Surviving the Era of Zero Junior Developer Hiring

Summary

• In the entry-level job market that rapid AI progress is bringing, writing code and doing a first-pass review of it will be an utterly common skill.
• What the market will actually find scarce is the ability to get AI running inside a real organization and prove that it works, but the entry-level door into that work is very narrow.
• So students should go to a small organization now, actually run a system there, and build up concrete evaluation and improvement records as evidence.

Vanishing Graduate Hiring and a Shaken Major

According to a Korea Financial Newspaper report on August 14, 2026, Naver held no open recruitment for new-graduate developers in the first half of 2026. Naver had run open recruitment every year from 2019 to 2025 and, in 2021, hired a record of about 900 people across two rounds; skipping the first-half round was a first in seven years. No schedule for second-half 2026 hiring had been announced either. Kakao is in a similar position.

The employment crisis facing new engineers is following the same path outside Korea. In a report the Federal Reserve Bank of New York published in February 2026 using 2024 American Community Survey data, computer science had an unemployment rate of 6.99% among recent college graduates aged 22 to 27, fourth highest of 74 majors. Computer engineering, at 7.78%, was second highest of all majors. The numbers from the Stanford Digital Economy Lab, in a paper revised in August 2026 that analyzes US payroll data through June 2026, are more direct. Employment of 22- to 25-year-olds entering occupations highly exposed to AI was 19% lower relative to the employment trend of their peers in occupations with low AI exposure. No such gap was observed among experienced workers.

The job-search formula of the generation just before, memorizing algorithms, grinding problem sites and practicing whiteboard coding for coding tests, has changed too. In a June 2025 engineering blog post, Canva said it had replaced its traditional computer science fundamentals screening for backend and frontend engineers with an interview in which candidates must use AI, while still assessing fundamentals. Meta decided to let candidates use AI during coding tests because that is closer to the environment its engineers will work in (404 Media, July 2025). Nexon replaced the coding test for new game programmers with an AI interview and an assessment of AI use that grades how candidates write prompts and improve the results (Bloter, 2026-08-26). Algorithm tests have not disappeared, though. Samsung held its SW competency test offline on October 11, 2026. Grep, which runs the Programmers platform, says companies hiring developers keep the coding test and add its AI assessment on top (Digital Insight, 2026-10-12). And Google CEO Sundar Pichai said in June 2025 that Google would bring back at least one in-person interview round to make sure the fundamentals are there. Algorithm skill has turned from an edge into the baseline, and a second skill is now graded on top of it: using AI and checking whether its output is right.

Seniors and experts each have advice. Develop the judgment to evaluate what AI produces. Become an expert in a particular domain. Learn planning and leadership. But this advice only points at where to end up; it does not tell a student fresh out of university how to get there.

That is why this essay looks at the market of 2029 rather than today. Students now in their first to third year of university will, after military service and their remaining semesters, look for their first job sometime between 2028 and 2030. Preparing for today’s entry-level market means arriving at graduation ready for a market that has already become the past. So we need to forecast the 2029 labor market from the trend in AI models’ task capability measured by METR and from data on how companies are actually adopting AI, and work backward to what to do now to get there.

What computer science students entering the workforce between 2028 and 2030 need is not a vague destination but an evidence-based reading of how the market is changing and a concrete path they can start on now.

An Abundance of Code Writing and First-Pass Review

Writing code yourself and finding its bugs will not be a differentiator for a new engineer in 2029. The speed and range at which machines write software have begun to outpace the human learning curve.

According to the time-horizon measurements of METR, an AI model evaluation organization, an early Claude Mythos Preview model measured in April 2026 completed tasks that take a human expert about 17.4 hours with a 50% success rate, and tasks that take about 3 hours with an 80% success rate. The task length METR measures has doubled roughly every 89 days since 2024, and roughly every 131 days when measured from 2023. If you simply extend the 80%-success figure of 3.1 hours at a doubling every 131 days to June 2029, the length comes to about 1,400 hours. That is about 35 weeks of work at 40 hours a week. METR itself warns that extrapolating to tasks lasting months carries large uncertainty, so this figure should not be taken as an absolute completion time. But the direction is clear. By 2029, development work that is well specified and self-contained will mostly be done by AI.

Length of tasks AI can complete

Inside big tech, this shift has already happened. By Anthropic’s own count, Claude wrote more than 80% of the lines of code merged into its internal codebase as of May 2026. At the fintech company Stripe, more than 1,000 agent-written pull requests containing not a single line of human-written code are merged every week, and all of them are reviewed by human engineers. Inside OpenAI, Codex does the first-pass review of every internal pull request. The labor of writing implementation code and doing first-pass review is moving to machines, and the final verification that remains falls to experienced people who know the system’s internals deeply.

The Gap Between Model Capability and Real Deployment

The ability of models to write software is soaring, but systems actually running inside real organizations are still rare. The gap between technical performance and working deployments is not closing.

In a McKinsey report published in August 2026, based on a survey of 1,719 people from May to June 2026, the share who said AI had any impact on EBIT was 37%, little changed from a year earlier. The share of high-performing companies also stalled at about 6%. The share saying they had scaled AI across the enterprise rose from 38% to 44%, yet real business results stood still. In a National Bureau of Economic Research (NBER) Digest on a survey of about 6,000 executives in four countries from November 2025 to January 2026, about 70% of firms said they used AI, but the effect on employment and productivity remained small.

AI adoption vs. real returns

The gap widens as companies get smaller. In the McKinsey survey, the share of companies with revenue of $1 billion or more reporting that they were scaling AI agents jumped from 27% to 40%, while it stayed at 22% among companies with revenue under $1 billion. In Eurostat data extracted in December 2025, 55% of large European enterprises used AI, against only 17% of small ones.

Benchmarks in which software operates screens and operating systems to do real work show what the failures look like. OSWorld 2.0, released in June 2026, measures 108 real workflows with a median human completion time of about 1.6 hours. On the leaderboard updated on October 15, 2026, the best model on the current task version (v2.1), Claude Opus 5, completed 44.3% of them, fewer than half. The researchers recorded clear causes of agent failure. Agents forgot constraints, missed new information that arrived mid-task, guessed at moments when they should have asked the user, and skipped verification. By contrast, on the earlier benchmark OSWorld-Verified, which evaluates relatively short tasks, the best model scored 86.0%, above the human baseline of 72.36%.

AI companies have begun putting large sums into closing this gap. In May 2026, OpenAI launched the OpenAI Deployment Company with more than $4 billion. Its initial core staff were about 150 experienced forward deployed engineers (FDEs) brought in through an acquisition. An FDE is an engineer an AI company sends into a customer organization, spending most of their time on site rather than at headquarters, to connect AI to the customer’s own data, internal tools and work processes and make it actually run. In October 2026, Anthropic announced a $100 million investment to train 10,000 forward deployed engineers by the end of 2027, stating that people who can make AI work inside a real business have become the “hardest talent to find” in the market. As Epoch AI anticipated in its outlook for 2030, real-world deployment and social impact may lag considerably behind gains in model performance itself.

What a Forward Deployed Engineer Actually Does

Getting AI running in the field is not a matter of vague communication skills or customer-service technique. It is engineering work: logically decomposing someone else’s messy work, physically connecting an AI system to that organization’s data and tools, and proving with numbers that the system works properly.

Three steps to put AI to work

Colin Jarvis, who leads OpenAI’s forward deployed engineering organization, said in an August 2026 newsletter interview that his engineers spend about 50% of their working time at customer sites. He explained that the organization’s value lies not simply in writing code but in actually solving the customer’s problem. The working process the OpenAI Deployment Company has published is concrete. It diagnoses the customer’s actual work, selects a small number of high-priority core workflows, and then goes inside the organization to design, build, test and deploy systems that connect the model to that company’s own data, internal tools, control systems and business processes.

Forward deployed engineers still spend most of their working hours writing code. Among the 1,000 FDE postings Bloomberry analyzed, the most common type, about 60% of postings, describes the job as 70–90% coding. Meanwhile, Hamel Husain and Shreya Shankar, who write on AI evaluation, report that on the projects they themselves worked on, 60–80% of development time went to error analysis and building evaluations. As AI writes a larger share of the code, what separates one engineer from another is the ability to judge whether the system’s output is right. Coding takes most of the hours, but evaluation is what sets engineers apart. IBM’s chief human resources officer, Nickle LaMoreaux, said in a February 2026 interview that IBM is tripling entry-level hiring, including software developers, and stressed that new developers still write and test code but, rather than just maintaining existing products, meet customers directly, gather feedback and build new products that did not exist before.

The more the middle step of implementing code is automated, the more both ends expand: the front end, where you break down and define a problem in someone else’s messy environment, and the back end, where you evaluate whether the system works safely to requirements. Neither someone who only talks well nor a developer who only types code to a given spec can do this work.

What Remains and the Limits of Automation

System integration and deployment will also gradually be affected by automation tools. In a 2025 report, the venture capital firm Andreessen Horowitz (a16z) argued that AI tools now let implementation and integration work be done far more efficiently than before, and that some of it can be fully automated. The startup Sierra promotes its Ghostwriter tool by saying people no longer need to write integration code by hand. Sierra has not, however, published concrete data on deployment time or staffing.

But some work remains even as tools advance. It is exactly where agents still failed on more than half of the tasks in OSWorld 2.0: grasping hidden work context, asking the user instead of guessing under uncertain conditions, and rigorously verifying what the system produces. METR has pointed out that even if the number of human interventions an AI system needs is halved, the amount of human labor each intervention takes may actually grow, because the failures become much more complex. It added that for tasks that demand extreme reliability and are hard to verify, automation needs a success rate of 98% or more to have real economic value.

The share of work that must pass through human hands may shrink, but each intervention gets much harder. This narrow, demanding territory goes only to people who have run systems in real environments.

What a New Graduate Must Prove in a Market Built for Experience

The entry-level door is narrow, because the hiring market is reorganizing around experienced workers. According to Indeed Hiring Lab, entry-level positions requiring 0–1 years of experience made up 4.5% of US software development job postings in the first quarter of 2026, the lowest of the occupations studied. The Korean market is the same. According to a joint Saramin–Jumpit survey announced on August 27, 2025, new-graduate postings were only 0.8% of all developer postings, yet new-graduate job seekers accounted for 29.5% of applications, the largest share of any experience level.

Junior developer postings vs. applications

Forward deployed engineering also looks experience-heavy on the surface. When Bloomberry analyzed 1,000 FDE job postings, among those that stated an experience requirement, 12% targeted 0–2 years while 60% required 3–5 years. On October 29, 2026, I went through the public careers pages of eight major companies myself: OpenAI, Anthropic, Databricks, Scale AI, Cohere, Decagon, Harvey and Ramp. Of 277 FDE-related postings, none carried an entry-level title. A Databricks posting stated outright that it was not aimed at internship, new-graduate or entry-level applicants.

One might expect that once experienced engineers are all deployed and there are still not enough people, companies will finally turn to new graduates. But the shortage is being filled first by retraining experienced staff and converting them into FDEs. As noted earlier, OpenAI’s deployment company started by absorbing experienced people at once through an acquisition. The program through which Anthropic announced it would train 10,000 FDEs by the end of 2027 also starts with engineers currently working at Accenture, Bain, Capgemini, Commonwealth Bank of Australia, Deloitte, McKinsey, Morgan Stanley and Novo Nordisk. They go through a 12-week residency in which they are trained by leading a real project at their own organization (Anthropic, October 2026). In Korea, Samsung SDS announced in a September 8, 2026 press release that it would train about 200 FDEs within the year. In Bloomberry’s analysis of 100 LinkedIn profiles of current FDEs, 45% had previously been software engineers. The more routine coding AI absorbs, the larger the pool of experienced developers who can be converted. The hope that experienced workers will run out and it will then be new graduates’ turn is unlikely to come true any time soon.

So where should new graduates get in? In Bloomberry’s analysis of 1,000 FDE postings, 12% of the postings that state required experience target 0–2 years, and these come mostly from companies newly building FDE programs. And 58% of all FDE roles are at growth-stage companies with 11–200 employees. Palantir, the company that created the FDE role, was not among the eight companies counted above. On its public hiring board on October 29, 2026, 38 of its 108 FDE-related postings (forward deployed software and infrastructure engineers and deployment strategists) were New Grad or Internship roles, covering commercial and government clients in the US, the UK, France and Australia.

Palantir’s posting for a new-grad forward deployed software engineer defines the role clearly: “As a New Grad, you will be entrusted with responsibility on day 1. You will not be handed a ticket queue. You will not be shielded from the customer, the chaos, or the consequences.” Look beyond the frontier labs, which absorb experienced engineers, and the entrances open to new graduates come into view: FDE programs with their own new-grad and internship tracks, like Palantir’s, and growth-stage companies just starting to build FDE teams. What these entrances ask of applicants matches the evidence this essay recommends building: having taken responsibility in front of a real customer.

The era of new hires who only processed tidy tasks someone else had prepared in a ticket queue is over. What separates new-graduate candidates at this narrow door is evidence that they have already actually done this work.

How to Start a Student Residency

The portfolio students usually build to prove themselves is a polished, smoothly running service for users or a fun app. But as METR’s trend shows, a well-specified, self-contained project like this is a task AI will mostly handle by 2029, and it no longer distinguishes job seekers. Getting an AI system running inside a real organization, on the other hand, is where today’s agents are weakest (the OSWorld 2.0 failure causes described above: forgetting constraints, guessing instead of asking, skipping verification), and it is where AI companies are paying heavily for people.

The realistic way to train for this is a student residency. Before graduating, find one or two real organizations outside computer science, identify a recurring problem they have, connect an AI system directly into their work, and actually run it for several months. Just as Anthropic adopted the medical residency model for training forward deployed engineers, students too need to create for themselves a setting where they run and verify a system on a real case.

Small organizations such as a university administrative office, a neighborhood shop, a nonprofit or a small business make an excellent practice ground. Korea had 8,499,552 small and medium-sized enterprises as of 2024 (Ministry of SMEs and Startups, 2024 Basic Statistics on SMEs), and the AI adoption rate of small European enterprises, at 17%, lags far behind that of large ones (55%). But do not mistake them for employers or a promising paying market. Most small organizations do not feel any need to adopt AI, and fewer are spending money on outside help. In a 2024 survey by the Korea Federation of SMEs, only 5.3% of 300 Korean SMEs had adopted AI, and 80.7% of non-adopters said their business did not need it. In Eurostat data too, among small enterprises using AI, the share whose AI was developed or modified by external providers fell from 34.06% in 2021 to 25.79% in 2024, while 56.3% used ready-made commercial software as is.

So do not expect small organizations to hire students. The one decisive thing they offer a student is access. They have real, untidy problems, people who will let a student in, and no experienced FDEs competing for the work. The concrete records built there are evidence that a résumé or a toy project cannot show, and they set an applicant apart at the narrow entry-level door.

Still, knocking on a real organization’s door and being let in is the hardest barrier for a student. That is why governments and universities should run matching programs that connect students with small organizations. The precedents abroad have two things that current student–small-business matching in Korea lacks: they give the host organization real funding and support, and they keep the student inside the organization for months under academic supervision.

In the UK’s Knowledge Transfer Partnerships (KTP, Innovate UK), a graduate employed by a university works inside a company for 12 to 36 months, typically with half a day a week of supervision from an academic, and the government covers 67% of project costs for small and medium-sized firms. In an evaluation carried out for Innovate UK (SQW, 2023), participating firms showed significant growth in employment and turnover compared with similar non-participating firms, but no significant effect on productivity, and 62% of the associates surveyed were offered a job by their host company.

An evaluation of Canada’s Student Work Placement Program (ICTC, 2023) likewise found that only 24% of employers said they would have hired their student without the wage subsidy, and the biggest barrier cited was the time managers and staff spend training and supervising students. It shows that hosting a student costs a small organization money and time, and a support system must make up for that.

aivancity, an AI school in Paris, runs an “AI Clinic” modeled on law school clinics: under faculty supervision, students spend 6 to 9 hours a week, as part of their course, doing pro bono AI work for small and medium-sized enterprises, associations and nonprofits. In Korea, the Ministry of Science and ICT’s 2026 AI-centered university program selects 10 universities with a total of 25.5 billion won, supports each for up to eight years with up to 24 billion won, and lists industry AI transformation among its main tasks alongside company-linked industry projects and internships (as reported by Digital Daily). But existing matching stays short and focused on marketing, like Naver’s “SME Facilitator,” which worked with five universities to support local businesses’ online marketing for one semester. If a system is set up that funds host organizations and keeps students on site for months under supervision, this kind of residency becomes a path many students can take, not only the few who managed to get into an organization on their own.

Hamel Husain, a machine learning engineer with more than 20 years of experience who has worked at Airbnb and GitHub, and Shreya Shankar, a computer science researcher with a PhD in EECS from UC Berkeley who will join Carnegie Mellon University as an assistant professor in 2027, lay out a practical procedure for evaluating AI. Ordinary software always gives the same output for the same input, so you can write tests for the cases you anticipate in advance; an AI system answers the same request differently each time, and it is hard to list beforehand where it will go wrong. So their method starts not by imagining hypothetical test cases but by collecting the failures that actually occurred in the field. Suppose, for example, that a small nonprofit introduces an AI that drafts replies to donors’ inquiry emails. The student reviews the drafts the AI actually wrote side by side with the original emails and records the errors that really appeared: a donor’s name written wrong, a date for issuing a donation receipt made up, or the way to cancel a recurring donation explained using an outdated rule.

• Review at least 100 of the system’s execution records (traces).
• Annotate and analyze at least 30 of them yourself.
• Group the failures by type and count how often each occurs.
• Set clear pass and fail criteria for each failure type.
• Leave the final judgment of the system’s quality to one practitioner in the domain who understands the real users.

Once each failure type has a check (for example, names can be matched against the donor list in code, and whether a policy explanation matches the current rules can be judged by another AI), the student fixes the system and runs the same checks again. The change in the pass rate then becomes the record that the system actually works inside this organization.

There are three artifacts to hold at the end of this process. First, a requirements specification that breaks the organization’s messy problem down into code-level specs. Second, a failure-type taxonomy and the evaluation dataset built to check against it. Third, an improvement log that tracks usage data accumulated over the real operating period and the system errors found and fixed.

The single principle that supports this practice is fully understanding the code you deploy. In a January 2026 Anthropic experiment in which 52 participants learned a new programming library, the group using an AI assistant scored only 50% on a quiz, below the 67% of the group that wrote code by hand. The largest gap was in debugging. However, a small number of participants who had the AI generate code and then kept asking follow-up questions until they understood how it worked scored high. This is a tendency seen in a very small sample, but the implication is clear. Only by mastering machine-written code, rather than being content to churn it out quickly, can you fix the complex errors that arise in the field.

What an Engineer in 2029 Needs

What you should put in front of a recruiter in 2029 is not an impressive GitHub contribution graph or a list of polished toy projects. A record of implementing tidy tasks pulled from a ticket queue by a deadline also loses its value.

A new engineer’s market price is set somewhere else. Have you taken someone else’s messy real-world problem, broken it down logically, gotten an AI system running in that person’s unstable environment, and proven with numbers that the system works correctly? Someone with this ability is no longer a worker processing a ticket queue but an engineer who solves field problems to the end. The investment has to start with finding that field now.

References

1. METR, “Time Horizons” (accessed 2026-10-29)
   https://metr.org/time-horizons/

2. METR, “Time Horizon 1.1” (2026-01-29)
   https://metr.org/blog/2026-1-29-time-horizon-1-1/

3. METR, “Clarifying limitations of time horizon” (2026-01-22)
   https://metr.org/notes/2026-01-22-time-horizon-limitations/

4. Anthropic, “When AI builds itself” (2026)
   https://www.anthropic.com/institute/recursive-self-improvement

5. Stripe, “Minions: Stripe’s one-shot, end-to-end coding agents” (2026-02-09)
   https://stripe.dev/blog/minions-stripes-one-shot-end-to-end-coding-agents

6. OpenAI Forum, “How OpenAI engineers use Codex to tackle big projects with rigor” (2025-12-04)
   https://forum.openai.com/home/blogs/how-openai-engineers-use-codex-to-tackle-big-projects-with-rigor-2025-12-04

7. McKinsey, “The state of AI in 2026: On the road to ROI” (2026-08)
   https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai

8. NBER Digest, “Global Evidence on Business Use of AI” (2026-05)
   https://www.nber.org/digest/202605/global-evidence-business-use-ai

9. Eurostat, “Use of artificial intelligence in enterprises” (data extracted 2025-12)
   https://ec.europa.eu/eurostat/statistics-explained/index.php?title=Use_of_artificial_intelligence_in_enterprises

10. OSWorld 2.0 (accessed 2026-10-29)
    https://osworld-v2.xlang.ai/

11. OSWorld-Verified leaderboard (accessed 2026-10-29)
    http://osworld-v1.xlang.ai/

12. OpenAI, “OpenAI launches the OpenAI Deployment Company” (2026-05-11)
    https://openai.com/index/openai-launches-the-deployment-company/

13. Anthropic, “Claude Frontier Academy” (2026-10-02)
    https://www.anthropic.com/news/claude-frontier-academy

14. Epoch AI, “What will AI look like in 2030?” (2025-09-16)
    https://epoch.ai/blog/what-will-ai-look-like-in-2030

15. Engineering Leadership, “Inside OpenAI’s forward deployed engineer…” (2026-08-15)
    https://newsletter.eng-leadership.com/p/inside-openais-forward-deployed-engineer

16. Hamel Husain & Shreya Shankar, “LLM Evals: Everything You Need to Know (FAQ)“
    https://hamel.dev/blog/posts/evals-faq/

17. Charter, “Why IBM is ‘tripling’ entry-level hiring as AI reshapes jobs” (2026-02-11)
    https://www.charterworks.com/ibms-nine-year-ai-journey/

18. a16z, “Trading Margin for Moat: Why the Forward Deployed Engineer Is the Hottest Job in Startups” (2025-06-04)
    https://a16z.com/services-led-growth/

19. Sierra, “Agents as a Service” (2026-03-25)
    https://sierra.ai/blog/agents-as-a-service

20. Indeed Hiring Lab, “The Labor Market Is Tilting Toward Seniority” (2026-07-23)
    https://hiringlab.indeed.com/2026/07/23/the-labor-market-is-tilting-toward-seniority/

21. Etoday, “Saramin–Jumpit release developer hiring market report” (in Korean, 2025-08-27)
    https://www.etoday.co.kr/news/view/2500510

22. Bloomberry, “I analyzed 1,000 forward deployed engineer jobs” (2025-11-18)
    https://bloomberry.com/blog/i-analyzed-1000-forward-deployed-engineer-jobs-what-i-learned/

23. Databricks, “AI Engineer – Forward Deployed Engineering” job posting (2026-10-28)
    https://databricks.com/company/careers/open-positions/job?gh_jid=8546367002

24. Palantir, “Forward Deployed Software Engineer, New Grad – Commercial” job posting
    https://jobs.lever.co/palantir/e500bcf3-19d8-4d3c-b340-4d76e4a55b40

25. Anthropic, “How AI assistance impacts the formation of coding skills” (2026-01-29)
    https://www.anthropic.com/research/AI-assistance-coding-skills

26. Korea Financial Newspaper, “Neither Naver nor Kakao holds open recruitment for new developers” (in Korean, 2026-08-14)
    https://www.fntimes.com/html/view.php?ud=202608141444496947094b7542a8_18

27. Federal Reserve Bank of New York, “The Labor Market for Recent College Graduates” (2026-02)
    https://www.newyorkfed.org/research/college-labor-market

28. Brynjolfsson, Chandar, Chen, “Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence”, Stanford Digital Economy Lab (revised 2026-08-12)
    https://digitaleconomy.stanford.edu/app/uploads/2026/08/Canaries_August2026.pdf

29. McKinsey, “The state of AI in 2025: Agents, innovation, and transformation” (2025-11)
    https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai

30. Samsung SDS, Real Summit 2026 press release on its enterprise AI transformation strategy (in Korean, 2026-09-08)
    https://www.samsungsds.com/kr/news/sds20260908.html

31. Ministry of SMEs and Startups, 2024 Basic Statistics on SMEs (in Korean, 2026-08)
    https://startuprecipe.co.kr/wp-content/uploads/2026/08/260901_korea_000000999293535.pdf

32. Metro Seoul, “Nine in ten SMEs do not use AI” (in Korean, 2024-11-04, Korea Federation of SMEs survey)
    https://www.metroseoul.co.kr/article/20241104500414

33. Eurostat, “Artificial intelligence by size class of enterprise” (isoc_eb_ai, 2021, 2024)
    https://ec.europa.eu/eurostat/databrowser/view/isoc_eb_ai/default/table

34. SQW, “Knowledge Transfer Partnerships Evaluation: Final Report”, Innovate UK (2023-08)
    https://www.ukri.org/wp-content/uploads/2023/10/IUK-23102023-KTP-Evaluation-Final-Report-FINAL-Aug-23.pdf

35. ICTC, “The Impact of Work-Integrated Learning on Student Success and the Canadian Economy” (2023-08)
    https://ictc-ctic.ca/media/1009/download

36. aivancity, “The AI Clinic” (accessed 2026-10-29)
    https://aivancity.ai/en/la-clinique-de-lintelligence-artificielle

37. Digital Daily, “Ministry of Science and ICT to select 10 new AI-centered universities with 25.5 billion won” (in Korean, 2026-01-18)
    https://www.ddaily.co.kr/page/view/2026011616242132410

38. Naver, “Naver runs ‘SME Facilitator’ program with five universities to support small businesses’ digital marketing” (in Korean, 2025-08-29)
    https://www.navercorp.com/media/pressReleasesDetail?seq=33247

39. Canva Engineering, “Yes, You Can Use AI in Our Interviews. In fact, we insist” (2025-06-11)
    https://canva.dev/blog/engineering/yes-you-can-use-ai-in-our-interviews

40. 404 Media, “Meta Is Going to Let Job Candidates Use AI During Coding Tests” (2025-07-29)
    https://www.404media.co/meta-is-going-to-let-job-candidates-use-ai-during-coding-tests/

41. Bloter, “Nexon drops the coding test, hires developers who solve problems with AI” (in Korean, 2026-08-26)
    https://www.bloter.net/news/articleView.html?idxno=671869

42. Samsung Newsroom, “Samsung holds its H2 2026 Global Samsung Aptitude Test (GSAT)” (in Korean, 2026-10-25)
    https://news.samsung.com/kr/삼성-2026년-하반기-삼성직무적성검사gsat-실시

43. Digital Insight, “Grep’s AI competency assessment replaces PT interviews at commercial banks” (in Korean, 2026-10-12)
    https://ditoday.com/?p=154678

44. Lex Fridman Podcast #471, “Transcript for Sundar Pichai: CEO of Google and Alphabet” (2025-06-05)
    https://lexfridman.com/sundar-pichai-transcript

45. Palantir, Lever job board (accessed 2026-10-29)
    https://jobs.lever.co/palantir