▲ 109 points
back
70 comments
Against better judgement, I ran it on my repo. Poking at the "Ask anything about the report" bot, it penalizes me heavily for not having "positive tracked outcomes", as I don't report back to Claude what the outcome is. It wanted me to do this:
"At the end of research-heavy sessions, add a tiny closure note:
Decision: Use CacheMountStore with registry/local/GHA backends.
Why: GHA cannot expose the same content.Ingester path, local import has discovery issues, registry can resolve by tag.
Proof / current artifact: See files X, Y, Z. Subagent found A, B, C.
Next action: Implement interface in package N. Do not add new cache-mount flags yet. Use mode=cache-mount on existing cache flags."
There's no need for this, Claude is not my task tracker.
It also penalized me for:
- Using Claude to introspect a codebase as throwaway work and ` not close with a crisp “acceptable / risky / copy this / avoid this” decision this session.`. No action was needed, it wasn't an actionable session!
- Using Fable for code review. `Several BuildKit subagent sessions produced research reports, but the transcript does not show whether those findings became implementation choices.` Yes, Fable launches subagents to code review. Valid findings get turned into fixes or WONTFIX.
- `The Cloudflare CI upload failure got narrowed well, especially after you supplied the exact timestamp and challenged the /v2 routing recommendation, but the session ended without a chosen next diagnostic owner or action this session.` There's no need to pick a next diagnostic owner when I'm a solo founder with no employees.
I had the exact same experience. Claude is not my validation engine, I do not tell it when it has completed. I verify that in the code/project and end the session.
I also got this,
> A useful next habit is to end each correction with a concrete acceptance test, owner artifact, or stop condition: “write it into PROGRESS.md,” “make nix run .#bench fail until this is real,” “rerun this exact command,” or “do not proceed until these two choices are explicit.”
> You already do this well in the biggest penance sessions. Apply it to the smaller ones too.
Which I have found to be counterproductive in my personal work. Current models can generally infer acceptance tests of this level of granularity (not true for larger project-level prompts, but those don't produce good enough code for me yet -- even with specific acceptance criteria).
I also got penalized for using claude in read-only mode for the same validation reason?
> For read-only work, end with one of:
“turn the top finding into a PR-sized plan”
“mark these as accepted/rejected/deferred”
“write a cleanup checklist”
“give me the exact command I should run safely”
“stop, no action recommended”
It's a little sad to be honest, I would actually enjoy a product that helped me improve prompting + ai usage.
Yeah, I've found similar issues with even basic tasks like 'draft an email response' which will then populate a md file tagged as draft. Claude will then chastise me a few days later for not sending said draft. I've had to take real steps to avoid NannAIy in the last month that haven't even been a problem before in similar workflows.
This tool might be useful for AI tool beginners who need some guidance, but using it as a scoring tool looks like a mistake. It’s just going to penalize people for having different ways of working.
I don’t know how much it’s actually used to rate people. Rating anyone based on how they work instead of the results produced is an age-old mistake though.
It's used for YC applications as an optional portion of the founders profile. I'm not sure how it's weighted.
> Based on Paxel’s own site, 1.2 million+ coders have so far uploaded their reports to YC.
I checked the Paxel website and it says this:
> So far, 1,543,553 sessions have been uploaded and analyzed.
The count is for sessions, not coders. I assume the tool uploads a lot of sessions from each person who uses it.
That’s a large number, but it’s not a million different people. I am surprised that so many people think it’s a good idea to download a run a program which gathers up their coding sessions and submits information about them.
"They trust me."
I feel so tempted to post the next two words that famously follow that utterance.
I could be mistaken, but isn't Paxel a tool that YC *itself* built to understand how founders/applicants apply AI? If so, this feels less scary than some of the comments (i.e. not 3rd party).
Also, not trying to take shots, but should the title be "I got into YC Startup School by hacking it" instead? Isn't that different than the main YC program?
All in all, you did them a solid by finding and responsibly disclosing. Nice job.
So let me get this straight... technical founders are providing limitless access to their IP, and this is supposed to be a positive indicator that they would be a good founder? Thats wild
`cd ~/path/to/your-project && curl -fsSL https://paxel.ycombinator.com/upload.sh | bash` is the command from Paxel's website. So you're literally just piping whatever the hell they want into bash. It's long, lots of comments, and honestly would be time consuming to audit properly.
Supposedly this is what it does from the opening comment
# Paxel upload script
# ===================
#
# What this does (up to 17 steps):
# On your machine
# 1. Check Docker is installed and running
# 2. Sign you in (browser-based device auth)
# 3. Pull or build the Paxel Docker image
#
# Inside the container — file bodies stay local; only aggregate metrics +
# metadata (paths, commit numstat, session events) are uploaded
# 4. Discover projects and sessions (Claude Code, Codex CLI, Cursor)
# 5. Read your git history
# 6. Parse transcripts
# 7. Summarize each session (cloud Haiku via YC proxy)
# 8. Group git commits by session
# 9. Group sessions into multi-day work streams
# 10. Extract steering traces
# 11. Extract decision exchanges (cloud Haiku)
# 12. Redact code before upload (regex pattern redaction)
# 13. Link decisions to outcomes
# 14. Analyze code quality (L1 deterministic)
# 15. Score episodes across 5 axes (cloud Haiku)
# 16. Assemble your report
# 17. Upload redacted summaries + scores to the server
#
# Then: opens your results in the browserNo.
It uploads only the AI coding agent transcripts in the directories you explicitly run the upload script in.
For the longest time, the YC application included the question "What was your greatest (non-computer) hack?". They have always liked people who think of ways to work around existing systems. So it's no surprise that their response was positive. Also, all the principals are ex-founders, mostly engineers, who totally understand hacking culture.
Sadly, it looks like they took that question off the application though.
That question is still there, it lives in the founder profile.
Except this was a computer hack. So technically wouldn't qualify.
Why does the author keep mentioning that a one line bash command that downloads a bunch of crap is “great UX”?
Am I misreading this, or is ycombinator running a tool that transfers IP from an applicant’s machine to OpenAI and then transfers the generated summaries into their own servers. And that tool has some “webapp security 101” level vulnerability in it.
Feels like a disaster waiting to happen.
It doesn’t upload code (it’s not meant to, anyway); it uploads coding agent transcripts. It’s meant to be a fun app like Spotify wrapped - that’s how it was described when presented internally.
AI can do no wrong.
Maybe this is another one of Garry's 30k LOC/day vibecode specials that he's so proud of
> I uncovered Y Combinator was scoring 100k+ founders around the world through Paxel
Alright, my ears are wide open. Tell us more, how did YC use the private submission data from thousands of founders to score them? They fed some 3rd-Party AI all personal data to score who should get an interview? I can't be the only one here seeing a bad news story unfolding in real time...
YC is currently polarizing against companies that are rejecting the philosophy of giving (a single user) AI full read access to company data. They find that resistance frustrating and the complain publicly about it, calling such an attitude outdated "so 2024".
https://youtu.be/B246K_G7mHU?si=UVJei9Jpk7ZvREnH&t=1483
https://youtu.be/B246K_G7mHU?si=ikHZD4MwxA5JtXLj&t=1716
I understand that YC and startups in general have less incentive to worry about security because they don't have much to lose. But I think the big winners will care about security from the start, because it's a bottleneck at the top, what would whatsapp be if they vibecoded security from day 1? It would never have been able to fix its reputation after the fact.
They then switch the topic and talk about token spend and how one can spend millions or billions of dollars in tokens, so it's a bit clearer what their incentives are.
I don't want to be negative and personal, but this is business, I listen to CEOs like Garry Tan and Satya Nadella, and I feel that I'm on an opposing team, my personal bet is to skip this cycle and wait for the next cycle of CEOs. I'm focusing a lot on security on this cycle, hopefully that will pay off.
sure….
> YC wanted me to use something called Paxel on my computer as part of the application.
> I should run a script, a very easy-to-use cURL one-liner that installed something on my computer and analyzed every line of code I’ve written with a coding agent, compile a report, and upload it to YC’s servers.
Yikes! I hope this is NOT the future of hiring.
the sad thing is that paxel sounds like it would be a pretty fun tool to run locally and see what my LLM assisted code gets classified as. instead, to borrow cory doctorow's framing, it is designed to do something to me rather than for me.
It can't possibly be an inevitable future. Something this would be super easy to game.
From their perspective, it's code running on an untrusted third person/adversary's computer... If only last 40 years of computer history could be any guide as to what might go wrong here ..
Everytime you think this it will be. Let's go back to analog.
Don't worry, it doesn't analyze your generated code, it just analyzes your source prompts, the thing that presumably generates the code. If only there was a word for text that generates a computer program.
paxel my gstack until i submit
its for G brain types
you wouldn’t understand
wtfffff
what do you think it says if you don't use AI
So if you built something without the use of an LLM, you’re automatically disqualified now? What a world.
Don't read this horror story before bed time if you care about information security.
Perhaps running the script outside a heavily sandboxed system should trigger an automatic rejection.
surprised running paxel doesn't disqualify you, the program sounds sus
Oh, I remember that thing, it wanted to see my prompt and my code in exchange for telling me my horoscope.
Some hands you have to fold, and I'm folding this one. Patiently waiting for the "let's give all our data to a single AI user" bubble to bust with some massive exploits.
> I’d also disclosed it in private through email 12 days earlier to no response. But publicly at least the process works.
lmao “I got into the YC Startup School by publicly embarrassing them”
This is how you end up with gstack. When you believe too hard in secret metrics because when people know about metrics it stops working you inflict psychosis chasing correlations that don't mean anything.
> I found that this year, YC wanted me to use something called Paxel on my computer as part of the application.
> I should run a script, a very easy-to-use cURL one-liner that installed something on my computer and analyzed every line of code I’ve written with a coding agent, compile a report, and upload it to YC’s servers.
This is ridiculous. Any kind of application (job, startup school, whatever) should not do this. Honestly, wtf is YC thinking with this?
> (what) is YC thinking with this?
They want people who do what they say and don't challenge them?
“Got into YC” is being used a bit loosely here. They got invited to startup school, which is a two-day event hosted by YC.
I got into HN front page by hacking it (the post title)
Fake it until you make it, baby!
paxel and how it works sounds so dumb.
so i’m not at all surprised at how it is used and by whom.
Paxel seems like a tool that would cause brand damage to YC if it were better known.