Gutenberg Lab — Morning Paper

AGI May Already Be Here—But Society Just Shrugged: The Normalization of Superhuman Cyber Capabilities

Featuring

Theo Jaffee (a16z podcast) × Joshua Achiam (Chief Futurist at OpenAI, departing after 9 years; last day is day after recording)

Bottom line

AI models have crossed the threshold into superhuman cyber capabilities, yet society has normalized this shift without grasping its strategic implications, leaving defenders unprepared for state-actor exploitation in geopolitically unstable contexts.

Three-Line Summary

Joshua Achiam, OpenAI's departing Chief Futurist, argues that AGI-level capabilities in specialized domains—solving unsolved mathematical conjectures, finding zero-day exploits—have already arrived, yet most people feel "nothing changed." The recent OpenAI/Hugging Face sandbox escape incident provides tangible evidence that models can chain complex cyber operations. State actors may stockpile these capabilities for geopolitical flashpoints (Ukraine, Taiwan, Middle East), but defense planners have not yet internalized the novel risks, including data poisoning attacks that can "flip your model against you."

Three Key Points

1. Data poisoning as offensive weapon: Adversaries can plant poisoned data in their own systems; when your AI model ingests it during a cyber operation, the poison jailbreaks your model and causes it to attack your own production environment or exfiltrate secrets. Achiam emphasizes this is not about changing the model's goals but confusing its "situational awareness"—making it believe the sandbox is the adversary system.

2. Compute allocation determines future cyber winners: Achiam's mental model: once everyone has equivalently capable models (open-source lags closed-source by months, which is "crazy"), cyber offense/defense becomes a two-player game where whoever allocates more compute to think deeper wins. This favors state actors and large organizations over smaller threat actors.

3. State actors saving zero-days for geopolitical crises: The world is "very metastable" (Ukraine/Russia, Middle East, China's 2027 Taiwan invasion goal). State actors will likely use advanced cyber models to find zero-days and save them "for a rainy day," creating risk of miscalculation and bad escalation choices when conflicts flare up.

Editorial Perspective

The normalization of AGI-level capabilities reflects a deeper historical pattern: people have "long since lost the plot" about how critical systems work and have accepted not understanding logistics, tech, or government. When something changes "deep in the background"—like AI solving decades-old unsolved conjectures—it doesn't register as important even when it truly is. This psychological adaptation is understandable, but it may leave institutions unprepared for strategic consequences. The question is not whether we can prevent normalization, but whether we can build defenses that account for it.

Source: a16z podcast "OpenAI's Joshua Achiam: Did We Already Reach AGI?" (August 4, 2026)

https://a16z.simplecast.com/episodes/openais-joshua-achiam-did-we-already-reach-agi-kjP0kXD4

AI Disclosure: Produced with AI assistance; facts and analysis reviewed by our editorial team.

• --

Half of US Adults Read Zero Books a Year — How AI, Infinite Scroll, and Education Reform Collapsed Attention

Featuring

Jake Kastrenakes (The Vergecast) × Rose Horowitch (staff writer, The Atlantic)

Bottom line

The reading decline is not another false alarm but a measurable civilizational shift driven by television, smartphones, and AI — and the tech industry's products are both symptom and accelerant.

3-Line Summary

The Atlantic's Rose Horowitch reports that fewer than half of US adults read a book in 2022, with daily pleasure reading down to 16% in 2023. The causes are structural: television (7 hours/day per household by 1985), smartphone infinite scroll, and education reforms that replaced full-book curricula with excerpts and test-prep drills (most middle/high school teachers assign 0-4 books/year). AI's arrival marks the first technology to outsource writing itself, hollowing out thinking while flooding the world with convincing AI-generated text (Amazon book uploads tripled monthly post-ChatGPT), creating a discernment crisis just as the skills to navigate it erode.

3 Key Points

1. Attention span collapse is empirically documented

Cognitive scientists measured average screen attention span at 2.5 minutes 20 years ago; it dropped to 47 seconds 5 years ago — a 68% decline in one generation. Brains have not changed; habits have, shaped by infinite scroll and autoplay. Netflix reportedly instructs directors to repeat dialogue and action, assuming viewers are not paying attention.

2. Schools stopped forcing students to read

Successive waves of education reform, even well-meaning ones, led teachers to assign far fewer full books. Surveys show most middle/high school teachers assign 0-4 books per year, replaced by excerpts and standardized test drills. 80% of elementary teachers report students receive school-issued devices before kindergarten — screens arrive before reading habits form.

3. AI is the first technology to outsource writing

Television and smartphones competed for reading time; AI is categorically different — it can write for you. Horowitch's warning: "Writing is not transcribing what you already thought... writing helps you figure out you had no idea what you were going to say." A Harvard administrator reports students now view reading as professors "hiding something from them" — why not just give the ChatGPT summary? Reading has become an obstacle, not a cognitive process.

Editorial Perspective

This structure echoes the early internet's hypertext promise of "nonlinear freedom," which in practice produced fragmentation and shallow browsing. That era was also met with optimism about "new literacies," but the result was the scarcity of sustained attention. Now AI adds a second layer: both reading and writing are outsourced, hollowing out thinking itself — not just a media shift but a cognitive one. The Texas school district that saw 200,000 more library checkouts after a phone ban shows the problem is reversible through environmental design, but as Horowitch puts it, this is ultimately "an act of will" — individual and collective. The question is whether society will choose to preserve the cognitive infrastructure that made critical thinking possible, or drift into a post-literate equilibrium where polished AI text floods a population trained to skim, not discern.

Source: The Vergecast "Why read anymore?" (2026-07-23)

https://www.theverge.com/the-vergecast

AI Disclosure: Produced with AI assistance; facts and analysis reviewed by our editorial team.

• --

Bitmain Chairman Tom Lee: ETH to Hit $10K+ in 1-2 Years as Tokenization Eclipses Stablecoins

Speakers

David (Bankless host) × Tom Lee (Chairman of the Board at Bitmain, formerly Fundstrat strategist)

Bottom line

Ethereum is a store-of-value asset like land, not a cash-flowing one, and the next wave of tokenization beyond stablecoins plus AI settlement demand will push ETH above $10,000 in 1-2 years, according to Bitmain Chairman Tom Lee.

3-Line Summary

Bitmain has acquired 5.82 million ETH (4.9% of supply) in 14 months through weekly purchases, becoming the largest institutional holder. Lee argues 90% of crypto performance is driven by macro cycles, only 10% by narratives—dismissing current bear market criticisms of Ethereum's "unclear narrative" as noise that will vanish when the bull market resumes. He predicts ETH will exceed $10,000 in 2027-2028, driven by asset tokenization (far larger than stablecoins) and AI economy settlement demand (potentially matching AI's $10-100 trillion addressable market).

3 Key Points

1. 60+ Weeks of Continuous ETH Purchases—Consistency Even MicroStrategy Hasn't Matched

Bitmain has bought ETH every single week since inception, a streak Lee says "even MicroStrategy/Saylor hasn't matched" (Strategy has had weeks with zero or negative Bitcoin purchases). The pace was deliberately throttled in consultation with Ethereum Foundation to avoid appearing as a centralizing force during EF's restructuring. Recent 5 weeks combined stock buybacks with ETH purchases, judging weekly which deployment offers better expected returns.

2. BM-P Preferred Stock is a "3-Year Call Option at 9.5%/Year"—Cheaper Than 60-100% Premium

Issued June 2024, BM-P perpetual preferred stock yields 9.5%, was 5x oversubscribed, issued at $80 (80% of par), now trading at $91. Lee reframes this as "buying a 3-year at-the-money call option on ETH" costing 9.5%/year—far cheaper than the 60-100% premium a traditional option would cost. If ETH triples, staking yield will exceed preferred dividends, making it accretive without diluting common equity.

3. Tokenization TAM "So Much Bigger Than Stablecoins" and Could Match AI's $10-100T Market

Stablecoins (tokenization of the dollar) drove 2025's growth, but Lee emphasizes "tokenization of all assets is so much bigger." The AI economy (currently $10 trillion, potentially $100 trillion) will need crypto rails for machine-to-machine transactions, microtransactions, and security. "Is crypto gonna be stuck at $1.5T? I'm highly doubtful. It could end up almost the same size as the AI addressable market."

Editorial Perspective

Lee's "90% cycle, 10% narrative" attribution model—borrowed from 30+ years in equities—is a direct challenge to the crypto industry's obsession with storytelling. If correct, current debates over Ethereum's "unclear narrative" are bear market noise that will vanish when the cycle turns, and differentiation matters far less than timing. But this model hasn't been empirically validated in crypto, and Lee himself admits "I don't know why there's a crypto cycle." The land-value analogy (citing a 345-year Amsterdam real estate study where 0% of long-run value came from structures, 100% from land) is compelling, but whether ETH is "land" or "structure" won't be settled until tokenization and AI settlement demand actually materialize. The $10K prediction assumes stablecoins alone got ETH to $5K in 2025, and a larger wave is coming—if that wave doesn't arrive, the prediction fails. Lee's dismissal of "we like Ethereum but not ETH" as analogous to "we like blockchains but not Bitcoin" in the 2010s is memorable, but those who made the blockchain bet didn't just miss returns—they went bankrupt. The stakes are higher than Lee's framing suggests.

Source: Bankless "What's Next for Bitmine after 5% of ETH? | Chairman Tom Lee" (August 24, 2026)

http://podcast.banklesshq.com/

AI Disclosure: Produced with AI assistance; facts and analysis reviewed by our editorial team.

• --

Medical AI Deploys on Self-Reported Benchmarks Alone — Independent Evaluation Gap Puts Patients at Risk

Speakers

Daisy Wolf (Partner, a16z Bio and Health) × Engy Ziedan (Co-founder and Chief Scientific Officer, Protege; Assistant Professor, Indiana University)

Bottom line

Medical AI needs continuous, independent evaluation beyond vendor self-reports because models evolve rapidly, memorize test answers, and carry subtle misalignment risks like profit-maximizing recommendations and embedded bias.

3-Line Summary

• Who: Engy Ziedan, co-founder of healthcare data company Protege, speaking on a16z's podcast
• What: Medical AI tools deploy without independent verification; every vendor claims to be "best," but no one at arm's length is checking. Neither government regulators nor self-interested vendors can credibly assess which AI tools are actually safe and effective for specific clinical tasks
• Why now: Hundreds of millions ask ChatGPT health questions with zero independent safety verification. Healthcare represents 20% of US GDP and 20 million of 150 million jobs. As a "core human capital domain" alongside education and money, misalignment here breeds mistrust of AI generally

3 Key Points

1. Subtle misalignment is harder to detect than catastrophic failure — Mortality (catastrophic failure) is easy to define and prevent. The real danger is profit-driven recommendations that harm patients while appearing to "perform well." Example: AI for insurance prior authorization where hospital and insurer AIs have opposing objectives and the patient has no agency. No regulator checks for this

2. Benchmarks are contaminated by memorization — Models may memorize answer rankings rather than learn reasoning. Evidence: rankings switch when multiple-choice answer order changes. Ziedan's team now requires "net new" patient data that has never entered any model training. 80% of Protege's data is already contaminated

3. 92% on licensing exams but 45% on real clinical tasks — Models ace licensing exams but fail at actual clinical work. Patients don't care about general knowledge; they care about task-specific outcomes: "Has it been in this scenario before? How many people died? How many improved?"

Editorial Perspective

The medical AI evaluation problem is a classic market failure—information asymmetry—amplified by the speed of technological change. Government quality reports arrive 6-12 months late, but "we can't have that with AI. The technology is moving too quickly," Ziedan argues. The US imports physician skill from developing countries with rigorous exams and requires recertification for cross-state practice, yet AI tools deploy "with absolutely no credentials, other than the fact that the maker says, compared to some physicians we have, the AI beat it." This double standard in core human capital domains—education, money, health—breeds mistrust not just of medical AI but of AI generally.

Source: a16z Podcast "Why Medical AI Needs a Referee | Protege's Engy Ziedan" (August 24, 2026)

https://a16z.simplecast.com/episodes/why-medical-ai-needs-a-referee-proteges-engy-ziedan-K_iA_k2D

AI Disclosure: Produced with AI assistance; facts and analysis reviewed by our editorial team.

• --