AI literacy is the real career filter now, because everyone claims to use AI and almost no one can prove it. The claim is free. The proof is rare. In a market where every resume lists the same three tools, the people who get hired are the ones who can show they understand what those tools actually are. The rest get filtered out, often without ever knowing why.
This is the part that does not get said out loud. We have moved past the moment where "I use AI" was a differentiator. A year ago it might have set you apart. Today it sets you in with everyone else. When a phrase appears on every resume, it stops carrying information. So hiring managers have quietly moved the bar. We are no longer impressed that you use AI. We are watching how you talk about it. And how you talk about it reveals, in a few sentences, whether you understand the machine you claim to depend on.
Last month I taught a live session to a cohort at one of India's better engineering programs. Working professionals, young designers, a few sixteen-year-olds who had grown up with these tools. Smart room. Halfway through, a participant referred to Higgs Field as a model. I stopped and said, as gently as I could manage, that if he had said that in an interview with me, I would have ended it on the spot. He looked surprised. I was not trying to embarrass him. I was trying to give him the one piece of feedback nobody else in his life was going to give him before it cost him a job.
Because it has cost people jobs in front of me. I interview candidates. I have sat across from developers who told me, with full confidence, that they use Perplexity as their large language model. Or Copilot. I had one candidate tell me he uses GitHub Copilot as his LLM. I said, "let us talk later," and I closed the interview. Not out of arrogance. Out of math. If a developer does not know that Copilot is a wrapper and not a model, I cannot trust their judgment on the harder calls, and the harder calls are the entire job.
People hear that and think I am being harsh. I understand why. But put yourself on my side of the table. I am not hiring someone to type prompts. I am hiring someone whose decisions will compound over months. Which model to use for a workload. When a tool's limits are about to hurt the product. Whether a vendor's claim is real or marketing. Every one of those calls rests on a foundation, and the models-versus-wrappers question is the cheapest, fastest test of whether the foundation is there. It takes one sentence to find out. I would rather find out in minute three than in month three.
The story behind the rejection
Here is what was going through my head in that room, and in those interviews.
Perplexity does not have its own model. Copilot does not have its own model. They are wrappers. They sit on top of models built by someone else and add a useful interface, some retrieval, some routing. That is genuinely valuable work. I use both. But calling them an LLM is like calling the steering wheel the engine. It tells me the person has never looked under the hood, and the job I am hiring for is mostly hood.
Higgs Field is a clearer case. It is not a model at all. It is a platform that aggregates other models, Kling, Kimi, Qwen, GPT, nano, and lets you reach them, including through tool connections into something like Claude. Knowing that one fact, that Higgs Field orchestrates models rather than being one, is the difference between being taken seriously by a developer and being quietly written off. I have watched both outcomes happen in real time.
I have twenty-nine years in this industry. Long enough to have run my own restaurant in 2000, long enough to have hired across many roles. What I have learned is that depth shows up in language. When you develop an eye for AI, you develop a language for AI. And that language is exactly what surfaces in interviews and in conversations with your seniors. You cannot fake it for sixty minutes. The vocabulary gives you away in the first five.
The uncomfortable part is how common the gap is. There are people calling themselves AI engineers and forward-deployed engineers who do not know the basics. We have a phrase for this in Hindi. Andhon mein kaana raja. In the land of the blind, the one-eyed man is king. A lot of self-described AI experts are running on one eye and a confident tone. That works until they meet someone with two. And the market is filling up with people who have two, faster than most one-eyed experts realize. The window where confidence alone could pass for competence is closing.
The three-signal literacy test
So what am I actually testing for? Not trivia. I do not care if you can recite parameter counts. I care about three signals, and they map to how serious practitioners actually work.
- Models versus wrappers. Can you tell me which tools have their own model and which sit on top of someone else's? Claude, GPT, Gemini are models. Perplexity, Copilot, Higgs Field are not. This is not pedantry. It tells me whether you understand where capability comes from and where it can break.
- The model stack. Do you rely on one model, or do you run a stack? Every serious individual and company I know runs a stack now. They do not bet everything on Claude. They know each model's strengths and each model's gaps, and they pick the right one per task.
- Operating past the chat box. Are you using these tools through their default chat window, or are you working at a deeper level? If your entire experience of AI is typing into a chat interface, you are scratching the surface. The people doing serious work are often not even using the front end.
Signal 1 Signal 2 Signal 3
Models vs Wrappers Model Stack Past the Chat Box
| | |
+--------+---------+---------+----------+
| |
Surface user Real practitioner
(claims to use AI) (can prove it)Let me take the stack point further, because it is the one most people miss. Claude is a powerful model, but it is not an omni model. It is largely a text and reasoning and math and coding model. It can read images but it cannot create them. GPT can create images. GPT does not create video. Gemini does video. None of this is a knock on any of them. It is just literacy. If you know each model's deficiencies and efficiencies, you stop blaming the tool and start picking the right one. Picking the right model for the task is itself a skill, and it is a hireable one.
The person who runs a stack stops asking "is this model good or bad" and starts asking "good for what." That shift sounds small. It changes everything about how you work. You write a long technical brief and reach for the reasoning model. You need a visual and switch to the one that draws. You want a short clip and you move to the one that handles video. You stop forcing one tool to do work it was never built for, and then you stop being disappointed when it does that work badly. In an interview, this is the answer that separates the two candidates who both said they use AI. One of them can tell me which model they reach for and why. The other just has a favorite.
The third signal is where most of the real ceiling sits. People hit a wall, fire off a single prompt, get a mediocre result, and conclude the tool is overhyped. The tool is fine. The usage was at level one. Think of it as a ladder. Using Claude Code on the Claude desktop app is level two. Level four or level five is a completely different way of working, and the output gap between them is enormous. The limitations people complain about almost always come from low-level usage, not from the model. I wrote a full field guide on these levels, and the pattern repeats everywhere I look. The people who say AI did not work for them were usually standing on the bottom rung.
Here is how I hear it in an interview. A candidate at level one talks about AI the way a passenger talks about a car. It took me here, it was slow today, the traffic was bad. A candidate higher up the ladder talks about it the way a driver does. They describe what they set up, what they connected, what they automated, where they let the tool run and where they kept their hands on the wheel. The vocabulary is the giveaway again. Passengers describe outcomes that happened to them. Operators describe systems they built. I am always hiring the operator, and the operator is almost never the person who only uses the chat window.
None of this requires you to be a researcher or to train a model. It requires curiosity and a refusal to stop at the front door. The front door is the chat box. It is comfortable, it is well lit, and most people never walk further in. The whole house is behind it.
Three things to do this week
You do not need a new tool or a budget to close this gap. You need a week and some honesty about where you actually are.
- Audit your own vocabulary. Write down every AI tool you use and label each one as a model or a wrapper. If you cannot label it confidently, look it up today. The act of sorting them is the literacy. Do this before your next interview or your next conversation with a senior, because that is where the labels get tested.
- Build a two-model stack. Pick one task you do weekly and run it through two different models. Notice where one is sharper than the other. By Friday you should be able to say, in one sentence, which model you reach for and why. That sentence is what a hiring manager wants to hear.
- Climb one rung past the chat box. Take the tool you use most and learn one capability beyond the default chat window this week. A project, a connected tool, a coding interface, anything that moves you off rung one. You will feel the difference in output immediately, and you will stop blaming the tool.
What to Read Next
While you are here, the back catalogue has more on this: