EAIDaily — 2026-09-15
Daily English AI brief focused on AI coding and embodied intelligence. Eight curated items for Tuesday, September 15, 2026.
1. OpenAI: ~10,000 concurrent agents cracked Navier–Stokes in 88 hours — and triggered a provenance firestorm
What happened: OpenAI disclosed that an unreleased internal model (training began Aug 28, “significantly more capable than GPT-6 Astra”) produced a proof related to the Navier–Stokes existence-and-smoothness problem — one of the seven $1M Clay Millennium Prize Problems, unsolved by humans for ~90 years. After hearing rumors that a competitor had solved Millennium problems, OpenAI pointed the system at the six remaining open ones. By the final stage, ~10,000 agents were running concurrently, exchanging ~2.7M messages and burning ~130B output tokens; the solution landed Sep 5 after ~88 hours, with Lean formal verification taking another 17 hours. Compute costs ran “emphatically in the millions of dollars” (CRO Mark Chen) — outside estimates $10M–$22.5M, roughly 1,000× OpenAI’s earlier math results. The proof shows a class of initially stationary fluids under smooth forcing develops a singularity (velocity → infinity in finite time). OpenAI says it will not claim the prize. The catch: hours before the announcement, NYU mathematician Tristan Buckmaster and Anthropic researcher Levent Alpöge published 245 pages of their own AI-assisted work on closely related Euler blowup questions — a route they had pursued for a year using OpenAI’s own Codex and Claude. Buckmaster accuses OpenAI of racing down his unpublished path and alleges researcher Sébastien Bubeck offered him unlimited compute plus sole authorship on condition his collaborator’s name be dropped (“All I had to do was throw Levent under the bus”). OpenAI denies seeing their work pre-publication, but conceded it “cannot rule out that de-identified data derived from their usage of our products helped improve our models” — while a spokesperson later said it was “categorically impossible” for Buckmaster’s recent Codex prompts to have influenced the system, including training.
Why it matters: This is the first Millennium-scale claim executed by an agent swarm, proving cluster execution is now a research instrument — but it exposes three structural fault lines at once. (1) Auditability: the formal output is machine-checkable, yet the search process, intermediate reasoning, and path selection are completely invisible, so priority and credit cannot be established. (2) Prompt privacy: if your prompts to a frontier lab’s tools can improve the very model that competes with you, every researcher doing unpublished work inside Codex or Claude now carries an unpriced exposure — a novel IP conflict with no existing framework. (3) Market transmission: the same alignment of Amodei/Altman/Musk on “pacing” sent global compute stocks down sharply on Sep 14 (SK Hynix −6%+, Intel/Micron/AMD −4%+). Terence Tao’s summary — “It’s like having machines that can lift weights for you at the gym” — captures the field’s ambivalence: capability is real, provenance is broken.
Source signal: phys.org (AFP), The Verge, MIT Technology Review, NYT, Times of India, data-today.net, The Agent Times — Sep 12–15; cross-verified against Chinese coverage (Weibo/AI Daily Sep 13–15).
2. Real-SWE benchmark: the best coding agent fails ~60% of the time on real private enterprise code
What happened: Specific Labs (YC Fall 2025) published Real-SWE on Sep 12, running frontier coding agents on tasks licensed from private production codebases never published online — including a 200K-user consumer events app and a fintech platform that has processed 100K+ bank statements. Setup: 8 model-and-harness combinations × 10 tasks × 8 attempts = 640 scored rollouts. The leaderboard: Claude Fable 5.1 (Claude Code) 38.8%, GPT-6 Astra (Codex CLI) 33.8%, Gemini 3.8 Flash (Gemini CLI) 31.2%, GLM 5.3 28.8%, Grok 4.6 and Muse Spark 1.3 tied at 23.8%, Kimi K3 18.8%, GPT-5.6 Sol 16.2%. The same models score 70–90% on SWE-bench Verified. Median reference fixes touch 11 files (~2× public benchmarks); the environment spans Postgres, Mongo, Redis, Kubernetes, AWS, Slack and more. Six of ten tasks fell below 15% resolution; on an analytics stream reducer, all models went 0-for-64. Failure modes are diagnostic gold: missed requirements (36.7%) and integration errors (34.7%) lead — the agents wrote working code for the wrong thing. Cost per run ran $2.50–$6.96, and grinding longer didn’t help (71.4% of runs under 10 minutes failed vs 73.4% of longer ones).
Why it matters: The 30–50 point gap between public benchmark scores and private-codebase performance is the distance between a demo and a deployment — “can write code” and “can do a software engineer’s job” are different claims. Specific’s design (scoring model+harness as a unit, in unfamiliar production conventions) implies enterprise selection criteria should shift from leaderboard totals to failure-mode fit: repos dense with implicit business rules fail on unverified assumptions, not on syntax. Two caveats keep this directional rather than gospel: the code stays private so results are not independently reproducible, and Specific Labs itself sells enterprise workflow data to AI labs ($100K–$1M per their pitch) — a benchmark proving models struggle on private code is also an advertisement for private code as training data.
Source signal: onthewire.ai, runtimewire.com, dev.to, thevalue.engineering, worldprogramming.org — Sep 12–15.
3. Nex AGI open-sources Nex-N2.5-Pro: computer-use agent SOTA moves to self-hostable open weights
What happened: Nex AGI open-sourced Nex-N2.5-Pro, a 397B-parameter MoE agentic model with 17B activated parameters, scoring 87.4 on OSWorld-G computer-use grounding — above Claude Opus 5’s 76.8 — while running on a single 8×H100 node. It lands the same day Scale AI released MCP-Atlas, an open benchmark measuring end-to-end tool-use reliability against real MCP servers, on which even frontier models fail a large share of tasks.
Why it matters: Open weights + single-node deployability moves the GUI-agent/computer-use SOTA threshold from closed frontier labs into the self-hostable range — a direct substrate candidate for RPA replacement, agent products, and privacy-sensitive enterprise automation. Combined with MCP-Atlas’s finding that strong reasoning does not equal stable tool use, the picture is consistent: the frontier is splitting into a capability race above and an integration-reliability race below, and open models are now competitive in the grounding layer that computer-use agents actually need.
Source signal: quidproquo AI Daily, AI Daily (CN) Sep 15 — cross-confirmed in two independent digests.
4. The agent data layer becomes the battleground: Alibaba’s PolarDB Agentic Data Foundation and ByteDance’s OpenViking
What happened: Two Chinese cloud giants shipped infrastructure for the layer beneath the model on the same day. Alibaba Cloud launched PolarDB Agentic Data Foundation, restructuring the database into a unified layer for agent memory, context, and execution state. ByteDance Volcengine open-sourced OpenViking, a self-evolving context database for agents that lifts task success rates by 6.87–11.87 points on tau2-bench with the same LLM. The same signal runs through the rest of the ecosystem: Sourcegraph launched Agentic Batch Changes (one engineer runs code migrations across hundreds-to-thousands of repos via agents), and every one of today’s trending GitHub agent projects competed on governance, not intelligence — Tencent’s teamai-cli (push→review→pull sync of skill/rule/MCP configs across a team’s Claude Code/Codex/Cursor), OpenBot’s policy gate, AgentVerse-OS’s per-project containers, agent-launcher’s unified multi-CLI management.
Why it matters: The biggest pain point for production agents is not model intelligence but long-horizon state management, cross-session consistency, and fleet governance. Both Alibaba and ByteDance are explicitly betting that the next performance delta comes from the context/memory data layer rather than parameter counts — and the +7–12 point tau2-bench gain without changing the model is the first hard evidence that this thesis pays. For coding-agent teams, the moat is migrating from “which model” to “what wraps the model”: config sync, permission gates, isolation, auditability. Cloud pricing is following the same commoditization logic — DeepSeek-V4.1-Flash joined Alibaba Cloud’s Token Plan from $6/month while GPT-6 Astra landed on Amazon Bedrock, extending multi-model cloud supply.
Source signal: quidproquo AI Daily, AI Daily (CN) Sep 15; AWS weekly roundup; GitHub Daily Digest.
5. World’s first robot-specific SIM card and tariff: operators formally join the embodied AI stack
What happened: At an event in Hefei, Leju Robotics together with China Telecom Anhui and Huawei launched “Jushen Yilian”, the first 5G-A communication package purpose-built for embodied robots, plus China’s first robot-dedicated SIM card, and unveiled a “5G-A Embodied Intelligence Application Incubation Base.” The tariff is designed around robot traffic profiles: Standard ¥599/month (300 Mbps peak uplink, 200 GB high-speed uplink data), Enhanced ¥999/month (500 Mbps uplink, 500 GB, dynamic network slicing); network side delivers 500 Mbps peak uplink, 20 ms@95% latency, and business-priority scheduling. The launch demo: 19 Kuafu humanoids with custom 5G-A modules patrolling 8 main park routes in parallel, streaming HD video and sensor data to Leju’s data platform for cleaning, labeling, and model-iteration feedback.
Why it matters: Robots could always bolt on an IoT SIM — the change is that a carrier has now re-engineered its network and pricing around robot data profiles, formally recognizing humanoids as a new class of “industry terminal.” Multi-robot concurrent operation has been the quiet scaling bottleneck (what happens when dozens of robots share uplink at a real site?), and uplink-heavy tariff design plus a closed data-feedback loop (“patrol → collect → label → retrain”) are the infrastructure preconditions for robots that get smarter the more they work. The operator — absent from every previous embodied value-chain map (body, reducer, motor, dexterous hand, chip, training ground) — is a genuinely new entrant, and China is again first to productize the layer.
Source signal: AI Daily (CN) Sep 15; Leju/China Telecom Anhui announcements.
6. The humanoid price curve breaks: Unitree G1+ at ¥95K, LanChong’s modular “Xiaobai” at ¥9,800
What happened: Unitree announced the G1+ humanoid (detailed Sep 14–15): ~35 kg, 25-DoF standard configuration (EDU version expands to 45 DoF with an optional 7-DoF dexterous hand), a new 2-DoF neck, shoulder/waist motors with peak torque up 110% and heat at same torque down 72%, binocular 2MP cameras + wide-angle monocular + 3D LiDAR, 5 head touch sensors, 6-mic array with sound-source localization, dual 10W speakers, external DC power port for continuous duty, and a 9,000 mAh quick-swap battery (~2 h). Standard edition: ¥95,000 including tax. Days apart, LanChong Robotics (蓝虫具身, a Xi’an Jiaotong University incubated startup) opened orders for its modular Mantis Standard “Xiaobai” — chassis, arms, waist, and head are all standard modules that reconfigure into different forms — starting at ¥9,800.
Why it matters: Two thresholds crossed in one week: a credible full-featured humanoid holding under ¥100K, and a modular entry ticket under ¥10K. Cost is a harder scaling variable than specs — only when unit price reaches the zone where SMEs and developers buy in volume does the industry move from expo performance to real work, and only volume deployment generates the training data the whole field is short of. The contrast frames the year: on the same day the news cycle carried 10,000 agents solving a Millennium problem, the embodied ticket to participation dropped to laptop prices. The competition question has decisively shifted from “can it be built” to “can it be cheap, reliable, and usable.”
Source signal: Sina Finance/Robot Global News Sep 15 (Unitree full spec sheet), AI Daily (CN) Sep 15 (LanChong).
7. Figure 03 enters commercial deployment at BMW Spartanburg — humanoids move from pilot to production logistics
What happened: BMW Group formally confirmed that the next-generation Figure 03 humanoid has commenced operations at Plant Spartanburg (Hall 52, expanded to assemble X3 variants and the electrified iX5). It follows the 10–11-month Figure 02 pilot, where a fleet helped produce 30,000+ BMW X3s by loading sheet-metal panels in the body shop. The new deployment shifts from rigid pick-and-place to logistics sequencing: autonomously sorting unsorted components from bulk containers into sequencing trolleys, which tugger trains and Smart Transport Robots then deliver to the line “just in sequence.” Figure 03 adds tactile sensors on every fingertip, palm cameras, soft safety components, foot-integrated wireless inductive charging, and speech-to-speech audio — upgrades driven by hard-won failure data (the Figure 02 pilot identified forearm and dynamic-wrist cabling as the primary failure points under 10-hour shifts). The platform was validated by a 200-hour continuous autonomous sorting marathon at Sunnyvale: ~250,000 packages, zero mechanical breakdowns. In the same week, UBTECH began first deliveries of its UWORLD U1 home-companion humanoid (13K+ pre-orders).
Why it matters: This is the industry’s cleanest “pilot → commercial re-entry” datapoint: the task moved from the structured body shop to high-variability assembly logistics, which is where rivals (Boston Dynamics’ Atlas kitting routines, headed to Hyundai’s Metaplant from 2028) see the commercial gateway for humanoid labor. The redesign loop is the real story — failures measured on a real production line directly became the next generation’s architecture, and BMW’s “don’t rebuild the factory, fit the robot to it” path shows humanoids entering human-designed facilities with near-zero capital retrofit. Distinguish maturity levels when reading the week’s headlines: Figure/UBTECH are delivering to customers; several rivals remain at internal-trial or announcement stage.
Source signal: Humanoids Daily, BMW Group announcement via bmwhelper.com, techcapsules.com — Sep 14–15; cross-verified against Chinese auto-industry coverage.
8. Data and touch become the new moats: JD’s 10M-hour flywheel and AGIMind’s full-stack tactile sensor
What happened: Two stories define where embodied AI’s real scarcity now sits. JD.com (announced Sep 9 at JDDiscovery 2026, still anchoring the week’s agenda) launched its Physical AI Acceleration Plan targeting six “world’s firsts”: the largest embodied data collection center — 10M+ hours of real-scene video within 2 years — plus 80+ RoboBases nationwide within 5 years, a robot-parts industry alliance to cut costs for 100 OEMs, ¥10B of retail resources by 2028, and JD Logistics purchasing 3M robots, 1M autonomous vehicles, and 100K drones over 5 years. The context: household-task success is 89.4% in standardized simulation but collapses to ~12% in real homes, and as of early 2026 the world held only ~500K hours of compliant real physical-interaction data (Guosheng Securities) — against an estimated ~1B hours needed for a human-level embodied brain, a ~200× gap. Meanwhile AGIMind (灵感机器人), a Shenzhen team working on visual-tactile sensing since 2016 (HKUST/USTC roots), published its Lux series full-stack visual-tactile sensors on Sep 15: ultra-miniature imaging plus dense force-field sensing that captures force, contact deformation, texture, micro-slip, and geometry simultaneously — part of a full chain spanning flexible e-skin, six-axis force sensors, data collection/processing platforms, teleoperation, and fine-manipulation systems. At CIFTIS, enterprises also opened 100K+ hours of human behavior data (15K scenarios) and demonstrated a unified domestic robot electronic architecture merging “brain + cerebellum.”
Why it matters: The binding constraint of embodied AI has migrated from brains to data and contact. JD is the demand-side giant converting its supply-chain scenes into a proprietary data flywheel — the “picks and shovels” position in a gold rush where everyone else is short of ore. Tactile sensing is the next capability ceiling (precision assembly, flexible grasping — squeezing a strawberry without bruising it), and full-stack sensor-to-data-platform plays are bets that whoever owns contact data owns manipulation. Germany’s leading roboticist Wolfram Burgard crystallized the geopolitical read this week: China’s robot fleet and scenario density give it the fastest data-collection loop in the world, and data sharing — not model parameters — is the decisive bottleneck. Valuation logic follows: order-backed deliverers and upstream data/components suppliers are being priced above concept-stage body makers.
Source signal: China Securities Journal, Cailianshe/Toutiao (Sep 9–10, resurfaced Sep 15); Sina Finance Sep 15 (AGIMind); CIFTIS coverage via AI Daily (CN) Sep 15.
Today’s throughline
Capability ceilings, cost floors, and the layers beneath the model were all put on the table the same day. On the AI coding side: 10,000 agents can climb a 90-year-old mathematical wall in 88 hours — yet the strongest coding agents still fail 6 of 10 tasks in an unfamiliar private codebase, and the industry’s energy is flowing into the layers that make agents governable and reliable (context databases, policy gates, config sync, private-code evaluation). On the embodied side: a carrier built a SIM card for robots, the price of admission fell below a laptop, Figure’s second generation graduated from pilot to commercial logistics — while JD and AGIMind made clear the durable moats are real-world data hours and the sense of touch. The field’s center of gravity has moved from “how strong” to “how controllable, how cheap, how deployable.”
Compiled Sep 15, 2026 · Sources cross-verified via WebSearch + multi-outlet WebFetch · Selection: 8 of ~30 scanned items