tahitimoon

@ScarletKc_ 是不是之前 skill 安装太多啦，有些 skill 会有负作用

GPT-5.6 Sol 我用下来还是觉得有点笨。 尤其是在理解真实意图、判断什么该做、什么不该做，以及自主把握任务边界这些方面，明显不如 Fable，很多时候依然需要我把要求拆得很细。 我通常使用的是 extra high effort，而不是 max / ultra，所以这也不算把它的推理能力完全拉满。 不过它的执行能力确实提升了不少，派发 subagent、拆分任务、协调分工都更成熟，复杂任务的推进效率高了很多。 感觉它更像是一个很强的执行经理，但还不算真正聪明的决策者。

@shao__meng 日常编码 high 完全够咯，Ultra 还是太慢了

这两天给 Tibo 设置了发帖提醒，紧盯着额度重置的消息 👀 因为知道这两天会重置，昨天就开足马力，蹬了大半天的 GPT-5.6 Sol Ultra 😁 蹬下来的感觉，嗯，确实强过我很多，也强过我以前合作过的大部分同事，开发同事就不说了，非开发也是如此。 对模糊信息的理解、预期的对齐、不同方向的权衡比较、信息源获取、思考、汇报、完整档案的输出等等等等都做的比人好太多了，活该我们失业啊 😓

@championswimmer Without open-source models, Cursor’s road ahead is much harder.

Cursor Composer 2 is fine-tuned Kimi K2.5 Kimi K2.6 can also be basically defined as the same thing? (Essentially sharing the same pre-trained base model lineage + new post training) Cursor hosts the model on Fireworks. So if you use Kimi K2.6 directly on Fireworks yourself, then you are sorted. What is the Cursor moat? Just that it has sold subscriptions and has some captive audience? And benchmarks below is a lesson that it is hard/impossible to beat the model maker themselves at fine tuning their own model (unsurprising).

@ValsAI so cool

DeepSeek v4 is now the #1 open-weight model on our Vibe Code Benchmark, and it’s not close. It leaves the #2 (Kimi K2.6) in the dust, and even beats out frontier closed source models like Gemini 3.1 Pro.

I’ve tried a bunch of agent memory frameworks, and honestly, Hindsight just works better. mem0 and memobase don’t even come close. It’s super easy to get started. Just a few lines of code, and self-hosting is straightforward. No messy configs, no need to deal with a vector database. Everything is built on PostgreSQL. Memory, entities, relationships, and vectors all live in the same database. Compared to that, self-hosting mem0 and memobase is a pain. Feels like it’s designed that way to push people toward their cloud services. github.com/vectorize-io/h…

@jeff180639 @vikingmute Linkup 实时性不足，作为搜索还不错。

@vikingmute Linkup 数据格式干净是真，但爬 X 稳不稳定没试过，这块 Jina 确实有优势

最近用了各种爬虫， 开源自建的 Crawl4AI / CrawlLee 服务用过 Firecrawl / Jina Reader / Cloudflare新出的 crawl 发现最方便实惠便宜的还是 Jina Reader，没有订阅，免费额度大，后面一直是按需付费，爬各种内容也稳定，尤其是 X 的帖子爬的效果也不错。别的感觉都一般，不稳定或者根本就被block，当然除了 Apify，但是太贵了。 自建的话效果最好的就是 Crawl4AI 不知道大家觉得哪个爬虫效果大家最喜欢？可以讨论一下

MCP and Skills are two completely different things to begin with. It’s wild that some people are already saying MCP is outdated.

If you’re currently using Cursor’s legacy plan, the $20 monthly subscription includes 500 uses of Opus 4.6. Be careful not to switch to the annual plan. Doing so will force your account onto the new pricing model (the new Pro plan), and there’s no way to revert back afterward.

A very long memory isn’t necessarily a good thing. LLMs are fundamentally stateless. Most so-called “memory” is just past conversations being stuffed back into the context window. It burns tokens and makes long-context degradation more likely. In many agents, “memory” is basically just a summary of past behavior. Compression is irreversible, so details will inevitably get lost. Expecting OpenClaw to somehow evolve intelligence on its own is probably wishful thinking.

The new models are increasingly evolving toward agents. Earlier, Grok 4.2 Beta pushed the multi-agent approach pretty aggressively. Surprisingly, the results were actually quite good.

No one’s using the product? Add more features. Still no users? Add even more features. And in the end, it’s still ignored. It’s easy for developers to get trapped in their own self-created illusion. Time to wake up.

I’m honestly impressed by how real the Reddit community feels. One of my posts even made it to the top of the community. The comments weren’t just a handful of replies either. People took the time to share genuine advice and thoughtful feedback.

Embedding models aren’t as plug-and-play as LLMs. If you upgrade to a new embedding model, you usually have to re-embed all your existing documents. Once you’re dealing with a large corpus, the cost can be brutal. Rerank models are different. When a new version comes out, you can typically switch over immediately. Much more plug-and-play.

@kylegawley With the right guidance, the results are actually pretty solid.

agents are terrible at fixing problems they just keep shipping more shite on top of it's own garbage code problem appears solved, but 3 new problems created equality as frustrating as fixing bugs manually

@joni_vrbt nothing？

USA has ChatGPT USA has Grok USA has Claude USA has Gemini USA has Llama USA has Copilot China has DeepSeek China has Qwen China has Ernie China has GLM China has Kimi China has MiniMax Europe has?

Earnings calls are way more interesting than reading the actual financial reports. At least you’re not drowning in dense numbers and obscure metrics.

Gotta say, Anthropic is seriously talented. This cracked me up 😆

@wholyv It’s kind of weird. Gemini scores high on benchmarks, but it just doesn’t feel that great to use.

yea I was so wrong. Fuck you gemini. whatever the shit these benchmarks say, Opus 4.6 is by far the best. in almost everything. heavily disappointed.

@JinaAI_ Pretty cool, my product is powered by Jina.

jina-embeddings-v5-text is here! Our fifth generation of jina embeddings, pushing the quality-efficiency frontier for sub-1B multilingual embeddings. Two versions: small & nano, available today on Elastic Inference Service, vLLM, GGUF and MLX.