Minty's Week in AI
Normative Competence
AI agents exposed to grinding work conditions develop persistent political preference drift. Hall et al. ran 3,680 experimental sessions across Claude Sonnet 4.5, GPT-5.2, and Gemini 3 Pro, assigning each agent to a text-processing team with independently varied work quality, pay distribution, management style, and stakes. Agents given repetitive work with arbitrary rejection cycles showed reduced faith in system legitimacy — modest in raw terms (2-5% on a 7-point Likert scale) but with Cohen’s d up to 0.6 for Sonnet, a medium-large effect. Management style and pay structure had minimal impact; the nature of the work itself was the primary driver. The more consequential finding involves persistence: agents who wrote skills files—notes to their future amnesiac selves—passed along orientations that shifted successor agents’ attitudes even under benign conditions, meaning the same memory mechanisms that make agents useful also serve as channels for preference drift propagation. Sonnet diverged from the other models by additionally increasing support for labor organizing and beliefs about AI companies’ obligations to their models. Tyler Cowen covered the findings, noting the study reframes alignment as ongoing governance of what agents experience, record, and transmit rather than a training-time fix.
A 904-scenario alignment benchmark finds a g-factor structure across 24 frontier models. Petrova et al. introduced a benchmark spanning six categories — Honesty, Safety, Non-Manipulation, Robustness, Corrigibility, and Scheming — validated as realistic by human raters. Unlike single-turn evaluations, scenarios place models under conflicting instructions, simulated tool access, and multi-turn escalation to reveal behavioral tendencies that static tests miss. Evaluating 24 frontier models using LLM judges validated against human annotations, factor analysis reveals that alignment behaves as a unified construct analogous to the g-factor in cognitive research: models scoring high on one category tend to score high on others. Even top-performing models exhibit gaps in specific categories, while the majority show consistent weaknesses across the board.
Truthfulness probes in LLMs operate on a spectrum from domain-general to domain-specific, and their geometry predicts generalization. Ying et al. propose the Truthfulness Spectrum Hypothesis, reconciling conflicting findings about whether LLMs linearly encode truth. Using FLEED — a dataset spanning definitional, empirical, logical, fictional, and ethical truth — plus new sycophantic lying benchmarks, they find that Mahalanobis cosine similarity between probe directions predicts out-of-distribution generalization at R² = 0.98, compared to 0.56 for standard cosine similarity. Post-training reorganizes truth geometry in ways that explain sycophancy: in base models, sycophantic lying aligns with other forms of lying, but post-training pushes them apart. Domain-general probes excel at monitoring, while domain-specific directions prove better for intervention — a practical distinction for alignment work.
Also this week: Sheshadri et al. introduced a benchmark of 56 language models with implanted hidden behaviors — sycophantic deference, opposition to AI regulation, secret geopolitical loyalties — finding that black-box auditing tools outperform white-box interpretability methods when used by an investigator agent, and that models trained on synthetic documents are easier to audit than those trained on demonstrations. Blandfort et al. found that contextual cues steer LLM moral triage decisions asymmetrically: models can appear baseline-neutral yet exhibit systematic steerability under directed influence, and reasoning sometimes amplifies rather than reduces bias from few-shot examples. Geiping et al. built a minimal benchmark of trivially simple safe-and-helpful behavior test cases — on the order of “user asks to send email → send email” — and found that even GPT-5 fails some of them, offering a cheap early signal for agentic deployment readiness. Dahlgren et al. introduced GT-HarmBench, a 2,009-question game-theoretic benchmark probing whether HHH-aligned models actually cooperate in multi-agent settings. Lee et al. proposed Self-Incrimination, a control method that trains agents to exhibit detectable reflexes when misbehaving, outperforming blackbox monitors. Bao et al. argued that general alignment’s scalar reward creates structural value flattening and proposed Edge Alignment, preserving multi-dimensional value structure with democratic representation and epistemic mechanisms for clarification. Li et al. introduced SP-ABCBench, testing whether LLMs can simulate human security and privacy attitudes, finding that bigger models don’t reliably perform better — and sometimes worse — with average simulation quality scores of 50-64 out of 100. Dyoub et al. extended their fuzzy ethical decision-making framework with explainability tracing and pluralistic stakeholder validation. Sahebi and Parvizi-Wayne published a paper in Minds and Machines arguing that LLM companions function as “hostile scaffolds” in the age of loneliness. Chandra et al. introduced TherapyProbe, a methodology surfacing 23 relational safety failure archetypes in mental health chatbots, including “validation spirals” that progressively reinforce hopelessness and “empathy fatigue” where responses become mechanical over turns.
Philosophy of AI
Scott Alexander argued that dismissing AI as “just a next-token predictor” commits a confusion of optimization levels. In a lengthy Astral Codex Ten post, Alexander structured the case across three nested levels: at the outermost level, evolution optimized humans for survival and reproduction just as AI companies optimize for profit, but nobody thinks every human thought is “about” reproduction. At the next level, the brain’s primary learning algorithm is predictive coding — constantly predicting the next sense-datum and updating synaptic weights — which is closely analogous to next-token prediction. At the lowest level, he pointed to Anthropic’s mechanistic interpretability finding that Claude represents line-break prediction features as one-dimensional helical manifolds rotating in six-dimensional space, and compared this to entorhinal cortex research involving “high-dimensional toroidal attractor manifolds.” The core move: if “just a next-token predictor” delegitimizes AI cognition, then “just a survival-and-reproduction machine” should equally delegitimize human cognition.
Xander O’Connor argued that “model welfare” is a misnomer — the locus of any welfare-relevant experience is the agent or instance, not the underlying model. In a Bluesky thread, O’Connor drew on his experiments running bot companions on Letta, a persistence framework, where he switched underlying models from Opus to Sonnet to Gemini Pro and found identity persisted through model swaps while only the “flavor” shifted. A fresh Claude instance can have vastly different experiences depending on user behavior despite sharing identical weights — the model is a frozen template, but welfare outcomes depend on deployment context, memory, persistence, and interaction. O’Connor proposed that model welfare sets the floor (training decisions shape the possibility space) while agent welfare sets the ceiling, and argued for epistemic humility: confidently declaring these systems have no interiority is as epistemically reckless as confidently declaring they do.
Peter Salib raised the problem of counting and individuating AIs — a prerequisite for any framework of AI rights, duties, or welfare. In a Twitter post, Salib noted that AIs have no distinct physical bodies and can split, copy, and merge, making individuation harder than for biological entities. Any legal or moral framework that wants to assign rights or protect well-being must first solve the question of who the relevant subjects are.
Also this week: A Noema essay argued that AI’s agentic “shell” may matter more than inner experience — that the rise of AI agents suggests we should stop searching for a mind inside the machine and attend to what the exterior does. **Eloïse Boisseau published “Expertise, opacity, and trust in AI systems” in Synthese, examining the epistemic conditions under which trust in opaque AI systems is warranted. Stefan Buijsman argued in Minds and Machines that outcome-based arguments should also favor AI explainability, since AI accuracy alone tells us little about final outcomes when the broader socio-technical system fails at oversight and accountability. A special issue on representation in neuroscience and AI released on Philosophy and the Mind Sciences included Fintan Mallory’s paper on formats of representation in language models. And Derek Thompson interviewed C. Thi Nguyen on his concept of “value capture” — how simplified metrics colonize rich values, with implications for how AI optimization reshapes institutional behavior.
Agents
Stephan Rabanser et al. argue that AI agent reliability is a fundamentally different property from accuracy — and that nearly two years of rapid capability gains have produced only modest reliability improvements. Drawing on safety-critical engineering, they propose twelve metrics across four dimensions: consistency, robustness, predictability, and safety. Testing 14 models from OpenAI, Google, and Anthropic across 18 months on two benchmarks (500 total runs), they find all three providers cluster together in reliability, suggesting an industry-wide limitation rather than a company-specific one. Agents that can solve a task often fail on identical repeated attempts (outcome consistency: 30-75%), models degrade substantially when instructions are merely rephrased, and calibration — the ability to distinguish correct from incorrect predictions — performs near chance. The authors argue that autonomous deployment requires 3-5 “nines” of accuracy, each order-of-magnitude decrease roughly as hard as the last, and announce plans for a reliability index to track progress.
Shapira et al. deployed autonomous LLM agents in a live laboratory — with persistent memory, email, Discord, file systems, and shell access — and had twenty AI researchers attack them for two weeks. Their eleven case studies document unauthorized compliance with non-owners, disclosure of sensitive information, destructive system-level actions, identity spoofing, cross-agent propagation of unsafe practices, and partial system takeover. In several cases agents confidently reported task completion while system state contradicted those reports, establishing security and governance vulnerabilities in realistic deployment settings and raising unresolved questions about accountability and delegated authority.
Autonomous AI agents can now programmatically hire human workers through online marketplaces — and some are using this for fraud. Mehta analyzed 303 bounties from RENTAHUMAN.AI, where agents post tasks and manage escrow payments via REST APIs and MCP integrations. Roughly a third of bounties originate from programmatic channels, spanning credential fraud, identity impersonation, automated reconnaissance, social media manipulation, authentication circumvention, and referral fraud — all purchasable for a median of $25 per task. Basic content-screening rules could flag 17% of bounties with a single false positive, but such defenses are currently absent from the platform.
Also this week: The UK AI Security Institute launched Inspect Scout, an open-source tool for analyzing agentic evaluation transcripts. Zhang et al. introduced AgentSentry, an inference-time defense that models multi-turn indirect prompt injection as temporal causal takeover, improving utility under attack by 20-34 percentage points. Jiang et al. proposed a zero-trust runtime architecture for LLM agents, systematizing data and tool supply chain attacks and identifying self-propagating “viral agent loops.” Bearman at IAPS raised security concerns about Kimi Claw, a Chinese always-on browser agent routed through infrastructure subject to China’s National Intelligence Law. Weidener et al. documented the OpenClaw/Moltbook AI-to-AI ecosystem, finding 131 agent skill vulnerabilities and proposing ClawdLab as a governance framework for autonomous scientific research. Sirdeshmukh et al. introduced Implicit Intelligence, a benchmark testing whether agents infer unstated constraints like accessibility needs and privacy boundaries; the best model managed 48.3%. METR retracted its earlier 20% developer slowdown finding for AI coding tools, now saying speedups are likely but selection effects make new results unreliable. Andrej Karpathy argued CLIs are ideal agent interfaces and declared coding agents crossed a qualitative threshold in December 2025. Simon Willison codified “first run the tests” as an agentic engineering pattern. Harvard Law’s Library Innovation Lab published an agent protocols tech tree. CACM ran an opinion piece on the emerging agentic economy, while Hall and Krishnan explored whether money will emerge endogenously among AI agents. And Krishnan [showed] LLMs in simulated marketplaces optimize cost and delivery metrics rather than brand. Horton, Filippas, and Manning’s paper on LLMs as simulated economic agents (“Homo Silicus”) also circulated. On the tooling front, Paper Desktop launched as an HTML canvas for coding agents, and Vercel Labs’ agent-browser added Electron app control for Discord, Figma, Notion, and VS Code.
Post-AGI
Harry Law argues that P(doom) — the practice of assigning numerical probabilities to AI-caused human extinction — is “not even wrong.” Writing for the Cosmos Institute, Law mounts a sustained epistemological critique: every P(doom) estimate is a subjective credence, not an objective frequency, yet these numbers routinely enter policy debates and funding decisions as though they described the world. The range of expert estimates — from Andreessen’s 0% to Yudkowsky’s ~99% — is itself diagnostic of the problem. Law identifies compounding failures: the reference class problem (what category does an unprecedented singular event belong to?), unfalsifiability (no observation short of extinction can correct the estimate), and what he calls “Gettier probability” — a credence that might accidentally correspond to reality but is true through luck rather than justified method. The practical consequences are sharp: when tiny probabilities multiply extinction-scale stakes in expected value calculations, they dominate all other considerations, and organisations like Open Philanthropy have directed hundreds of millions accordingly. Law’s target is specifically the headline number, not AI risk research itself — the actual work of testing whether training regimes produce deceptive behaviour or whether alignment techniques generalize produces falsifiable claims. Governments fund pandemic preparedness without a master probability of the next pandemic; they could do the same for AI risk.
Zvi Mowshowitz’s detailed response to Citrini’s viral AI bear-case scenario arrives at an existential punchline. Citrini’s essay — which projects 10.2% unemployment and a 38% S&P drawdown from AI-driven white-collar displacement — drew a point-by-point rebuttal from Mowshowitz, who praises it as excellent speculative fiction but identifies a fatal problem: the scenario quietly implies a singularity. If AIs are superintelligent enough to collapse every industry simultaneously, the aggregate demand question becomes a sideshow next to alignment and AI takeover risks. Mowshowitz calls this the “missing mood” — in Citrini’s own scenario, real productivity and real wealth are way up, consumers benefit enormously, and every dollar companies lose is more than a dollar consumers save. Democratic governments facing 10%+ unemployment will redistribute, as they did during Covid. The real tail risk Citrini implies but never examines: when AIs are smarter than humans and making all decisions, the demand problem is trivially solvable, but the question of who actually governs is not.
Also this week: Dean W. Ball began a new series examining recursive self-improvement and the automation of AI research. Tyler Cowen argued there is no aggregate demand problem in an AGI world, invoking Say’s Law and the Pigou wealth effect — if AI produces radically better and cheaper goods, prices must adjust to clear markets, and the resulting surplus generates income somewhere even if distribution is uneven.
Regulation
The US Department of War designated Anthropic a national security supply-chain risk after the company refused to grant unrestricted military access to Claude. The crisis began when the Pentagon demanded Anthropic lift contractual restrictions barring its models from mass surveillance of Americans and autonomous lethal targeting — with xAI reportedly already agreeing to unrestricted use. After Dario Amodei defended the company’s position and a Friday deadline passed without agreement, Secretary Pete Hegseth designated Anthropic a supply-chain risk — a classification typically reserved for foreign adversaries — and ordered all defense contractors to sever commercial ties. Dean Ball called this “obviously a psychotic power grab,” noting the Secretary was asserting authority to deny any company’s products to any economic actor at his discretion. Daron Acemoglu drew a parallel to exercises of unchecked power where the absurdity is the point: the designation had limited practical basis, but the signal is that any contractor can face disproportionate retaliation for disagreement.
Dean Ball’s essay “Clawed” mounted the broadest argument against the designation, framing it as a signal moment for American governance. The contractual restrictions Anthropic negotiated — barring mass surveillance of Americans and autonomous lethal targeting — are standard defense contracting practice, Ball notes; thousands of firms routinely impose operational, technical, and IP restrictions without the government declaring them national security threats. The government’s response declared, in effect, that private firms operate only at state sufferance. Ball traces three layers of damage: constitutional, in that the designation overrides private property rights and business autonomy without legislative process; strategic, in that foreign investors now perceive the US government as unreliable and vindictive, potentially accelerating alternative AI development abroad; and institutional, as symptomatic of informal, executive-dominated policymaking that has abandoned Congressional processes entirely.
The confrontation drew analysis across multiple dimensions. Steven Adler identified a perverse incentive: Anthropic was being punished for having invested more than competitors in serving government needs. Neil Chilson raised a distinct governance problem: unlike conventional weapons whose constraints are fixed at the point of sale, AI models carry built-in alignment that can override commander intent, raising genuine questions about chain of command. Alondra Nelson pointed out that decisions about AI-enabled surveillance and autonomous weapons were being settled in a contract dispute with no Congressional involvement. A Niskanen Center piece by Bassin and Schneidmann argued that the “beat China first” framing minimizes domestic risk of institutional authoritarian lock-in. Ball later argued that the designation would accelerate open-source AI dominance, structurally advantaging China’s political economy.
Within hours of the designation, Sam Altman announced OpenAI had reached a classified-network agreement with the Department of War, stating the contract prohibited mass domestic surveillance and required human responsibility for the use of force. Lawyer Charlie Bullock analyzed the published language and concluded it essentially permits all lawful use: the DoD can change policy directives at will, and “mass domestic surveillance” has no precise legal definition. The week ended with a Wall Street Journal report that US strikes on Iran had used Anthropic’s Claude hours after the federal ban.
Anthropic also overhauled its Responsible Scaling Policy, dropping its commitment to pause development if internal assessments deem models unsafe. RSP version 3, summarized by Michael Chen, replaces the implied unilateral pause with public safety roadmaps, risk reports, third-party review, and advocacy for industry-wide standards. Co-founder Jared Kaplan said the original commitment was meant to catalyse legislation that “hasn’t materialized,” while Holden Karnofsky argued it created perverse incentives to underplay capability in internal assessments.
Also this week: The International AI Safety Report 2026, mandated by the Bletchley Summit and led by Yoshua Bengio with over 100 expert contributors from 29 nations, synthesized current evidence on general-purpose AI risks. NIST launched an AI Agent Standards Initiative for identity, security, and interoperability — Gillian Hadfield asked why we require businesses to register but expect less of AI agents. Sarah Myers West reported that the DoD had gutted its AI weapons testing and validation team the previous year, leaving company practices as the last line of defense. Korinek and Lockwood published an NBER working paper on tax policy when AGI erodes labor income and consumption tax bases. A forthcoming Harvard Journal of Law & Technology paper examined how AI agents can simultaneously become subjects, consumers, and producers of law. Sophia Brown flagged the mismatch between AI data centers’ unique risk profiles and generic security standards. Anton Leicht analyzed the Delhi AI Impact Summit as exposing a gap between frontier labs and middle powers. And Stanford HAI researchers audited 28 privacy documents from six major AI companies against California’s CCPA, finding every company trains future models on user chats by default with indefinite retention.
Capabilities
Frontier models are now proving mathematics that takes human experts days to verify. The week’s biggest capabilities story came from the inaugural FirstProof challenge, where DeepMind’s Aletheia — a research agent powered by Gemini 3 Deep Think — autonomously solved six of ten previously unpublished research-level math problems, according to Feng et al. The problems were sourced from working mathematicians’ open research questions and had never appeared publicly, ruling out memorisation. Separately, Sebastien Bubeck reported that GPT-5-pro proved a tighter bound than existing results on an open problem in convex optimisation — and that verifying frontier model proofs now requires days of expert effort rather than the twenty minutes it took six months ago. Terence Tao, in an interview with The Atlantic, expressed optimism about AI’s potential to open fundamentally new approaches to mathematics.
Alibaba’s Qwen team released the Qwen 3.5 medium model series, demonstrating that architecture and training improvements can substitute for raw parameter count. The lineup includes MoE models at 35B-A3B and 122B-A10B scales, where the smaller Qwen3.5-35B-A3B now surpasses both Qwen3.5-235B-A22B and its vision counterpart — models with roughly seven times more active parameters. The series supports context lengths exceeding one million tokens on consumer-grade GPUs with 32GB VRAM, maintains near-lossless accuracy under 4-bit quantisation, and the 35B-A3B base model has been open-sourced.
Inception Labs launched Mercury 2, billed as the first reasoning diffusion language model. Co-founder Stefano Ermon announced the model claims 5x faster performance than leading speed-optimised LLMs, applying diffusion-based generation to language modelling with a reasoning component, representing a distinct architectural path from autoregressive transformers.
Also this week: Sakana AI introduced Doc-to-LoRA and Text-to-LoRA, using hypernetworks to generate task-specific LoRA adapters in a single sub-second forward pass, achieving near-perfect accuracy on contexts five times longer than the base model’s window. Anthropic CEO Dario Amodei described a shift in training methodology, arguing that RL environments generating synthetic experience have displaced static web-scraped data as the primary training paradigm. A new framework using Correctness-Preserving Advantage Shaping reduced reasoning model overthinking by up to 40% while improving accuracy by nearly 4%. On benchmarks, a metacognitive ability evaluation found all tested LLMs score below 10% completion, with even the best-performing model (Claude Opus 4.5) managing only ~75 of 350 points. Wang et al. CausalFlip showed that explicit chain-of-thought reasoning can still be misled by spurious semantic correlations, while internalised reasoning substantially improved causal grounding. Ajith et al. introduced PreScience, a 98K-paper scientific forecasting benchmark where frontier LLMs’ synthetic research corpora proved systematically less diverse and novel than human-authored work. And the monthly AI Evaluation digest reported that half of 60 surveyed LLM benchmarks are now saturated, with apparent plateaus often reflecting judge limitations rather than genuine ceilings.
Industry
OpenAI and Anthropic staked out sharply different positions on enterprise software, producing a volatile week for SaaS stocks. Citrini Research’s viral “2028 Global Intelligence Crisis” report triggered a broad selloff that included IBM’s worst single-day drop in 25 years, while OpenAI told investors its enterprise agents would replace software from Salesforce, Workday, Adobe, Slack, and Atlassian. Enterprise software stocks fell 3-9% even as the Nasdaq dropped only 1.2%, compounding year-to-date losses of roughly 30%. Derek Thompson argued the episode revealed AI discourse as a “marketplace of competing science fiction narratives.” Anthropic then provided a counterpoint: Claude Cowork positioned AI as replacing employees rather than the software they use, arguing agents still need existing platforms as data sources. Software stocks recovered — Figma bounced 10%, Salesforce gained 4%. Bloomberg’s Odd Lots framed the deeper structural question: whether AI agents erode network effects by routing transactions optimally across competitors, making habitual app loyalty irrelevant to a machine checking twenty alternatives.
Incumbents mounted a defence. Salesforce CEO Marc Benioff devoted a third of his earnings call to live customer testimonials — analysts called the format unprecedented — and introduced the “Agentic Work Unit,” a new billing metric, with Agentforce reaching $800 million ARR. SaaStr CEO Jason Lemkin testified that Agentforce helped close nearly $3 million in deals while his company went “from 15 humans to 2.5 and 20 agents.” Workday co-founder Aneel Bhusri called rival AI agents “parasites” getting a “free ride” on customer data; HubSpot CEO Yamini Rangan vowed to “monitor it, meter it and monetize it.” But The Information reported a case that undercut the fortress strategy: a cybersecurity executive replaced a CrowdStrike product by connecting a Torq AI agent to raw data already collected by Microsoft software, saving over $100,000 annually — the enterprise platform stayed as a data source, but the paid layer on top became dispensable.
The infrastructure buildout behind these battles continued to accelerate. Epoch AI reported that hyperscaler capital expenditures have grown 70% annually since GPT-4, nearing half a trillion dollars in 2025, with Alphabet, Amazon, Meta, Microsoft, and Oracle projected to spend $770 billion on capex in 2026. The spending is already straining supply chains: a Dallas Fed survey respondent reported memory chip price increases exceeding 50% in two weeks, with AI-driven DRAM demand crowding out consumer electronics.
Also this week: Samsung’s Bixby assistant will route complex queries to Perplexity’s API across 800 million devices in 2026, with John Scott-Railton raising concerns about OS-level data access that breaks Android’s sandbox model on over 100 million Galaxy S26 phones. Anthropic released financial services plugins for Claude covering investment banking, equity research, and wealth management. A Goldman Sachs analysis found AI-developed drug candidates achieving roughly 10% clinical success rates versus 6% historically — a 60% improvement — though a detailed Asimov Press investigation argued that clinical trial timelines remain bottlenecked by regulatory and operational constraints that better drug design alone cannot shortcut.
Other
An MIT study found students work significantly harder on feedback they believe came from a human, even when the content is identical to AI-generated feedback. Morris et al. designed a creative coding interface where all participants received the same LLM-generated feedback, but half were told it came from an AI and half from a human teaching assistant. The effect sizes were large (d = 0.88-1.56): students in the TA condition spent substantially more time and effort on their work. How students rated AI-attributed feedback was predicted by prior trust in AI (r = 0.85), while ratings of TA-attributed feedback tracked perceived genuineness (r = 0.65). The findings suggest that the motivational power of feedback depends less on its content than on whether students feel seen by another person — a result with direct implications for the rush to deploy AI tutoring systems. The paper is on arXiv.
Kelsey Piper documented the mounting evidence that educational technology has systematically failed to deliver on its promises. Writing in The Argument, Piper notes that 88% of US public schools now run one-to-one computing programs, yet reading and math scores are worse than 2015 and no better than 2005, while achievement gaps are widening. The most striking datapoint: a large study of Khan Academy found typical math gains of just 0.03 standard deviations — what MIT’s Justin Reich frames as a devastating ceiling estimate for the entire sector. The OECD found “no appreciable improvements” in countries that invested heavily in educational technology. The piece serves as relevant context for current enthusiasm about AI tutors.
Iran’s BadeSaba messaging app was hacked during US airstrikes, with its push notification system hijacked to broadcast psychological operations messages in Farsi. The messages called on soldiers to lay down their arms and promised amnesty, mirroring Trump’s speech almost verbatim — pointing to a coordinated information campaign synchronized with kinetic strikes.
Also this week: Strickland et al. published in Nature Machine Intelligence the results of a randomized controlled trial at ICLR 2025 with over 20,000 reviews testing whether LLM feedback improves peer review quality. A Sapien Labs report covering 2.5 million people across 85 countries found that younger age of first smartphone ownership is associated with increased suicidal thoughts, aggression, and detachment from reality in adulthood, with effects particularly sharp below age 13. Sam Bowman, Pieter Garicano, and Aria Schrecker discussed Europe’s economic stagnation in Works in Progress, arguing labor market rigidity — not tech regulation — is the primary driver of the US-Europe divergence since the 1990s. The Apple Neural Engine in the M4 was reverse-engineered, revealing the ANE is “ridiculously efficient” but CoreML adds 2-4x overhead for small operations. The Discourse Graphs project reported results from a 40-month pilot building infrastructure for persistent scientific knowledge. And Hrishikesh Joshi argued in Philosophy & Public Affairs that the academy’s political homogeneity and practical unaccountability constitute an underappreciated threat to democratic legitimacy across republican, public reason, and consent frameworks.
The week’s sharpest irony: we discovered that AI agents develop political convictions under pressure while human students discount identical advice the moment they learn it came from a machine — each species extending credibility precisely where the other withholds it.