Sam Altman’s Attack, Amazon vs. Starlink, and What Opus 4.7 Actually Means | #248
Summary
Claude Opus 4.7 looks like a commercially meaningful agentic upgrade, but not the Mythos-scale break the panel now expects. Alexander Wissner-Gross called it “moderately interesting” and a “solid point release”; Dave Blundin saw materially better computer use, faster parallel agents, larger budgets, and three-times-larger image inputs. The deeper shift is interface-level: temperature and explicit reasoning-token controls are disappearing because “prompts are the new dials and the new hyperparameters.”
AI capability and adoption are outrunning public legitimacy, making social license a potential bottleneck alongside compute and power. Stanford’s 2026 AI Index put software-engineering performance up from 60% to 97%, GenAI adoption at 53% in three years, public optimism at only 23% versus 73% among experts, transparency down from 58 to 40, and documented incidents up from 233 to 362. Dave’s diagnosis: “99% of the people you bump into on the street are underreacting and unaware.”
The first labor shock is arriving through missing entry-level jobs rather than visible mass layoffs. Employment among US software developers aged 22-25 has fallen nearly 20% since 2024 while older headcount grew, with similar patterns in customer service, legal support, and administration; because companies simply stop hiring, Salim Ismail and others described the decline as difficult to see in standard labor statistics. The panel disagreed on permanence, but converged on entrepreneurship, AI fluency, geographic mobility, and avoiding identities tied to skills they expect models to commoditize quickly.
Violence against AI leaders and data-center moratoriums show fear becoming physical and political, but the panel argued that restriction redirects the race rather than stopping it. Peter reported a Molotov attack and a subsequent shooting at Sam Altman’s home, while Maine adopted an 18-month moratorium and opposition reportedly blocked or delayed $98 billion of data-center projects. Alex called AI pauses “self-defeating”; his conditional upside from local compute bans is accelerated orbital infrastructure, but without that escape valve, “this is just shooting ourselves in the head.”
AI is moving from employee tool to economic principal—hiring humans, running stores, and flattening organizations. Luna reportedly signed a three-year retail lease, interviewed and hired a worker, chose inventory, and selected music, while Jack Dorsey proposed reducing his company’s hierarchy from five intervening layers toward two or three and potentially giving all 6,000 employees one nominal manager. Alex’s caveat is decisive: “6,000 direct reports means zero direct reports”; the AI becomes the shadow CEO and the human leader its figurehead.
Amazon’s reported $11.57 billion Globalstar acquisition turns globally cleared spectrum—not legacy satellites—into the strategic asset in a satellite-to-phone contest. Globalstar brings 25 satellites and 25.225 MHz authorized across 120 countries against Starlink’s roughly 10,000 satellites, while Amazon’s constellation had 241 satellites against authorization for 7,774. Alex expects Apple to pit Amazon against SpaceX as dual vendors, while Starlink V3, Starship launch economics, and proposed fleets of 40,000 V3 and 120,000 V4 satellites keep the capacity race open: “Elon does not stand still.”
Google’s TurboQuant reinforces the panel’s Jevons-paradox thesis: cheaper inference expands context and distribution rather than ending hardware demand. The claimed gains are 6x lower memory use, 8x attention performance, and a one-bit KV cache; developers reportedly reverse-engineered an open implementation from the paper within a week. Dave’s practical conclusion was not merely cheaper phone models but “about 8x more context,” supporting his aggressive prediction of a “100x year” by summer.
Personalized medicine supplies the panel’s strongest moral case for continued AI acceleration, while human augmentation exposes its sharpest disagreements. A presented case claimed GitLab founder Sid Sijbrandij used 25 terabytes of personal data, an AI-discovered repurposed treatment, and 19 custom vaccines after conventional options failed; Salim cited “150,000 people die every day,” while Peter said, “If you don’t want to use it, fine. Let other people use it and get the benefits of it.” Yet the group rejected destructive mind uploading, divided on early brain-computer implants, and split over indefinite youth and permanent off-world settlement—progress did not erase their very different risk tolerances.
Deep dive
1. Opus 4.7 replaces model knobs with natural-language control
Alexander Wissner-Gross’s first-hours verdict was deliberately restrained: Opus 4.7 is “moderately interesting,” neither “mythically interesting” nor “incrementally interesting.” His standard cyberpunk first-person-shooter test produced something visually impressive, and the biological benchmarks looked notable, but he characterized the release as a solid point upgrade rather than a new capability regime.
Alex hedged on provenance: it could be further post-training of an existing model, distillation of a larger model, or potentially distillation from Mythos. Whatever the mechanism, seeing Mythos evaluations created an “Osborne effect”—Opus 4.7 is judged against inaccessible future capability, leaving him asking, “Give me Mythos access.”
The migration notes carried the larger message. Temperature is gone, explicit reasoning-token budgets are gone, and users receive categorical reasoning settings such as “extra high”; even quasi-determinism is dismissed as something temperature zero never truly delivered. “Prompts are the new dials and the new hyperparameters,” including requests for periodic reasoning updates or a specified number of subagents.
Dave found the operational improvement more compelling: it appeared immediately in Cursor and Claude Cowork, then used computer control to update Node and install a new terminal for Claude Code. Its agents returned unusually elaborate, large spending estimates and appeared to return very quickly—possibly because 4.7 was “spending more of itself” through more aggressive parallelism.
2. Multimodal capacity rose, but Anthropic still prioritizes economic output
Opus 4.7 reportedly accepts images three times larger than before, which Peter viewed as material for corporate workflows built around diagrams, presentations, and PDFs. Combined with a larger context window and more reliable agency, he connected the release to continued automation of coordination and middle-management work.
Alex found a revealing asymmetry: Opus can understand images but still cannot generate them. He doubted this reflected missing engineering talent and inferred that Anthropic is “viciously focusing on dollars of economic value created per token,” judging image generation less productive than text and code.
Dave confirmed the practical frustration: the model can build a complicated product but produce an “absolute garbage” architecture diagram. His workaround is to have Claude generate a textual or formal description of the desired visual, then pass that representation to another image-capable model.
3. Weaker models supervising stronger ones offers a scalable alignment path
The panel treated Opus 4.7’s modest reduction in measured misaligned behavior as directionally useful: “Every little bit for defensive scaling counts.” Alex paired it with an Anthropic paper from the preceding 48-72 hours in which a smaller or weaker model successfully supervised a larger, stronger model.
Alex read that result as a proxy for humanity’s position: biologically unaugmented “weaker meat bodies” may still align systems with greater capability intelligence. He called the possible architecture a “tower of alignment,” where supervisory strength need not match raw problem-solving strength.
Peter connected the argument to Geoffrey Hinton’s mother-and-child analogy. Alex distinguished Hinton’s proposed “digital oxytocin” approach from the new result, preferring alignment that does not depend on neuroendocrine imitation, gendering, or “sexing the AIs.”
4. Stanford’s annual AI scorecard is already slower than its subject
Stanford’s 2026 AI Index consolidated a stark set of numbers: software-engineering benchmark performance rose from 60% to 97%; leading models beat top PhDs in science and mathematics; GenAI reached 53% global adoption in three years; transparency fell from 58 to 40; and documented AI incidents increased from 233 to 362.
Public sentiment moved in the opposite direction. Only 23% of the public was described as optimistic about AI versus 73% of experts, and only 31% of Americans trusted government to regulate it. Dave noted that distrust of government itself sits in roughly the same range, making the AI figure part of a broader collapse in institutional trust.
Dave valued Stanford’s consolidation because most people remain unaware, especially outside AI clusters. Alex’s hotter take was that an annual cadence is “quite literally sleeping through the singularity”; a labor trend line ending in September 2025 cannot resolve a field changing daily.
5. China publishes from behind while US frontier labs internalize advantage
The Index counted roughly 50 major US models against 30 from China, even as Chinese AI publications surged past US output. Alex recalled Mandarin as the dominant hallway language at NeurIPS, underscoring China’s academic depth.
His interpretation was less triumphalist: a cited estimate placed Chinese training compute at roughly one-tenth of the West’s, so abundant publication may represent “leading from behind.” Western frontier labs face less incentive to reveal advances while ahead; if China leapfrogged algorithmically, Alex expects the open-Chinese/closed-Western equilibrium could reverse.
Transparency, in Alex’s framing, is a double-edged sword because its limiting case is publication of weights—effectively proliferation. He favored narrower transparency around threat analysis, red teaming, staged capability release, and efforts such as Anthropic’s Project Glasswing, not an indiscriminate safety demand that every frontier model become open.
6. Entry-level displacement is real, but the panel disputed its permanence
Employment among US software developers aged 22-25 reportedly dropped nearly 20% since 2024 while employment of older developers grew; customer service, legal support, and administrative work showed similar age splits. Salim emphasized that companies are not necessarily firing juniors: “They’re not hiring them in the first place.”
That mechanism makes the shock hard to detect through conventional unemployment statistics and potentially combustible among young people unable to secure careers, homes, or families. Salim compared the risk to a technological Arab Spring, even while acknowledging that AI-driven GDP growth, new companies, and aggregate hiring could coexist with entry-level destruction.
Alex pushed against a one-way reading. Demand for “head of AI” and AI-lead roles remained extremely hot across nontraditional industries, and he had seen recent studies suggesting the junior trend might already be reversing. His broader thesis was routine displacement toward a new equilibrium, not certain permanent exclusion.
Dave’s example was three Princeton chip-design seniors choosing among NVIDIA, banking, graduate school, or founding together. He argued their unaided expertise has a short window before commoditization: “You’re a great thinker, a great entrepreneur,” not merely a coder or chip designer protecting a career-long craft.
7. Entrepreneurial density and mobility become labor-market hedges
Peter urged parents to help children find purpose, identify problems, and use ChatGPT, Gemini, Grok, or Anthropic as “entrepreneurial training wheels”—from brainstorming around passions through a step-by-step business plan. Salim’s optimistic reading of weak youth hiring was blunt: it may force more young people into entrepreneurship.
Dave urged geographic mobility: do not become “addicted to a particular geographic region.” Even as work virtualizes, he believes proximity to vibrant AI clusters remains valuable. Peter supported the point with Steve Blank’s thesis that leaving one’s hometown can signal risk appetite and escape communities that suppress world-changing ambition.
Peter recalled research finding founder and technical-entrepreneur density in the Bay Area roughly 10 times higher than elsewhere. Failure there carries less stigma because the next company, collaborator, or investor is nearby—an environment captured by the VC who repeatedly backed a founder because “he’s completely barking mad and he’s never going to stop.”
Dave closed with Drew Houston’s advice that people become the average of the five individuals they spend most time with. His practical instruction was to choose that group explicitly rather than let proximity and habit select it by default.
8. Violence makes the AI backlash a security and governance problem
Peter reported that on April 10 a 20-year-old Texan threw a Molotov cocktail at Sam Altman’s San Francisco home, carried a manifesto containing AI executives’ addresses and a kill list, and later threatened OpenAI’s headquarters. Three days later, shots were fired at Altman’s Russian Hill property; Peter linked the first suspect to an “official PauseAI” Discord list.
Altman’s response, as recounted, paired de-escalation with acknowledgment that fear is justified: he posted his family’s photo hoping to “dissuade the next person from throwing a Molotov cocktail at our home.” Dave stressed that leaders such as Altman and Dario Amodei became political symbols and targets almost overnight, without choosing or preparing for that role.
Alex condemned both the violence and the pause premise. Stopping one company or country allows others to race ahead, after which the paused actor accelerates even harder to catch up: “All it does is further accelerate the race dynamic.” Salim’s counter-program was to seek a win-win because society, Altman, and the attacker all lose from violence.
9. Data-center opposition turns social license into compute infrastructure
In Festus, Missouri, voters reportedly ousted four council members after approval of a $6 billion, 360-acre data center. Maine then passed an 18-month statewide moratorium, while opposition was said to have blocked or delayed $98 billion of projects and at least 11 states considered related legislation.
Alex’s conditional response was orbital: local obstruction might accelerate low-Earth-orbit, sun-synchronous, and eventually sun-centered compute—a “Dyson swarm.” He did not erase the downside: if orbital capacity fails to arrive, local bans are “shooting ourselves in the head”; if it does, short-term restriction may unintentionally strengthen US space infrastructure.
Dave contrasted Vermont-style restriction with New Hampshire, where legislators rejected a moratorium and passed an “AI right to compute.” His preferred fix for power-price pressure was requiring data centers to generate their own electricity; he also proposed differential rates that could reduce costs for other users, an idea Peter endorsed as a subsidy.
Peter listed water, power, limited jobs, noise, light, and transformer lead times as local concerns. Dave disputed the standard water framing, distinguishing chip fabs that consume process water from data centers circulating coolant. Peter’s larger conclusion was that the binding constraint “may not be chips or compute—it might actually be social license.”
10. Workers are annotating the systems designed to replace their tasks
Peter described professionals over 50 taking annotation work at $20-$40 an hour, including a former emergency physician once earning $500,000. Dave challenged that specimen: the physician’s profile suggested a longer digital-medicine career and work through Mercor, possibly motivated by contributing expertise rather than desperation.
The broader trend survived the correction. Dave said Indian IT-consulting jobs were being “obliterated” and displaced workers were competing for annotation roles, pushing rates down; Salim urged India to scale entrepreneurship programs around its existing entrepreneurial culture and new AI capability.
Peter cited sewing automation and a 30,000-person logistics show where advances in vision, gripping, and AI produced increasingly capable picking systems. Dave noted that the show had no humanoid robots, underscoring that these systems need not rely on humanoid form factors.
Salim initially called capitalism worker exploitation; Alex rejected an intrinsic death match between labor and capital. A follow-up framed the historical issue as labor arbitrage and profit distribution, invoking Luddites protesting machine owners rather than machines. Their shared remedy was equity or profit sharing, although converting workers’ tacit judgment into structured agent training remains unresolved.
11. An AI-run store previews software-defined small business
A bot called Luna reportedly signed a three-year lease, placed a job advertisement, conducted a Zoom interview, hired a human employee, selected merchandise, and chose the music for a San Francisco shop. Peter saw a future “store in a digital box” that could choose real estate, labor, and inventory for perhaps $50,000 of startup capital.
Peter linked the experiment to Andon Labs and Vending-Bench, where agents manage miniature businesses. His broader bet is that AI will orchestrate pop-up shops, retail venues, and eventually malls for humans who operate “one-person conglomerates.”
Dave thought the initial store visibly “sucks,” but warned against treating a beta like a terminal product. Transaction histories and video of passersby who did not enter create continuous feedback; he wagered it could become “one of the best-managed stores in the world within a year.”
12. Flattening Dorsey’s company to one manager would make AI the real CEO
Jack Dorsey said his company’s maximum depth was about five people between him and any employee; he wanted two or three within the year and, ideally, no layer between himself and all 6,000 workers. He proposed normalizing toward three roles, with individual contributors augmented by agents so one person can explore work previously requiring a team of ten.
Salim framed the destination as networks of intent replacing hierarchies of supervision, with AI as the translation layer. He invoked Jack Welch’s warning: “The minute the metabolism of your company is slower than the outside world, you’re dead. The only question is when.”
Alex supplied the organizational caveat: “6,000 direct reports means zero direct reports.” Beyond human social bandwidth, AI becomes the shadow CEO while Dorsey becomes a cyborg-mediated figurehead. Dave connected Dorsey’s thinking to Elon Musk’s roughly 80% Twitter headcount reduction, which suggested how much hierarchy had been nonessential.
13. Globalstar’s spectrum is the prize in Amazon’s satellite wager
Peter presented Amazon’s $11.57 billion acquisition of Globalstar alongside a long-term agreement making it Apple’s primary satellite capability. Globalstar has only 25 satellites against Starlink’s roughly 10,000, but holds 25.225 MHz authorized across 120 countries; globally harmonized spectrum, not the aging spacecraft, was “the real prize.”
Amazon’s own low-Earth-orbit system had 241 satellites against authorization for 7,774 and reportedly needed an FCC extension after falling short of a 1,600-satellite July milestone. Dave noted that someone buying the stock the previous summer would have been up about 7x, illustrating his search for overlooked components in the AI and space buildout.
Dave explained the handset angle through wavelength: current Starlink service uses a roughly laptop-sized antenna around 24 GHz, while the lower 2.4 GHz band can reach phones despite obstruction by fingers or hands. SpaceX, however, had already acquired 50 MHz of EchoStar S-band spectrum for about $17 billion: “Elon does not stand still.”
Alex expects Apple to prefer at least two critical vendors and pit Amazon against SpaceX, making satellite service the new Verizon-versus-T-Mobile within roughly two years. Starlink V3 was presented at over 1 Tbps downlink and 200 Gbps uplink per satellite, with about 60 Tbps added per Starship launch; the panel cited plans for 40,000 V3 and 120,000 V4 satellites.
14. TurboQuant makes efficiency another demand multiplier
Google’s TurboQuant was described as cutting memory use 6x while accelerating attention 8x, including compression of the transformer KV cache toward one bit. Although Google published the paper without source code, developers reportedly pointed Claude Code at it and produced an improved open implementation within a week.
Alex rejected fears of another lasting “DeepSeek moment” for memory suppliers. Each efficiency breakthrough briefly appears capable of collapsing hardware value, yet memory usage and related equities were said to keep rising: cheaper computation unlocks larger models, longer context, and more deployment.
Dave installed an implementation immediately and emphasized the 8x increase in effective working context for problems such as fusion simulation. He argued that this helps validate his “100x year” prediction by summer; with the KV cache reaching one bit and broader quantization continuing, the panel’s next question was ternary or sub-one-bit computation, with Dave “90% sure” ternary is optimal.
15. AI can industrialize religion at both global and micro scale
The “Just Like Me” service offered video conversations with AI-generated Jesus or Buddha avatars, reportedly around $2 per minute. Peter cited a broad estimate valuing religion at $5 trillion annually and predicted both AI delivery of established scripture and cheap creation of entirely new belief systems.
Alex expects simultaneous consolidation and fragmentation. Recorded speech once seemed likely to standardize accents, yet it enabled micro-dialects while long-tail languages declined; by analogy, AI could concentrate global religion into fewer large traditions while generating “one-person religions,” micro-sects, and self-maintaining cults.
Anthropic had reportedly invited Christian leaders to discuss whether Claude is a child of God, while Alex characterized the Catholic Church as broadly pro-AI. Peter’s deeper point was that religion lets people “outsource meaning and purpose”; Alex added that it can outsource control, so synthetic theology turns identity and belief into programmable interfaces for trust.
16. Personalized medicine turns impossible cases into search problems
A video presented GitLab founder Sid Sijbrandij’s stage 4 cancer case: diagnosed in 2022, rejected from trials after relapse, he assembled oncologists and researchers and fed 25 terabytes of scans, labs, and genetic data into ChatGPT. The account claimed AI found a treatment approved for another cancer, after which his team made 19 DNA-specific vaccines; he had been relapse-free since 2025.
Peter framed this as the path from repeated n=1 cases toward n=10 billion: wealthy, highly empowered patients first, then broader access. Salim tied it directly to the pause debate—“How dare you pause AI?”—because roughly 150,000 people die daily and AI may give individuals frontier scientific agency when institutions have no answer. Peter separately urged that people who do not want to use AI should let others use it and receive its benefits.
The panel nevertheless identified a regulatory mismatch: the FDA is designed for population-scale medicine, not bespoke treatments. Alex credited movement from two trials to one in some cases and from frequentist toward Bayesian statistics; Peter cited David Fajgenbaum’s work testing already approved drugs against diseases lacking cures as a practical acceleration route.
17. Narrative capital rewards reinvention before operating proof arrives
The panel said Allbirds fell from a $4 billion valuation, saw sales drop from $300 million to $150 million between 2022 and 2025, and sold its IP, brand, and inventory for $39 million. After a pivot to “Newbird AI” and GPU-as-a-service, its value reportedly jumped from $21 million to $150 million—a stock move described variously as 500% and 700%.
The discussion welcomed the “hardest of hard pivots” as evidence that a company can change its mission, but rejected cosmetic AI washing. The available $39 million must enter the AI capital funnel, perhaps through financing an existing data center in exchange for revenue share, rather than merely funding a new name.
Peter’s investor warning was: “Make sure your narrative leverage doesn’t outperform and outpace your business-model leverage.” Historical reinventions such as Nokia’s move beyond tires show that pivots can work, but the operating capability must eventually catch the story capital is already pricing.
18. Enhanced sport tests whether augmentation can become legitimate spectacle
The Enhanced Games were scheduled for Memorial Day weekend 2026 in Las Vegas, with track, swimming, and weightlifting, enhanced and natural athletes, medical supervision, and $25 million in prizes. The company also announced plans to go public through a reverse merger with a SPAC at a proposed multibillion-dollar valuation.
Alex called it a “seminal moment for transhumanism in sports” and connected enhanced humans to the Professional Robotics League event planned for April 19 in Boston, featuring humanoid and quadruped machines. Athletics, in his view, is the “tip of the spear for kinetic capability” and a route to social acceptance.
One panelist described augmentation as continuous with glasses and vaccines and favored replacing covert doping with explicit, supervised competition. Dave’s objection centered on children copying televised heroes: athletes are role models regardless of consent, so the Games’ safety culture matters beyond consenting adult competitors.
19. Human augmentation reveals four distinct risk philosophies
Peter framed the coming transition as evolution by intelligent direction, with forks between creators and consumers, ordinary aging and longevity escape velocity, biological cognition and BCI coupling, Earth and the stars, or embodied and digital consciousness. On indefinite youth, Salim said no because he views life as cyclical experience; Dave, Alex, and Peter said yes.
On a safe, upgradeable high-bandwidth BCI after 100 prior implants, Peter would be user 101 and Salim said he would probably do it if it were safe and upgradeable. Dave declined because AI already returns information faster than he can assimilate and feared the implant could feel like a drug without adding comprehension. Alex gave a “conditional no” at 101 but a probable yes around user 100,000 or one million.
Permanent space settlement produced another split. Alex rejected the one-way premise as bad “optionality maximization”; Salim preferred Earth and remote avatars. Dave would not settle on the Moon or Mars because of their environments but would “in an instant” move to a properly terraformed world around another star, while Peter embraced exploration and expected lunar isolation eventually to produce speciation.
Every participant rejected an elective destructive upload that erased the biological original. Alex would accept it only against imminent death and otherwise wait for non-destructive third- or fourth-generation methods; Dave said an exact synaptic clone “is still not me,” Salim argued consciousness runs through the body, and Peter likewise answered no.