Shuo Xiao

GPT-5.6’s reasoning improvement is obvious. Sol feels much stronger logically, but also noticeably slower in daily use. Better intelligence is great — but latency still shapes the workflow.

@Eulertimistic @noname19890604 @MaxForAI 这个结论部分正确，部分有问题，GLM5.2对比GPT5.6和Fable是有代差的。和GPT5.5和opus4.8，在常规工作和编程领域是同一水平线接近的，这个结论可以从openrutor的使用统计得到，OpenRouter 上 GLM-5.2 已经累计 3.25T tokens，没有人单纯为了低价而去用无法满足自身需求的模型。

@BuildWithShuo @noname19890604 @MaxForAI 国内模型天天跑分震天响，实际用过的都用回 GPT 和 Claude 了

Anthropic在回复记者请求其置评关于中国政府和相关企业禁用Claude Code的时表示： 被建议卸载其旗舰产品Claude Code的中国用户本来就不应该使用它。

@noname19890604 @MaxForAI 光打嘴炮就太low了，大家都是成年人，拿数据看，都是可以查到的资料。 GLM 系列从 2025 年 7 月的 GLM-4.5 对标 2024 年 6 月发布的 Claude 3.5 Sonnet，到 2026 年 6 月的 GLM-5.2 与同年 5 月发布的 Claude Opus 4.8 在 coding、agent 等场景形成直接比较，时间间隔由约一年缩短至数月

@BuildWithShuo @MaxForAI 梦里啥都有

This brings back a lot of memories. Around 2013, I was working on voice interaction. Back then, the hardest problems were speech recognition, far-field audio, noise reduction, wake words, and making conversations feel natural. More than a decade later, voice AI is finally becoming the interface many of us imagined. The biggest lesson for me is that technology revolutions take much longer to become real products than we expect. But that journey is never empty. Every stage creates new opportunities — call centers, in-car assistants, meeting notes, dictation, and real-time translation all became meaningful businesses long before the technology was “perfect.” Great products aren’t built by waiting for perfect technology. They’re built by finding the right use case at the right time.

@pankajkumar_dev Looking forward to

GPT-5.6 Sol Leaks: Coming Tomorrow - OpenAI has confirmed GPT-5.6 Sol Ultra is coming to Codex, alongside Sol, Terra, and Luna. - GPT-5.6 Sol is also launching on Cerebras, delivering up to 750 tokens/second, fastest frontier model deployments yet. - Expect significantly higher usage limits than GPT-5.5. Cerebras access will likely be more expensive than standard inference. - If everything stays on schedule, GPT-5.6 should begin rolling out tomorrow.

As a technical founder, Fable 5 feels like a warning: closed AI is becoming luxury infrastructure. My bet: within 5 months, as Huawei Ascend-based training clusters come online, Chinese open-weight models will reach ~95% of the practical performance of the best frontier models available at that time — at 1/6–1/10 the cost. Frontier models were built on decades of open web data. The value captured from that open web should flow back to users and builders — not just compound into higher margins. I’m pro open, auditable, affordable AI stacks.

The moment you raise venture capital, the game changes. You also have to ask: How fast can this grow? How big can this become? How will investors get a return? That pressure can break the rhythm of a small but healthy product. My current belief: Build something that can survive without funding first.

I’m not anti-VC. Venture capital is a powerful tool for the right kind of company. But not every product should become a VC-backed company. Some products are better as small, profitable businesses. Build something useful. Make money. Maintain it. Or eventually sell it. That can be a great outcome.

I probably met 500+ investors in my first 6 years as a founder. Looking back, I think I misunderstood fundraising. I used to think raising money was almost the same as building a “real” startup. It wasn’t. Here’s what I learned after 10 years of building products: