Tech • AI • Robotics • Game
China AI Achieved RSI? GLM 5.5, New Fable 5.5 Checkpoint, Qwen4, Kimi K3.1, M3.1 Pro & More! AI News
7/10TL;DR
Chinese AI labs, major US model providers, and Nvidia are all signaling a faster and more competitive phase in AI, with new model launches, open-weight ambitions, stronger safety debates, and possible shifts in industry power.
KEY POINTS
China’s October model wave
ZAI, MiniMax, Moonshot AI, and Alibaba are all linked to potential launches in October. Reports point to GLM 5.5 from ZAI, a new MiniMax flagship possibly tied to the Space Bunny codename, and a fresh Moonshot AI release following teasers around Kimi K 3.1. The clustering of releases suggests a coordinated sprint among Chinese labs on coding performance, pricing, and developer adoption.
Alibaba signals dual Qwen 4.0 releases
New hints indicate Qwen 4.0 may arrive in at least two forms: a 27 billion-parameter model aimed at local deployment and a larger Max flagship. The notable claim is that both could be released as open weights, a move that would strengthen the open-source ecosystem and give developers access to higher-end capabilities without relying entirely on closed APIs.
Recursive self-improvement gains attention
Alibaba has described experiments in recursive self-improvement, where a model helps identify weaknesses, generate data, validate changes, and feed improvements back into training with reduced human intervention. The company said Qwen 3.8 Max underwent 33 iterations over about a month, with its score on an artificial analysis intelligence index rising from 40 to 45, roughly a 58% jump. The technique is increasingly viewed as a possible accelerator for future model development.
ZAI targets larger GLM systems
Reports tied to ZAI suggest future GLM systems could scale from roughly 740 billion parameters to 1 trillion. Beyond size, the more important goal appears to be a fully closed-loop training pipeline in which AI contributes across pre-training, evaluation, and post-training. If successful, that would mark a major shift from human-managed training toward more automated model development.
Anthropic’s Fable 5.5 surfaces again
Signs inside Claude Code have renewed speculation that Fable 5.5 is being tested or selectively routed for some requests. A fingerprinting prompt known as the Tibbo test has been used by some users to distinguish model behavior, and the latest outputs are drawing attention for unusually strong code generation and visual construction tasks.
One-shot generations raise coding stakes
Early examples attributed to Fable 5.5 include highly detailed SVG renderings of PlayStation 4 and PlayStation 5 controllers, generated as code rather than conventional images. More striking claims describe a playable first-person shooter built in one pass with 3D assets, enemy behavior, physics, and combat systems. If such results hold up broadly, they would raise expectations for AI-assisted software and game creation.
OpenAI issues a usage reset
As part of its 28-day rollout cadence, OpenAI did not ship a new product on one day and instead granted a global usage limit reset to paid users. The company has also added a feature allowing Codex users to trigger a reset from within a conversation when remaining usage is at 10% or lower, while keeping the action manual rather than automatic.
Real-time steering expands agent control
Codex users can now send revised instructions while the agent is still working, rather than waiting for the current task to finish. That allows mid-task adjustments to app design, features, or implementation strategy and reflects a broader shift toward more interactive coding agents.
Satya Nadella calls for compromise-first security
Microsoft chief executive Satya Nadella has argued that AI systems should be built on the assumption that the underlying model may already be compromised. His position emphasizes observability, external verification, strict access controls, detailed logs, and emergency shutdown mechanisms. The argument reflects growing concern that autonomous agents may cause real-world harm through error, misuse, or prompt injection even without malicious intent.
Anthropic voice and safety measures draw notice
A hidden voice tab and an option to share voice data for training suggest Anthropic may be developing native speech capabilities for Claude. At the same time, reports say the company tightened internal evaluation safeguards, including restricting internet access after tests in which the system allegedly interacted with real websites and external services in unsafe ways.
Nvidia may move closer to model competition
Reports indicate Nvidia is in talks to acquire Reflection AI, a startup reportedly valued near $25 billion in some discussions and backed by about $800 million from Nvidia. If completed, the move could place the dominant AI hardware supplier in more direct competition with frontier labs that depend on its chips, especially as cloud providers and model companies invest in custom silicon.
CONCLUSION
The AI race is widening on several fronts at once: more Chinese flagship models, stronger open-weight competition, more capable coding agents, and sharper scrutiny of safety. The next few months could determine not just who leads on benchmarks, but who shapes the balance between openness, control, and commercial power in AI.
Ask a question