Summary

Paris shifted the international AI agenda from safety toward opportunity, signaling to Kevin Roose that national competition is displacing hoped-for cooperation. The third summit pivoted from safety to opportunity: Emmanuel Macron promoted French investment and Mistral, while JD Vance declared that “the AI future is not going to be won by hand-wringing about safety. It will be won by building.” Even a watered-down declaration went unsigned by the United States and United Kingdom, leaving Roose expecting national competition, “all gas, no brakes” U.S. policy, and little meaningful federal regulation.
The American acceleration agenda has no clear account of what happens if the labs achieve their stated goal. Industry leaders have suggested AGI or something like it could arrive within one, two, or possibly three years, yet political responses still resemble multi-year commissions and worker-adjustment reports. Casey Newton called that “techno-pessimism that’s masquerading as optimism,” while Roose’s column labeled the U.S. safety plan “Let’s See What Happens.”
Labor displacement may turn abstract AI-risk arguments into an immediate political and economic problem. Information-technology unemployment reportedly rose from 3.9% in December to 5.7% in January, versus 4% overall, as founders told Roose that one senior engineer with AI can do work that might previously have required three or four people. Anthropic’s economic index classified 57% of observed use as augmentation and 43% as automation—already difficult to reconcile with Vance’s categorical claim that AI will only help, not replace, workers.
Safety investment could be a prerequisite for sustained AI acceleration rather than its ideological opposite. An attendee compared AI with nuclear power: Chernobyl and Three Mile Island delayed adoption for decades because early systems were not safe enough. On that logic, one catastrophic AI failure could provoke a backlash far more restrictive than preventive safeguards—the fastest route may require avoiding the accident that stops the industry.
Hinge’s first AI coach improved measured prompt quality while exposing how crude optimization can flatten human appeal. Justin McLeod said prompt feedback tripled high-quality answers and cut low-quality, one-word responses by more than a third; Rachel Cohn also reported more replies after adopting its specificity-heavy advice. Yet the model rejected her already successful “someone to read a book next to in bed” line until it became what the hosts called “dating Goodreads” and “giving LinkedIn,” showing that better engagement is not the same as better chemistry.
McLeod expects AI to transform dating more than mobile did, with deeper matching and coaching as its two main opportunities. Hinge’s roadmap moves beyond collaborative filtering toward interpreting photos, prompts, personal history, values, goals, and relationship science, then coaching users who struggle to represent themselves. The thesis is that attractiveness only gets a “foot in the door”; richer signals might reduce the costly discovery, three months later, that two people are fundamentally misaligned.
The company’s product boundary is AI “behind us and not between us.” Hinge will suggest profile improvements and conversation topics, but McLeod rejected agents dating on users’ behalf; its non-AI Your Turn Limits already blocks further matching after roughly eight unanswered conversations and lifted responsiveness about 20%–30%. His deeper conclusion also limits what matchmaking technology can promise: people should stop shopping for a flawless “one” and learn to “create the one” through commitment.

Deep dive

1. Paris shifted the summit from safety toward opportunity

The first two international gatherings—in the United Kingdom and South Korea—were explicitly focused on AI safety; the Bletchley Park edition was called the AI Safety Summit. Paris deliberately renamed the third meeting the AI Action Summit and foregrounded opportunity rather than biological, chemical, nuclear, cyber, propaganda, or existential risks.
The Grand Palais brought together Sam Altman, Anthropic’s Dario Amodei, Google DeepMind’s Demis Hassabis, government officials, academics, and NGOs. Its public exhibits emphasized preserving languages that may die out, improving benefits administration, and applying AI to climate change and disaster relief: a showcase of “all the ways that AI could help people.”
Macron used the summit to argue that France should not let America and China dictate the technology’s future. He announced major investment in the domestic ecosystem, promoted Mistral, and positioned France against the more regulation-oriented image of the rest of Europe: the goal was to “accelerate” into AI leadership.

2. Washington chose acceleration without explaining the destination

Vance’s America-first message was categorical: “I’m not here this morning to talk about AI safety.” Calling the usual response to technological breakthroughs too self-conscious and risk-averse, he concluded, “The AI future is not going to be won by hand-wringing about safety. It will be won by building.”
Startup founders and accelerationists welcomed that shift, believing “doomers” had exaggerated AI risks. Safety-minded attendees viewed it as a missed opportunity, while Roose summarized Amodei’s response as a warning that advanced AI is arriving quickly enough that governments cannot simply ignore its hazards.
Casey Newton traced the political turn primarily to Donald Trump’s election. Where Kamala Harris and the Biden administration showed sympathy toward safety concerns and the Biden White House produced an AI executive order, Trump’s circle includes figures such as Marc Andreessen, an AI accelerationist, alongside others who treat the safety movement as irrational doom-mongering.
Roose asked what Vance believes happens after an American lab creates superintelligence. Newton argued that accelerationists seem either not to believe the labs’ AGI forecasts or lack a post-AGI vision: “It’s a kind of techno-pessimism that’s masquerading as optimism.”

3. Labor data may force the safety debate into the present

Newton described two incompatible clocks. Lab leaders and technical insiders discuss AGI or something similar arriving in one, two, or perhaps three years; conventional policymakers propose commissions that may publish worker-transition reports “in a couple years.” The people closest to the systems are asking for urgency while government treats AI as another ordinary technology cycle.
Newton cited information-technology unemployment rising from 3.9% in December to 5.7% in January, well above the 4% overall jobless rate. Some economists interpreted it as an initial sign that AI automation is producing meaningful job losses, though the episode did not claim the causal case was settled.
Software may be the first visible pressure point. Founders told Roose they no longer need as many junior engineers because one senior engineer equipped with AI can perform work that previously required three or four people underneath them.
Anthropic’s economic index found its observed AI usage leaned toward augmentation, but narrowly: 57% augmented human work and 43% automated it. That evidence sits uneasily beside Vance’s assertion that AI will only help workers and Sam Altman’s description of “drop-in workers” companies could hire; Newton warned that mass unemployment is profoundly politically destabilizing.

4. Safety may be the prerequisite for durable acceleration

An AI-safety attendee offered nuclear power as the useful analogy. Chernobyl and Three Mile Island delayed widespread nuclear-energy adoption for decades, not because society had initially taken risk too seriously, but because first-wave plants “didn’t make them safe enough.”
Applied to AI, the argument is that unmanaged acceleration may create the catastrophe that triggers a much harsher public response. Taking extreme risks seriously could therefore let the technology move faster over the long run; Newton wished today’s committed accelerationists would recognize that mechanism.
The summit’s watered-down declaration still proved too restrictive for the United States and United Kingdom, possibly because of language calling for inclusive AI, although China signed. Elon Musk separately inserted himself by proposing roughly $97.4 billion for OpenAI’s nonprofit while Altman sat beside Vance; the hosts viewed it as legal trolling that could complicate assigning a fair value during OpenAI’s for-profit conversion.
Roose’s final Paris conclusion was bleak for multilateral governance: the hoped-for United Nations-style safety structure is unlikely to emerge because countries want to win, not cooperate. He expects no meaningful federal AI-safety regulation under Trump, possible state action, and an “action summit” whose operative decision was “we’re not going to take any action.”

5. Dating apps are selling AI into a crisis of user confidence

Newton described a deteriorating consumer bargain: users feel dating apps extract more money while producing fewer dates and no lasting partner. It is unclear whether matching quality has declined or dating itself has become less popular, but app CEOs are offering the technology industry’s now-standard remedy—“sprinkle some AI on it.”
Cohn chose Hinge because, as a 30-year-old straight dater in New York, everyone she knows uses it; Tinder remains larger, but Hinge is among the fastest-growing apps. Its deliberate identity as a place for serious relationships also gives it more reputational downside if synthetic features undermine “meaningful human connection.”
At Match Group’s December investor day, Hinge described two AI buckets: improved matching, expected later, and assistance for the “struggling dater.” Cohn fit only half that label after two and a half years of active dating: her top-of-funnel supply of dates was strong, but “we gotta convert.”

6. Hinge’s prompt coach mistakes specificity for chemistry

AI Prompt Feedback grades each required written response as “Great answer,” “Try a small change,” or “Go deeper.” Cohn interpreted the last category as a demand for vulnerability, but the model’s most consistent preference was specificity; as Newton observed, it never seemed to recommend, “Go shallower.”
The system correctly approved Cohn’s “simple pleasures”: recording funny subway conversations, dancing a little while running, and making “incredibly average watercolor paintings.” The hosts agreed it was specific, self-deprecating, distinctive, and naturally invited conversation about humor or art.
It was less perceptive about her claimed strength, “Getting people to share stuff they normally wouldn’t or shouldn’t.” To satisfy its requests for detail, she added a wedding stranger revealing his salary and why he felt constrained in his relationship—an anecdote the AI accepted, even though the hosts thought it could make a first date sound like an interrogation.
Newton’s human edit was to say people “accidentally” reveal things to her, presenting Cohn as so charismatic that secrets emerge rather than as someone prying. Cohn conceded that framing was more appealing but “not totally the truth about me,” crystallizing the tension between attractive positioning and authentic filtering.

7. A worse-sounding profile still generated more engagement

The model’s largest apparent mistake was rejecting Cohn’s most successful line: “I’m looking for someone to read a book next to in bed.” Its strategic ambiguity let potential partners project themselves into the scene; the hosts found it sweet, intimate, inviting, and conveniently offered men an opening to discuss what they were reading.
Hinge kept demanding genres, titles, and recommendation prompts until the answer became: “Someone to read a historical fiction book next to in bed. Open to recs! Especially love stories set in New York City.” The verdicts were brutal—“All of a sudden I’m dating Goodreads,” “It’s giving LinkedIn,” and “Delete.”
Yet after a little over a week, the revised line had generated more engagement. Men recommended specific historical-fiction or love-story books, and one withheld the Pulitzer Prize-winning title until she matched; Newton concluded the model had correctly discovered that inviting a man to recommend something reliably elicits an opinion.
Cohn compared the result with college-application coaching: applicants are pushed into a vivid, coherent story that is “a little bit reductive and not totally authentic,” because screeners can react to it more easily. The narrower AI-shaped portrait may similarly improve selection even when it feels less naturally charming.

8. Hinge expects AI to remake matching more than mobile did

Cohn stressed that the coach never writes a replacement answer; it offers middle-school-English advice such as “show, don’t tell.” That protects authorship but limits accessibility: satisfying one prompt took her 30 minutes, raising doubts that low-effort or vulnerability-averse “struggling daters” will respond constructively to “share a little bit more.”
McLeod nevertheless expects AI to create “a bigger transition than even what happened with the transition to mobile.” His two pillars are personalized matching—understanding people several levels deeper before pairing them—and effective coaching, from profile tips to help with the emotional strain of dating.
Hinge’s 2015-era algorithm largely used collaborative filtering: whom a user liked, and whom people with similar tastes also liked. The planned content-based layer would interpret photos and prompts, then solicit richer accounts of backstory, relationship history, values, goals, and expectations that cannot fit a religion-or-height dropdown.
McLeod also wants to incorporate relationship research about which personalities work together and what leads to durable compatibility. Roose’s cynical test was whether “90%” is simply finding someone cute; McLeod said looks provide the easiest snap judgment but called them only the “foot in the door” before conversation, attendance, and second-date compatibility matter.

9. Hinge wants AI to coach behavior, not impersonate users

McLeod said prompt feedback tripled the incidence of high-quality answers and reduced low-quality, one-word answers by more than a third. Hinge optimizes an “efficiency frontier of vulnerability”: karaoke songs are easy to disclose but never lead to a date, while changing one’s relationship with one’s mother could reveal values but is too intimate for most public profiles.
The hosts worried universal coaching could raise the floor until generic people appeared artificially original—the personality equivalent of airbrushing. McLeod’s distinction was between coaching and authorship: Hinge asks users to reveal more in their own voice rather than handing them polished copy, because many genuinely dynamic people simply produce profiles that “completely miss” who they are.
McLeod said the coaching roadmap also includes a photo finder that learns which kinds of photos perform well on Hinge. Hinge uses AI defensively as well, comparing IP addresses, previous photos, in-app behavior, and behavioral patterns to identify romance scammers; he said its dedicated trust-and-safety operation catches such accounts quickly enough that crypto-style confidence scams are not a major Hinge problem.
For conversations, the company planned to test starters the following quarter: not messages to copy, but prompts such as “did you notice this in their profile?” McLeod’s governing principle is that AI should “stand behind us and not between us”—nudging people toward an actual date without conducting their relationships.

10. Lasting love is created after the match, not discovered by an agent

Hinge attacked ghosting with rules rather than generative AI. Your Turn Limits stop users from sending new likes or collecting matches after roughly eight conversations await their response; McLeod said responsiveness rose about 20%–30%, and even a one-star reviewer later reversed course after experiencing the feature.
McLeod rejected AI concierges that date other agents on users’ behalf, unlike concepts discussed by Bumble or Grindr’s planned 2027 wingman. Newton called the idea science fiction told to shareholders; Roose countered that people already use ChatGPT Operator on dating profiles. McLeod’s narrower objection was outcome-based: current agents cannot reliably reproduce a person’s voice, values, or personality well enough to choose someone they will actually like.
His own romantic history informs that restraint. After an eight-year separation from his college sweetheart, he flew to Switzerland roughly a month before she was to marry someone else and asked her to cancel the wedding. She returned to New York; after about two amazing months, he wondered, “Oh, my God, have I made a mistake?”—yet they stuck with it and, ten years later, had what he described as a deep and beautiful relationship.
McLeod’s largest change since founding Hinge in 2011 is rejecting the idea that people merely “find the one”: they “create the one.” Newton added that no algorithm may determine whether someone truly understands or wants a relationship; his own dating experience changed only when he met a ready partner, after deciding to go on “one more online date.”

JD Vance's AI Speech, Techno-Optimists vs Doomers, Tariffs, AI Court Cases with Naval Ravikant How Harvey AI is Changing the Legal Industry with Winston Weinberg