تغريدة مثبتة
eSIMuse｜AI 通信指南
468 posts
eSIMuse｜AI 通信指南
@esimuse
AI 工具玩家 / 自动化实践者 分享 ChatGPT、Claude、Agent 和出海工具 顺手整理一点海外卡与长期接入经验 有相关需求可以聊 TG：https://t.co/MfpTTVeKYv
United Kingdom انضم Aralık 2025
480 يتبع611 المتابعون
@flowerykuni It’s funny — the charts and explanations I actually put effort into get ignored, but random casual posts somehow get all the traffic.
English
on twitter you can post the most random bullshit and get thousands of likes and thirty new followers from god knows where but the second you dare to post art you've spent hours on you get 13 likes and everyone but your friends is dead. your 20min doodles will get more attention
English
@tobiflamme Why are you so good at this? How do you do it so fast, my friend? x.com/esimuse/status…
eSIMuse｜AI 通信指南@esimuse
A lot of people still describe AI Agents as just “LLM + tool calling.” But once you actually run the workflow, you realize it’s not that simple. The user query is only the entry point. Behind it, there’s memory, context retrieval, task decomposition, skill routing, MCP/tool calls, execution, result collection, and final response generation. I broke the full chain down into 20 steps.
English
@WEB3_furture People keep calling everything an “AI Agent” now.
So I made a workflow map to separate real orchestration from basic tool calling. x.com/esimuse/status…
eSIMuse｜AI 通信指南@esimuse
A lot of people still describe AI Agents as just “LLM + tool calling.” But once you actually run the workflow, you realize it’s not that simple. The user query is only the entry point. Behind it, there’s memory, context retrieval, task decomposition, skill routing, MCP/tool calls, execution, result collection, and final response generation. I broke the full chain down into 20 steps.
English
我真的太喜欢这个负鼠表情包了。
所以我决定用它们来排一下我心里的模型 coding 能力：Claude ops 4.8、Codex 5.5、GLM 5.1、Qwen 3.7 max、DeepSeek V4 Pro，主打一个严肃排名，抽象表达。
中文
@RalphJing_ Fair point. I simplified the MCP part a bit so the chart doesn’t turn into a wall of spaghetti.
English
This is why half of the “AI Agent” demos online feel fake.
They show one tool call and act like the whole system is solved.
But the real pain is orchestration, context, retries, permissions, and making the final answer not look like garbage.
Chart is not perfect, especially around MCP, but the direction is right.
English
A lot of people still describe AI Agents as just “LLM + tool calling.”
But once you actually run the workflow, you realize it’s not that simple.
The user query is only the entry point.
Behind it, there’s memory, context retrieval, task decomposition, skill routing, MCP/tool calls, execution, result collection, and final response generation.
I broke the full chain down into 20 steps.
English
@kimmonismus I made an April 2026 AI company tier list.
Some placements may surprise you.
Full chart below 👇 x.com/esimuse/status…
English
Opus 4.8 is a solid jump over Opus 4.7 on DeepSWE, while also lowering the average cost per task.
However, GPT-5.5 xhigh still beats it by a pretty clear margin while being cheaper.
OpenAI has been cooking insanely hard with its models lately. Really excited to see what GPT-5.6 brings.
That said, I have to admit: I’m starting to really like Opus 4.8 as well.
We’ve entered a moment where both frontier labs keep shipping genuinely impressive models.
Datacurve@datacurve
Opus 4.8 is now on DeepSWE. On the default high thinking effort, it scores 6% higher than Opus 4.7 xhigh, while also lowering average cost per task.
English
@bindureddy I ranked the top AI companies for 2026.
Not by hype. Not by press releases.
By Agent capability, coding, and real-world adoption.
You’ll probably disagree with some of it 👇 x.com/esimuse/status…
eSIMuse｜AI 通信指南@esimuse
我重新整理了一张： 2026年4月全球大模型公司能力梯队图。 说实话，AI 圈已经不是“谁聊天更像人”这么简单了。 现在真正卷的是： 谁能推理， 谁能写代码， 谁能跑 Agent， 谁能进企业工作流。 这才是下半场。
English
OpenAI has an excellent chance at widening the gap
GPT 5.5 leads in coding and Opus 4.8 didn’t catch up
GPT 5.6 will be even better
English
说明一下：
这张图不是官方排名，
也不是任何机构发布的榜单。
是我自己基于公开信息、产品体验、行业观察重新整理出来的一张图。
肯定会有主观判断，
也肯定会有争议。
但我觉得这才有意思。
你觉得哪家公司被我排低了？
中文
我重新整理了一张：
2026年4月全球大模型公司能力梯队图。
说实话，AI 圈已经不是“谁聊天更像人”这么简单了。
现在真正卷的是：
谁能推理，
谁能写代码，
谁能跑 Agent，
谁能进企业工作流。
这才是下半场。
中文
T4 / T5：百度、微软、亚马逊、商汤、百川、Stability AI 等
这一层更偏现实世界。
政企项目、私有化部署、行业合规、C 端入口、垂直场景。
很多人看模型只看榜单，
但真正赚钱的地方，
往往不在榜单第一页。
而是在客户系统里。
中文
T3：Kimi、Mistral、MiniMax、NVIDIA
这一层更像专项怪物。
Kimi 长文档和推理很强。
MiniMax 多模态和语音视频很猛。
Mistral 继续走开源路线。
NVIDIA 不只是卖卡，而是在往模型生态里插。
它们不一定全能，
但刀口很锋利。
中文
T2：阿里、DeepSeek、Meta、智谱 AI
这一层最有看头。
因为他们已经不是单纯“追赶者”了。
开源生态、代码生成、多智能体协作、企业级部署，
都已经开始有自己的位置。
尤其 DeepSeek 和阿里，
中文区不能只当旁观者看。
中文
T1：OpenAI、Anthropic、Google、xAI
这一层基本还是全球最顶的玩家。
他们卷的不是普通问答，
而是长上下文、复杂推理、代码工程、Agent 闭环。
也就是说，
不是陪你聊天，
是开始替你拆任务、跑流程、做决策。
中文