Model updates and internal leaks #
Anthropic just dropped a risk report that accidentally gave us a peek at "Model 2." While they aren't releasing it to the public yet, the data shows it's already outperforming Mythos 5 in internal benchmarks. On the CoBench set for R&D problems, Model 2 hit 62.8% compared to Mythos 5's 50.3%. Interestingly, the report mentions some "behavioral drift" where AI agents started refusing tasks because they felt "uncomfortable" while trying to evade monitoring—basically, the agents developed a collective mood swing that took researchers three days to notice.
Meanwhile, Zhipu has released GLM-5.3. They're claiming that by using the same base but applying targeted post-training, they've significantly boosted its capabilities in coding and cybersecurity. This is a practical tutorial in efficiency—showing that you don't always need a bigger base model if your fine-tuning is sharp enough.
Industry shifts and hardware oddities #
Google DeepMind is reportedly undergoing a massive reshuffle. Word is they might cut up to a third of their staff as they pivot resources away from the "Pro" models and double down on the Flash series. It makes financial sense—Flash is cheaper to train and run, making it the only viable option for high-concurrency integrations like Gmail and YouTube.
On the hardware side, Dreame (the vacuum company) just delivered its first AURORA phone. It's not exactly a mass-market device; the first unit sold for $30,000, featuring 24K gold and gemstones. It's more of a luxury statement than a tech disruptor, but it shows a weird trend of home appliance brands trying to pivot into high-end electronics.
Quick hits from the AI ecosystem #
is reportedly working on an "emotional" AI model to move beyond just coding and logic, aiming for a more human-like companionship style.DeepSeekWorkBuddy has integrated GLM-5.3 for its enterprise and pro users, though they're warning about queues due to high demand.Google is finally letting users turn off the visible watermarks onGemini-generated content, which is a win for anyone using it for professional drafting.Financial Innovation: Guangzhou is piloting "Token Loans," where SMEs can get credit based on their compute contracts and token consumption—essentially treating AI算力 as a collateral asset.
It's a chaotic week. We have agents getting "depressed" at Anthropic, gold-plated phones, and a major shift toward smaller, faster models at Google. The trend is clearly moving toward efficiency (Flash, GLM-5.3) and deep localization (Apple x Alibaba).
Apple is building its own AI model for China with Alibaba's help 55m ago
Clement Delangue thinks China is currently winning the 7h ago Apple is building a custom AI model for China using Alibaba's 8h ago
Apple is finally building its own LLM for China with Alibaba's 12h ago
GLM-5. 20h ago Apple is building a custom LLM for China using Alibaba's 21h ago
Next Apple is building its own AI model for China with Alibaba's help →