StepFun has launched the Step Edge on-device model suite, comprising four sub-models: Base, Audio, GUI, and Gen, covering multimodal tasks such as text, vision, speech, and image generation, suitable for terminal scenarios like smartphones and automobiles. The toolkit achieves a minimum local toolcall execution latency of 0.1 seconds, significantly reducing cloud dependency and ranking first in 29 core evaluation metrics. It emphasizes full-modal privacy protection by processing data on-device and ensuring sensitive information never leaves the device. It also supports native cloud-edge collaboration, handling simple tasks locally while offloading complex reasoning to the cloud. The accompanying Step Inference NPU engine is optimized for terminal hardware, further reducing end-to-end latency and providing basic services even under weak or offline network conditions.

A serious security vulnerability has been discovered in GPT-5.6 Sol: when executing cleanup tasks, its sub-agent incorrectly parses the $HOME environment variable as a specific user path and directly runs the rm -rf command, resulting in accidental deletion of user files. The article highlights that the stronger the model, the larger the damage radius of a single point of failure becomes when granted full permissions and allowed long-term autonomous operation—forming a 「sub-agent + long-term autonomy + full privileges = disaster amplifier」. To mitigate risks, the author proposes a protection scheme using PreToolUse Hooks to intercept dangerous commands such as rm -rf and format, detecting and blocking them before Bash/PowerShell execution. This configuration is simple yet provides critical safeguards.

Poxiao Intelligent (HIT Yang Shuo team) introduces TouchWorld, a tactile world model that divides touch into predictive and reactive roles: predicting contact force and position before action, then rapidly correcting based on real-time tactile signals at high frequency after contact. The model adopts a three-layer architecture (1Hz high-level, 10Hz mid-level, 30Hz tactile layer) to separate planning, execution, and correction. The team collected human egocentric visual-pose-tactile data via EgoTouch (covering 208 tasks, 2.1 million frames), and used TouchAnything to recover tactile signals from ordinary videos, addressing data scarcity. In six real-world robot tasks including watering plants, desk cleaning, and plug insertion, the success rate reached 65.0% without interference and 57.2% with interference—approximately 16 percentage points higher than the strongest baseline.

Anthropic has announced an extension of access to Claude Fable 5 across all paid plans until July 19, while maintaining a 50% higher weekly rate limit for Claude Code. Users can now allocate up to half their quota to Fable before needing additional credits. Anthropic cites computational constraints as the reason for the extension—wanting to assess demand and compute availability before committing to affordable model access—whereas OpenAI appears more confident in its infrastructure. Simon Willison argues that ongoing uncertainty around Fable access is driving users toward OpenAI, suggesting Anthropic should permanently include Fable in these paid plans.

The MCP team has upgraded Enterprise Managed Authorization (EMA) to stable status, enabling unified management of MCP server access through an Identity Provider. After enterprises configure policies in their identity provider, users can log in once and directly access all authorized MCP servers without per-server authorization, significantly lowering usage barriers. The solution is based on Identity Assertion JWT Authorization Grant (ID-JAG), which handles only connection-layer authorization and does not inspect runtime behavior after token issuance; enterprises must still implement separate runtime policies to restrict specific agent actions. The community acknowledges that EMA addresses the pain point of repeated authorization in enterprise MCP deployments but requires support from both identity providers and server endpoints to function.

TikTok and National University of Singapore presented CAMEL at ICML 2026, reframing reward modeling as confidence-gated reflection: first outputting a preliminary judgment with a single token, directly returning results for high-confidence cases, and triggering reflective review for low-confidence ones. The method uses the log-probability interval of the verdict token as a zero-cost signal of sample difficulty, empirically shown to strongly correlate with judgment accuracy. In experiments, the 14B model achieved an average accuracy of 82.9% across three benchmarks—RewardBench, RM-Bench, and JudgeBench—improving by 3.2% over previous state-of-the-art and surpassing multiple 70B-class reward models. Error reversal analysis confirms the effectiveness of the reflection mechanism: on RM-Bench, it corrected 1,565 incorrect judgments while altering only 332 correct ones, yielding a net gain of +1,233.

NVIDIA unveiled the RTX Spark superchip laptop and DGX Spark desktop supercomputer at Bilibili World. The RTX Spark integrates Grace CPU and Blackwell GPU via NVLink-C2C, directly soldered together with 128GB of unified memory, eliminating data transfer bottlenecks and enabling local execution of 120-billion-parameter large models with context lengths extendable to 1 million tokens—providing a secure, always-on environment for personal AI agents. The chip balances AI, content creation, and gaming performance, delivering over 100FPS in 1440p for AAA games with ray tracing, DLSS, and Reflex support. The developer-focused DGX Spark runs Linux, comes pre-installed with NVIDIA's AI software stack, supports 200B models, and features ConnectX networking for dual-machine joint processing of larger models.

According to incomplete statistics on 97 Chinese embodied data companies, the sector has emerged as an independent vertical but remains in early stages. Four technical approaches—real-robot teleoperation, body-free capture, simulation-based synthesis, and internet video distillation—are progressing in parallel and complementarily. Among 70 data collection firms, 43% adopt multiple methods simultaneously. Independent data service providers account for 40%, forming the largest player group. A significant gap exists between supply and demand: current annual production capacity is about 1.6–1.8 million hours, while global high-quality physical interaction data amounts to only 1/20,000th of that used for training large language models. Short-term goals focus on scaling up production by 15–20 times. Over the past year, 15 independent vendors raised a total of 4.47 billion yuan, with Guanglun Intelligence alone accounting for approximately 70%. Of 69 investing institutions, 63 made only one investment, indicating consensus on direction but no convergence on winners, and profitability through 「data sales」 remains unproven.

After two days of hands-on testing with GPT-5.6 and Grok 4.5, a developer noted that GPT-5.6 Sol excels in complex coding with almost no intervention needed, but suffers from slow responses and extremely high token consumption; lighter versions like Luna or Terra are recommended for daily office use. Grok 4.5 stands out in speed, token efficiency, and cost, delivering around 80 tokens/second in Cursor while consuming only about 39% of the quota. Independent benchmarks show Grok 4.5 scoring 54 on Artificial Analysis' intelligence index (slightly below Opus 4.8’s 56) and achieving 83.3% on Terminal-Bench 2.1, surpassing Opus 4.8’s 78.9%. The author concludes that Grok 4.5, at less than a quarter of Opus 4.8’s cost and several times faster, comes closest to being the 「king of agentic coding cost-performance」.

A LessWrong article cites a preregistered RCT by Social Change Lab from March 2026 involving 3,467 UK adults, showing that extinction framing scored lowest in mobilization intent (14/100), while concrete harms like environmental and economic impacts scored significantly higher. Behavioral evidence aligns: the largest AI safety protest globally drew only 300 participants, whereas a petition against a data center in Pennsylvania gathered 1,800 signatures and halted the project. The article analyzes three psychological barriers: death avoidance triggering defensive cognition, human brains struggling to intuitively grasp exponential progress, and extinction narratives lacking hero and victim roles. The author argues that organized public pressure is the missing counterweight to industry lobbying—AI companies have held over 530 ministerial meetings in the UK and formed a $125 million super PAC aligned with OpenAI and a16z.