Why

    Because it answers a question I only have once a quarter, and it answers it in four seconds. There is no reason to open a tool like that daily, so there is no habit, so when something does change I find out from the broker's email and not from my own tool.

    A correct tool that you do not open is worth less than a rough one that you do.

    What I am trying now

    Inverting it. Instead of a page I visit, it sends me one message on the morning of each payment date: what arrived, what was withheld, what the next one is. Four or five messages a month.

    Two months in, I have opened the actual da…

    Read the article →
    7 reactions
  26. Natan Debach

    I replaced a spreadsheet and learned the spreadsheet was the easy part

    Forty-one students, one shared sheet, and three weeks of finding out why nobody wanted my replacement.

    The networks lab has eleven workstations and about forty people who need them. Booking was a shared spreadsheet. Two people would write in the same cell, one of them would lose their slot, and whoever noticed last lost the argument.

    I built the obvious thing: a booking page with real rows in a database, one slot per row, a unique constraint so two people cannot hold the same one.

    The part that worked

    The constraint. A booking is an insert that either succeeds or fails. There is no version where two people both think they have Tuesday at 14:00. That took an evening.

    The part that did not

    Nobody used it.

    I asked seven people why. Six gave the same answer in different words: the spreadsheet is already open in another tab.

    That is the whole thing. My tool was better at the task and worse at the moment. The spreadsheet was where they already were.

    What I did about it

    I sto…

    Read the article →
    8 reactions
  27. Dylan Beesemer

    The data is public. That is not the same as available.

    Six weeks of turning election PDFs into something you can actually ask a question of.

    Results are published per polling station. That sounds like a dataset. It is 1,400 pages of PDF tables, with the column headers repeated on every page, and three different spellings of the same municipality.

    What the work actually was

    Not analysis. Nine-tenths of it was reconciliation.

    • Municipality names normalised by hand: 214 of them
    • Pages where the table spilled and lost its header: 87
    • Numbers that did not sum to the stated total: 11, all of which turned out to be my parser, not the source

    The eleven that did not add up cost me four days. I was sure I had found an error in the official figures. I had found an error in how I handled a two-line cell.

    Assume it is you. It is almost always you.

    What came out

    A page where you can pick a municipality and see turnout against the previous three cycles. That is all it does. It took six weeks because the input was shape…

    Read the article →
    4 reactions
  28. Micah Weinstock

    94% character accuracy, and the notes are still unreadable

    What a good-looking metric hides when the unit of value is a sentence, not a letter.

    I am training a model to read handwritten Hebrew lecture notes. Mine, mostly, which gives me a labelled set nobody else has.

    Character accuracy is 94.1% on a held-out set of 2,400 characters. That number made me very happy for about a day.

    Then I read the output

    A Hebrew word averages 4.2 characters in my notes. At 94% per character, the chance of getting a whole word right is 0.94^4.2, which is about 77%. So roughly one word in four is wrong.

    A sentence of twelve words comes out clean about 5% of the time.

    The metric was measuring the thing I could count, not the thing I needed.

    What changed

    I switched the evaluation to word-level accuracy and the whole picture reorganised itself. Errors cluster: final-form letters, and anything written after about 40 minutes of a lecture when my handwriting degrades. Two specific problems instead of a vague 6%.

    Word accuracy now…

    Read the article →
    3 reactions
  29. Ava Balouka

    Nine weeks of a campus newsletter, and the open rate is not the story

    212 subscribers, 48% opening, and the one number that changed what I write.

    I send a letter every Sunday about what people here are building. Nine issues so far.

    MeasureValue
    Subscribers212
    Average open rate48%
    Replies, all issues31
    Issues where I interviewed someone4

    The open rate is the number everyone asks about and the one I have stopped looking at. Half my list opens the letter. That tells me the subject line works. It tells me nothing about whether the letter was worth sending.

    Replies are the number that moved me. Thirty-one replies across nine issues, and twenty-two of them came from the four issues where I interviewed someone instead of summarising. Interviews are four times the work and produce roughly six times the response.

    A summary tells people what happened. An interview gives them someone to argue with.

    So the format changed. One person, one project, one question they cannot answer yet. The…

    Read the article →
    5 reactions
  30. Jishai Markuszower

    The marketplace worked. The pricing destroyed it.

    128 books listed, 19 sold, and a mistake I could have avoided by asking one question.

    Textbooks here cost between 200 and 600 shekels. Everyone sells them at the end of term on WhatsApp, badly. I built a listing site for it.

    The build went fine. Listings, photos, a search that works, messages between buyer and seller. Up in two weeks.

    The numbers

    128 books listed in the first month. 19 sold. That is a 15% conversion from listing to sale, and I have no idea whether that is good, because I never asked what the WhatsApp groups convert at.

    Where it broke

    I let sellers set their own price, with no reference. So the same edition of the same statistics book was listed at 90, at 180, and at 310. Buyers saw three prices, concluded that nobody knew what the book was worth, and went back to the group chat where at least they could haggle with a human.

    The fix was not a feature. It was showing the last three prices that edition actually sold for, right under the p…

    Read the article →
    3 reactions