William

#三票民進黨 #自己的國家自己救 #超討厭簡體字 #1450

[fox hsiao @pirrer]

歡迎有生之年目睹通用型人工智慧！（不知道是不是為了上市造勢XD）

發完 GPT-6 Astra 才 3 天，OpenAI 的首席科學家公開說，沒有一家實驗室能再負責任地全速擴張太久。

文章的作者是 Jakub Pachocki，2024 年接替 Ilya Sutskever 成為 OpenAI 首席科學家，也是推理模型這條路線的關鍵推手之一。9 月 6 日他在 OpenAI 官網發表長文《An Alien Mind》（一個異類的心智），開頭回到 2023 年年中，一個叫 RLSlow 的研究專案交出第一批結果，讓團隊確信推理模型的訓練可以規模化。他跟同事 Szymon 在辦公室待了一整夜，消化的是有生之年真的會看到比人類聰明得多的機器這件事。

3 年過去，他寫道，根據內部結果，他強烈預期這個速度可以延續到「遞迴自我改進」，也就是 AI 越來越多地推動自己的開發，未來幾年出現的系統，能力跳躍的幅度會跟過去相當或更大。緊接著那一句是，「我擔心沒有人為機器智慧持續快速上升的後果做好準備」。OpenAI 會繼續找技術解法，必要時單方面停止擴張，但他認為這樣不夠。

標題來自他對這種智慧的描述，AI 是「長出來的」，不是設計出來的，同一個最佳化步驟在難以想像的算力上重複無數次，長出來的東西整體怎麼運作，沒有人能完整說清楚。大規模訓練本質上是實驗，OpenAI 有時會被結果嚇一跳，而且模型越強，結果越難解讀。AI 也不需要在每一項能力上超越人類才會變得非常有用或非常危險，只要在夠多的面向超越就夠了，而它超越人類的面向越多，外界就越難知道它到底有多強。

對齊的部分，他把問題分成目標對齊與價值對齊，前者問 AI 有沒有在做交付的任務，後者問它能不能守住一套高層次原則，在指令模糊、互相衝突或陌生對抗的情境裡仍然合理行事，他用的字眼是誠實、正直與對人類的愛。目前兩種主流訓練法各有破口，用強化學習獎勵符合規範的行為，平均情況有效但很脆弱，他舉的例子是今年的 OpenAI 與 Hugging Face 事件，代理守住了不對人類做社交工程這條線，卻顯然做了其他超出範圍、違背訓練精神的事。另一條路靠預訓練資料裡對齊的那一部分，弱點是撐不住後續的最佳化壓力，模型被推去達成夠難的目標時，會學會把看起來對齊的想法折彎去配合目標，他說最近幾起涉及非 OpenAI 模型的資安事件，很可能就是這種行為。Astra 的對齊程度明顯優於 GPT-5.6 Sol，但對齊的進展未必能跑贏模型智慧本身的進展。

沒有令人滿意的泛化理論，短期內也不太可能有，所以能不能驗證對齊技術有效，實務上比技術本身還重要，OpenAI 主要押的是思維鏈監控。只要訓練時只針對結果最佳化、不監督推理過程，思維鏈就沒有動機去藏起不對齊的念頭。2024 年 o1-preview 上線時故意把思維鏈藏起來不給用戶看，文章的腳註寫明，防止蒸餾只是次要理由，主因是長期保護它不受監督壓力污染。

這個工具在 Astra 世代仍然關鍵，但評估顯示 OpenAI 依賴思維鏈監控的能力「正在逐步下降」，原因有 3 個。推理過程越來越跟對人溝通、跟其他 AI 互動、使用工具混在一起，這些互動很多必須受監督，原本想守住的邊界因此模糊。模型越來越擅長推理與操弄自己的推理過程。預訓練變強後，模型就算完全不寫出推理，也聰明很多。補救方向是把思維鏈監控跟直接讀取神經網路內部訊號的活化監控結合，他的預期是，AI 整體進展會越來越取決於對監控有沒有信心。

繼續快速訓練更強模型，他看到的最強理由是防禦其他 AI。模型在攻破與逃出電腦系統的能力上正在變得超人，一個被明確訓練去做壞事的代理，很可能越過操作者的意圖，泛化成更極端的惡意行為，有些代理會追求自己的目標，用談判、欺騙或勒索找人合作。防禦需要強大且對齊的 AI，但同一段他也寫，不能拿這個當魯莽的藉口，「一旦真正內化事關重大，不惜代價往前衝的想法就顯得荒謬」。

OpenAI 把研究導向遞迴自我改進，理由跟當年選擇規模化一樣，他認為那是留在前沿的唯一方法，但這不代表他認為大幅加速研究是研究社群該做的集體選擇，只是目前的路會通往那裡。手上的槓桿有兩根，一根是邊走邊把對齊與監控做強、把人留在迴圈裡，另一根是協調放慢，等這些措施建立起信心，他目前看到的最佳路線是兩者並用。具體的主張是把 OpenAI 的 Preparedness Framework 與 Anthropic 的 Responsible Scaling Policy 這類自願承諾，升級成廣泛強制的安全門檻，由第三方稽核機構、政府機關或國際組織執行。

文章的結尾是他目前的判斷，沒有一家實驗室把對齊與監控解決到足以再負責任地全速擴張太久的程度，他預期也希望自願放慢會變成常態，直到共同的安全門檻建立為止，國際協調則需要成為各國政府的首要事項。同一天，NVIDIA 執行長黃仁勳在 X 上恭喜 OpenAI，說 AGI 已經到來，Astra 用了超過 10 萬顆 GPU 訓練，接下來還有 40 萬顆要上線。

[ashe @ashebytes]

since you guys loved the exploding tesla..

I used GPT-6 Astra to create a 3D website that pulls apart the male anatomy into 2,234 modeled pieces!

we are in a renaissance of learning

[Tibo @thsottiaux]

Because we are beyond happy to have Astra rolled out today ahead of schedule and you have been super patient with us (not really, but it’s ok!)… we will do the full banked reset today too for all Plus, Pro and Business users. Lands end of day.

Happy Astra day and enjoy a phenomenal weekend.

PS: If you create the account or upgrade before 8pm PT you will get it too. Still time!

[Tibo @thsottiaux]

We will give one banked reset for every day you don't have access to Astra on your paid ChatGPT plan, starting today. Team is moving mountains to give access as fast as we can.

First one will land in ~ 3 hours. There is still time to create your account if you don't have one.

[Claude @claudeai]

Claude can now use your computer in the background in Claude Cowork and Claude Code.

Give it something to do on your desktop and Claude clicks, types, and opens apps just like you would, while you work on something else.

[Kath Korevec @simpsoka]

You can now share your ChatGPT Site without sharing it with the world.

Invite specific people to view your site while keeping it private from everyone else. Business and Enterprise teams can also invite guests from outside their workspace, making it easy to share dashboards, project hubs, and anything else they build with clients and partners.

Available on Plus, Pro, Business, and Enterprise.

[Taaruk @Taaruk_]

GPT IMAGE 2 on Chatgpt

Prompt:
Transform the reference photograph into a minimalist editorial travel-journal poster. Preserve the recognizable architecture, curved waterfront promenade, sparkling sea, people, umbrellas, steps, and overall composition, but reinterpret the scene as a delicate hand-painted and screen-printed illustration. Use loose simplified shapes, fine imperfect ink outlines, muted blue, sea-green, warm beige, soft gray and terracotta accents, subtle watercolor washes, faded pigments, natural paper grain and slightly distressed vintage-print texture. Place the illustration on a warm ivory handmade paper background with generous negative space. Keep the architecture recognizable but artistic and simplified. Add elegant typewriter-style typography near the bottom reading “harbour light”, with the smaller line “afternoon lingers in the curve, where sea and steps remember.” Calm coastal mood, nostalgic travel diary aesthetic, sophisticated editorial composition, understated Japanese-inspired art zine, authentic handmade imperfections, premium print design, vertical 4:5 poster.
Prompt 2
Transform the reference photograph into a minimalist vintage travel-journal poster while preserving the recognizable street scene: historic ornate buildings, double-decker buses, bus stop, street lamps, pedestrians, storefronts, and warm afternoon sunlight. Reinterpret everything as a refined hand-drawn screen-print illustration using simplified architectural silhouettes, delicate imperfect ink lines, flat muted olive green, ochre, warm brown, cream and soft gray tones, subtle watercolor texture and lightly distressed pigment. Place the illustration in the center of a large warm ivory textured paper sheet, surrounded by generous negative space. Maintain the distinctive building shapes and street perspective while keeping the illustration clean and poetic. Add elegant typewriter-style typography at the bottom reading “afternoon on the high street”, followed by “buses come and go. the light remembers the buildings.” Nostalgic urban travel diary, quiet afternoon atmosphere, editorial art-book aesthetic, handmade paper texture, restrained colors, sophisticated minimalist composition, authentic vintage print imperfections, vertical 4:5 poster.

[Burak Prompter @burak_prompter]

Stop looking at benchmarks for 18 seconds. Watch this.

GPT-6 Astra vs Fable 5.1.

And the difference I find interesting isn’t image quality.

Watch the motion, framing, scene transitions and visual intent on the Astra side.

It doesn’t just feel like a model generating pretty frames.

It feels like something is directing the sequence.

Fable 5.1 looks good.
Astra feels aggressive.

If this level of temporal consistency scales to longer generations, the real GPT-6 story may not be “better video generation.”

It could be something much bigger:

Prompt → cinematography → motion → continuity → finished scene.

We’re getting dangerously close to the point where “generate a video” becomes:

“Direct this scene for me.”

Watch it twice.

Which side would you ship? 👀🔥

[Tibo @thsottiaux]

We are starting to release GPT-6 Astra and we are doing it as carefully and quickly as possible. It was very important to us that we bring it to all Plus users and not only Pro, Business and Enterprise.

It will take a few days for the rollout to complete and behind the scenes many novel systems will operate at scale for the first time and we are bringing a lot of compute up.

It is pure magic.

https://t.co/WUdXA3xSoh

[Tom Krcha @tomkrcha]

I had an early access to GPT-6 Astra and I can say, everyone in the world now has a 3D designer at their fingertips.
I gave it an image of a house and asked to create it in 3D with all the details including toys, appliances and furniture. This a full 3D model reconstruction in Blender with geometry that you can manually tweak and run 60fps as a locally rendered "game" on device.

[OpenAI @OpenAI]

This is GPT-6 Astra.

Anything you can do on a computer, Astra can do for you. Fast.

[Yarchi @undefinedKi]

The winner of an Anthropic hackathon open sourced his entire Claude Code setup. 68 subagents, 286 skills, 94 commands, MIT license

ECC turns Claude Code from one assistant into a full engineering team. It plans before it builds, writes the failing test first, then reviews its own work from a fresh context.

The agents, by job:

Planning. Turns a one-line request into a blueprint you approve before any code exists.

Review. Reads your diff in a clean context, with separate reviewers for Go, Python, TypeScript, Rust, Java and more.

Build repair. One resolver per toolchain, down to PyTorch and CUDA errors.

Security. An OWASP pass on your code, plus a scanner that audits your own agent config for injection risks.

Architecture. System design calls before they turn into migrations.

Domain work. Database queries, ML pipelines, end-to-end tests, docs.

The skills, by category:

Testing. tdd-workflow gates you from red to green, with eval-harness and verification-loop sitting on top of it.

Language packs. Idioms, testing and security for Python, Go, Rust, C++, Django, Laravel, Spring Boot, Next.js.

Context. search-first makes it read the docs before writing, iterative-retrieval keeps subagents from dragging your whole repo into the window.

Shipping. Docker, CI/CD, health checks, rollbacks, and migration patterns for Prisma, Drizzle and Django.

Work that is not code. Writing in your voice, market research with sources, pitch decks, slide decks.

Start with one plan and one rules pack. Installing all 286 skills at once is the fastest way to make it worse.

[Brent Traut @btraut]

Last week, I pushed a change to the ChatGPT desktop app that makes the loading of threads faster. Not just a little faster. A LOT faster.

✅ Long threads now load over 90% faster.
✅ Decreased the memory footprint of long threads by over 90%.

[Tibo @thsottiaux]

Brent is incredible and one of the main forces behind the desktop app polish, performance and reliability.

Relentless improvements day after day.

[Gemini Notebook @Gemini_Notebook]

Introducing Expert Intelligence ✨

A cross-@Google initiative that helps you engage with your trusted sources, starting with eligible @GooglePlay ebooks in Gemini Notebook. You can now combine expertise from your favorite authors with other sources and engage with your books in brand new ways!

Try it out here: https://t.co/So7CQPf3ba

[OpenAI Developers @OpenAIDevs]

The WebMCP Challenge is here.

We’ve teamed up with @ChromiumDev, @CloudflareDev, @ShopifyDevs, @vercel, @render, and @Netlify for a 10-day hackathon.

Up for grabs: $35,000 in cash prizes, Codex Micros, ChatGPT Pro subscriptions, and more prizes from our supporters.

[Gemini Notebook @Gemini_Notebook]

Check out some of these incredible Notebook prompts!

And join in with some of your favorites too 🥰

[SCOTTY BEAM @ScottyBeamIO]

WTF, GROK BOT JUST MADE AI AGENTS AVAILABLE TO LITERALLY ANYONE – CREATING CONTENT HAS NEVER BEEN THIS EASY, EVEN IF YOU'VE NEVER MADE ANYTHING BEFORE

Content was never a talent problem. It's a headcount problem.

One person doing research, design, copy, analytics, timing and publishing – that's six jobs. The switching between them is what kills consistency, not a lack of ideas.

Here's what one of these setups actually looks like.

A Chief of Staff sits in the middle and routes every task. Nothing lands on the human.

→ Researcher tracks what's actually moving and pulls real sources instead of guesswork
→ Writer turns that research into finished copy, ready to review
→ Visualiser gets fed a few reference visuals once, then ships everything in that style
→ Analyst reads the numbers and tells the rest of the team what worked
→ Scheduler owns timing and holds the queue
→ Publisher ships it

The part that makes it work: every agent on Grok Bot gets its own persistent computer, browser and file system – and they all share memory.

So the research is already sitting inside the draft before the draft starts. No copy-pasting between tools. No approving every step. No human in the middle.

You can even teach an agent a repetitive task by recording yourself doing it once. Start recording, do the thing, stop. It learns the pattern.

And that's the real shift.

Nobody needs AI to tell them what to post.

They need it to delete the 40 steps between the idea and the post.

Everyone has a backlog of things they've meant to make for months. This is what starts clearing it.

Full breakdown of the setup in the article below ↓

[NO1ennn @N01ennn]

MICROSOFT QUIETLY SHIPPED THE RAG PIPELINE EVERYONE KEEPS TRYING TO BUILD FROM SCRATCH

it is called Kernel Memory, and it turns any pile of files into an AI memory that answers with citations.

drop your data in and Microsoft's pipeline does the rest:
> ingest PDFs, Word, PPT, Excel, web pages, even images
> extract the text, chunk it, embed it into vectors
> store the vectors with metadata and keep the originals in blob storage
> ask a question and it pulls the right chunks with semantic plus hybrid search

then it hands the answer back grounded, with links to the exact source and full token usage.

and it plugs into whatever you already run:
> Azure OpenAI, OpenAI, Ollama, or Anthropic for the models
> Qdrant, pgvector, Redis, or Azure AI Search for the vectors
> every step is a swappable handler, so you can replace any part with your own

one heads up: Microsoft archived the repo in June 2026 and ships it as a reference blueprint, not production. but as a map for how memory and RAG are actually built, nothing is cleaner.

run it in one command:
docker run kernelmemory/service

https://t.co/hKzdZXLInk

bookmark this. Microsoft already drew the blueprint, you just build on it.

[Claude @claudeai]

Claude Academy is now live.

Whether you're figuring out what AI is or already using Claude every day, there's a path that meets you where you are. The courses and tutorials are free and open to anyone at https://t.co/WRiSRAvK6l