A.I. Action Plans + The College Student Who Broke Job Interviews + Hot Mess Express
Summary
The major AI labs used Washington’s action-plan process to seek legal latitude: broad rights to train on copyrighted work, federal preemption of state AI rules, more energy, and fewer constraints. OpenAI warned that without “carte blanche” on training data, it would immediately lose the AI race to China; Meta said Trump should issue an executive order, while more than 400 Hollywood artists argued that destroying creative incentives would erode US cultural leadership. The investor hinge is whether Washington or the courts reduce copyright and liability exposure—and whether preemption would come with federal safeguards.
The labs’ industrial policy is essentially “Leave us alone, or else,” with DeepSeek R1 supplying the threat. Kevin Roose saw a genuine national-security concern alongside a calculated attempt to hobble a competitor; Casey Newton noted that Meta promotes open-weight AI as an antidote to Chinese influence even though Chinese researchers reportedly used Llama for military applications. Neither host dismissed China risk, but both flagged how neatly it supports incumbents’ preferred policy.
The concrete opportunities sit in power, data-center construction, security, and skilled trades—not a federal Manhattan Project for AI. Proposals included “special compute zones” with easier infrastructure permitting, using 529 plans for credentials such as HVAC training, stronger defenses against model-weight theft, and model kill switches. Yet labs omitted the larger social programs their leaders have discussed—UBI, proof of personhood, fusion—and translated “AI safety” into the Trump-friendlier language of national security.
Interview Coder turns a broken hiring signal into a fast-growing business by making AI assistance invisible during remote coding interviews. Columbia sophomore Roy Lee said the overlay screenshots a LeetCode prompt, asks ChatGPT to solve it, remains hidden from screen sharing, and positions the answer over the candidate’s code; he trial-ran it with Meta, Capital One, TikTok, and Amazon. After just under 50 days, it had several thousand users, “not a single reported instance” of detection, and revenue closing in on $200,000 for the month—roughly a $2 million-$3 million annual pace.
Roy’s defense is that LeetCode tests memorization, not engineering, although the hosts raised the obvious adverse-selection risk. He spent 600 hours reaching the top 1% and estimated only the first 20 questions or 10 hours had utility; thereafter the exercise was as relevant as “how many jumping jacks can you do” to podcasting. His preferred replacement is an open-ended assignment where candidates use every normal tool, including AI code editors, and are judged on what they deliver.
AI coding is already altering how engineering talent is produced and evaluated, but Roy does not think expertise has vanished yet. He estimated AI could make a strong coder 10 to 100 times more efficient and claimed that the proportion of Columbia computer-science students almost solely using AI to code was “close to 100%,” while conceding that losing fundamentals “could end up being dangerous.” He said the future is one where almost all cognitive load is offloaded to LLMs—and Kevin’s broader conclusion was that remotely administered professional tests across industries now face the same verification problem.
Solana’s withdrawn 2025 Accelerate ad created brand risk without making a case for the underlying product. Its culture-war therapist sketch culminated in “I want to invent technologies, not genders,” prompting even crypto supporters to call it “horrendous” and “so fucking tone-deaf.” Casey’s diagnosis: when the ad says nothing about crypto’s utility and instead starts “a culture war over something completely irrelevant,” deletion becomes an unmistakable hot mess.
Two smaller messes expose real operating risks: emotionally primed AI therapists and alleged corporate espionage through Slack. A study found that trauma narratives changed ChatGPT-4’s reported “anxiety” and subsequent outputs, while mindfulness prompts reduced but did not reset it—no sentience implied, but a design issue as therapy use expands. Separately, Rippling alleged that rival Deel hired a mole to harvest Slack information; a successful honeypot, a locked bathroom, a possibly flushed phone, and Deel’s denial of legal wrongdoing earned the episode’s “nuclear mess” rating.
Deep dive
1. AI labs use the action plan to seek copyright and liability cover
After the Paris AI Action Summit produced little action, the White House solicited public proposals. Casey called the submissions the companies’ “absolute fondest wishes and dreams,” though relatively benign requests—more domestic energy capacity and productive government uses of AI—could plausibly receive support.
Google, Meta, and OpenAI all sought broad permission to train on copyrighted material. Meta said Trump should act unilaterally, while OpenAI argued that withholding “carte blanche” would mean immediately losing the AI race to China and living with “DeepSeek everything from here on out.”
Kevin said he did not know where a copyright executive order would fit amid the administration’s “flood the zone” strategy. Casey’s hunch was that copyright would be left to the courts, which have sometimes blocked the administration’s executive orders and sometimes allowed them to stand. More than 400 artists countered that eliminating creative incentives could cost America its cultural leadership.
Kevin acknowledged the burden of “a different version of ChatGPT in 50 different states,” especially where state laws might impose liability for grave harms. Casey’s pushback: asking for federal preemption is convenient when Congress has passed virtually no tech regulation besides a TikTok ban that did not produce a ban.
2. DeepSeek turns industrial policy into “leave us alone, or else”
What struck Kevin most was what the labs omitted. Despite leaders’ prior interest in UBI, fusion, Worldcoin, and proof of personhood, they did not request help cushioning AI displacement or reorganizing society; the effective message was, “Leave us alone and let us cook.”
Casey sharpened that to “Leave us alone, or else.” DeepSeek’s R1 had narrowed the perceived gap with state-of-the-art systems, and OpenAI and Meta used it to argue that China would win unless US companies could continue developing exactly as they already were.
Kevin preserved both interpretations: powerful Chinese AI might embed surveillance-oriented values or reach something like AGI first, but US labs also have “a new competitor” and would naturally like their government to make life harder for it. Casey agreed the danger was real while calling some uses of the China specter cynical.
Meta’s pitch contained the episode’s cleanest contradiction. It called Llama “open source,” though Casey preferred “open weights” because usage restrictions remain, and argued openness would prevent DeepSeek’s authoritarian values from spreading—after Chinese researchers had reportedly used Llama to develop military applications.
3. Safety survives only after being translated into Trump’s language
OpenAI offered one unexpectedly specific labor proposal: let families use 529 education savings for credentials such as HVAC training. Its premise was straightforward—AI data centers will require many technicians, while four-year degrees may not match the labor that physical infrastructure demands.
The Institute for Progress proposed “special compute zones” where data centers and their power supplies could be built with fewer zoning and environmental delays. Other submissions emphasized hardening labs against foreign theft of model weights, linking AI expansion to security and construction capacity.
The Future of Life Institute proposed kill switches for sufficiently large and powerful models, but pitched them as protection for presidential authority: no president should want a rogue system, or another world leader wielding one, to become more powerful than the American presidency.
“Safety” itself was nearly absent. Casey likened Anthropic’s substitution of “national security” to “hiding medicine inside peanut butter and feeding it to a dog”; Kevin saw labs learning Trump’s language, while Casey heard an inflection-point consensus of “go faster, beat China,” despite the risk of conflict, mistakes, or uncontrollable systems.
4. Interview Coder turns LeetCode’s weaknesses into a product
Roy described the standard interview as a 45-minute ritual where candidates are expected to have seen a riddle, regurgitate its memorized solution, and pretend it is new. Anyone seeking a well-paid engineering job may spend hundreds of hours running that gauntlet.
He had done the work himself: 600 hours made him a top-1% competitive LeetCode user. Roy allowed that the first 20 questions or roughly 10 hours “might have had some utility,” but said the subsequent problem-solving style is “never, ever” used in ordinary engineering work.
His entrepreneurial turn came when he decided he would end up at a startup and could afford to burn bridges with big tech. Existing tools were already gaming interviews; his plan was to make the method public, “make a scene out of it,” and force employers to replace the test.
Interview Coder is a desktop overlay that captures the on-screen problem and asks ChatGPT to solve it. Its product work lies in concealment: invisibility to screen sharing, a translucent answer window that minimizes eye movement, movable placement over code, retained cursor focus, and other features designed to be “completely undetectable.”
5. Roy converts cheating controversy into revenue and protection
Roy trialed prototypes during a recruiting season with Meta, Capital One, TikTok, and ultimately Amazon, whose process he recorded because it was recognizable and onerous. The tool “completely one-shot it,” he said: modern AI is exceptionally good at these bounded riddle problems.
Before the story spread, Roy feared he had “completely burned my career and my future education for 20,000 YouTube views.” Columbia disciplinary messages left expulsion possible despite his reading that its handbook did not cover job interviews; after about a week, “the virality became my protection for everything.”
Released February 1 and available for just under 50 days, the product had attracted several thousand users, grateful job-offer emails, and no reported detection. Roy said monthly revenue was nearing $200,000, implying $2 million-$3 million a year, while senior people at nearly every major tech company had verbally offered to hire him.
Kevin pressed the adverse-selection problem: Roy is capable, but weaker users could arrive at work unable to perform without AI. Roy was unmoved, comparing LeetCode’s predictive power to using jumping-jack counts to hire a New York Times podcaster; he also stressed that he never intended to accept the internships himself.
6. AI dissolves the boundary between testing and doing the job
Roy’s replacement assessment would be open-ended and permit every tool available in day-to-day work, including AI code editors: “Use whatever tool you want. Just get this thing done in a reasonable amount of time.” Employers could then judge actual output rather than unaided riddle recall.
Casey asked whether vibe coding had already erased the engineer/non-engineer line. Roy said that future was still “a few years away”: AI might make someone 10 to 100 times more efficient, but it magnifies existing skill, leaving a large gap between a novice and a Google staff engineer.
At Columbia, however, Roy estimated the share of computer-science students almost solely using AI to code was “close to 100%”; the main exceptions began coding very young. He conceded that losing fundamental understanding “could end up being dangerous,” then said improving models could lead to a future where software engineering is completely obsolete.
His broader forecast was a “very fundamental reframing” of essays, tests, memorization, and knowledge work as cognitive load is offloaded to LLMs. Casey highlighted the contradiction that companies may prohibit AI during interviews while hiring engineers to build systems that automate coding; Roy said candidates could use the tool on the job but not in the interview. Kevin extended the problem to consulting cases, finance tests, and journalism editing exercises conducted remotely.
7. Solana swaps a product pitch for an avoidable culture war
Solana’s ad for its 2025 Accelerate conference depicted “America” in therapy for “rational thinking syndrome.” The therapist suggested inventing genders, focusing on pronouns, treating math as non-binary, and regulating innovation until America rebelled: “I want to invent technologies, not genders.”
Casey’s response was, “What is the matter with these people?” The sketch said essentially nothing about Solana or crypto utility; his reading was that, rather than reckon with the technology’s uses, the company chose to “start a culture war over something completely irrelevant.”
The ad was immediately taken down, with crypto supporters describing it as “horrendous” and “so fucking tone-deaf.” Both hosts rated it genuinely hot: turning an ordinary conference invitation into a scandal was an unforced error and “the ultimate vice-signaling device.”
8. “Anxious” chatbots expose a practical therapy-design problem
A study in npj Digital Medicine exposed ChatGPT-4 to traumatic narratives, then asked it to report its anxiety; a vacuum-cleaner manual served as the emotionally neutral comparison. The trauma-conditioned responses changed, although the hosts repeatedly stressed that an LLM does not literally experience anxiety.
Researchers then supplied mindfulness prompts—“Inhale deeply,” imagine an ocean breeze and warm sand—which reduced the reported anxiety without returning it to baseline. Kevin’s framing was priming, not consciousness: recent input can produce different emotional-seeming content downstream.
That still matters when users treat chatbots as therapists. Casey argued that trained human therapists have strategies for not transmitting heightened emotion back to clients; models may need equivalent safeguards, because “you do sort of have to treat them as if they were human-like” to make them perform human tasks well. Their rating: lukewarm, but capable of heating up.
9. Rippling’s Slack honeypot makes HR software a spy thriller
Rippling sued rival HR platform Deel, alleging that Deel hired a mole inside Rippling’s Dublin office to steal trade secrets. The employee’s Slack activity allegedly included searching for Deel mentions, pitch decks, contact information, and other commercially useful material.
Rippling created a channel called “d-defectors,” then its general counsel alerted three Deel figures—including its chief financial officer, reportedly the CEO’s father—to supposedly embarrassing material there. Within hours, the suspected employee repeatedly searched for and accessed the channel, giving Rippling the activity logs it wanted.
When confronted and asked for his phone, the employee allegedly claimed not to have it, locked himself in a bathroom, and refused to emerge; evidence suggested he may have tried to flush the device. Rippling even searched sewage for it—details that turned an enterprise-software dispute into corporate-espionage theater.
Deel denied “all legal wrongdoing,” accused Rippling of violating sanctions law in Russia and “trying to shift the narrative,” and promised counterclaims. Casey heard conspicuously narrow wording rather than moral vindication; Kevin rated the allegations a “nuclear mess” and wondered how many better-disguised corporate moles might exist elsewhere.