The firewall separating consumer AI from credentialed expertise in healthcare and finance is fracturing, as OpenAI’s ChatGPT Health integrates electronic health records via b.well partnerships alongside Apple Health, MyFitnessPal, and Peloton for personalized explanations of lab results, doctor-visit prep, insurance comparisons, and longitudinal pattern detection—all siloed from general chats to assuage privacy fears. This caps two years of collaboration with 260+ physicians, targeting the 230 million weekly health queries (5% of global prompts, 40+ million daily users) amid surging dissatisfaction with ballooning costs, while JP Morgan ditches external proxy advisors entirely for Proxy IQ, an in-house platform parsing annual meetings to guide portfolio decisions. OpenAI’s bold thrust—echoed in week-long ads depicting AI aiding eczema checks, workout squeezes, and condition breakdowns—signals commoditization of "health ally" roles, but US-only records and iOS tethering expose rollout frictions; simultaneously, Elon Musk pegs AI at automating 50% of white-collar tasks today (offices before factories), compressing timelines to "all work" in five years per his unfiltered calculus.

Yet this vertical push amplifies tensions: while empowering patients (e.g., trend-spotting in sleep/activity), it risks eroding leverage in labor negotiations as cognition/empathy moats evaporate, per David Shapiro’s thesis that work’s moralization masks its role as systemic bargaining chip—necessitating new levers like decentralized sentiment boycotts.

ChatGPT’s dominance eroded 22 points in 12 months (86.7% to 64.5%) as Gemini quadrupled to 21.5% via Android ubiquity and app integrations (YouTube summaries proving sticky for users shifting 40% from ChatGPT), while Grok tripled to 3.4% on X’s 250M daily users fueling xAI’s "maximum truth-seeking" ethos—capped by a $20B raise vaulting it past Anthropic as second-most-funded lab. xAI’s unique substrate (real-time X data, tent-pitching culture, Tesla rides post-36-hour shifts) prioritizes attention over raw params, contrasting Google’s arsenal (TPUs since 2015, Transformer invention, DeepMind/$400M buy, $2.7B Noam Shazeer repurchase) that revived via Sergey Brin’s return and Gemini’s cross-app shove. This velocity—OpenAI hemorrhaging 20% share yearly—underscores distribution as the new moat, with Elon conceding "within 3 years all knowledge work WILL be done by AI," yet bubble warnings frame it as central-bank-fueled froth channeling printed capital.

Such fragmentation hardens open-weight challengers: South Korea’s state-backed trio now tops Hugging Face trends, proving national open-source plays can pierce US-China duopoly.

Andrej Karpathy’s nanochat miniseries v1 validates Chinchilla-optimal regimes at toy scale—sweeping d10-d20 models (0.5M batches on 8xH100 sans accumulation) for $100 (~4 hours), hitting equal N/D exponents (~0.5) with 8-token horizon constant, benchmarking via CORE scores against GPT-2/GPT-3 to forecast "turn the dial" extrapolation confidence, complete with public scaling_laws.sh and miniseries.sh for <$100 GPT-2 parity. This dovetails NVIDIA’s Jensen Huang touting 10x efficiency leaps Hopper-to-Blackwell-to-Rubin, enabling robotics’ trifecta (AI factories + Omniverse sims on RTX Pro + Thor/Orin brains with safety OS)—advice favoring verticals (EMS assembly, surgical bots) over horizontal generality amid enabling tech convergence.

Paradoxically, while cheap pretraining democratizes (matching GPT-2 soon <$100), agency costs stickiness: a $100K MRR SEO bot chews $99.5K across OpenAI/$18K, Anthropic/$17K, Gemini/$1.5K APIs plus infra, netting $437 profit yet affirming founder autonomy.

LTX-2 unleashes full open weights/training code/distilled models for local RTX inference—native 4K audio-video sync, lip-sync stability, LoRA-fine-tunable "scientist mode" (hypothesis-evidence-conflict-revise loops)—positioning as production stack versus toys, while MiroThinker 1.5-30B "search agent" tops 1T-param baselines on BrowseComp-ZH (66.8% vs Kimi-K2’s 62.3 at 30x params/cost). Complemented by Reachy Mini’s CES-fueled preorder surge (~90-day delivery for builder bots) and "everything is a file" agent hacks (grep dirs for dynamic context), this surge—echoed in Claude Code’s 30-min "right way" tutorials and cc/codex data-exploration wins—evaporates cloud dependency, but cybersecurity lags ("calm before storm") as breaches loom. Open ecosystems now lead: China’s Qwen eclipses all US/EU downloads/finetunes, per Interconnects’ 8-plot ecosystem tracker.

"The correct way to think about LLMs is... a family of models controlled by a single dial (the compute you wish to spend)" — Andrej Karpathy (full post)

These threads converge on post-labor inevitability: Shapiro’s FUD amnesia prophecy holds as hindsight renders trajectories "super obvious," yet deliberate redesign averts sleepwalking into leverage voids.