OpenAI Builds Humanoid Robots, DeepSeek V4.1 Cuts Costs and Speeds Up: Three Shifts in the AI World in One Day

The AI industry moves at a breathless pace. In the past day alone, OpenAI confirmed it will build its own humanoid robots, DeepSeek released a new model that halves token consumption, and Nvidia boosted Kimi K3 inference speed by 1.68x—three pieces of news pointing to the same conclusion: competition among large models is shifting from a "parameter arms race" to a two-front battlefield of "engineering efficiency + deployment form factor."

OpenAI Enters Robotics: Altman Only Wants the "Brain"

When asked on a podcast whether he would personally build humanoid robots, Altman's answer was just three words: "We will." His reasoning: the world is designed for humans, so a humanoid form makes sense.

This could easily be read as hardware ambition, but Altman's focus is actually on the software side. He explicitly stated that the robot's appearance and limbs are "minor details," and what really matters is developing the "brain" that drives the robot. In other words, OpenAI wants to build a general-purpose control system for embodied intelligence, not compete with Boston Dynamics or Tesla Optimus on mechanical structure. This positioning aligns naturally with its model capabilities, and also explains why the project is described as "still a long way from becoming a formal product"—Altman expects to first deliver an impressive demo prototype, and robots actually walking the streets will take several more years.

For China's embodied intelligence sector, this is a signal: the real barrier may not be in body manufacturing, but in generalized decision-making capabilities in the physical world. On the same day, Hangzhou-based embodied brain company Gewu Tongzhi announced the completion of nearly 100 million yuan in angel funding, focusing on a "physical inversion world model" that enables robots to infer implicit physical information such as contact and force from visible actions—an approach that echoes OpenAI's "brain-first" strategy.

Embodied AI refers to intelligent systems that integrate perception, decision-making, and action capabilities within a physical entity. Unlike large language models that only process text symbols, embodied AI requires models to complete the perception-reasoning-action loop in real physical environments and cope with uncontrollable factors such as lighting, friction, and object deformation. The "brain-first" approach means companies direct resources toward cross-embodiment general policy models (such as vision-language-action models, VLA), then adapt to different hardware forms with minimal data; the challenge lies in the scarcity of real-world data and the simulation-to-reality gap (sim-to-real gap).

DeepSeek V4.1 Flash: A "Driven" Iteration That Halves Token Usage

DeepSeek^Ad updated to version V4.1-Flash-1017, with the web, app, and API all switching simultaneously. The company disclosed two key changes: first, through self-feedback reinforcement learning training, the model's ability to build complete projects from scratch has significantly improved, with higher completion rates on "wishful" programming tasks like game development and 3D modeling; second, token usage in most scenarios dropped by over 50% compared to the previous version.

Halving tokens is no less significant than capability improvements. For developers, this means direct bill reductions; for DeepSeek itself, it's inference cost optimization. Combined with the newly opened Search and Fetch interfaces (both at 5 yuan per thousand calls), DeepSeek is packaging "model capabilities" as orchestratable engineering components rather than a mere chat entry point. The flood of comments like "so driven" and "super kill line" in the comment section reflects market recognition of the high-frequency iteration pace.

Worth noting is the organizational adjustment the same day: DeepSeek integrated its Harness team with the Web, App, and Platform teams into a "Product Department," led by former Harness head Cui Tianyi. Technical iteration and product organization advancing in parallel shows that its competitive strategy has expanded from single-point model breakthroughs to building a productization system.

Self-feedback reinforcement learning refers to training methods where, after generating results, the model uses itself or a judge model to score and reflect on outputs and update its policy accordingly. Common forms include rejection sampling, self-consistency checks, and verifiable rewards (such as whether code passes unit tests). Compared to preference data that relies on human annotation, it can use "whether the task is actually completed" as a reward signal, making it particularly suitable for scenarios with automatically verifiable results like programming and 3D modeling. Token usage directly determines inference bills: producing fewer intermediate steps and redundant explanations means lower unit costs for equivalent tasks—which is why "capability improvement" and "cost reduction" can happen simultaneously.

Nvidia's Two-Front Strategy: Accelerating Competitors' Models

Nvidia CentML engineer Benjamin, through an upstream vLLM optimization, achieved 423 tok/s/user inference speed for the 2.8-trillion-parameter Kimi K3 on GB300 NVL72—a 1.68x performance improvement. The core method is large-scale Megakernel fusion—writing decode-specific kernels in Cute DSL, with the entire Megakernel containing 12,000 lines of code.

The subtlety here is that Nvidia is helping others' models run faster. As a compute supplier, it needs to prove the efficiency ceiling of its hardware on mainstream open-source models to consolidate ecosystem stickiness. The report also mentions that Nvidia is developing optimizations for vLLM on models like DeepSeek V4.1—the deeper the optimization, the harder it is for customers to leave the CUDA ecosystem.

Another piece of news reveals another side of Nvidia: executive Jeff Fisher denied rumors that the RTX 5090 has been discontinued, acknowledging limited production capacity and coordinating restocking. Consumer graphics card prices have risen sharply due to supply-demand imbalance, mirroring the hot data center business.

Megakernel refers to merging the many fine-grained operators in the decoding phase (attention, normalization, activation, matrix multiplication, etc.) into one or a few GPU kernels for execution, reducing kernel launch overhead and memory round-trips. During token-by-token decoding of large models, each step involves small computation and is memory-intensive; traditional implementations are often constrained by "launch overhead exceeding computation itself," and fusion can significantly improve single-user throughput. Cute DSL is a CUDA template abstraction layer in Nvidia's CUTLASS system, used for writing high-performance tensor operation kernels; vLLM is an open-source high-throughput inference framework whose mechanisms like PagedAttention are widely adopted, so upstream optimizations quickly benefit the entire ecosystem.

Robot Repair: An Underestimated New Employment Track

JD.com announced a set of noteworthy figures: it has reached after-sales repair partnerships with nearly 100 robotics companies including AgiBot, YunShenchu, Zhongqing, and Songyan Power, and plans to create over 100,000 robot repair engineer positions in the next five years, while promoting the transition of frontline logistics workers such as couriers.

As robot shipments rise, repairs will inevitably become essential. The value of this news lies in revealing an easily overlooked link in the embodied intelligence supply chain—the service network. The more complex the hardware, the heavier the aftermarket; whoever establishes the maintenance system first will control the last mile of robot deployment. Compared to OpenAI's "brain" approach, JD.com takes a typical scenario-side entry—the two don't conflict but rather form two ends of the supply chain.

Robot aftermarket refers to maintenance, spare parts, refurbishment, training, and operational services after the complete machine is sold. Its economic characteristics resemble the automotive 4S system: the more complex and expensive the hardware, the larger the share of service revenue. Humanoid robot joint modules, reducers, force sensors, and batteries are high-wear components, and fault diagnosis requires understanding both mechanical and software logs, so repair engineers need interdisciplinary training. This link also determines the hidden costs of scaled deployment—if annual maintenance costs per unit are too high, customers' total cost of ownership (TCO) will weaken purchasing willingness, making service network density a prerequisite for shipment volume.

Other Developments Worth Watching

Security alert: Open-source AI programming tool OpenCode disclosed a security incident where a legacy system being decommissioned suffered unauthorized access, affecting approximately 1.4% of OpenCode Console users. Third-party AI service API keys (including OpenAI and Anthropic) stored in workspaces may have been leaked. Some users reported that their monthly $20 API quota had been exhausted. OpenCode is negotiating with multiple model service providers for refunds on anomalous charges but has not yet committed. The risk of hosting keys on third-party platforms has once again been brought to the fore.

Chip landscape statement: Intel CEO Chen Liwu (Lip-Bu Tan) said that US and Chinese chips each have advantages and disadvantages, and the gap is not as large as imagined. He also emphasized that the US government's minority stake does not affect the company's governance independence, and the restructuring is 80% complete.

Automotive and consumer electronics: Tesla's new-generation Roadster, dubbed a "flying car" due to stunts like mid-air flips, has drawn industry questions about its practical use cases and regulatory classification. Netizens joked, "If it were a domestic manufacturer, they'd be criticized into delisting in minutes." iPhone Duo pre-orders are booming, with some consumers paying over 5,000 yuan extra just to be the first in their social circle to use one.

Musk's new business: Tesla is investing resources to develop Hagglebot, a negotiation bot for commercial scenarios built on Grok. Musk claims it can improve enterprise cost efficiency by over 5%, saving up to $6 trillion annually based on global GDP. This is a software bot, positioned differently from the Optimus physical robot.

From OpenAI's "brain" declaration to DeepSeek's halved tokens, to Nvidia accelerating competitors' models, the keywords of AI competition are shifting from "bigger" to "cheaper, faster, and more deployable."

One-line Summary

OpenAI builds robot brains, DeepSeek halves tokens, Nvidia accelerates Kimi K3—competition shifts to efficiency and deployment.

Paragraph Summary

The article reviews three key AI industry developments from the past day and offers a unified conclusion: competition among large models is shifting from a parameter arms race to a two-front battlefield of "engineering efficiency + deployment form factor." OpenAI confirmed it will build humanoid robots, but Altman emphasized the focus is on the "brain" driving the robot rather than the mechanical body, echoing Gewu Tongzhi's physical inversion world model approach in China; DeepSeek released V4.1-Flash-1017, improving from-scratch project building through self-feedback reinforcement learning, cutting token usage by over 50% in most scenarios, adding Search and Fetch interfaces, and integrating its product department; Nvidia engineers used Megakernel fusion to optimize vLLM, boosting 2.8-trillion-parameter Kimi K3 inference speed by 1.68x to consolidate the CUDA ecosystem. Additionally, JD.com plans to create over 100,000 robot repair positions in five years, and OpenCode suffered a key leak security incident, supplementing the service and risk sides of the supply chain.