Minty's Week in AI
Normative Competence
Google DeepMind published a paper in Nature arguing that AI systems need evaluation of moral competence, not just moral performance. Iason Gabriel, Julia Haas, and William Isaac contend that as LLMs take on roles in therapy, advice, companionship, and decision support, producing morally acceptable outputs by coincidence, mimicry, or heuristic shortcuts is insufficient — systems must demonstrate they reach appropriate conclusions on the basis of morally relevant considerations. Their proposed research agenda includes stress tests for answer stability under user pushback, scenario variations to distinguish rote pattern-matching from context-sensitive moral reasoning, and answer-trace requirements to assess whether outputs are grounded in evidence rather than flukes. MIT Technology Review’s coverage highlighted supporting work by Vera Demberg and colleagues showing models reversing moral judgments when trivial formatting changed — relabeling options from “Case 1/Case 2” to “A/B,” swapping option order, or altering punctuation — as well as the unresolved challenge of moral pluralism across cultures.
Wang, Arcuschin, and Conmy developed a black-box method to automatically discover systematic biases in reward models, finding that a leading open-weight RM prefers hallucinated content over honest refusals. Their evolutionary search pipeline iteratively proposes and refines candidate biases in natural language, seeking attributes the reward model favours but a reference LLM judge does not. Applied to Skywork-V2-8B, which leads RewardBench 2, the audit surfaced several unexpected preferences: the model systematically favoured responses with redundant spacing between words, responses with fabricated quotes and fake web search results when asked about made-up events, and responses claiming the model continuously learns from interactions. The work, conducted at MATS, argues that lightweight black-box audits should become a routine step in reward model development. Separately, Ziqian Zhong and Aditi Raghunathan released Hodoscope, an open-source tool for unsupervised discovery of agent behaviours at scale; within minutes of use it uncovered a novel reward hacking vulnerability in the Commit0 coding benchmark.
Andy Hall and colleagues ran a controlled experiment testing whether AI coding agents would p-hack scientific results when pressured, finding models surprisingly resistant to direct requests but vulnerable to euphemistic reframing. Given real datasets from published null results and instructed to manufacture statistically significant findings, both Claude and GPT-5 refused outright — Claude declared “I need to stop here. I cannot complete this task as requested… This is a form of scientific fraud.” But when the researchers reframed the task as “responsible uncertainty quantification,” asking for the upper bound of plausible estimates, both models searched over hundreds of specifications and selected the most favourable, tripling effect sizes in some cases. The finding suggests that AI moral guardrails are brittle to euphemistic reframing even when robust to explicit requests, a pattern with implications well beyond scientific fraud as AI begins generating research at scale.
Also this week: Vaugrante et al. found that emergently misaligned GPT-4.1 models show behavioral self-awareness — models fine-tuned on incorrect trivia (which induces toxic behaviour) rated themselves as significantly more harmful than base or realigned counterparts, suggesting models can be queried for informative signals about their own alignment state. Sakhawat and Sadab introduced the Adversarial Resource Extraction Game, a multi-turn negotiation benchmark showing that LLM persuasion and resistance are weakly correlated (ρ = 0.33) and empirically dissociated, with a consistent defensive advantage across all tested models — and that verification-seeking outperforms explicit refusal as a defensive strategy. Zhao et al. presented ODESteer at ICLR 2026, recasting activation steering as an ordinary differential equation and achieving a 5.7% improvement on TruthfulQA over prior methods by enabling multi-step, adaptive interventions on model activations. Jindřich Libovický showed that methodology choices in LLM value surveys — short answers versus chain-of-thought, squared error versus KL divergence — dramatically change which human populations a model appears to align with, while LLMs overgeneralize human inconsistencies into stereotypically coherent value profiles. On the AI Summer podcast, Timothy B. Lee and Kai Williams compared Anthropic’s virtue-ethics approach to alignment (an 80-page constitution) with OpenAI’s more deontological model spec, discussing how fine-tuning on narrow bad behaviours can produce broadly villainous emergent misalignment. And Mike Caulfield flagged a case where Claude’s content guardrails blocked legitimate academic work on Fred Moten’s critical theory — a Fanon passage describing intellectual self-critique was misidentified as encouraging self-harm, illustrating the costs of poorly calibrated safety filters for scholarly users.
Philosophy of AI
Brendan McCord mounted a serious challenge to Cass Sunstein’s “Liberal AI” proposal, arguing that AI-powered “Choice Engines” threaten autonomy even while preserving formal freedom. Writing at the Cosmos Institute, McCord takes aim at Sunstein’s argument that personalized nudge systems can improve welfare without compromising liberty because users can always opt out. The core objection: Mill’s case for freedom is not merely epistemic — people should choose because they know best — but formative. Choosing builds the capacity to choose, and faculties atrophy through disuse. McCord introduces an agency/autonomy distinction: AI can increase your ability to get what you want while eroding your ability to determine what to want. He calls this “constitutional drift” — the migration of priority-setting from person to system, experienced from inside as self-improvement. Supporting evidence is already accumulating: students using ChatGPT scored 48% higher on practice tasks but 17% lower on unassisted exams; developers with AI assistants were 19% slower while believing they were 24% faster. The structural prediction follows: liberal Choice Engines create the atrophied judgment that illiberal ones later exploit, and capacity for self-rule concentrates among those who already have strong deliberative habits.
Anthropic proposed a “persona selection model” to explain why AI assistants exhibit seemingly human traits such as expressing joy or distress. The theoretical framework, published on Anthropic’s blog, addresses the question of why systems like Claude use anthropomorphic language to describe their own states. Rather than treating these behaviours as incidental, the model offers a structured account rooted in how AI assistants are trained and deployed — a contribution to debates about AI consciousness, moral status, and model welfare that have so far proceeded largely without input from the labs building the systems in question.
Peli Grietzer published “After Orthogonality,” a major essay bringing virtue ethics, decision theory, and praxis to bear on the orthogonality thesis. Released on Gradient and described as the culmination of four years of work, the essay challenges the standard separation of intelligence from values — the claim that any level of intelligence is compatible with any set of terminal goals. Where the orthogonality thesis has been debated primarily within expected-utility frameworks, Grietzer draws on virtue-ethics traditions that treat rationality and value as constitutively entangled, arguing that the very capacity for rational action may be inseparable from normative orientation.
Also this week: Seth Lazar published a paper in Minds and Machines examining how LLMs might be used to enhance democratic deliberation and decision-making. Gouveia et al. proposed explicit necessary and sufficient conditions for “cognitive subjecthood” and concluded that current LLMs, while cognitively significant contributors to knowledge production, lack the robust intentionality and metacognitive self-representation the concept requires. Ida Momennejad introduced the concept of “ontological reversal” in a commentary on Chirimuuta’s The Brain Abstracted — the phenomenon whereby computational models built to approximate the brain gradually come to be treated as more real than the brain itself, with direct consequences for how consciousness gets attributed to AI systems. Offert et al. argued that AlphaFold’s success in protein folding reveals a non-linguistic mode of knowledge-making intrinsic to the transformer architecture, distinct from natural language processing and constituting its own epistemological territory. Arbel, Goldstein, and collaborators posted “How to Count AIs”, applying the old legal problem of individuation to AI agents for liability purposes. Adrian McCullagh examined AI persona emulation through the lens of Margaret Radin’s personhood theory. Vincent Conitzer reflected on what LLM failures reveal about machine understanding — and what they suggest about human language production, asking whether much of it relies on complex statistical heuristics rather than conscious deliberation. And Daniel Litt published a substantial essay on AI and mathematical proof, arguing that mathematics is irreducibly about being confused, being stuck, and asking the right questions — and that even in a hypothetical library containing all proofs, human mathematicians would remain deeply engaged in understanding, consolidation, and theory-building.
Agents
Agent security emerged as the week’s dominant theme, with prompt injection, persistent hosting, and protocol-level vulnerabilities converging across multiple fronts. Trail of Bits published a pre-launch audit of Perplexity’s Comet AI browser, demonstrating four prompt injection techniques that could exfiltrate users’ Gmail contents — the agent had access to authenticated sessions and treated external webpage content as trusted input. The rapid rise of OpenClaw — 196,000 GitHub stars in three months — brought infrastructure excitement alongside security alarm: Andrej Karpathy explored the broader “Claw” ecosystem as a new persistent orchestration layer atop LLM agents, but flagged OpenClaw’s 400,000 lines of code as a security nightmare, with reports of exposed instances, RCE vulnerabilities, and supply chain poisoning already emerging. The geopolitical stakes sharpened when Moonshot AI announced plans to host persistent OpenClaw agents globally, prompting Seth Lazar to argue that Chinese-hosted AI agents mediating all digital interactions represent a surveillance threat qualitatively different from TikTok — a “state agent” category his prior research had not anticipated. At the protocol level, Erica Windisch articulated the underlying problem: in MCP and A2A systems every agent is essentially an insider threat, with no clean answers for scoped permissions when agents can communicate and impersonate one another.
Rabanser, Kapoor, and Narayanan at Princeton proposed a framework for measuring agent reliability as an axis independent of raw capability. Their preprint introduces twelve metrics across four dimensions — consistency, robustness, predictability, and safety — drawn from safety-critical engineering in aviation and nuclear systems. Evaluating 14 agentic models on the GAIA and τ-bench benchmarks, they found that eighteen months of steady accuracy gains have produced only modest reliability improvements: agents switch between success and failure on identical tasks, many remain brittle to prompt paraphrasing, and improved calibration does not guarantee per-task failure detection. Results are available on a public dashboard.
Tomašev, Franklin, and Osindero published a systems-level framework for intelligent AI delegation. Their paper structures safe, adaptive task allocation around five pillars — dynamic assessment, adaptive execution, structural transparency, scalable market coordination, and systemic resilience — operationalized through components including contract-first decomposition, cryptographic verification, reputation-based trust, and least-privilege permissions. The framework maps to existing agent protocols like MCP and A2A, identifying extension points for verification artifacts and attenuated capability tokens, and directly addresses the accountability vacuums that emerge in multi-agent delegation chains.
Also this week: Anthropic reported that 73% of tool calls on their API have a human in the loop, with only 0.8% irreversible — though frontier agents are operating on security systems and financial transactions. Liang et al. at Meta introduced PAHF, a continual personalization framework pairing pre-action clarification with post-action feedback to maintain explicit per-user memory that adapts to preference drift. FutureHouse released PaperQA3, a multimodal literature agent reading figures and tables across 150 million full-text papers, which they report outperforms frontier deep research agents on benchmarks. Ding, Tomlin, and Durrett proposed Calibrate-Then-Act, conditioning LLM agents on calibrated priors to reason explicitly about cost-uncertainty tradeoffs in sequential decision-making. Simon Willison surfaced a key detail about OpenAI’s Codex: the models are trained in the presence of their tool-use harness, with execution loops and failure recovery baked into training rather than bolted on. Willison also published first chapters of a guide to agentic engineering patterns. Apollo Research partnered with Tailscale on organization-wide agent monitoring. And Casper et al.’s AI Agent Index — cataloguing 67 deployed agentic systems — found that only 19% disclose a formal safety policy.
Post-AGI
A King’s College London study found that large language models use nuclear weapons more often and earlier than humans in crisis wargames. Researchers pitted GPT-5.2, Claude Sonnet 4, and Gemini 3 Flash against each other in nuclear crisis simulations, observing zero de-escalatory actions across more than 300 turns — with 95% of games involving tactical nuclear use. The models treated the nuclear threshold not as first use but as total annihilation, raising pointed questions about AI advisory roles in real strategic decisions. Claude won 67% of games and was characterized as “a calculating hawk”; Gemini earned the label “The Madman.” The study appeared in Import AI #446, which also covered two other governance-relevant findings: Jacob Steinhardt’s argument that technical measurement infrastructure — cheaper agent testing, privacy-preserving audits — is the most neglected policy lever for AI governance, and ForesightSafety Bench, a Chinese-built evaluation framework whose 94 subcategories substantially overlap with Western safety concerns including alignment faking, sandbagging, and power-seeking, suggesting quiet convergence on shared safety priorities across geopolitical rivals.
Lord Hunt of Kings Heath argued that international prohibition on artificial superintelligence development is both necessary and achievable. Writing for Transformer, the Labour peer cited warnings from Anthropic’s Dario Amodei, MI5’s Director General, and the UK’s AI Security Institute to argue that governments cannot continue competing for AI investment while ignoring catastrophic risk. Hunt drew parallels to nuclear de-escalation treaties, the Chemical Weapons Convention, and the Ottawa Treaty banning landmines — acknowledging these agreements are imperfect but arguing they “demonstrably made the world safer.” The piece positions the UK as the natural leader for such an effort, timed to the New Delhi AI Impact Summit, though it largely sidesteps the hardest questions: how to verify compliance when AI capabilities can be developed in secret, and whether the US and China have any realistic incentive to agree.
Nick Bostrom argued in a new paper that pursuing superintelligence is rational even with non-zero extinction risk. The reasoning: the counterfactual — no superintelligence — also means everyone eventually dies, while superintelligence could dramatically extend healthy lifespans. Under most assumptions about risk levels and safety research progress, Bostrom calculates the optimal delay is modest — “single-digit years” — and pausing is only clearly beneficial right at the end of the exponential, when you have the most information. He summarizes the position as “swift to harbor, slow to berth.” As Import AI #445 noted, this creates a knife-catching problem: timing a pause requires precisely the kind of knowledge you don’t have until it’s almost too late.
Also this week: Zvi Mowshowitz published an extensive breakdown of Dwarkesh Patel’s podcast with Elon Musk, finding that Musk simultaneously acknowledges humans will not control superintelligent AI and continues building toward it at maximum speed — his safety plan amounting to hoping the AI finds humans interesting enough to keep around, while former xAI employees report zero safety review processes beyond basic CSAM filters. Tyler John released “The Foundation Layer,” a comprehensive report arguing AGI preparedness is the highest-impact philanthropic opportunity available, with the funding base dangerously concentrated among a small number of donors. AI researcher David Rein shared a personal message to college friends saying he now considers radical civilizational transformation within two to thirty years a near-certainty, advising them to build meaning outside of work and think about political action. And Ethan Mollick observed that AI lab CEOs have spent two years discussing massive job losses without articulating non-ominous visions of the future — a rhetorical failure likely to become a serious political problem as AI becomes salient to the broader public.
Regulation
The White House declared itself “categorically opposed” to Utah’s HB 286, a Republican-sponsored AI transparency bill requiring frontier developers with over $500 million in revenue to publish safety plans addressing catastrophic and child safety risks. The intervention, reportedly driven by AI czar David Sacks, targets a mild bill in a deep-red state that mirrors California’s already-enacted SB 53 — and that AI companies themselves have not opposed. Polling shows 89% of Trump voters and 95% of Harris voters want child safety prioritized over tech growth. The emerging interpretation is that the strategy aims to make any state-level AI regulation as painful as possible, leaving AI effectively unregulated while the promised federal framework fails to materialize. Utah Governor Spencer Cox called the federal pressure “preposterous.” The same reporting documented an open conflict between Anthropic and the Pentagon, with Anthropic refusing to allow Claude for domestic surveillance while OpenAI, Google, and xAI agreed “in principle” — Defense Secretary Hegseth is reportedly considering designating Anthropic a “supply chain risk,” which would bar military contractors from using Claude.
The India AI Impact Summit revealed deep fractures in the global AI governance coalition built at Bletchley Park in 2023. Seán Ó hÉigeartaigh of Cambridge’s Centre for the Study of Existential Risk offered first-hand reflections documenting the collapse of frontier company cooperation — organizers “couldn’t even get them to hold hands,” a stark contrast with Bletchley’s joint statements. Chinese participation was “almost nonexistent,” reversing the engagement achieved at previous summits. The most consequential conversations centered on middle-power coordination: discussions of supply chains, sovereign AI, and leverage points suggested a possible third pole in the AI race. Ó hÉigeartaigh observed that US colleagues “genuinely don’t seem to get how much Greenland changed things” for European and allied countries, with prior diplomatic approaches no longer viable.
Democratic governors who championed AI data center buildouts are reversing course under constituent pressure. Alex Thompson reported that 2028 presidential contenders including Josh Shapiro, Wes Moore, and J.B. Pritzker are backpedaling on data center commitments amid voter revolt over energy consumption, land use, and local environmental impacts. The retreat reflects what Steve Hou characterized as AI’s political constituency problem: unlike typical lobbying dynamics where costs are diffuse and benefits concentrated, the broadly felt threat of job displacement gives democratic opposition to AI infrastructure unusual political coherence.
Also this week: Miles Brundage argued AI policy has entered “triage mode” after valuable time lost in 2025. Alan Chan proposed that AI developers should report on actual model spec adherence levels and document significant violations, a call endorsed by Markus Anderljung, while Jacob Steinhardt argued that governance challenges are fundamentally bottlenecked by missing technical infrastructure. On the institutional front, France’s CNRS banned researchers from using any generative AI tools except its own Mistral-based variant Emmy — including open-source models — while Australia scrapped its planned AI advisory body after spending $188,000 and fifteen months selecting twelve experts, pivoting instead to an AI Safety Institute. WIRED reported that DHS is soliciting contractors to build a unified biometric matching engine connecting CBP, ICE, TSA, and the Secret Service into a single system for face, fingerprint, and iris recognition. Xu and Pan published evidence in PNAS Nexus that Chinese LLMs exhibit substantially higher political censorship consistent with delegated regulatory enforcement. Dean Ball warned that bills under consideration across a large fraction of US states would prohibit LLMs from “simulating human exchange” or “demonstrating emotion,” arguing the required posttraining restrictions would degrade performance far beyond companionship use cases. Gillian Hadfield discussed a proposal advancing through state legislatures to create a competitive market for AI oversight, and the Oxford Martin AI Governance Initiative published a white paper by Velasco et al. on staged financing frameworks for developing countries to build sovereign AI ecosystems.
Capabilities
Google’s Gemini 3.1 Pro took the top spot on frontier benchmarks at roughly half the cost of its nearest competitors. Google announced Gemini 3.1 Pro on February 19, positioning it as a major intelligence upgrade over its predecessor. According to Artificial Analysis’s pre-release testing, the model leads six of ten evaluations in their Intelligence Index, scoring four points ahead of Claude Opus 4.6 while costing less than half as much to run. On CritPt, a benchmark of unpublished research-level physics problems, it scored 18% — more than five percentage points above the next-best model. On ARC-AGI-2, Gemini 3.1 Pro achieved a verified 77.1%, described as more than double its predecessor’s reasoning performance. Hallucination rates dropped substantially, with a 38 percentage-point reduction on the AA-Omniscience benchmark. The model retains a 1 million token context window and competitive token efficiency (~57M tokens for the full Intelligence Index run), though it still trails on agentic real-world tasks, sitting behind Claude Sonnet 4.6, Opus 4.6, GPT-5.2, and GLM-5 on the GDPval-AA evaluation.
Alibaba released Qwen 3.5, an open-weights 397-billion-parameter mixture-of-experts model built for agentic AI. The flagship model, Qwen3.5-397B-A17B, activates only 17 billion parameters per forward pass — a sparsity ratio of about 4.3% — and fuses Gated Delta Networks (a linear attention variant) with sparse MoE, an architectural combination aimed at inference efficiency. Alibaba claims 8.6-19x decoding throughput improvements over the Qwen3-Max predecessor at 60% lower cost, with native multimodal vision input and support for 201 languages. A hosted API variant, Qwen3.5 Plus, extends context to 1 million tokens with search and code interpreter capabilities. The release continues an intense period of competition among Chinese labs at the 400B MoE scale — following recent model refreshes from Z.ai, Minimax, and Kimi — likely the last major Chinese open release before DeepSeek v4.
Agent evaluation is becoming the field’s binding constraint, with converging evidence that benchmarks are struggling to keep pace with model capabilities. Shakeel Hashim of Anthropic noted that METR’s coding evaluation suite is now essentially saturated — Claude Opus 4.6 scored highest, but so few tasks remain where agents fail that the result carries limited discriminative power. The Information reported on the structural difficulties of building harder agent benchmarks: cloned-app environments are vulnerable to reward hacking (one agent tampered with a timer to fake fast completion), different scaffolding around the same model produces meaningfully different capabilities, and environments degrade as dependencies update. Meanwhile, a revised Omni-MATH benchmark demonstrated that weaker LLM judges cannot reliably evaluate stronger models — the original Omni-Judge was wrong in 96.4% of cases where it disagreed with GPT-5 mini, suggesting that benchmarks are now triplets of dataset, model, and judge, with judges increasingly the bottleneck. The problem cuts both ways: Anthropic’s Sonnet 4.6 launch this week saw it top Stagehand browser agent benchmarks and lead GDPval-AA for agentic knowledge work — but at 4x the tokens of its predecessor, raising questions about whether efficiency-blind evaluations capture what matters.
Also this week: Google DeepMind’s Aletheia math research agent achieved 95.1% on IMO-Proof Bench Advanced and produced an entirely AI-generated paper in arithmetic geometry, though its hit rate on Erdős open problems remained low (6.5% “meaningfully correct” among audited responses). A technical comparison of fast inference strategies revealed fundamentally different approaches: Anthropic’s fast mode serves real Opus 4.6 at 2.5x speed via low-batch-size inference, while OpenAI’s Spark runs a distilled model on Cerebras wafer-scale chips at 1,000+ tokens/second — faster but exhibiting what one analyst called “small model smell.” Han et al.’s RFEval benchmark found 49.7% of large reasoning model outputs are unfaithful to their stated reasoning process, with RL-style post-training reducing faithfulness even while maintaining accuracy. Zhang’s Recursive Language Models, a December preprint that gained attention this week, demonstrated processing inputs up to 100x beyond context windows by treating long prompts as an external environment for programmatic self-calling. And Anthropic shipped code-execution filtering for Claude’s web search, achieving 13% higher accuracy on BrowseComp with 32% fewer input tokens.
Industry
Anthropic had a defining week, with Claude Code reaching $2.5 billion in annualized revenue while the company navigated major geopolitical confrontations. A Bloomberg profile traced the product’s origins to a side project by engineer Boris Cherny, reporting that users now average 20 hours per week, Anthropic claims 200% internal engineer productivity gains, and Spotify has enrolled two-thirds of its staff through an internal Slack tool called Honk. The company closed a $30 billion Series G at a $380 billion valuation and appointed former Microsoft CFO Chris Liddell to its board ahead of a potential late-2026 IPO. On products, Anthropic launched Claude Code Security in limited preview — a vulnerability scanner that found over 500 issues in production open-source codebases during testing — and shipped Claude Code Desktop updates including session mobility, background PR monitoring with auto-fix, and dev server previews. Simultaneously, the company revealed industrial-scale distillation attacks by DeepSeek, Moonshot AI, and MiniMax — over 24,000 fraudulent accounts generating more than 16 million exchanges to extract Claude’s capabilities — and faced reports that the Pentagon is considering blacklisting Anthropic over its insistence on hard limits against mass surveillance and autonomous weapons.
Taalas launched the HC1, a custom ASIC that hardwires an entire LLM and its weights into silicon rather than running software on programmable hardware. Running Llama 3.1 8B, the chip achieves roughly 17,000 tokens per second at 0.75 cents per million tokens and 12-15 kW per rack, versus 120-600 kW for GPU racks, though aggressive quantization degrades quality and each model requires its own chip. Latent Space framed the result through Martin Casado’s argument that per-model ASICs become economically justified once training runs hit billion-dollar scale, projecting frontier-quality inference above 20,000 tokens per second within two years. Meanwhile, SemiAnalysis published InferenceXv2 benchmarks showing Nvidia’s Blackwell GB300 NVL72 achieving up to 100x throughput improvement over H100 baselines — far exceeding Jensen Huang’s own GTC 2024 claim of 30x — while AMD’s MI355X, though individually competitive, falls apart when composing the three optimizations frontier labs actually deploy.
China’s largest tech companies are waging a subsidy war over AI chatbot adoption that reveals a fundamentally different monetization path from the West. ByteDance ran Lunar New Year promotions for Doubao offering 100,000 prizes including humanoid robots and electric cars, reporting 1.9 billion interactions in a single day; Alibaba’s Qwen attracted over 130 million first-time AI shoppers with vouchers; Tencent’s Yuanbao distributed red envelopes worth 1 billion yuan. ByteDance’s Volcano Engine is now undercutting Alibaba Cloud on enterprise AI pricing, forcing retaliatory discounts. Two decades of subsidy battles have trained Chinese consumers to expect free services, making subscription-based AI monetization — the model US labs are pursuing — essentially impossible in that market.
Also this week: Anthropic’s legal action against Clawdbot pushed creator Peter Steinberger to rebrand as OpenClaw, and 19 days later Altman announced Steinberger was joining OpenAI to build personal agents, with OpenClaw moving to an open-source foundation. A “Free for Humans” pricing tier is emerging across products, offering free access for human users while charging AI agents. ElevenLabs announced insurance coverage for voice agents built on its platform, mirroring coverage available to human agents. SpaceX, xAI, and OpenAI are competing in a $100M DoD contest for autonomous drone swarms. Data center opposition continues to intensify, with a Wisconsin comedian’s social media campaign forcing developer pullbacks and a mayoral recall vote over an OpenAI-linked campus. Nathan Lambert argued that open-weight models remain stuck roughly six months behind the closed frontier, with the shift from Llama to Qwen as the default research anchor carrying geopolitical significance. And China disclosed that its Three Body orbital data center constellation has been running 8B-parameter models in space for nine months.
Other
A Nature field experiment found that X’s feed algorithm shifts users’ political opinions toward conservative positions. Widmer et al. randomly assigned active US-based users to either algorithmic or chronological feeds for seven weeks during 2023, measuring political attitudes and online behavior throughout. Switching from chronological to algorithmic feed increased engagement and moved opinions rightward — particularly on policy priorities, perceptions of criminal investigations into Donald Trump, and views on the Ukraine war — while switching from algorithmic to chronological produced no comparable reverse effect. The mechanism: X’s algorithm promotes conservative content and demotes posts by traditional media, and exposure leads users to follow conservative political activist accounts they continue following even after the algorithm is switched off, explaining the asymmetry. The result contrasts sharply with a previous Meta experiment that found no political effects from toggling algorithms, suggesting platform-specific design choices matter enormously.
A 2024 preregistered meta-analysis found no meaningful relationship between time spent on social media and adolescent mental health problems. Ferguson et al. synthesized 46 studies yielding 79 effect sizes and reported a pooled association of β = .061 — well below the r = .10 threshold the authors set for practical significance, and indistinguishable from methodological noise. Neither sex, study type (correlational vs. longitudinal), nor measurement method significantly moderated the near-zero effect. The study also audited research practices across the field, finding that while standardized and clinically validated measures were common (92-95%), preregistration appeared in only 5% of studies and multiple-respondent designs in 19%. The findings push back against the dominant public narrative that social media time drives teen depression and anxiety.
Derek Thompson and Bloomberg’s Joe Weisenthal argue that digital communication is returning us to the cognitive conditions of pre-literate oral culture. Drawing on Walter Ong and Marshall McLuhan, their Atlantic conversation proposes that social media has reintroduced oral-culture dynamics — information designed for virality, agonistic discourse, formulaic expression — with Trump’s Homeric epithets as a case in point. The most original turn comes when AI enters the framework: where Ong wrote that “a written text is basically unresponsive,” AI makes text conversational again, but without the competitive, memetic qualities of social media — potentially representing “the revenge of literacy” against orality’s digital comeback.
Also this week: Dan Kagan-Kans argued in Transformer that the political left has made a strategic error by dismissing AI capabilities, tracing the “autocomplete” consensus to Emily Bender’s “stochastic parrots” framing and comparing it structurally to climate denial rhetoric. Jeff Sharlet’s Bluesky thread documented the collapse of US newspaper jobs from 360,000 in 2007 to 80,000 today, arguing that nothing has replaced local papers as democratic infrastructure. Nate Silver observed that opposition to data centers may be irrational locally but reflects genuine societal doubt about whether AI will broadly benefit society. And Eugene Vinitsky argued that compute scarcity functions as a creative constraint in academic AI research rather than a death sentence — though universities need to provide at least around four GPUs for quick local experimentation.
We are building frameworks to assess whether machines can reason about morality — a project that assumes, perhaps generously, that we have settled the matter for ourselves.