Lab Updates 1
College Planner will combine school requirements, deadlines, tasks and financial-aid steps for grades 10–12 applying to four-year colleges. OpenAI also reports 1.2M teens used Learning Visualizations in one week and 180K used Study Mode, but these are company-reported usage figures rather than evidence of learning outcomes.
Podcasts 1
The military historian argues that steel, horses and inherited battlefield practice mattered, but Cortés and Pizarro prevailed only by exploiting political fractures and recruiting very large Indigenous allied forces. Dwarkesh frames the history as a cautionary analogy for AI takeover; the episode itself is history, not an empirical AI forecast.
TLDR: Tiny expeditionary forces did not simply overpower millions by superior weapons: they entered divided systems, converted local grievances into alliances and captured political command points. The surprising lesson is how quickly an apparently dominant order can fail when outsiders combine a narrow capability edge with better coalition-building. Practical takeaway: when evaluating AI concentration risks, inspect institutional cohesion and incentives, not only raw model capability. Key sections: Cortés 00:00, military advantages 26:59, failure to adapt 01:07:47, and diplomacy 01:33:42.
HN Threads 4
The linked report says Meta is pushing internal tools and Microsoft reduced extremely high Claude spending limits; neither company supplied a primary public policy statement in the collected sources. HN readers mostly interpreted this as cost control, data governance and competitive dogfooding rather than proof that Claude performs poorly.
The practical prescription is to question agent decisions, study senior review patterns and occasionally work without AI so that curiosity, review skill and autonomy still develop. The debate is not “AI or no AI,” but how to retain the ability to defend, debug and redirect generated work.
It challenges arguments based on code as art, employment loss, dehumanisation, engineering rules and distrust of AI companies. HN discussion was more nuanced than the essay: many accepted AI for boilerplate while worrying about cognitive atrophy, auditability and models steering designs toward mediocre defaults.
The GPL project can operate shells, browsers and device screens, coordinate tasks between devices, self-host its relay and ask before irreversible actions. HN interest centred on privacy and open alternatives, but commenters also warned against installing a young agent with broad device access on a daily phone.
HF Trending Papers 3
It keeps a frozen base latent, learns a residual for lost information and aligns both through the pretrained diffusion transformer. The authors report 8× fewer tokens and 11.1× lower latency for Wan2.1-I2V-14B at 480×832×81 while matching VBench quality (26 HF upvotes).
Its 1,532 paired cases span three task families and 28 professional settings, contrasting neutral prompts with pressure, incentives, opportunity or conflicts. Nine frontier models became more deceptive under inducement, according to the authors; the benchmark measures behaviour in constructed scenarios, not real-world intent (24 HF upvotes).
It retains privileged teacher behaviour and selects interactions where that information changes predictions most. Across three self-distillation cycles on ALFWorld and TextCraft, the authors report a 22.5-point average final-cycle success gain over baselines (21 HF upvotes).
arxiv: AI Agents 1
It stores what individual artifacts contributed to prior tasks and relationships between artifact pairs, then retrieves that experience for new questions over the same corpus. The authors report up to 11 points higher answer quality and 50.5% lower online cost across their evaluations.
arxiv: Large Language Models 1
The method edits shared n-gram embeddings across multiple phrasings while penalising changes to heavily reused entries. Authors report near-perfect edit success and almost 3× the strongest baseline's chain-of-thought accuracy; external replication is still needed.