AK

technology is the arms race. (de)generative ai and emerging tech. random thoughts and news.

[Anthropic @AnthropicAI]

Our internal data shows Claude is accelerating AI development���a possible path to recursive self-improvement, or AI autonomously building a more capable successor.

It’s happening faster than we thought, and the implications deserve greater attention. https://t.co/OVVPJO7VQx

[mrinank @MrinankSharma]

Today is my last day at Anthropic. I resigned.

Here is the letter I shared with my colleagues, explaining my decision.

[Pedro Domingos @pmddomingos]

AI career:
1. Get PhD from top school
2. Work for DeepMind/OpenAI/Anthropic for 6 months
3. Quit and start your own neolab with billions in funding
4. Fail to achieve AGI but retire rich

[alex zhang @a1zhang]

Much like the switch in 2025 from language models to reasoning models, we think 2026 will be all about the switch to Recursive Language Models (RLMs).

It turns out that models can be far more powerful if you allow them to treat *their own prompts* as an object in an external environment, which they understand and manipulate by writing code that invokes LLMs!

Our full paper on RLMs is now available—with much more expansive experiments compared to our initial blogpost from October 2025!

https://t.co/x47pIfIkTb

[MLB @MLB]

now pitching: Takashi Murakami! 🌸

[unusual_whales @unusual_whales]

Ikea says it plans to expand its price cuts into 2024 across all of its markets globally, aiming to dial back price increases that were introduced in 2022, because it says that "global inflation has eased," per CNBC.

[unusual_whales @unusual_whales]

99% of Americans live near a Mexican restaurant, per Axios.

[unusual_whales @unusual_whales]

Here we go again?

[unusual_whales @unusual_whales]

Aspartame, the sweetener in Diet Coke, is now classified as "possibly carcinogenic to humans," according to the WHO.

An adult would have to drink more than nine to 14 cans of aspartame-containing soda daily to exceed the limit and potentially face health risks.

[AK @AshuxK]

@emollick It’s very simple. Humans follow other humans.

[Ethan Mollick @emollick]

We know how social networks die. Slowly, and then all at once. Folks start to go, lowering the cost-benefit ratio for remaining members, setting up a cascade where departing users prompt their friends to go, too. Its likelier when there are alternatives.

I hope Twitter holds on.

[BayAI Circle-e/acc @GPTDAOCN]

AI Agents即将大爆发，OpenAI下一步

本文编译自 Lilian Weng 的个人博客，Lilian 现在是 OpenAI 的 Head of Safety Systems，之前还领导过 OpenAI 的 Applied AI 团队。

AI Agent 被认为是 OpenAI 发力的下一个方向。

OpenAI 的联合创始人 Andrej Karpathy 在近期的一次公开活动上提到“相比模型训练方法，OpenAI 内部目前更关注 Agent 领域的变化，每当有新的 AI Agents 论文出来的时候，内部都会很兴奋并且认真地讨论”，而在更早之前，Andrej  还评价 AutoGPT 是 Prompt Engineering 下一阶段的探索方向。

Lilian Weng 的这篇 Blog 可以说是目前 AI Agent 领域优质论文的系统综述，她将 Agents 定义为 LLM、记忆（Memory）、任务规划（Planning Skills）以及工具使用（Tool Use） 的集合，其中 LLM 是核心大脑，Memory、Planning Skills 以及 Tool Use 等则是 Agents 系统实现的三个关键组件，在文章中，她还对每个模块下实现路径进行了细致的梳理和说明。到今天，构建 AI Agent 的工具箱已经相对完善，但仍需要面对一些限制，例如上下文长度、长期规划和任务分解，以及 LLM 能力的稳定性等。

从今年 3 月 AutoGPT 推出后，Generative Agent、GPT-Engineer、BabyAGI 项目的爆发将 LLM 地叙事代入了新的阶段，从“超级大脑”到真正有可能成为Lillian 在自己的 Twitter 中也认为“This is probably just a new era”。

Agent System 是什么

在以 LLM 为驱动的 AI Agents 系统中，LLM 是代理系统的“大脑”，并需要其他几个关键组件的辅助：

1. 规划（Planning）• 子目标和分解：AI Agents 能够将大型任务分解为较小的、可管理的子目标，以便高效的处理复杂任务；• 反思和细化：Agents 可以对过去的行为进行自我批评和反省，从错误中吸取经验教训，并为接下来的行动进行分析、总结和提炼，这种反思和细化可以帮助 Agents 提高自身的智能和适应性，从而提高最终结果的质量。

2. 记忆 （Memory）• 短期记忆：所有上下文学习都是依赖模型的短期记忆能力进行的；• 长期记忆：这种设计使得 AI Agents 能够长期保存和调用无限信息的能力，一般通过外部载体存储和快速检索来实现。

3. 工具使用（Tool use）• AI Agents 可以学习如何调用外部 API，以获取模型权重中缺少的额外信息，这些信息通常在预训练后很难更改，包括当前信息、代码执行能力、对专有信息源的访问等。

[Ethan Mollick @emollick]

Big new AI thing: Microsoft Bing (which uses GPT-4 in Creative Mode), accepts images as input.

The results are impressive. I fed it a meme, it could understand context & read text! A new dimension of AI use just opened up. So expect a flood of AI Twitter influencer threads...

[AK @_akhaliq]

GPT Engineer

github: https://t.co/ILUvtZHzKD

Specify what you want it to build, the AI asks for clarification, and then builds it.

GPT Engineer is made to be easy to adapt, extend, and make your agent learn how you want your code to look. It generates an entire codebase based on a prompt.

[unusual_whales @unusual_whales]

BREAKING: The International Monetary Fund is working on a platform for central bank digital currencies to enable transactions between countries.

[Ethan Mollick @emollick]

The key to creativity is the smell of cinnamon*

*Not really, the real finding is cooler: Folks were told to smell a vial of cinnamon & 50% were told it was a creativity drug. The placebo boosted creativity! Why? People needed permission to be creative & reasons to feel competent

[AK @_akhaliq]

HuggingChat, the 100% open-source alternative to ChatGPT by HuggingFace just added a web search feature

Link: https://t.co/vd0EZDusvY
GitHub Repo: https://t.co/FbKyUQCdPV

[Nick Frosst @nickfrosst]

The term “hallucination” for large language models seems really misleading.

All ungrounded generations are hallucinations, just sometimes you get lucky and what comes out of the model is true

This is why retrieval augmented generation is so crucial for making this stuff useful

[Ethan Mollick @emollick]

This is a discussion of why it is so hard to build production software on LLMs: they are slow, weird, make errors & are inconsistent.

Thinking of AI as software results in disappointment. It is usually more helpful to think of AI as acting like a person (even though it isn’t).