AI Wire

NVIDIA's Korea AI factory blitz

NVIDIA used the day to stack up a chain of Korean partnerships that read like a national industrial plan. SK hynix signed a multiyear deal to co-develop next-gen memory for NVIDIA platforms from Vera Rubin to Jetson Thor, and to apply Omniverse, CUDA-X, and PhysicsNeMo to fab digital twins and semiconductor design (@nvidia). NAVER will scale its GAK Sejong data center on the NVIDIA DSX platform, starting at 55 megawatts with a stated path to gigawatt scale for sovereign and physical AI (@nvidia).

LG Group will build an AI factory on the same DSX stack for robotics, autonomous driving, data center tech, and GPU cloud — Jensen Huang met Chairman Koo Kwang-mo to lock it in (@nvidia). Doosan is expanding across DSX, MGX, robotics, power, and electronics materials (@nvidia), and Jensen also sat with Hyundai Motor Group leadership on mobility and physical AI (@nvidia). The throughline: Korea is being wired as a vertically integrated NVIDIA territory — memory, factories, robots, cars — well before the Vera Rubin generation lands.

AI coding productivity gap & autonomous agent skepticism

A new MIT/FT-circulated study reframes the "AI coding boom" as a shipping bottleneck: code volume up ~300% and commits up 180%, but software releases only +30% (@garymarcus). The paper argues the production funnel — review, integration, test, packaging — is the weak link, and that app marketplaces show more new apps without more total usage (@garymarcus). This is consistent with two recent NBER/SSRN drafts on writing-vs-shipping productivity effects across AI tool generations (last30days, nber.org; last30days, papers.ssrn.com).

Gary Marcus paired this with a "Meta-Agent Challenge" result showing today's agents are still weak at building agents end-to-end without humans in the loop, concluding "we aren't close to RSI, regardless of the hints IPO-bound Anthropic tried to drop last week" (@garymarcus). Read together, the day's signal is that capability per token is rising faster than throughput per organization.

Long-running agent orchestration patterns

Operators are converging on a "design the loop, not the prompt" stance. Sam Altman announced a 100-day Codex program giving one standout user per day a 10× usage bump to see what extended autonomy can produce (@sama). Anthropic's Boris Cherny offered a five-point recipe for running Opus autonomously for hours or days: auto-mode permissions, dynamic workflows orchestrating many subagents, /goal or /loop nudges, cloud Claude Code, and proper environment setup (@bcherny).

Philipp Schmid framed the same shift bluntly — "you shouldn't be prompting coding agents anymore. You should be designing loops that prompt your agents" — and pushed "subagentmaxxing," where oversight is itself delegated to nested agents (@_philschmid). OpenAI's drop of "dozens of real-world workflows" (inbox triage, PR review, Figma-to-code, bug triage, deploys) is the productized version of that same idea (@gdb).

Hugging Face / open-source model momentum

Open weights had a loud day. Gemma 4 MTP officially merged into llama.cpp, enabling fast Gemma 4 QAT + MTP local setups (@huggingface), and Ivan Fioravanti showed a Reachy mini robot running Gemma 4 E4B QAT, Parakeet TDT 0.6B v3, and Qwen3-TTS 1.7B locally in near real time (@huggingface). A community "Super Gemma 4 26B Uncensored GGUF v2" build claims 0/100 refusals and 90% faster prompt processing on 16–22 GB VRAM (@huggingface) — a useful capability signal, and a reminder that uncensored local models are now a click away.

The "American open source is back" framing came from the observation that 9 of the top 30 models on Hugging Face's front page are NVIDIA's (@huggingface, @clementdelangue, @_akhaliq). Separately, all 15 CVPR 2026 paper finalists are now browsable with GitHub and HF artifacts attached (@_akhaliq).

Geopolitical & governance tensions in AI

Jeremy Howard flagged a "striking shift toward Chinese models by American AI startups since the start of the year" (@jeremyphoward) — a quiet but meaningful supply-chain story. On the US side, Gary Marcus highlighted reporting that xAI is "a mess," with Elon Musk firing most original hires after unrealistic deadlines (@garymarcus), and two Danish pension funds (AkademikerPension and Pædagogernes Pension) banned SpaceX investments citing S-1 losses outside Starlink and Musk's ~85% voting control (@garymarcus).

Reliability bit teeth too: Notion disabled Anthropic's Opus 4.7 and 4.8 in its picker after degraded performance, rerouting traffic to other providers (@clementdelangue) — a rare public deprecation by a major distribution partner.

AI theory, alignment & meta-commentary

Jeremy Howard resurfaced Geoffrey Hinton's distillation results showing robustness even when training and target distributions barely overlap, and tied it to the on- vs off-policy distillation debate (@jeremyphoward). Roon argued the field "got insanely lucky that LLMs really are what it is" and that a growing "mutual conditional pause" consensus on the eve of RSI is healthy (@tszzl). Ethan Mollick's complement: as implementation gets cheap, hoarding genuinely unusual ideas gets more valuable (@emollick).

The Bottom Line