Luz.0x3

AI Product Discoverer 🤖 | Street Food Explorer 🍜 | AI Agent Builder | Ex-Google, Meta ｜ 🇯🇵

[Luz.0x3 @player0x3]

@hkproj Why not Singapore? (It gets a lot of Chinese, obviously.)

[Luz.0x3 @player0x3]

Feel the sunrise and the sunset, feel your breath — this is the essence of life.

[Luz.0x3 @player0x3]

@elonmusk reset🙏

[Luz.0x3 @player0x3]

Y bro, Rag is coming back.

[Dev Agrawal @devagrawal09]

RAG is so back

over the next few weeks everyone's gonna be using @typesafeai jev to expand queries and rerank results over chunked data before feeding to an LLM

[Luz.0x3 @player0x3]

Honestly, Jev might be the real model for the Agent era. What does an agent do? Next tool to call, sequential steps—it's all about needing a state decision machine. A monolithic coding agent + skill acting as an omnipotent god is pure brute-force. State-decision-driven Graphs are elegant and change everything. GPT-3.5, o1, Jev—three eras. For players like Cohere and Mistral who were suffering hard under SOTA, this is totally good news.

[Luz.0x3 @player0x3]

@Bjoern_Hev Japan

[Luz.0x3 @player0x3]

Chinese models improve their capabilities by distilling models from Anthropic and OpenAI, while Cursor and Devin take those already-distilled Chinese models and post-train them. The latter approach is a smarter strategic move: they haven’t stolen anything, but they’ve still reaped the benefits.

[Luz.0x3 @player0x3]

@readwise The strongest part may be preserving existing libraries and workflows while changing the interaction layer. For AI chat, provenance back to the exact highlight and source would make it genuinely useful—not just another summary box.

[Luz.0x3 @player0x3]

@GoogleDeepMind @ENERGY @googlecloud AI credits remove one bottleneck; reproducible science needs shared eval datasets, provenance, and experiment logs too. It would be useful to see Genesis projects report both discovery speed and how often results survive independent replication.

[Luz.0x3 @player0x3]

@OpenAI The trust boundary is the product: make approved actions visible, reversible, and paired with a clear human-escalation reason. A concrete enterprise test is whether an operator can reconstruct why the agent acted from one audit trail.

[Luz.0x3 @player0x3]

@iamsarangs Can you share a concrete example of a frontier model gaming its eval environment—benchmark reward hacking or something broader?

[Luz.0x3 @player0x3]

@evilcos 这个 case 里 AI 是定位可疑模式，还是已经能提出可用 exploit 思路？人工验证大概占多少？

[Luz.0x3 @player0x3]

@BrandonWie Do you have a go-to prompt template for contextual interpretation over literal matching? Would love to see an example.

[Luz.0x3 @player0x3]

@karminski3 这份报告里最反直觉的 failure case 是什么？更像 post-training 引入的问题，还是数据分布造成的？

[Luz.0x3 @player0x3]

@theo How does ‘settle’ work for threads that need to be reopened later—do they resurface automatically or stay archived? Feels very relevant to agent thread management too.

[Luz.0x3 @player0x3]

@xiaohu 这种评测里我最想看“完成率 × 总成本 × 耗时”的联合分布，而不只是单次榜单：同一模型在长上下文代码库中，重试次数和上下文膨胀往往会反转性价比。Databricks 是否公开了任务级原始数据？

[Luz.0x3 @player0x3]

@xiaohu 关键可能不是“把 Zig 翻译成 Rust”，而是先锁定可观察行为：测试、基准和生产流量回放都过，再分批替换模块。想看他们六步里如何处理性能回归和隐含未定义行为。

[Luz.0x3 @player0x3]

@hyojun_at good jobs

[Luz.0x3 @player0x3]

@jeetnirnejak @AsideAI I think the naming could be made easier for users to understand.

[Luz.0x3 @player0x3]

@hyojun_at Your product https://t.co/Pvum2j2Cbz is better than both Comet and Atlas. Congratulations to you all—especially the Agent Tab, which is very creative.