Say what you will about Cluely,¹ the app that sits in on your Zoom calls and gives you instructions about what to say. Useful for sales calls, interview answers, or even a D&D session where you just really want to impress. It feels purpose built to go viral: from the slick marketing promos to the slogan "cheat on everything."²

You've probably seen an "auto-notetaker" join your Zoom calls before. They act as a virtual scribe to free up participants to focus on the material instead of taking the notes.³ Behind the scenes they're mostly just using Whisper and some LLM to handle the summarization.

This new wave of OS-integrated utilities is different:

1. They don't show up as a participant since they tie into the system audio APIs directly
2. They might not even show up if you share your screen since they try to hide from screenshare software

If everyone is doing their job right (the interviewee, this overlay software, the LLMs) - they'll answer each question perfectly and you'll have no idea. Maybe just some nagging feeling of "huh that answer was a bit too perfect."

It's becoming pretty expected that any interviews done virtually will be lossy. Even when people are not straight up lying about their identity, it's becoming way harder to get signal from people via remote interviews. Pedigree doesn't tell you much either. I've met many a university grad that can't code that well. I've met a lot of self-taught engineers that are rockstars.

This isn't really all that new. I heard a story ~5 years ago of some guy who applied to a company, absolutely crushed the take-home interview (near kernel optimized C code), and did pretty well on the systems design whiteboard interviews. He got the job and it soon became clear in his PRs that he had no idea how systems engineering actually worked. Turns out his partner worked at Apple in some low level llvm optimizer role.

But sometimes risks are introduced by making things easy, not just making them possible. Photoshop has existed for a long time to doctor photos if you're willing to invest the time and effort; lowering the barrier to doctored photo editing raises new risks. Risk profiles can change by virtue of scale. The same is becoming true for interviews.

The mouse always wins

The obvious response to AI-assisted interviews is to try to detect when someone is using AI assistance. Companies are already experimenting with browser lockdown software, eye tracking, and requiring candidates to stream their entire desktop during technical interviews.

But there's only so much you can do when your interviewees control the end device. If they control the drivers that are recording the screen or recording their video, they're eventually in fuller control than the lockdown software. In a cat and mouse game where the cat is up against an environment they don't control, the mouse wins.⁴

AI plagiarism detectors are increasingly unreliable: the risk of false positives is pretty high, and most cautious students can mitigate that anyway via rewriting strategies, changing prompts, or even using secondary models. All the plagiarism detectors have to go off-of are pattern matching. They look for telltale signs of AI generation like unusual word choice or specific stylistic fingerprints. Going multimodal makes it even harder.

Plus, what signals are they even looking for? When someone is using Cluely, they're typically not reciting AI text verbatim⁵. They're getting suggested talking points, clarification on technical concepts, or code as output. Trying to classify this output on any kind of binary scale is pretty fraught and the training/evaluation sets almost certainly don't exist yet.

Is the result that we just have to get physical again with everything? I've anecdotally seen this a bit more recently: some companies are trying to solve this by moving back to in-person interviews, even in early rounds, or doing a socratic deep dive on what was actually written during a take home to see if candidates really understand it.

Certainly we're going to place more emphasis on reputation. Public reputation online, private reputation within your industry, within past people that you've worked with, etc.

This has always played a role in the Valley. Second time founders get way more venture interest than first time ones. Experience and connections count for a lot. But for the average software engineer looking for a job, there's a reason why Cracking the Coding Interview and Leetcode exist. It used to be true that if you can get through resume screening, crush it during a series of interviews, you could wind up with 8 remote job offers. Not so any more.⁶

Bootstrapping reputation

If you already have a strong professional network, this shift probably sounds incredible. You've worked hard - you've built a strong reputation. Your connections can vouch for your work, provide warm intros, and offer actual references. When more things hinge on your own reputation, people are going to elevate the people that make them look good as peers. Which means not advocating for people that are just mediocre. Kudos, you've made it.

If you're early in your career or changing industries, this change hurts. Perhaps even locked-out-of-the-job-market hurts. You don't have the established relationships that can propel you into the interview cycle and ideally onto an onsite. You're competing both with AI-optimized resumes and people cheating on the remote interviews.

When companies have limited budget and limited headcount, expending the energy to try to sort through unknown applicants or fallback on reputation becomes a tougher calculus.

This creates a bifurcated job market. For senior roles at competitive companies, it's increasingly about who you know and who can personally vouch for your work. For entry-level positions, companies are moving toward more extensive practical assessments.⁷ Showing initial out of the box thinking and familiarity with new tools (extensive Agent use, social media virality, etc) become more important than skills or perhaps even grit alone.

Reputation markets

Reputation systems naturally tend toward winner-take-all dynamics. They're maybe the ultimate network effect.

Once someone has established credibility in a network, it becomes increasingly easy for them to maintain and extend that credibility. People are more likely to recommend people they've already seen others recommend. We're seeing that right now with the $200M pay packages for top AI researcher while some recent CS grads are unable to land even an intro interview.⁸

The signaling costs are changing too. With the interview process potentially compromised, the signal value shifts to who referred you and how strongly they're willing to vouch for your work. This is creating new markets for reputation brokerage. Professional referral services & paid introduction platforms are emerging to help people build the connections they need. I see these ads online all the time.

I imagine some of these help - but more because of random chance than deliberate construction. Most people with an existing network don't feel the need to play that game, which leads to an isolated sphere of the reputation brokers. I'd personally discount any recommendations that come out of networks that are pay to play.

Hiring - like matchmaking - is better when it's casual, no stress, and organic. A rising tide situation.

There's an interesting parallel here to academics. I think that world has always been more reputation-based than the tech industry. Your PhD university, who your advisor was, and which senior researchers will write recommendation letters often matters more than your performance in job talks. Knowledge work might be converging toward this model too.

Historical precedents

Let's turn back the clock a bit. This isn't the first time tech has disrupted trust capital.

Photography in the 1800s: Initially increased trust in visual evidence - "seeing is believing" became more literal. But it didn't take long for photo manipulation techniques to emerge, forcing courts and newspapers to develop new verification standards.

Transition from handwritten to typed documents: Handwriting analysis had been a primary method for verifying document authenticity. Typewriters made documents more professional but they also made forgery easier. Legal systems had to adapt by developing new authentication methods and changing what types of evidence were considered reliable. That's one of the reasons why we still need to wet-sign documents.

Interview Coaching: Companies responded to professional resume writing services by developing more behavioral interviewing techniques, introducing practical skill assessments, and placing greater emphasis on reference checks.

Communities on the web have also gone through multiple iterations even in recent memory. Forums seem so quaint by modern standards. They all had an unbelievable depth of knowledge and community participation. They were also all largely honor-based systems - they could self police, even anonymously, in a largely decentralized way. As they scaled, reputation systems emerged giving us user ratings, karma points, and verified accounts.⁹.

Verification systems

All these historical instances added some verification regimen over time. That regimen is different for each medium. Some of it are developing skills (ie. professional image manipulation experts) and some of it is cultural (non-anonymous social platforms). What could that look like for interviews?

Unfortunately I don't have a ton of good answers here aside from the obvious. Pair program. Do take home projects actually targeted towards your company challenges. Live debug a legitimate issue. Most of that is just banal advice.

However to add an LLM angle: When interviewers have a ton of context on the interview question, versus just some esoteric answer, they'll be more likely to engage and challenge interviewees on their answers. If your interviewee always responds with a "you're right let me change my approach," that smells a lot like an LLM alignment to me.

Aside from that, deep interest is also hard to fake over a long period of time. Making a few random test apps on Github is one thing. Diving deep into some rust internals to showcase the Arc strengths and weaknesses shows something else. Even if some or all of it was actually vibe coded it shows some signal about their relationship with engineering.

It seems clear that we're heading toward a more portfolio oriented world where individual reputation matters more than institutional affiliation. Instead of "I work at Google," it becomes "I'm the person who built X, collaborated with Y, and delivered Z results."

Maybe recommendations get more systematized. I was brainstorming on X the other day about how social media could switch up their game theory incentives. The same is probably true for this high reputation future. Right now reference checks are largely throwaways. Does the dynamic change when you literally have to stake some part of your reputation on other people?¹⁰

The reputation-first future

We all seem to accept that companies need to invest in marketing - even vibe based marketing¹¹ - to build interest in what they're offering. The labor market is a market just like the goods market. So the same rules should apply. But if you're not a content creator, you probably haven't put much consideration into doing the same.

Most people just don't want to be doing work shit outside of work. And marketing yourself is certainly work. This has meant that most industries haven't have to spend much time filtering for people that actually care. People will self-bucket into the groups that want to do extra work and the ones that don't. When doing that extra work is just a prompt call away, I imagine that will change.

I haven't yet really figured out a way that we can help to mitigate this economic disruption for people just starting out their careers. Start building a reputation early, I suppose? Look for people that you can get to know personally / hack around casually before you try to get into a more formal arrangement? Ironically most of this advice overlaps with the advice of how to find a good cofounder. It just hasn't been necessary to land a job.

The bigger question is whether this shift ultimately leads to better hiring outcomes. In theory, reputation-based systems should be more accurate since they're based on actual work performance rather than interview performance. But they're also more susceptible to bias, network effects, and manipulation by those with resources to game the system.

If you have any brilliant ideas here, please, do let me know. Millions of college grads will thank you.