Tech • AI • Robotics • Game
Gemini 4, Genie 4, Mistral 4, and the AI bubble bursts in China. What a coincidence...
9.4/10TL;DR
Google has taken the lead in top-tier AI rankings with Gemini 4 Carbon, while DeepMind unveiled the world model Genie 4, Mistral strengthened its open-weight position, the Netherlands is exploring AI-driven public-sector reform, and China’s AI selloff looks more like a controlled deflation than a collapse.
KEY POINTS
Gemini 4 Carbon tops the field
Google launched Gemini 4 Carbon less than three weeks after Gemini 4 Argon, and the new model now leads independent intelligence rankings. On the Artificial Analysis index it posts a score around 59, ahead of rivals such as Claude Opus 5, and it also dominates preference-based chat comparisons on LMArena. The model is priced at $2 per million input tokens and $10 per million output tokens, matching Argon for now, though that pricing appears promotional and could rise.
Fast retraining is becoming the new competitive weapon
The jump from Argon to Carbon appears to come from aggressive large-scale reinforcement retraining rather than a wholly new base model. That reflects a broader industry trend already seen in Chinese labs, where a model can gain sharply in quality within weeks by using AI-assisted evaluation, feedback loops and industrialized post-training. The result is a much faster release cadence and shorter cycles between major performance gains.
Cybersecurity access is tightly restricted
Gemini 4 Carbon is not broadly available yet and is currently being prioritized for cyberdefense use cases. Its strength lies in scanning very large codebases, identifying vulnerabilities, proposing fixes, testing them and preserving long-context coherence better than most competitors. That same capability raises obvious dual-use concerns, which helps explain the restricted rollout and the emphasis on prompt-injection resistance, lower hallucination rates and safety checks.
Genie 4 moves beyond video generation
DeepMind’s Genie 4 is not a standard video generator but a world model that renders interactive environments in real time. It updates the scene based on user input, including movement controls and text prompts, while maintaining continuity in the environment. The new version is markedly smoother than earlier releases and can change conditions on the fly, such as adding snow, wind or other dynamic events during interaction.
A step toward physical understanding and robotics
Genie 4 is designed to model aspects of the physical world, including object behavior, spatial continuity and basic cause-and-effect. It can also generate audio in context, though still imperfectly. Its strategic significance is in robotics: paired with a reasoning model such as Gemini 4, it can simulate tasks, environments and edge cases at scale, creating synthetic training grounds for future robots without requiring constant real-world trials.
Mistral releases open weights and improves rapidly
Mistral has now published the weights for Mistral Large 4, after initially offering it only through its own infrastructure and API. The model is massive at roughly 1 trillion parameters and requires serious hardware, including around eight Nvidia H200 GPUs for full-scale deployment. In return, companies can run a frontier-class European model on their own infrastructure rather than relying entirely on American platforms.
French open AI is gaining ground
The newly published version of Mistral Large 4 is reportedly much stronger than the API version released just weeks earlier, especially in coding agents and some cybersecurity tasks. That improvement again points to heavy post-training in a very short time. The release gives France and Europe a notable open-model contender that challenges leading Chinese open systems and begins to pressure major US players.
The Netherlands is testing AI-led state reform
Dutch policymakers are examining how AI could help modernize public services over a 10-year horizon, with goals including greater reliability, simplification and better use of public money. The approach is not framed as wholesale replacement of workers, but as a tool to optimize workflows, reduce friction and support better decisions. The scale discussed suggests significant potential budget effects, though with major social and labor implications.
China’s AI slump may be a controlled reset
China’s STAR 50 tech index has dropped sharply, with roughly 40% erased over about four months, and robotics names such as Unitree have come under pressure. Yet the decline may reflect a deliberate correction of overheated valuations rather than a systemic retreat from AI and robotics. New listings were heavily oversubscribed, pushing prices to levels disconnected from current profits, and authorities appear willing to cool speculation while still backing strategic builders.
Capital is still flowing to Chinese AI infrastructure
Even amid the selloff, major Chinese players continue to secure large funding. DeepSeek has reportedly raised about $15 billion, while Huawei obtained roughly $2.3 billion in bank financing for computing capacity and data-center expansion. That suggests Beijing is steering the sector away from speculative excess and toward industrial capacity, semiconductors, robotics and compute infrastructure.
CONCLUSION
The latest AI cycle is being shaped by speed, scale and post-training efficiency as much as by new architectures. The main divide is no longer simply between American and Chinese labs, but between speculative hype and durable industrial capability.
Ask a question