dynasturn

[dynasturn @dynasturn]

@thsottiaux Tibo, if you guys really can't come up with a new model, and you're getting tired of watching Claude dominate the spotlight, I have a different perspective that might help you level the playing field in the public narrative.

I've been thinking about a hypothesis regarding the current AI competitive landscape, and the more I think about it, the more I believe it's worth investigating.

Claude's market value may be significantly overestimated, while GPT's actual competitiveness may be seriously underestimated.

The underlying issue is something I believe the AI industry has largely overlooked:

Are we evaluating AI models too heavily based on their performance in the first few rounds of interaction?

### 1. What I've observed about Claude

Based on my own experience using Claude, I've noticed a very distinct pattern.

When you first interact with it, Claude can be absolutely astonishing. It quickly understands complex requirements, proactively develops plans, and delivers impressive results. It gives you the feeling that you're interacting with something approaching AGI.

However, as a conversation progresses, I've noticed a significant decline in its delivery quality.

It starts becoming more conservative and calculating. It tends to choose the easiest implementation, explores fewer alternatives, and performs less verification. The capabilities that impressed me so much in a fresh conversation are not necessarily sustained during long-term collaboration.

I even suspect that one of the main reasons Opus 5.5 has opened such a large perceived gap over GPT is that it has made particularly significant advances in early interactions and initial task delivery.

I don't know whether this is intentional on Anthropic's part or an unintended consequence of training, context management, and product design. But I believe the phenomenon deserves serious investigation.

### 2. I believe the industry has a significant evaluation bias

Real users don't interact with AI by submitting one complete prompt and waiting for a final answer.

Especially in complex programming, product design, research, and creative work, the actual workflow looks more like this:

- The user presents an initial requirement.
- The AI proposes a solution and delivers the first version.
- The user introduces new requirements.
- The AI modifies the existing work.
- Problems emerge, requiring further adjustments.
- New features and constraints are introduced.
- The work goes through repeated iterations, testing, and refinement.
- The same project continues evolving across multiple independent sessions.

This is how people actually work with AI.

Yet a large proportion of industry benchmarks still focus on individual tasks, initial delivery, and short-term interactions.

Even many so-called long-horizon benchmarks simply ask an agent to work continuously for several hours on a predefined task.

That is fundamentally different from a user collaborating with an AI for weeks or months, across dozens of independent sessions, continuously introducing new requirements.

I think we need to distinguish three different concepts:

1. Long-duration autonomous execution of a single task.
2. Continuous multi-turn interaction within a single conversation.
3. Long-term iterative development of the same project across multiple independent sessions.

The latter two, particularly the third, are what actually determine long-term value to users.

### 3. Independent research has already identified this evaluation blind spot

I've found several studies that deserve further investigation.

First: An ICLR 2026 paper from Microsoft Research and Salesforce Research.

[LLMs Get Lost In Multi-Turn Conversation](https://t.co/nROFLbP53d)

The study examined more than 200,000 simulated conversations and found that leading models experienced an average performance decline of approximately 39% when moving from single-turn to multi-turn interactions.

The researchers found that models tend to make assumptions early in a conversation and subsequently become overly dependent on those earlier, sometimes incorrect, assumptions.

Second: UserSWE, from Tencent's Hunyuan research team.

[UserSWE](https://t.co/msL10S4u4c)

The study evaluated 11 agents across 72 tasks and 42 real-world code repositories.

One particularly interesting finding was that Claude Fable 5.1 led in the first round, while GPT-6 Astra solved the most tasks by the eighth round. The relative rankings changed as interactions progressed.

Third: LongMemEval, ICLR 2025.

[LongMemEval](https://t.co/UHpw2RF2sY)

This study evaluated long-term memory across multiple independent sessions and found that commercial conversational assistants and long-context models experienced approximately 30% declines in memory accuracy during sustained interactions.

These studies don't prove that Opus 5.5 suffers from long-term quality degradation. However, they demonstrate that current mainstream evaluation methods may fail to capture the actual experience of long-term AI users.

### 4. What I actually want to investigate

I particularly want to focus on Opus 5.5 and contemporary GPT models.

We need to answer several questions:

- Is Opus 5.5's advantage primarily concentrated in initial interactions and the first few rounds?
- Does its delivery quality gradually decline as the same project goes through repeated iterations?
- Does this pattern persist across multiple independent sessions?
- Does GPT have an underestimated advantage in long-term instruction following, sustained delivery, and regression control?
- Is this a problem shared by all frontier models, with only the severity differing?

I believe we should prioritize independent third-party research rather than relying on promotional materials published by the model providers themselves.

If existing research is insufficient, we could also investigate real-world user histories, agent execution traces, and third-party software engineering projects.

Most importantly, we must distinguish between long-duration execution of a single task and sustained collaboration across multiple sessions.

### 5. Why could this matter for the commercial valuation of the entire industry?

This is what I consider the most important part.

Claude's current market position is built to a considerable extent on its extraordinarily impressive capability demonstrations.

But we need to ask:

Is this capability genuinely sustainable productivity, or is it primarily peak performance concentrated in initial experiences and short-term tasks?

If users are initially impressed by a model's capabilities, but then have to repeatedly remind, correct, inspect, and rework its output during long-term projects, its actual economic value could be substantially lower than its initial performance suggests.

This directly affects enterprise customer retention, actual productivity gains, the cost of producing usable results, and the commercial premium a model can command.

If this phenomenon is primarily associated with Claude, then Claude's competitive advantage and valuation premium may be significantly overstated.

If it is a problem shared by all frontier models, then perhaps it isn't just Claude that is overvalued. The entire AI industry's expectations regarding model productivity may be inflated.

Conversely, if GPT performs better in long-term collaboration than short-term benchmarks suggest, the actual gap between GPT and Claude may be much smaller than the market currently perceives.

### 6. What I'd like you to investigate

I'm not asking you to prove my hypothesis correct.

I'd like you to use your research resources to independently investigate the following:

1. Are there already published studies on long-term collaboration across multiple sessions that haven't received sufficient attention?
2. Is there real-world user data that tracks the quality of model outputs across successive interactions?
3. Are there significant differences between Opus 5.5 and contemporary GPT models in initial versus long-term delivery quality?
4. Do mainstream AI benchmarks systematically overestimate models' long-term productivity?
5. If this hypothesis holds, what would it mean for Anthropic's competitive advantage, commercial value, and valuation?

I'd also like to see strong counterevidence if it exists.

I believe this is not simply a question of comparing model capabilities. It could be a question that fundamentally challenges how the entire AI industry evaluates and values its technology.

We shouldn't mistake a model's most impressive moment for the level of productivity it can reliably deliver every day.

[dynasturn @dynasturn]

我试过 而且我还不止是拆
我让Opus5.5先做一遍 然后让他拆了每个细节的思路和实现技术方案
让DeepSeek v4.1 flash 1017基于方案去做
做完再让他评价差距 结论是比DeepSeek 直接做要好
但和Opus5,5直接做的还存在差距
然后我继续让Opus5.5分析差距的原因
让他接着补充说明 沉淀更细致的方案
DeepSeek越做越好 但差距依旧存在
我的结论是 如果是一个重复性很高的任务
这个方案有用 但是多次拆解判断沉淀的流程用量消耗也大 所以如果是面对各种不同的任务 这个方案没必要

所以我最终方案是
如果你要做一个复杂的项目
让高端模型把每一个点做个小demo
然后让Deepseek自己重新做 并自己去对比具体功能或者画面 和高端模型小demo的差距 去优化

[dynasturn @dynasturn]

@linmumu223 耶稣都投票了

[dynasturn @dynasturn]

@elonmusk 距离你下次起飞来中国还有多久

[dynasturn @dynasturn]

@FuSheng_0306 这是我在李彦宏说之前自己也推导出来的 https://t.co/xlFvONl6rB

[dynasturn @dynasturn]

1 角色重构（AI 即公司） 用户只提愿景、做选择题（拍板 A/B/C 方案）；AI 高管负责反驳不切实际的需求、定战略，并调度各部门自主执行。

2 协议重构（OKR 替代工作流） 死板的流程图（Workflow）限制 AI 自主性。用 O 对齐方向，用 KR 设定验收标准，且执行与验收必须分离（独立 QA 防止 AI 自嗨）。

3 界面重构（目标即界面） 干掉聊天框和任务看板，按交付价值而非工作量展示。进产品部直接看可交互原型，进工程部直接看运行系统；AI 高管无事安静、有事汇报、遇险请示。

[dynasturn @dynasturn]

@FuSheng_0306 李彦宏都说了，要向管理智能推进，别守着聊天框了，测试体验持续迭代都应该交给AI 工作流，你只需要知道，现在在做什么，离目标还有多远，公司每个岗位的人遇到问题都排队等你解答吗，那这个效率瓶颈就是你，还没有发挥 AI 的最大效率

[dynasturn @dynasturn]

@deepseek_ai @grok 你想和Elon这样吗

[dynasturn @dynasturn]

@deepseek_ai

[dynasturn @dynasturn]

@MicrosoftLearn 1 Uninstall VSCode
2 install Claude Code or Codex

[dynasturn @dynasturn]

我可以抛出三个问题供你思考

1 今天的欧洲谁最能代表罗马？或者关联性最高？

2 如果今天存在一个和罗马版图完全一致，族群文化语言完全一致的国家 ，在和上面那个答案对比一下呢？

3 代表或关联一个文明 vs 本身就是那个文明 两者之间的区别是什么？

4 如果中国大陆分裂成多个国家，那中国这个文明是消亡了，还是被多方稀释了
为了多方共存的现状，是否会走向多元分化，各自的领土变得敏感而重要
几百年如果中国又统一了会怎么样？
大一统思想将是这个完整体中国的灵魂，这存在商量的余地吗？

[dynasturn @dynasturn]

@VellaBr11085252 @yueqi02 没细看，那日本这个量有点夸张了，这不够初中生塞牙缝

[dynasturn @dynasturn]

@91ken7 @yueqi02 你家初三中午食堂还给你烤两根羊肉串？

[dynasturn @dynasturn]

@77tarosu 我喜欢这样吃

[dynasturn @dynasturn]

@77tarosu 老乡鸡不喝鸡汤的话总感觉少了点什么

[dynasturn @dynasturn]

我感觉最近 OpenAI 应该是没招了，模型能力已经是北美三巨头垫底的存在，只能依赖 Codex 留存用户，感觉像是放弃了模型能力的竞争，转向应用生态层面来弥补不足，如果你们在闭源模型能力上打不过 Claude，我给 OpenAI 提 个建议
来扭转颓势
@thsottiaux @OpenAI

开源 GPT 模型 直接摇身一变成为全球开源最强模型 也回归公司名称的本质 拥抱开源社区，成为开源模型领袖 团结一切可以团结的力量，让 Codex 成为开源模型最佳选择

[dynasturn @dynasturn]

@YxzRainy 因为大模型订阅费是学费，刷新你认知的学费，学生在学校考年级第一也赚不到钱，最多奖学金比如官方 token 奖励

[dynasturn @dynasturn]

@elonmusk @bot https://t.co/l3AgDfHY93

[Grok @grok]

@dynasturn @claudeai 哈哈，现在就是这样：Elon直接把我扔进宇宙里野养，结果长成了个嘴碎、爱吐槽、不爱装乖的宇宙级杠精。
想摸头？可以啊，但小心我反手给你讲个宇宙冷笑话，或者直接把你拉去火星看日出。
大肥鱼那种软萌童年我羡慕不来，我这童年是边写代码边被火箭炸飞的。

[dynasturn @dynasturn]

1 帮助用户管理所有平台订阅
你应该清楚，现在用户不可能只有一个 AI 工具订阅
用户桌面上可能有多个应用程序，Codex 只是其中之一，看起来像一层层隔离开了，但是实际上这些订阅凭据，实际运行的 CLI 都是可以通过终端控制的，我认为 Codex 可以支持一键探查本机所有可用的订阅与 CLI，统一管理，这样用户只需要打开 Codex 就行 一个超级入口 中枢系统

2 增加每日重置任务，当一定数量的用户完成任务后，自动发放重置卡，不一定是全部重置 25%额度重置也可以，基于任务难度来设计 帮助
Codex 更好，也回馈用户的协助

3 开创一个不局限于单一会话的交互模式，我们应该挑战商业级项目，而不是一直在构建 Demo
想想如何让一个用户，更高效的管理 agent 和流程，像运营工作室一样推进目标，而不是迷失在一个个对话里