Students who want to work in artificial intelligence are usually told to collect more: another course, another certificate, another programming language, another summer program. Anthropic and OpenAI publish hiring guidance that points in a more demanding direction. Both organizations want evidence that a person can enter difficult territory, make progress, and explain what happened.
That is useful even if you are years away from applying. Anthropic currently says it does not offer internships, and neither company is presenting a checklist for high school applicants. The value is in the signal. Their guidance reveals what a strong technical portfolio should prove long before a résumé reaches either organization.
Anthropic puts direct evidence near the top
Anthropic is unusually plain about what it wants candidates to foreground. Its careers page says independent research, a thoughtful blog post, or a substantial open-source contribution belongs at the top of a résumé. It also says the company cares about what candidates can do, not where they learned it, and notes that many technical staff arrived without prior machine learning experience.
The interesting part is not that credentials are irrelevant. Many roles still demand deep expertise. The interesting part is that the company names artifacts that expose judgment. Independent research reveals whether you can define a question. A technical essay reveals whether you can organize an argument. Open-source work reveals whether you can understand an unfamiliar system, operate within constraints, accept review, and leave the codebase better than you found it.
Anthropic also describes a blurred boundary between research and engineering. Engineers contribute to research papers, while researchers build systems. For a student, the practical lesson is that a notebook full of ideas is incomplete and a repository with no intellectual question is incomplete. Strong work connects the question, implementation, evaluation, and explanation.
An open-book interview still tests fluency
Anthropic says technical candidates can look things up during live coding. That sounds forgiving until you consider what it makes easier to observe. If syntax recall is not the main obstacle, the interviewer can pay closer attention to problem framing, tradeoffs, implementation choices, testing, and debugging.
The company still expects familiarity with basic syntax and common libraries because constant searching consumes time. The deeper standard is not memorizing an entire language. It is knowing enough to preserve attention for the hard part. Students should practice with documentation available, then track whether they are using it to fill a small gap or to avoid understanding their own approach.
OpenAI defines high potential as a rate of progress
OpenAI's interview guide gives a compact definition of high potential: a person has demonstrated that they can ramp quickly in a new domain and produce results. That definition contains two tests. Learning speed without output can become endless preparation. Output without learning can become one narrow trick that does not transfer.
OpenAI also names collaboration, effective communication, openness to feedback, and mission alignment. Its skills assessments can vary by team and may include pair coding, take-home projects, or technical tests. The format may change, but the pattern is consistent: the work must survive interaction with another person. A candidate needs to build, explain, listen, and update.
Students can demonstrate the same pattern at an appropriate scale. Pick a domain you do not already know, establish what competent output would look like, and keep a record of the distance traveled. The final artifact matters. So does the change log showing why version four is different from version one.
Four artifacts that reveal more than a certificate
An evaluation memo
Choose a model or system, define a narrow capability, create a small test set, establish a baseline, and explain where the system succeeds and fails. Include examples that contradicted your first hypothesis.
A technical explanation
Write for an intelligent reader who does not share your context. Explain one difficult concept, one design decision, or one failure mode. Link claims to evidence and distinguish what you observed from what you infer.
An open-source contribution
Start with documentation, tests, or a tightly scoped issue. The value is not the size of the pull request. It is proving that you can read an existing system, follow its conventions, respond to review, and finish.
A replication with an extension
Reproduce a result from a paper, tutorial, or public benchmark, then change one meaningful variable. Record the environment, deviations, negative results, and limitations so another person could audit the work.
A 30-day evidence sprint
- Days 1 through 5: define the question. Read enough to identify a narrow, testable problem. Write down the claim you expect to be true and the observation that would change your mind.
- Days 6 through 12: build the baseline. Create the simplest working version. Save the environment, inputs, and initial results. A baseline gives later improvements meaning.
- Days 13 through 20: try to break it. Test edge cases, vary inputs, compare alternatives, and invite another person to challenge the setup.
- Days 21 through 26: revise. Choose one weakness that matters and improve it. Explain why that weakness deserved attention ahead of everything else.
- Days 27 through 30: publish the evidence. Clean the repository, write the memo, disclose limitations, and ask a knowledgeable reader to identify what remains unclear.
The sprint is intentionally too short for mastery. Its purpose is to produce a visible learning curve. A reviewer should be able to see the original question, the work, the failure, the feedback, and the improved result.
What weakens an otherwise strong AI portfolio
- A thin wrapper presented as research. Connecting an API to a polished interface can be useful engineering, but it does not become research without a real question and evaluation.
- Mission language copied from a careers page. Genuine motivation is specific. Explain the experience or problem that changed what you care about.
- Results with no baseline. A 90 percent score means little if the task, comparison, and failure cases are missing.
- AI-written technical prose. Tools can help inspect or edit, but the final explanation must expose your judgment. If you cannot defend a sentence, it is not evidence of your ability.
- A repository nobody else can run. Missing dependencies, private data, and unexplained setup steps turn an artifact into a claim.
The real advantage is legibility
Selective teams have limited time and imperfect information. A strong artifact reduces the amount they must guess. It shows how you choose a problem, how quickly you acquire missing knowledge, how you react when the evidence is inconvenient, and whether another person can understand or extend your work.
That is a better target than trying to look like an AI researcher as early as possible. Build something narrow enough to finish, test it honestly, and make the learning visible. When you eventually need a résumé, the strongest lines will already exist because the underlying work is real. For a broader framework, continue with the Selection Signals synthesis.
Primary sources
- Anthropic careers and hiring process - Official guidance on evidence of ability, technical interviews, backgrounds, and AI use.
- OpenAI interview guide - Official hiring philosophy, definition of high potential, and assessment formats.
- OpenAI careers - Current values, emerging-talent pathways, residency information, and open roles.