Pinned Tweet
With AI, one person can turn an idea into a working link in an evening.
Most ideas still never leave the notes app. What’s missing usually isn’t skill. It’s the first step, and people to take it with.
That’s what our One-Person Company cohort is for. You don’t need to code, and you don’t need the perfect idea. Just the wish to build something real, for real people, even if at first only one person uses it.
starc.institute/opc.html
#buildinpublic
How do you check if a professor in a paid research program is real?
A real faculty page only answers the first of four questions:
Identity: find them from the university’s home page, not a link the program gave you
Title: read the exact title, department and dates. A @UConn career guide warns that a title at one school can’t be assumed to mean the same at another
Activity: look for work from the last two or three years, and read one abstract
Involvement: contact them yourself, and get the hours in writing
Two things are easy to miss.
A university’s name doesn’t mean the university runs the program. @Penn’s policy lets faculty mention their affiliation in personal activities like consulting, as long as it doesn’t imply endorsement. Harvard Medical School is tighter, but the principle is the same: a professor can be real, and the school can have nothing to do with it.
And a .edu sender isn’t proof. A Detroit Free Press report cites a University of Michigan warning about scammers spoofing real professors’ email addresses. That report is about job scams, but its top tip from an @FTC official carries over: reach the person through contact details you found on your own.
We looked for an official guide to this exact question and didn’t find one, so the post pulls together what does exist.
Full post: starc.institute/blog/check-p…
2
68
Will admissions officers discount a project you built with AI?
We read the admissions AI policies at @Caltech, @NotreDame, @BrownUniversity and Yale, plus an AP report. None of them says how readers weigh an AI-built project. No published rule says it counts for less, and none says it counts the same.
What the rules do cover is what you submit:
Caltech’s guidelines are written about essays
Notre Dame’s rule names activities too
Brown covers any other submitted material, and allows AI only for spelling and grammar review
So keep two things apart: the project you built, and the words you write about it. Write those words yourself.
The most useful clue: AP reported that Caltech was launching an AI interview tool for students who submit research, with faculty reviewing the videos. The question behind it: “Can you claim this research intellectually?” That’s one school and it’s about research, but the test carries over.
Before an AI-built project goes on your application:
Read each school’s current AI policy
Write the description yourself
Name the tool in plain words
Lead with your decisions
Use only numbers you can show
Practice a two-minute walk-through
Keep your commits, notes and early versions
Treat required portfolios separately
Full post: starc.institute/blog/ai-buil…
More from @Starc_Institute.
1
25
What is a build in public cohort, and is it right for students?
A cohort gives you shared deadlines. Building in public means sharing the process while you’re still in it, not just the finished thing (@austinkleon: “Think process, not product”).
The question most programs skip is: public to whom? The group, the whole internet, or future employers?
Where the writing agrees:
Sharing creates a feedback loop: build, share what you learned, let the replies shape what’s next (@arvidkahl)
Alumni in a small cohort study felt more connected to the program and each other (@frontiersin)
Employers often look at candidates’ public profiles even when they aren’t pointed there (@LifeAtPurdue career resource)
And the pushback: when everyone builds in public, attention gets spread thin, and it isn’t required to succeed (@laike9m).
Before joining any build in public cohort, ask:
Who sees my posts by default?
Can I post anonymously, and take a post down later?
What exactly is due each week?
Who gives feedback, and is it a person?
How are numbers checked?
What happens if my project doesn’t work?
What does the program promise?
Most of the value comes from people reading your work every week, not from reach. Keep control of what goes public.
Full post: starc.institute/blog/build-i…
#buildinpublic
Follow @Starc_Institute for one question answered at a time.
What can a beginner actually build with AI in 48 hours?
Our answer: one page that does one job for one kind of person. Paste something in, get something useful out.
The people who build small tools with AI most often say the same thing from different angles:
@simonw has built more than 150 single-file “HTML tools” with AI: one file, no build step, a few hundred lines.
@AndrewYNg, in @DeepLearningAI’s The Batch, built a public-speaking practice prototype in a couple of spare Saturday hours by cutting scope hard: one audience member, a human choosing reactions, a simple 2D avatar.
@paulg’s test for ideas carries over: can you name who wants this right now?
What they mostly leave out, and what matters for a first build:
Skip logins, databases and payments.
Never paste an AI API key into a public page. Anyone can read it.
Don’t collect other people’s personal data.
Deploy early (@github Pages notes changes can take up to 10 minutes to go live).
Show it to 2–3 people who actually have the problem.
Write down what broke and how you fixed it.
Six one-page ideas (lab dilution calculator, schedule conflict checker, shuttle “next departure” page…) and a 48-hour plan:
starc.institute/blog/ai-one-…
More from @Starc_Institute
3
47
Should you put “vibe coding” on your resume as a student?
Our answer: list the project, not the buzzword.
The term started with @karpathy describing building with AI while you “forget that the code even exists.” Collins made it its 2025 Word of the Year, with a much broader meaning. @simonw argues it should mean building with AI without reviewing the code. On a resume, a recruiter could read it either way.
Meanwhile employers do want AI skills: NACE reports more than a third of entry-level jobs in its employer survey now require them. But NACE describes those skills as things like checking AI outputs, and Willison won’t commit code he can’t explain.
So:
Lead with the project: what it does, who uses it, a link.
Name the AI tools inside that bullet.
Say what you did yourself.
Real numbers only, even small ones.
Run the explain test: if you can’t walk an interviewer through it, learn it or cut it.
Keep the commit history and build log that back it up.
Full checklist, with before/after resume lines:
starc.institute/blog/vibe-co…
More from @Starc_Institute
5
111
Sophomore or junior: should you do research, or build a product? And how do you split the time?
There’s no single good answer online, just scattered advice from professors, career offices and @NSF. Read together, it points to one rule: pick by your next step, then go deep on one.
Professors who take undergrads expect real hours. One UC Berkeley professor asks for full-time in the summer and ideally 15+ hours a week during the semester.
Projects count, especially early. A @UWMadison computer science professor suggests most undergrads aim for project experience first, with research as a bonus. An @OhioState engineering career services post makes the same case for your first couple of years.
Summer is where research takes off. @NSF says most REU sites run in the summer, with deadlines mostly January to March, and applications usually need reference letters.
So: PhD next? Research is your main line. A job next? Your own product with real users. Not sure? Build now, apply for a summer REU junior year.
Give the main line about two-thirds of your spare hours. Protect grades first. Log your work every week; it becomes your interview material.
Full decision table and step-by-step: starc.institute/blog/researc…
3
92
How do you find your first 10 users on campus who aren’t your friends?
Friends try it, say it’s great, and never open it again. Strangers are harder, and nothing about finding them scales. What the sources agree on:
Go to them. “You can’t wait for users to come to you,” as @paulg put it in Do Things That Don’t Scale.
Start narrow: one course, one club, one lab section. @lennysan’s look at how big consumer apps got their first 1,000 users puts defining a narrow target user first.
Ask about the last time they hit the problem, not whether they like your idea. That’s the core of @robfitz’s The Mom Test (good summary by @brikis98). A @UVA I-Corps program says the same: discovery isn’t pitching.
End every chat with “who else deals with this?” That’s how you leave your friend circle.
Expect most people to say no. @NSF I-Corps asks startups for at least 100 interviews; you need far fewer, but the weekly habit is the same.
Check your school’s posting rules before you print flyers. Some dorms don’t allow commercial ads at all.
Count only people who come back and use it on their own.
Full two-week plan and table of where to look: starc.institute/blog/find-fi…
#buildinpublic
4
125
Can one person with no coding experience ship a web game with AI in a week?
We tried. Five days from idea to a public game, and nobody wrote a line of code by hand. An AI coding assistant wrote all of it.
What the humans actually did:
Said what we wanted. Our first brief got us a 30-day romance with an ending. "They're already together, no ending, come back every day" got us the real game.
Played it and found problems no one sees until they play.
Decided which of the AI's suggestions to keep.
Code wasn't the slow part. Images were. Image models draw exactly what you say and guess the rest:
"She makes a finger heart with her right hand and covers her mouth with one hand" → two right hands.
"Shoot back from the movie screen" → two people sitting with their backs to the screen.
The fix: write each image like code. Where each person is in the frame, what each hand is doing, where each one is looking. One-line prompts became 500–1,000 words, and redraws dropped sharply.
No server either. Static files on @github Pages, saves in the player's browser (localStorage, per @MozDevNet's docs), and a free counter for opens, Start presses and clicks through to us.
One legal note: @CopyrightOffice's 2025 report says prompts alone don't make AI output copyrightable. Don't build a business on owning AI art.
Full write-up, with the live counter: starc.institute/blog/build-w…
#buildinpublic
3
106
Do you need an LLC to sell an app you built as a student?
Usually not to start. Per the SBA, if you sell alone under your own legal name, you’re automatically a sole proprietor and don’t need to register anywhere. @stripe’s help center says you can sell without a separate business entity.
What you can’t skip: taxes. @irs says you report the income whether or not you get a Form 1099-K, and self-employment tax applies once net earnings reach $400.
Two cases where the answer changes:
On an F-1 visa, @studyinstates says starting a business counts as work. Talk to your international student office first.
Real liability risk (business contracts, sensitive data, a co-founder) is when an LLC is worth a look. Rules vary by state.
Checklist and sources: starc.institute/blog/student…
#buildinpublic
3
6
102
Can you use Copilot or Cursor on your senior capstone project?
Usually yes, but only within your instructor’s rules, and a syllabus that says nothing isn’t permission. At @Stanford, the default guidance treats AI like help from another person. An @OregonState engineering department’s guidance goes further: with no stated policy, only routine grammar fixes are allowed, and editor tools like Copilot and Cursor count “even when they feel more like autocomplete.”
Where AI is allowed, the written policies agree on three things:
disclose the tool and what it did
mark AI-generated code in the work itself
be able to explain and defend every line you submit
Sample policies from @CarnegieMellon’s teaching center range from a full ban to active encouragement. One says using AI without acknowledgment will be treated as plagiarism.
A safe split: AI can help you debug, explain unfamiliar code and draft tests. The design, the core algorithm and your results stay yours. Keep a short dated log as you go.
starc.institute/blog/ai-codi…
1
40
Anyone can type “500 users” on a resume. How do you prove your app’s user numbers are real?
Say exactly what you counted. In Google Analytics, total users, active users and new users are different metrics, and new users can even exceed active users. “500 users” could mean 500 people who loaded a page for two seconds.
Then keep evidence a stranger could check:
an analytics screenshot with the metric name and date range
a sign-up export with names and emails removed
a note of where each group of users came from
strangers counted separately from friends
@ericries calls total hits a vanity metric and says “metrics are people, too”: you should be able to trace a number to real people. Some tools, like @PostHog, let you share a dashboard by public link. And bought followers or views are fake indicators the @FTC’s fake-reviews rule bans for commercial use.
A small number you can defend in one sentence beats a big one you can’t.
starc.institute/blog/prove-p…
#buildinpublic
2
65
Your numbers are small and every build-in-public post you see has big ones. You can still write a build log people read, without rounding up.
One post, one thing. Five parts:
What changed this week, in one line.
The real number with its context: “Before: 4 of 15 testers finished sign-up. After: 9 of 14.” Say where it came from.
What a user actually said or did (no names without consent).
How you used AI, and what it got wrong.
What’s next, or what you’re stuck on, as a specific question.
Before posting: label visits, sign-ups and active users as different things. If a friend promotes it, they say they know you; the @FTC influencer guidance treats personal and family relationships as connections to disclose. And no bought followers or fake reviews: the FTC’s fake-reviews rule bans fake social influence indicators for commercial use.
@austinkleon’s Show Your Work! makes the case for sharing process, in small pieces. @googlesearchc’s helpful-content guide encourages showing first-hand evidence and saying how AI was used. For a student building with AI, that’s the most useful part of the log.
Full checklist:
starc.institute/blog/honest-…
1
1
80
Built a site with an AI coding assistant and it crawls on your friend’s phone? Measure before you change anything.
Run the mobile test in PageSpeed Insights, not just desktop. The lab test simulates a mid-tier phone on a mobile network.
Ask: is it slow for everyone, or only for people far from your server? Slow everywhere = the page is too heavy. Slow only far away = distance, so look at your server region and CDN.
Fix in order: resize images to the size they’re shown, add loading="lazy" to anything below the first screen, and delete libraries the assistant added that you don’t need.
Google’s own page experience docs say relevance still comes first, so fix speed for your users, not for a perfect score.
Full checklist, with the prompt to give your AI assistant for each step:
starc.institute/blog/ai-buil…
Sources: @ChromiumDev (web.dev Core Web Vitals and CDN guides), @googlesearchc (page experience docs), @MozDevNet (MDN lazy loading guide)
#buildinpublic
1
5
55
Starc retweeted
从今天开始,“小王子” 这个词会成为 AI 时代的一个符号
一个不具备技术背景的年轻人,通过 Vibe Coding 配合自学的3D建模,做出了大公司商业级的产品
16天,只需要16天。
- 16 天,你可以完成了专业团队2个月的工作
- 16天,你也可以与 a16z、Zara Zhang、硅谷顶级 Frontier Lab 公司创始人对话
我认为,@Saccc_c 的产品会是AI历史上浓墨重彩的一笔,也是 OPC 发展史的一条分界线
我们能够预见:大洋彼岸的美国科技媒体,12 个小时之后会被 “小王子” 三个字刷屏
The Little Prince
“小王子”三个字,自此将不再是一个童话故事
它会成为一个case,也会成为 AI 发展史上的分水岭
因为它真正开始让那些高傲的专业团队恐惧
也开始让每一个普通人深思
Building your first product with AI as a student: mobile app or web app first?
Our answer: web first, unless the idea truly needs the phone.
What the platforms’ own docs say:
Apple’s developer program is 99 USD per membership year for an individual, and every App Store app is reviewed. Guideline 4.2 says an app should be more than a repackaged website.
On Google Play, a new personal developer account needs a closed test with at least 12 testers for 14 days in a row before it can go to production. @GooglePlayBiz
A progressive web app can be installed and run offline from one codebase (via @MozDevNet’s MDN guide).
A web app is live at a link the day you deploy it, and every fix reaches users on their next page load. For a first product, getting to real users fast matters more than anything about the code.
Checklist + step-by-step: starc.institute/blog/mini-pr…
3
52
Non-CS major aiming for product manager jobs? A small product you built and launched can be your clearest proof, if it’s real: live, used by people who aren’t your friends, and explainable decision by decision. It won’t beat a hard CS degree requirement.
starc.institute/blog/non-cs-…
Today's breakdown of a paper worth thinking about
FlashMem: Distilling Intrinsic Latent Memory via Computation Reuse
arxiv.org/abs/2601.05505
Paper recap:
🌟Identifies a structural inefficiency in existing latent memory systems: they rely on segregated auxiliary encoders, which re-encode history and maintain disjoint KV caches.
🌟Proposes FlashMem, an intrinsic memory framework that reuses the backbone’s existing computation instead of introducing separate memory encoders.
🌟Treats the last hidden state as a sufficient summary of interaction history and uses it to seed memory generation.
🌟Introduces a Shared-KV Memory Consolidator that attends directly to the frozen backbone’s accumulated KV cache, avoiding redundant forward passes.
🌟Uses a projection-free cross-attention design with minimal new parameters and inherits weights from the backbone for stable alignment.
🌟Adds a parameter-free Cognitive Monitor that measures attention entropy and triggers memory consolidation only when uncertainty exceeds an adaptive threshold (85th percentile)
🌟Achieves up to 5× lower latency and near-Vanilla VRAM usage under long-context settings (64k tokens), while heavy baselines incur large memory and throughput penalties.
Automatic summary from real discussion:
🫧The central mechanism discussed was the Shared-KV Memory Consolidator. Conceptually, it synthesizes a memory token by directly attending to the backbone’s frozen KV cache rather than reconstructing history through a new encoder. This reframes memory not as an external retrieval module, but as a distilled view of what the model has already computed. The key question raised was: is this producing a “better KV,” or simply a KV that surfaces critical intermediate reasoning signals that might otherwise be diluted?
🫧Another focus was the Cognitive Monitor, which uses attention entropy as a trigger for consolidation. Instead of always injecting memory, the model intervenes only when internal uncertainty rises. This makes memory activation a conditional cognitive act rather than a fixed architectural addition. The entropy-based design avoids training a classifier, but it also raises a conceptual question about how robust entropy is as a universal proxy for confusion across model scales and task types.
🫧There was also discussion about architectural generality. Since the mechanism operates entirely within one agent and reuses its own cache, it resembles a lightweight alternative to RAG-style systems. However, this also invites a broader design question: if memory is simply distilled from existing KV states, could many alternative “better KV” constructions be proposed under similar logic? The paper presents one principled route grounded in hidden-state sufficiency, but the design space appears structurally open.
Thank you for the wonderful work. Yubo Hou, Zhisheng Chen, Tao Wan, Zengchang Qin
Follow us for more (no hype) real paper discussions.
4
287




