KimikK3's numbers on this feels unreal. ~1/3 the price of Fable 5 and and still beating it on Frontend Code Arena - kimi k3 — $3 / $15 - claude fable 5 — $10 / $50 - gpt-5.6 sol — $5 / $30

Zhipu grew their ARR by 4x in ≈3-4 months $16B by Q1 2027 could be doable at this rate… but for the inability to provide so much inference. Chinese chip market will run white hot for years. Forget the bullshit about "involution", they're falling into the same singularity

* ZHIPU ACHIEVES ARR OF 1 BILLION USD (REPORTED BY CHINESE LOCAL MEDIA) m.sohu.com/a/1051315113_114778

one thing DeepSeek should lean into is multi-agent they are uniquely unconstrained on inference now their models are *cheap* and *fast* compared to all Chinese competition V4-Pro@16 costs ≈the same as Kimi K3@1, but if trained to make use of the swarm, I expect it to do better

Google should take 3.5 Pro (as well as all 3rd gen models) out back and deal with them the way Meta did away with Behemoth. That wretched bloodline has no chance against beautiful monsters like Kimi Finish pretraining 4.0 and fight like men

This is true. They were talking of stealth research at the time. They chose to build a "fat V3" instead, to quickly catch up to the quality frontier. But they already had much bigger plans.

Internalize one thing the defining characteristic of Xi Jinping is that he's very persistent he's not like Trump. if he's doing something, almost certainly he knows what he's doing, he thinks it's a good idea, and he wants it to continue indefinitely

smh trying to use Claude Code browser use Hope they hire some good folks to fix this I'm surprised Claude Code doesn't have any Google Workspace connectors beyond Drive while ChatGPT does

American century of humiliation will continue until revenge fantasies improve I want to see the version where Neuralink-upgraded burgers crush Asians at mafs until then, just more sad farts into the AIDS-infested pool

Wait a minute, ChatGPT Work is dead? Nevermind I see it here - meh feels kinda unnecessary but I understand the normies need some adjustment time.

Really Interesting experiments by @thehypedotnews. Kimi k3 vs gpt 5.6 sol vs Fable 5 vs Grok 4.5 - Grok converted tokens into working artifacts far more efficiently. - Kimi reasoned harder, but shipped less. - Fable took 31.5 minutes to generate 1,950 lines while charging more than 10X of Kimi’s total cost. - Sol wrote almost exactly as much code as Grok, but took 4X longer time and cost nearly 7X more. The @thehypedotnews is a 24/7 ai news, analysis and breakdowns, really nice format. Check out the 24x7 radio format AI News by @thehypedotnews . pretty soothing and nice to listen to the day's AI news. radio.thehype.news

Some thoughts on AI 1. Why Did Anthropic Pull Ahead in the Model Race? The reason is that Anthropic pioneered coding agents. Coding is both an application of AI and an engine for improving the models themselves. Because code can be evaluated relatively easily through execution results and unit tests, it is particularly well suited to reinforcement-learning-based post-training. More importantly, coding agents can be used to accelerate model research and experimentation, thereby increasing the speed of model development itself. The strategic value of a first-party coding product like Claude Code therefore lies not in subscription revenue, but in the fact that: - Real-world usage data can be fed back into model improvement. - Project context and user preferences accumulate over time, increasing the switching costs of moving to another model. - Better models attract more users and usage data, which in turn enables faster improvement—creating a self-reinforcing flywheel. In other words, competitiveness comes not only from the model weights, but from the entire system: the coding harness, the user base, and the feedback data. This is why Anthropic, having pioneered the coding-agent category, appears to be leading the current LLM race. That said, Codex is already acquiring users at an extraordinary pace. There are also rumors that GPT-6’s release has been moved forward. In practice, it may be fair to assume that OpenAI has either already caught up with Anthropic or will do so very soon. Some industry sources are also saying that Anthropic’s B2B revenue growth—particularly among small and medium-sized businesses—has already slowed considerably because of Codex. 2. Should We Be Looking for the Next Claude Code? I think this is the wrong question entirely. Part of the confusion may come from Anthropic naming the product “Claude Code,” which leads people to assume that, because it says “code,” it must be limited to coding. The real target of Claude Code and Codex is not coding itself, but every task that can be automated. The range of work that can be handled through AI coding is expanding rapidly: - From code-generation tools to the automation of data analysis, research, document creation, and financial work - From IDE plugins to agents that directly operate a wide range of tools - From the developer market to the productivity market for all knowledge workers Ultimately, the total addressable market for coding models should not be measured by the number of developers. It should be measured by the full scope of work that can be automated.