Gio Paik

Making AI Lovable 🥰

[김상훈 @aaksgaa]

D4D | Deploy for Defense Hackathon APAC - SEOUL 🖐

​AI 엔지니어, AI 빌더, 개발자, 디자이너, 기획자, 창업가, 연구자, 군인과 제대군인까지, 다양한 배경의 빌더들이 함께 현실의 문제를 해결하는 APAC 최초의 디펜스 스타트업과 빌더 중심의 해커톤 - 24시간 진행됨

@Form_young
@GioPaik
#D4D

[Gio Paik @GioPaik]

Because a software engineer is not a role that just involves writing code.

[Gio Paik @GioPaik]

😂

[Ethan Mollick @emollick]

(Sorry, after seeing so many of these, could not resist):

🚨 BREAKING: Google just dropped a NEW paper that completely deletes RNNs from existence.

No recurrence. No convolutions. Nothing.
Just one mechanism. And it’s destroying every translation benchmark on the planet.

The title alone is a flex: “Attention Is All You Need”

Vaswani. Shazeer. Parmar. Uszkoreit. Jones. Gomez. Kaiser. Polosukhin.

8 researchers. 1 architecture. The entire field of NLP will never be the same.

Here’s why this is INSANE
→ LSTMs took DAYS to train. This thing trains in 12 hours on 8 GPUs. 🤯
→ 28.4 BLEU on English-to-German. That’s not an improvement. That’s a MASSACRE. They beat the previous SOTA by over 2 points.
→ English-to-French? 41.8 BLEU. At a FRACTION of the training cost of every model that came before it.
→ They called it the “Transformer.” The name alone tells you they knew.

But here’s the part nobody is talking about

👇
They threw out sequential processing ENTIRELY.

Every other model on Earth processes words one at a time. This thing looks at the ENTIRE sentence simultaneously and figures out what matters.

It’s called “self-attention” and it’s basically the model asking itself: “which words should I care about right now?”

Every. Single. Token. In parallel.

Do you understand what this means?

Training that used to take WEEKS now takes HOURS.

Models that couldn’t scale past a few layers? This thing stacks 6 encoders and 6 decoders like it’s nothing.

And the multi-head attention? 8 attention heads running at once, each learning DIFFERENT relationships in the data.

I’m not being dramatic when I say this paper just rewrote the rulebook.

RNNs are cooked. 💀
LSTMs are cooked. 💀
The future is attention.
And attention is ALL you need.
Follow for more 🔔

[#CVPR2027 @CVPR]

#CVPR2026 takes integrity and quality of the peer review process seriously. This year, ACs identified reviewers who did not meet expected standards, by abandoning the review process entirely or submitting reviews of egregiously low quality, including generation by LLMs.

1/2

[AI Dance @AI_Whisper_X]

前两天看到 Dario Amodei 在一个采访里说，continual learning已经 solved 了。
我当时就觉得这个说法挺大胆的。然后今天刷到 Tianle Cai（蔡天乐，Princeton PhD，在 Google DeepMind、MSR做过实习，毕业后加入了 bytedance）写了一篇长文回应这个观点，读完觉得他提了一个很好的框架，简单翻译一下并聊聊。

先说背景。Continual learning 这个词在传统ML里，大部分人把它等同于解决灾难性遗忘（catastrophic forgetting），做这个方向的人一直被这个狭义定义困住。Tianle自己做了好几年 continual learning，他说每次跟朋友解释自己在做什么，对方都一脸困惑。不是朋友怎么样，是这个概念本身就模糊，continual learning、test-time training、self-evolving、lifelong learning，一堆词互相纠缠，谁也说不清楚边界在哪。
他给出的重新定义我觉得是这篇文章最有意思的地方：continual learning 不是一个具体方法（a point），是一个方向（an arrow）。

大部分人想到 AI 技术演进，想的是一个个离散的方法：pretraining、SFT、RL、agentic context management……每个方法贡献一块能力，拼在一起组成系统。但 Tianle 说不对，你不应该把 continual learning 当成跟 SFT 并列的一个“方法”，它是所有这些方法背后共同指向的那个目标。
这个目标是什么？他借用了 METR 的 task horizon 概念：一个模型能可靠完成的任务的时间跨度。你可以把它理解成衡量模型能力的北极星指标，类似摩尔定律里的晶体管密度（虽然更难测）。所有技术演进，从 pretraining 到 SFT 到 RL 到 agentic context management，本质上都在做同一件事：把这个 task horizon 往外推。
然后他用了《创新者的窘境》里 S-curve 的框架来解释这个过程，这个类比也很到位。每一代技术刚出来的时候看着不如现有方案，但最终会超越它。我们已经反复见过这个模式了，最近一波就是 Anthropic 主导的 agentic coding 进展。

用这个框架回头看，每个时代的“continual learning”其实就是当时突破现有技术天花板的那个新东西。当我们只有 pretraining 的时候，SFT 就是那个时代的 continual learning，它让 base model 能从短 context 里“学习”然后回答问题。现在我们有了 RL，agentic context management 就是这个时代的 continual learning，让模型能压缩信息、做笔记、把记忆延伸到 context window 之外。
所以 Tianle 给出的定义是：“旨在突破现有技术可行 horizon 的一系列努力。”（The set of efforts aimed at breaking past the feasible horizon of current techniques.）

回到 Dario 那句话。Dario 说 continual learning solved 了，其实他说的是 context management 加上工程优化，已经能把 task horizon 推到几周甚至几个月了。从这个角度看，他没错：人类水平的 continual learning 可能确实“解决了”。
但 Tianle 的反问很妙：黎曼猜想还没解呢。很多超长 horizon 的任务还远远搞不定。

说到底，Dario 和 Tianle 其实不矛盾。一个在说”人类水平的 continual learning 够用了“，一个在说”但 AI 的 task horizon 还有很长的路要推”。 革命尚未成功，同志仍需努力，Tianle 原文用的就是这句 hh。

[Gio Paik @GioPaik]

X의 자동 번역 퀄리티가 너무 높아서 이제 굳이 어렵게 영어나 일본어로 글을 쓰는게 무의미하게 느껴진다... 다만 여전히 면대면 대화에서는 외국어로 소통할 수 있다는 게 큰 어드벤티지인데, 언젠가는 말을 통한 소통도 실시간으로 편하게 번역되는 시대가 오려나

[Gio Paik @GioPaik]

@seagetch Live2Dベースで何かプロジェクトをやってみようかと思ったこともありましたが、悪名高いライセンスのせいで断念しました。そろそろSDKくらいは、もう少し自由に使えるようにしてもいいのではないかと思うんですけどね…

[Gio Paik @GioPaik]

GenAI detox evening.
Sanitizing my brain with some 🍸

[Gio Paik @GioPaik]

Tried GPT-5.3 Codex Spark — and it was insane. Now I don’t have time to waste scrolling short-form videos while my agents are working.

[Gio Paik @GioPaik]

@seagetch そうですよね。歴史でも作品でも教訓を与えているのに、また繰り返されるのは残念ですね。

[Gio Paik @GioPaik]

Lately, one of my small joys has been installing Claude Code for people around me and showing them how to use it. At first, they’re like, “What is this…?” But within ten minutes, their eyes start sparkling like a kid who just got a box of Lego.

[Gio Paik @GioPaik]

Saw a movie "Sirat" today, and it was definitely one of the best movies I've ever seen. It changed the way I see war, life and death.

Highly recommended — but not for those sensitive to stress. The film is intense and shocking.

[電ファミニコゲーマー @denfaminicogame]

最新AI技術を詰め込んだAI VTuber「しずく」の音声合成が使用できる「Shizuku Lab」発表。しずくのキャラクターおよび音声合成を無償で提供
https://t.co/HHoCV3aMIY

Discordを中心にコミュティを運営し、メンバーは自由にしずくを利用した創作活動が可能に。参加条件は「しずくが好きなこと」

[Claude @claudeai]

Announcing Built with Opus 4.6: a Claude Code virtual hackathon.

Join the Claude Code team for a week of building. Winners will be hand-selected to win $100K in Claude API credits.

Apply here: https://t.co/SkEg8Py1l2

[Gio Paik @GioPaik]

@seagetch しかも、すごく速いです。 Reasoning effort mediumだと思って気楽に使っていたらextra-highだったのを見てびっくりしました。

[Gio Paik @GioPaik]

Tried codex5.3 for 5 hours to redevelop all my projects I worked on last week, and it was much more than I ever expected. I am quite sure that 2026 will be the last year companies hire people to write code, except for a few companies in finance and national defense.

[Sam Altman @sama]

GPT-5.3-Codex is here!

*Best coding performance (57% SWE-Bench Pro, 76% TerminalBench 2.0, 64% OSWorld).
*Mid-task steerability and live updates during tasks.
*Faster! Less than half the tokens of 5.2-Codex for same tasks, and >25% faster per token!
*Good computer use.

[Gio Paik @GioPaik]

I don't understand why do we have to make all that redundant .mds (e.g., .gemini, .codex, .cursor) for those "super smart" Agents.

[Alexander Embiricos @embirico]

📣 Open call to agent builders: Let's read agent skills from `.agents/skills`, so people don't have to manage separate folders per agent.

Today we pulled the trigger for Codex to read `.agents/skills`. Goal is to deprecate `.codex/skills`.

Pls like/tag/RT for momentum.

[Gio Paik @GioPaik]

What I added to https://t.co/SDsYoOd47Z was extremely simple:
- All commits must be atomic.

- Commit messages must include both the what and the why.

- Important decisions should be documented as ADRs and linked accordingly.

[Gio Paik @GioPaik]

After actively working with agentic coding, I’ve started to think that Git might actually be the tool for AI—the tool AI can leverage most effectively.