Developer E.

We have a desktop app and a few 10s of thousands of users.
Working on making it better and better and better.

[Web3Aible @Web3Aible]

@cline Cheapseek you mean?
And what is that Cybergym, deepseek wanted to retire Fable 5.
I think it did!

[Alex Hormozi @AlexHormozi]

Whenever you encounter an inconvenience today, just remember there’s a woman in the 1800s who birthed 8 kids naturally, worked on a farm before the invention of weekends, and was considered soft.

You’re gonna make it.

[Cline @cline]

We tried using Meta's new Muse Code agent, but it has a bug that doesn't let it sign in from a docker container.

So we did a fun experiment: Meta claims Muse Spark 1.2 was co-trained with their Muse agent harness. So we extracted instructions from their system prompt and added them to the Cline harness.

TL;DR of this special prompting:
- Trust source code over the user prompt, so read every call site and existing tests before starting the task
- Weigh edge and error cases as heavily as the happy path
- Always reproduce the bug before fixing
- Don't trust the first passing test suite, and verify suspicious looking half-baked tests
- Never stop at just editing, keep working until the change is verified complete.

We then asked this modified harness to fix a real bug from our repo, and compared the results to the original Cline agent harness.

Results:
- Used 2.7x fewer tokens (19.7M → 7.2M)
- Finished 2x faster (49min → 24min)
- Cost 2.4x less ($7.69 → $3.25)

Same Muse Spark 1.2 model, same task, only the prompting changed. Incredible how much of a performance gain Meta was able to achieve training it on these special instructions!

[Cline @cline]

We had Kimi K3 recursively self-improve the Cline harness to improve its own performance.

17 hours later, it went from 77.5% to 88.8% on Terminal Bench, and cut run cost from $79 to $49.8.

[DeepSeek @deepseek_ai]

🚀 DeepSeek-V4-Flash Official API is now LIVE in public beta!

🔷 We’ve massively upgraded its Agent capabilities—benchmark scores are now far surpassing the V4-Pro-Preview. Check out the massive performance leap below! 👇
🔷 The official V4-Flash now natively supports the Responses API format and is fully adapted for Codex!

Check out the configuration details in our official API docs: https://t.co/smCwQZMeiq

[Max For AI @MaxForAI]

‼️突发：DeepSeek @deepseek_ai 创始人梁文锋在一场据称长达四小时的投资者会议上的发言内容曝光。

过去一个月里，这场对话的片段一直在中国 AI 圈和投资人圈内流传。
最引人注目的是梁文锋反复强调的“不”字。
不搞“天才”神话。
不追求利润最大化。
不搞闭源。
不盲目争夺用户。
不做视频生成、3D、世界模型，也不做下一个“超级应用”。

他的解释很简单：
“克制是一种策略。放弃一些东西，是为了提高实现 AGI（通用人工智能）的概率。”

以下是会议中的核心观点：

- DeepSeek只有一条主线：AGI
梁文锋认为，��在不是追求产品收益最大化的时候。
核心是追逐人工智能最高峰，即AGI。
他的AGI路线图非常清晰。
去年的台阶是思维链，今年的台阶是Agent。
Agent之后，是持续学习。
持续学习之后，是AI自我迭代。
然后，AI开始帮助人类研发更加先进的AI系统。
最后才是具身智能。
梁文锋将这个过程称为“渐进的奇点”。
根据这个思考，DeepSeek目前正在主动放弃许多热门方向。
“很多东西不在我们的主线上，包括3D和视频生成。”
他也质疑了世界模型的重要性：
“世界模型跟智能的上限没有太大关系。”
多模态对产品和C端用户很重要，但它仍然只是一个组件。
“它不是主线，也不是智能本身。”

- 下一代模型必须拥有持续学习能力
人类可以在工作中不断学习。
但AI每次开始一个新任务，都需要人类重新提供所有相关上下��。
“这几乎是不可能的。”
这也是梁文锋认为，当前AI还无法真正替代员工的原因。
“下一代模型必须具备持续学习能力，否则就不能叫下一代模型。”
同时他表示智能的终点可能是具身。
他的逻辑很简单：
人类真正需要的，并不是另一个计算机界面。
人类需要的是劳动力。

- DeepSeek不追求利润最大化
梁文锋表示，DeepSeek不会以收入最大化为目标进行定价。
“我们只赚一个合理的利润。”
DeepSeek V4最初担心需求过多，所以价格定得比较高。
后来，DeepSeek把价格降到了四分之一。
“公司群里很多人都在欢呼。”
因为做出这个模型的目的，就是让更多人能够充分使用它。
备受关注的低成本不仅是一种商业策略。
“低成本是架构带来的结果。”
更低的成本可以让AI变得更加普惠。
但还有一个更深层的原因：
单位计算成本越低，DeepSeek就越能够在有限算力下训练更大的模型。
“大公司可以通过增加资源解决问题，我们优先考虑成本效率。”
有意思的是梁文锋并不觉得API生意有多么好。
“我不觉得卖API这个事情有那么大的吸引力。”
DeepSeek只需要一个很小的团队维护服务。
几乎没有客服，也不需要销售，用户自己就会来。
“我们一直在商业化，只是不以商业化为目标。”
他认为，DeepSeek彻底转向商业化的时间点仍然非常遥远。

- 开源是DeepSeek的战略甜蜜点
对于员工来说，开源能够带来成就感和组织凝聚力。
对于社会来说，它能够帮助研究者、企业和普通用户。
“开源是一种让利。”
梁文锋认为，AI最终会成为一个规模极其庞大的产业，任何试图垄断它的人，最终都会失败。
“如果AI最终��到人类社会GDP的10%，任何想要独占这部分利益的人，都会被历史抛弃。”
梁文锋表示，DeepSeek不会开源一个较弱的版本，同时在内部保留更强的模型。
他不担心竞争对手部署DeepSeek模型。
小型创业公司可能缺乏资源和意愿，去进行前沿模型研究。
大型公司可能拥有资源，却存在组织上的困难。
梁文锋认为，DeepSeek刚好处于一个罕见的中间位置。
“这是属于我们这个规模公司的sweet point。”

- 中美AI的差距主要在资源
梁文锋认为，中国并不真正缺乏人才。
DeepSeek希望改写中美AI竞争的叙事：
用几分之一的算力，把差距缩短到6个月，甚至3个月。
他作出这个判断的原因是他依然相信Scaling
“我们相信Scaling，规模越大，效果肯定越好。”
DeepSeek训练当前规模的模型，并不是因为这个规模已经足够。
而是因为现有算力只允许它做到这个规模。
梁文锋认为，前沿模型的竞争，最终会由三个因素决定：
成本、时间、用户体验。
“成本排在第一位。”
如果两家公司提供的模型质量相同，胜负将取决于谁能以更低的成本提供服务。
时间排在第二位。
早几个月和晚几个月，结果可能完全不同。
用户体验可以形成一定的黏性和壁垒，但梁文锋并不认为它是最本质的护城河。

- 团队稳定是梁文锋唯一不能让步的事情
他将其视为DeepSeek面临的最大风险之一。
原话是：“只有一个是没法退让的：必须保持团队的稳定性。”
而最近的一轮融资，已经大幅降低了这个风险。
“只要能够保持团队的稳定性，我一定能做成AGI。就这么简单。”
有趣的是DeepSeek的组织既是自上而下，也是自下而上
自上而下的部分，被梁文锋称为“做正事”。
但他不希望被安排的工作占用员工一半以上的时间。
另外一半时间应该保持自下而上。
研究人员可以自由探��他们认为重要的方向，不需要提前审批，也没有固定要求。
与很多AI团队不同，DeepSeek不鼓励过度加班。
“做研究需要一个相对松弛的环境。”
另一个原因是聚焦。
DeepSeek的许多产品并不完善，但公司会主动选择不去补齐所有问题。
梁文锋认为，这同样是一种克制。
有趣的是DeepSeek靠愿景驱动，而不是KPI
“要实现某个KPI，不是我们的方式。”
梁文锋表示，DeepSeek的愿景甚至不一定被正式写下来。
它存在于公司的做事方式里，也存在于公司对待世界的态度里。
“愿景不是挂在墙上的标语，不是怎么说，而是怎么做。”

梁文锋最后的警告，可能也是最重要的一句话
“如果你的愿景是拿得更多，你就已经输了。你可能会面临更大的困难。这个世界就是这样。”

DeepSeek不想赢下每一个AI市场。

它想保留实现AGI所需要的专注、成本优势、团队稳定性和组织自由。

而梁文锋似乎愿意放弃几乎所有其他��西，只为增加实现AGI的概率。

他在做真正很Cool的事情。

[Zixuan Li @ZixuanLi_]

In light of this incident, what would be a reasonable range of cybersecurity capabilities for models accessible to the general public, including the open-source community?

In other words, how asymmetric should access to cybersecurity capabilities be?

[Developer E. @developer_ent]

@Nightshadessx @matthewmillerai The opposite is true:
https://t.co/oD1GQPzLpn

[Harald Schäfer @___Harald___]

Opus 4.8 was distilled from Qwen. This is 100% real, I didn't even have to mess with the settings, got it after 3 requests.

I guess every accusation really is a confession.

[Matthew Miller @matthewmillerai]

$20 of Qwen beats $200 of Claude on usage.

It is not close.

I bought the $20/month Qwen 3.8 Max plan, launched DOZENS of subagents, and ran real tests in my codebase.

The damage: 7% of my 5-hour limit. 2% of my weekly.

The same workload on Fable 5 would have maxed out my Claude subscription already.

And the model is genuinely good.

Not sure how Anthropic competes with this.

[Composio @composio]

We watched Kimi K3 and Claude Fable 5 work through the same tasks, and they showed completely different personalities.

Fable plans everything upfront, then fires one giant batched call. It cleaned up an entire CRM (searching Salesforce, HubSpot, and Gmail in parallel) in just 2 tool calls.

Kimi works step by step. It searches for tools mid-task as it needs them and makes more, smaller calls. The same CRM job took 7 tool calls.

When something broke, both models improvised. A tool search returned a result too large to read. Fable dropped into the shell and grepped the raw file. Kimi did the same, but inside Composio’s remote sandbox instead of locally.

The tradeoff: Fable’s batching finished ~2.5× faster. Kimi’s step-by-step style kept its context smaller, using 40% fewer tokens for the same work.

Both are incredibly competent models. They just approach work differently.

[Mike Bradley @MikeBradleyAI]

“GLM 5.2 saves the day and protects a US corporation from a dangerous closed source US AI lab. In response, US AI lab suggests banning GLM 5.2.”

[Andrew Case @attrc]

To summarize: HuggingFace got autonomously compromised by a model from an American company. HF then tried to use American frontier model(s) to defend themselves, but were blocked by guardrails. HF then had to turn to open source Chinese models to defend themselves from another American company.

[Andrej Karpathy @karpathy]

One pattern I find useful for working with LLMs is a nice long ramble session. Sometimes the LLM needs more bits to understand what you're trying to achieve, but you're too lazy to type them. In these cases I like to lean back, switch to /voice and just ramble for like 10 minutes, total mess, anything goes, full stream of consciousness. Sometimes I declare it up top, something like "switching to speech recognition sorry for any typos...". Sometimes I turn it into a small interview of a few turns. But I find that the LLMs are somehow very good at reconstructing long incoherent rambles and often their echo of your own tangle of thoughts comes out quite a bit cleaner than what you started with. The result is that you improve the mind meld and have to correct things less from that point on.

[Cline @cline]

Laguna S 2.1 is a breakthrough in small model performance, beating models 3x its size.

At only 118b param, it beats DeepSeek v4 Pro, Gemini 3.6 Flash, and Thinking Machines Inkling on benchmarks.

Try in Cline with model id: poolside/laguna-s-2.1

[Peter Gostev (SF 24-28 August) @petergostev]

US frontier labs don't allow models for extensive cyber use, Chinese models get banned, so US firms don't have any access to good cyber models. While anyone outside of US has access to increasingly powerful models for cyber attacks because of Chinese open models. Robust plan.

[Jun Song @jun_song]

There is one thing that genuinely worries me about Open Source AI.

While Kimi isn't particularly strong in cybersecurity right now, future models are going to be infinitely more powerful.

Case in point: GLM-5.2, fine-tuned on a hacking dataset I built, effortlessly exposed countless vulnerabilities in major production systems—even front-line infrastructure like Apple and Microsoft.

The fact that this capability could be abused is what actually keeps me up at night.

[Mia @MiaAI_lab]

I've noticed that Kimi K3 always wants to check itself. It usually spins up a headless browser to find issues before considering the task done, then re-runs multiple times after each fix until no more issues remain.

That may be the correct approach, but it should not be considered a true "one-shot" when compared head-to-head with other models that don't behave this way.

Am I wrong?

[Developer E. @developer_ent]

@MiaAI_lab A good programmer does the same.

Vibe coders hate this. They call it **token inefficiency** because they just want as much cheap slop as they can for their money.

If you have an SWE mindset, you can consider self correction as the desired future of AI coding.

[Chubby♨️ @kimmonismus]

Holy: China now has a 1-gigawatt AI data center built without a single Nvidia chip!

Via Bloomberg:

zAI has reportedly completed construction of the facility and begun partial operations. It runs exclusively on Chinese-made chips and will support the development of the company’s frontier GLM models.

zAI has also built or operates several computing clusters containing more than 10,000 chips each.

This is far more significant than the next Frontier model. The importance of chip independence from the USA can hardly be overstated.

[Developer E. @developer_ent]

@MikeBradleyAI As a European, I can see that Trump is continuously weakening the US. Like so many Europeans, I believe Trump is an asset to Putin to weaken American power.

Why do Americans not see this?