An A.I. Assistant That’s Too Good to Be True?
Summary
Tech CEOs are now de facto political actors, but their calibrated response to the Minneapolis killings exposed how dependent industry leaders remain on Washington. Sam Altman said ICE was “going too far,” Dario Amodei called the events in Minnesota “horror,” and Tim Cook urged “de-escalation”; Casey Newton judged some statements “about the least that they possibly could,” while Kevin Roose noted that Chris Olah’s stronger post was reposted by Katie Miller as evidence of bias in AI companies’ models.
X and government media operations have fused policy, enforcement, and engagement into one feedback loop. A viral video alleging fraud in daycare centers helped trigger the Minneapolis operation, administration officials used X to justify it, and ICE deployed producers, editors, paid tools, and friendly influencers to package raids as content. Newton’s blunt assessment: X has “basically” become state media, while “winning on social media has become almost the entire point.”
AI-manipulated evidence creates value for political actors even when the alteration is obvious, because the payoff is the “liar’s dividend.” Images falsely portrayed Alex Pretti pointing a gun instead of holding a phone and civil-rights attorney Nakeema Levy Armstrong crying after her arrest; Vice President J.D. Vance reposted the latter despite an original image already being public. The chilling official response — “the memes will continue” — captures why durable labeling and watermarking rules may require legislation rather than voluntary platform courage.
Smartphones have become opposing weapons of state narrative and citizen accountability, yet redundant footage can still preserve shared reality. DHS characterized filming agents as doxing while protesters documented encounters and the administration brought its own influencers and cameras into operations — “phone-to-phone combat.” Roose fears video is no longer self-authenticating; Newton counters that multiple angles, journalistic verification, and even NRA criticism of the administration’s account show that “if enough people see it,” truth might survive.
Moltbot offers a compelling preview of an agent-native computer while remaining insecure, unreliable, and only marginally useful today. The local open-source assistant can connect to email, calendars, messaging, and other services, preserve memory in Markdown, and tailor a daily briefing — but Newton’s elaborate system worked only about 70% of the time and failed during the demonstration. The relevant signal is the direction, not the product: a trusted operating-system agent could displace many app interactions, but today’s version is often “a very complicated new widget that’s telling me the weather.”
The widening AI-adoption gap is becoming an experiment in whether frontier access produces genuine productivity alpha or merely risk and theater. San Francisco users are installing multi-agent swarms while conventional organizations still seek permission to use Copilot in Teams; institutional controls slow diffusion and can reduce exposure to prompt-injection attacks and compromised accounts. Andrej Karpathy called Claude Code and Codex “easily the biggest change” to his workflow in two decades, strengthening Roose’s concern that non-users could become less competitive if similar gains spread beyond coding.
The rapid-fire corporate stories showed AI pressure spilling into operations, hardware road maps, labor, and capital-market narratives. Amazon announced 16,000 layoffs after a calendar invitation about them was apparently sent in error under the name “Project Dawn”; Apple was reportedly developing an AirTag-sized AI pin for as early as 2027; Steak ’n Shake added $5 million in Bitcoin exposure; and SpaceX was weighing a mid-June IPO timed to Elon Musk’s birthday and a Jupiter–Venus conjunction. Meanwhile, LinkedIn’s proposed vibe-coding badges illustrated the risk that credentialing races ahead of demonstrated skill.
Deep dive
1. Tech chiefs are discovering that silence is also a political position
Roose and Newton begin with the fatal shooting of Alex Pretti, the second fatal shooting of a U.S. citizen by federal agents that month. Tech belongs in the story because its companies supply surveillance infrastructure, social platforms, and AI tools now entangled with state power.
Altman told OpenAI employees that ICE was “going too far” and that resisting overreach is part of loving the country. Amodei called the events in Minnesota “horror,” while Cook told Apple staff, “This is a time for de-escalation.”
Newton found the Altman and Cook messages weak and conspicuously designed not to irritate the White House. Roose dislikes obligatory corporate speech, but Newton’s counterpoint is structural: these leaders are “in their own way heads of state,” representing enormous user and employee constituencies.
The political cost is real. After Anthropic co-founder Chris Olah posted a heartfelt denunciation, Katie Miller argued that such views would become bias inside AI models; Newton therefore saw “some backbone” even in statements that said very little.
2. X has become both the administration’s newsroom and policy incubator
Newton distinguishes Minneapolis from the online reaction to Charlie Kirk’s assassination: both reward conflict, but only Trump has “the power of the state,” thousands of federal agents, and an administration skilled at converting spectacle into viral content.
The Minneapolis operation itself followed a video alleging fraud in daycare centers that went viral on X. For Newton, that means a social platform is not merely covering policy but setting the president’s agenda; Roose’s earlier prediction that X would become state media has “basically” arrived.
ICE reportedly maintains an internal content operation using paid social tools, producers, editors, and video makers. Officials have also brought right-wing influencers to operations, including Kristi Noem’s earlier Portland visit, blending law enforcement with the techniques of a brand campaign.
Roose observes that X turned a broadly comprehensible objection to agents shooting protesters into something resembling a 50/50 dispute, circulating conspiracies, manipulated footage, and deepfakes. Newton’s sharper conclusion: policy objectives can look secondary because “winning on social media has become almost the entire point.”
3. Synthetic evidence pays a liar’s dividend even when it fails to convince
One altered image made Pretti appear to point a gun at an agent when he was holding a phone. Another supposedly “enhanced” a blurry frame but, as Newton put it, “just made a bunch of stuff up,” introducing invented details rather than recovering evidence.
More consequentially, the White House published an altered arrest photo making civil-rights attorney Nakeema Levy Armstrong appear to be crying. DHS had already released the unaltered image, yet J.D. Vance reposted the fabrication anyway.
Newton invokes “nothing is true and everything is possible”: flood the environment with fiction until verifying reality becomes prohibitively expensive. The “liar’s dividend” then applies to authentic evidence too, because audiences know fabrication is possible and can dismiss anything as potentially fake.
Platform labels could help, but X now relies largely on Community Notes, which Newton considers useful yet unpredictable and often untimely. Roose’s case for congressional regulation is institutional consistency: labeling or watermarking rules should persist across administrations rather than leaving companies exposed to retaliation whenever they enforce them. He also warned that platform policy can help only so much if an administration is willing to lie and fabricate evidence.
4. Phone-to-phone combat is testing whether video can still anchor reality
Pretti held a phone when he was shot, while Renee Good’s partner filmed her fatal encounter. Their videos spread widely and, in the hosts’ account, shifted public opinion; Minnesota Governor Tim Walz urged residents to carry phones to document “atrocities against Minnesotans.”
DHS officials have described videotaping agents and publishing their images as doxing, with Tricia McLaughlin threatening prosecution for illegal harassment. The hosts’ pushback is categorical: filming public law enforcement is legal, and identifying agents acting for the state is not inherently doxing.
Masks raise the stakes. Newton wonders whether agents fear legally carried guns or the phones capable of connecting their conduct to their identities; meanwhile, the administration fields its own phones, influencers, talking points, photos, and clips to sell raids as effective enforcement.
Roose fears the artifacts themselves are becoming untrustworthy and that no canonical source of truth remains. Newton is more hopeful: the Pretti shooting had multiple angles, journalists verified them, few people claimed they were fake, and even the NRA criticized the administration’s portrayal. Roose still wonders how long that trust will hold as AI tools improve.
5. Moltbot packages the personal-agent dream inside an unsafe prototype
Moltbot, formerly Clawdbot, is Peter Steinberger’s free, open-source personal AI agent. It began as an attempt to use Claude Code through WhatsApp and promises a “little genie inside your computer” that can connect services and perform work through conversation. It is not affiliated with Anthropic’s Claude chatbot.
Installation resembles Claude Code: paste one line into a terminal and follow the setup. San Francisco tech enthusiasts bought Mac minis to give the agent a dedicated machine — partly enthusiasm, partly recognition that unrestricted installation on a primary computer is “insane behavior.”
The threat model is severe. A compromised Telegram account could theoretically expose the whole computer, while prompt injection could hide malicious instructions on a webpage; Newton disabled Telegram and limited integrations, but still connected email and calendar in what he admitted was an “undeniably risky” experiment.
Moltbot’s practical differentiators are local operation and persistent Markdown memories. Unlike Claude Code’s context-window compaction, it can revisit yesterday’s projects, find the relevant files, and make later changes without making every session feel like starting over.
6. The customized briefing works about 70%; reliability collapses on cue
Newton asked Moltbot to turn “good morning” into a briefing containing weather, important email, calendar events, overdue tasks, wrestling pay-per-views, new RuPaul’s Drag Race episodes, and Thursday previews of the week’s films. After several days, it worked about 70% of the time.
During the live demonstration, it took a long time and then failed. Newton encounters similar breakage daily, a contrast with his experience that Claude Code now generally completes a well-specified task; his explanation of Moltbot’s internals was pure vibe coding: “I have no idea, and I never will.”
His expectations moved “very far downward.” Security remains unresolved, and hours spent wiring services together can create the sensation of productivity without meaningful output — ending in that elaborate weather widget rather than a full-time digital employee.
Still, the prototype makes the future legible. One unverified example showed Moltbot using an ElevenLabs voice to telephone a restaurant after OpenTable failed; the hosts’ reason to watch is that technology can be “directionally correct, way too early,” barely working in January and becoming “pretty freaking good by November.”
7. AI adoption is splitting into experimentation, institutional delay, and resistance
Roose described a “yawning inside-outside gap”: San Francisco users put multi-agent Claude swarms in charge of their lives, while organizations elsewhere still seek approval for Copilot in Teams or debate whether AI may take meeting notes.
Newton sees the predicted bottleneck to AI diffusion. Real institutions have IT policies; startups full of “wireheads” install insecure GitHub projects immediately. The unresolved question is whether that freedom produces alpha or merely wasted effort, security incidents, compromised repositories, and fraudulent crypto promotions. Institutional restrictions may also limit exposure to those risks.
Roose worries that skepticism is hardening into an identity: either AI is transformative or it is fake and overhyped. Moltbot might be productivity theater, but coding already suggests a world where manual practitioners become less competitive than colleagues using capable agents.
Karpathy’s evidence carries weight: Claude Code and Codex were “easily the biggest change” to his basic coding workflow in two decades, arriving within weeks. Newton’s caveat is equally important: mainstream tools will sell themselves only when they help people earn, find, or keep work—not when improvement feels synonymous with replacing the work they are paid to do.
8. Operational failures now trigger immediate trust and governance questions
Amazon announced 16,000 layoffs; the day before, it appeared to mistakenly send employees a 5:00 a.m. invitation for “Project Dawn.” The hosts’ verdict was less about strategy than execution: avoid ominous science-fiction code names—and investigate which AI tool might have sent the calendar invitation.
TikTok’s transfer to a U.S. subsidiary with a group of majority-American investors coincided with a data-center outage, zero-view posts, and claims that ICE criticism or the word “Epstein” was being suppressed. Newton favored malfunction over censorship: during a move, “some of the dishware gets shattered.”
Amodei’s 38-page, 19,000-word “The Adolescence of Technology” catalogued AI dangers as a counterweight to “Machines of Loving Grace.” Roose thinks it drew less attention because risk warnings fit Amodei’s established persona; the earlier optimism was the surprising departure.
Two smaller stories captured AI’s widening trust deficit: a porn-quitting app exposed ages, masturbation frequency, and emotional responses, while an Alaska student was charged after eating roughly 57 AI-assisted images from another student’s exhibition as protest art.
9. Capital, hardware, and credentials are absorbing AI-era theater
Caroline Ellison left federal custody after serving about 14 months of a two-year sentence as the FTX saga was set to become a Netflix series, “The Altruists,” in the fall. Separately, Amazon’s “Melania” documentary received a private White House screening for roughly 70 guests, including Tim Cook, Eric Yuan, Lisa Su, and Mike Tyson.
Steak ’n Shake added $5 million in notional Bitcoin exposure, sending Bitcoin sales into its “strategic Bitcoin reserve.” The chain’s “burger-to-Bitcoin transformation” struck the hosts as another eccentric attempt to turn an operating business into a crypto narrative.
Apple was reportedly developing an AirTag-sized AI pin with cameras, microphones, a speaker, and wireless charging for possible release as early as 2027, potentially competing with OpenAI hardware. Roose called it vindication for Humane; Newton joked that it followed Apple’s Vision Pro formula of expensive hardware with an uncertain purpose.
SpaceX was weighing a mid-June IPO aligned with Musk’s birthday and a Jupiter–Venus conjunction, the first in more than three years. LinkedIn supplied the matching labor-market absurdity: proficiency levels assessed by Replit, Lovable, Descript, and Relay.app for the supposedly credential-worthy skill of “typing in a box.”