Puma AI Ghosted Its HN Applicants. That's Your Sourcing List.

On September 6, 2026, an engineer posting as boltzmann-brain walked into the comments of Hacker News' monthly "Who is hiring?" thread and said, in public, that they had applied to Puma AI in February with a handwritten, Puma-specific cover note, and never heard back. Two more engineers, Jose and Gleb, said the same thing under the same post. Puma's founder replied the next day and, to defend himself, quoted the generic emails he had been getting. He confirmed the inbox exists. He confirmed he reads it. He confirmed he does not answer.

That is not a PR problem. For any recruiter or founder hiring iOS, on-device ML, or browser-extension engineers this quarter, it is a name-attached outbound list with the qualification work already done.

Why the Puma AI thread is a sourcing list, not a story

The September 2026 "Who is hiring?" thread is a public receipt of high-intent applicants a YC-backed startup already ignored, which makes it the cleanest outbound pool you will see this month. The thread hit 254 points and roughly 397 comments in the first two weeks, and inside it Puma.tech posted for iOS/MacOS, Android, AI/ML, and browser-extension engineers, plus a "based social media intern," directing applicants to email careers+hn@puma.tech with a GitHub link and a note.

Then three named users publicly self-identified as ghosted. Founder Yuriy (HN handle html5cat, previously at Cloudant, Meteor, and Parse) replied in-thread by quoting two generic applicant emails, one signed "Best, Brian." The subtext of the defense is the story: months of Puma applications are sitting in an inbox Puma is not working. The candidates who wrote the good ones are still out there, still looking, and now publicly angry.

For anyone competing for the same profile, the comment tree is doing three jobs at once:

• Intent filter. These engineers self-selected into an on-device AI browser startup. They did not apply to 300 places on a whim.
• Effort filter. They wrote bespoke cover letters. boltzmann-brain explicitly says handwritten, Puma-specific, CV attached.
• Availability filter. They are complaining in September about a February application. If they were placed, they would not be posting.

A reply to boltzmann-brain, Jose, or Gleb today is warmer than any LinkedIn InMail you will send this quarter.

The scarcity math nobody is running

The reason the Puma pool matters more than a normal "candidates who applied somewhere" list is that the addressable pool for this exact profile is tiny, so every ghosted applicant is a large percentage of the market. In Refolk's index of professional profiles, the US shortage on this profile is stark:

Two hundred and fifty-three US iOS engineers with Swift as a current title. Sixty-seven US ML/AI engineers who list LLM work in a current title. That is the entire named pool for a Puma-shaped role in the United States before you even filter for interest.

67

Five ghosted HN applicants sourced against that 67 is roughly a 7% expansion of your addressable US pool. That is not a rounding error. That is the difference between filling the role in Q4 and rolling it into Q1.

Why the founder's reply doubles the list

Yuriy's in-thread defense is more useful than the ghosting complaint itself, because he quoted two real applicant emails to prove they were low-effort. The mechanism: to publicly rebut "you ignored careful applications," he had to open the inbox and paste from it. That confirms the inbox is intact, unarchived, and searchable. A competitor who reaches a Puma applicant and asks them to forward the original thread gets a de facto resume book with dates, cover letters, GitHub links, and take-home submissions from puma.tech/take-home-task.

You will not get all of them. You do not need all of them. You need three.

How to work an HN "Who is hiring?" ghost thread in one afternoon

Treat the thread like a queue, not a story: pull the names, enrich them, message them within 48 hours of their complaint, and pitch the one thing the ghoster cannot offer, which is a reply. Here is the actual sequence.

1. Pull every named applicant complaint from the thread. Search the September 2026 "Who is hiring?" post for words like "ghosted," "no response," "applied in," and "heard back." Note usernames and application dates.
2. Resolve usernames to real identities. HN handles frequently match GitHub handles, personal domains, or a linked profile in the user's HN bio. boltzmann-brain, Jose, Gleb are starting points; the profile page and their comment history usually finish the job.
3. Enrich with public signal. GitHub contribution graph, personal site, last three job changes on LinkedIn, any conference talks. You want to walk into the first message knowing what they build.
4. Message on the platform they complained on. HN's reply link plus an email through the profile beats a LinkedIn InMail from a stranger. Reference the specific complaint. Do not pretend you did not read it.
5. Pitch the reply, not the role. The differentiator against Puma is not equity or logo. It is a human answering within 48 hours. Say so.
6. Redirect to adjacent employers if you are not hiring. Perplexity, Arc / The Browser Company, Sesame, and Vercel all pull the same profile and appear as top employers in Refolk's iOS index. If you are a headhunter, that is your placement fee.

Steps 2 and 3 are the ones that eat the afternoon, and they are the ones Refolk collapses: describe "iOS engineers with Swift and CoreML who applied to browser-agent startups in 2026 and are active on HN," and get a ranked shortlist with GitHub, LinkedIn, and open-web signal already stitched together. That is the exact gap between reading a thread and actually sending outbound.

The repeatable loop: HN threads leak candidates every month

The Puma incident is not a one-off; the "Who is hiring?" thread runs monthly and has been leaking founder behavior at scale since 2018. HNHIRING has indexed 60,305 job ads from these threads going back to January 2018, and the September 2026 thread's roughly 397-comment volume is typical, not exceptional. That means the ghosting-and-complaining pattern reproduces every 30 days, on schedule, with a new set of founders and a new set of named applicants.

60,305

Two public tools already scrape these threads: HNHIRING at hnhiring.com and Kenny Tilton's whoishiring browser at kennytilton.github.io/whoishiring. They surface the postings. What neither does is close the loop from "employer posted" to "here are the specific named engineers who applied and did not hear back." That closure is manual and it is where the pool actually lives.

The ghosting complaint is not the drama. It is the qualification signal your ATS could never generate.

One data point from that same September 2026 thread makes the yield obvious. A different poster openly stated they received 216 applications in August and issued about 10 offers. That is a ~4.6% offer rate, meaning ~95% of high-intent HN applicants were unplaced within a month of applying. Multiply across the dozens of employers in a single monthly thread and the aggregate ghosted-but-qualified pool is larger than most sourcing tools' entire "AI engineer" segment.

Why YC branding makes the poach easier, not harder

Puma's post name-drops an investor list built to intimidate: Protocol Labs, Chris Larsen of Ripple, Anatoly Yakovenko of Solana, Illia Polosukhin of NEAR, Sridhar Ramaswamy of Snowflake, Jason Warner of poolside.ai (ex-GitHub CTO), Don Ho of Orange DAO, Evil Rabbit of Vercel, John Phamous of SF Compute. That prestige is exactly why careful engineers wrote careful cover letters. It is also why the ghosting stings enough that people posted about it publicly.

For a competing recruiter, that prestige asymmetry inverts on contact. A founder who replies within 48 hours to a Puma-ghosted engineer beats the YC brand on the one axis that matters to a jilted applicant: being treated like a person. You do not need a bigger cap table. You need a reply-all button and 48 hours.

What to actually say in the first message

Open by naming the specific complaint, name the specific role you have (or the specific competitor you would redirect them to), and offer a concrete next step with a date on it. Do not apologize on Puma's behalf. Do not editorialize about YC. Do not send a template.

A working template, generalized:

Saw your comment under the September "Who is hiring?" post about the Puma application in February. I run engineering at [company], we ship [thing] on iOS with on-device inference, and I would rather answer within a day than not answer at all. If you are still looking, I can put 30 minutes on the calendar this week with [name], the person who would actually work with you. GitHub link if you want to see what we ship: [link].

Three things this does that a normal cold outbound does not:

• Acknowledges the specific event. You read the thread. You are not scraping.
• Puts a real person on the calendar. Not "the team," not "a recruiter." A human name.
• Signals technical fit in one line. On-device inference, iOS. Not "exciting AI role."

A senior sourcer running this play against a single monthly HN thread, resolving handles to full profiles and enriching them in bulk, can send 30 to 50 warm, referenced messages in an afternoon. The reply rates on referenced outbound against a ghost-pool are structurally higher than cold, because you already know they wanted this kind of job.

The one filter to add before you send

Before you send anything, filter for people who have not silently taken a job in the six weeks since they complained, because a ghosted applicant in February may already be placed by September. Check three signals fast: LinkedIn current-title change since the complaint date, GitHub contribution pattern shifts (a sudden burst of internal-looking commits under a new org), and personal-site or resume updates. If any of the three moved, deprioritize.

This is another place Refolk shortens the loop: ask for "candidates who publicly complained about being ghosted by Puma AI on HN and have not started a new role in the last 90 days," and let the ranking sort out the ones who are still actually available. The alternative is 40 open tabs.

FAQ

Is it appropriate to message someone based on their public HN complaint?

Yes, as long as you reference the complaint honestly and pitch a real role or a real referral. Public comments on Hacker News are public. The line to avoid is pretending you did not read the thread, or using the complaint as leverage ("since Puma ignored you..."). Acknowledge, do not exploit. Recruiters and founders reach out based on public GitHub activity, conference talks, and blog posts every day; a hiring thread comment is the same class of signal, with more intent behind it.

How do I find these ghost threads reliably each month?

The "Who is hiring?" post drops on the first weekday of every month, pinned near the top of Hacker News. Aggregators like hnhiring.com and nchelluri.github.io/hnjobs index the postings but not the complaint threads underneath, so you still have to read the comments. Set a monthly calendar reminder for the second week of each month, which is when applicants who applied on day one have realized nobody is answering. That is when the named complaints tend to appear.

What if the ghosted applicants are not a fit for my role?

Broker the introduction to an employer who is a fit and treat it as a placement. Puma's profile (iOS, on-device ML, browser extensions) overlaps directly with Perplexity, Arc / The Browser Company, Sesame, and Vercel, all of which show up in Refolk's iOS and ML top-employer lists. A ghosted Puma applicant redirected to a hiring competitor within 48 hours is a legitimate agency placement, and you built the relationship on the strength of actually answering.

Does this work for non-engineering ghost pools?

The pattern works anywhere the applicant pool is small, self-selecting, and public. HN is the sharpest case for engineers because the comments name names, but design portfolios, sales candidate communities, and product manager cohorts show the same structure: high-effort applicants, tiny addressable pools, and public evidence of who got ignored. The playbook (pull names, enrich, reference the specific event, put a human on the calendar within 48 hours) transfers cleanly.