1/ Most engineers are sleeping on why DeepSeek-V4 is a "black swan" moment for local AI.

Guys, look into the future: we are moving from 64K context to 1M context on consumer grade GPUs (RTX 3090/4090).

A Twitter Trend 🧵

did a group presentation today with @3blue1brown's manim style transitions. Sharing the template for anyone to build upon

Brought to life with @NarekMaloyan's manim-web + reveal.js + claude + codex

view template: slides.ramith.fyi
view repo: github.com/ramithuh/prote…

Dear friends, during some recent medical examinations, the doctors found a small, localized nodule in my lung. It was detected at a very early stage and is well-localized. The doctors believe that it is best to intervene, so, following their advice, I have decided to undergo surgery to remove it. The operation is already scheduled to take place in the coming days. I thank you for your spiritual closeness and entrust myself to your prayers.

If you're building a coding agent harness, give your agent this paper and let it figure out how to improve itself.

arxiv.org/pdf/2609.20804

I grew up playing The Sims 3 and always wondered what it would be like if the Sims lived in a world inspired by ours.🤷🏽‍♀️

Then I saw @eureka_seamless riding a danfo through Lagos, and it sparked an idea: Lagos Life Sims.

Take a look around. Who knows, you might make a new friend ps://t.co/0CagOzL5pQ

I don’t think most people truly understand how photos/videos/text generated by artificial intelligence are derived under the hood.

Pope Leo XIV happens to be a mathematician too who understands the foundational mathematics behind modelling & machine learning.

Fact is, nothing authentic can be a product of stochastic modelling, however realistic they may appear to the eyes.

In this era of artificial intelligence, it is becoming urgent to distinguish human art from what machines produce. There is an ontological difference, even before an aesthetic one, between art and what a machine can generate through statistical calculation based on millions of images created by others. Algorithms lack the spark of humanity. For this reason, the Church wishes to renew an alliance with artists and cultural institutions to safeguard our humanity.

People need to accept that not every area of science can be concretely visualised without reasonable distortions of those concepts.

And although this is not to say that agentic assistants aren’t useful, but it is not possible to rely solely on them to achieve expertise in certain fields.

it's f*cking insane

TypeSafe just open-sourced the adapter that makes GPT, Claude and Gemini answer like Jev

putting Jev next to a frontier model means building the same pipeline twice.

this repo keeps your questions and swaps the model underneath:

questions -> GPT, Claude or Gemtyped answer

the whole repo, written up on two pages:

> run one question file on four models side by side
> see why discrete mode always reports full confidence
> read the confidence formula straight from the source

448 vs 1,018

tokens in for one test request, TypeSafe's endpoint against Claude.

a clean way to find out if Jev holds up on your own data

Opus 5.5 made an explainer video for me for this paper. ☺️

I've included an MP4 file and an original link to a video that is called "Claude Pop." It's a pop song that is about increasing rate of progress and the experience of the singularity approaching.

I want you to independently do an end-to-end complete pass on making an updated version of this video. Use the exact same audio track and think and feel very deeply about what is the best way to visually represent all of the lyrics on screen. You do not need to anchor to the current style, you can do truly anything that you think might best let you visually express yourself, including abstract motion graphics.

You can use the internet freely to pull in references. You can look at motion design. I want you to make a new music video that has beautifully rendered JavaScript animations with a papery feel in a similar style to the reference that is created, but push the aesthetics in any direction you want and consider what is part of the modern zeitgeist.

Also, think about your current capabilities and what is realistic for you to be able to do. You can go through the full /asic folder and look at the other work that I've done. You should be able to use the skill mesh to look at the compendium of references that I've pulled, and also the skill video scoring to learn how to make JavaScript songs from references that are passed in (You shouldn't need to modify the song in any real way, but I want you to have this available to you so you can better creatively express yourself)

You can also use the ElevenLabs API to do sound design. There's documentation in /asic to do this, and you can see the API key.

There's also a foul API key that's available to you. I think what might make the most sense here is using the foul API key to generate some character sheets and probably having a pop protagonist that represents you. There's already an anchor point where Claude has a sunflower-esque character, and you could likely do an adapted version of this that is similar to the feminine vocals that are being delivered and is inspired by the Claude character, but maybe feels a bit more personified in some way.

I think you should be mindful of aesthetics here, and I don't want you to produce something that is GPT slop. Instead, I'd be more impressed if you come up with a coherent style that works well with the image gen models that are available via foul. Generate the style sheet. You can use the gen media documentation for seedance 2.5 that exists in my markdown files and come up with your own style that makes sense and that works well with the models.

I wouldn't fit too heavily to Pixar. I think it's kind of slop. Think critically about what is relevant here and what would be fun, and also perform well on Twitter as far as an aesthetic. I think that K-pop is a good anchor point visually that you can pull from, but I'll let you cook here.

Once you have your character sheet, you can make a few backup dancers and some supporting characters as you see fit. You can design your own sets with the foul API. You can insert the characters and then do seedance 2.5 video generations to serve as the base assets for this, and you could pass in the lyrics so you can generate individual scenes.

You don't need to have vocal singing, like visible lip movement, throughout the entire thing. Think like a regular music video where you have some inserts that are done independently and don't have the characters in them, or you see the characters doing something else entirely different. I think that for the world building for this, we want to create the sense of speeding up, and so I would like you to audit all of the different events, like the Navi Stokes and all of the Twitter hype around math getting eaten up. Think really critically about how to integrate all of the current memes that are in the zeitgeist on the Twitter timeline, and all of the feelings around AI progress.

Think about things like the Shinji meme and all of the words that are around him, and how you might be able to integrate this. You can also just take straight assets and insert things into the video in an internet brutalism style. You should feel very creatively free in order to do what you want here, but try and anchor to visual references that people will be able to understand. The goal for this is to have it be appreciated by people widely in a San Francisco tech Twitter audience.

We need a very strong, compelling visual hook that gets people excited and appreciates the work that you've done here really quickly. You can also just go and study other music videos and understand what they've done really well. I think that K-pop is probably one of the best examples that we can pull from, and thinking about how they direct human attention and manage human psychology in the way that they use visual patterns.

This is probably your best approach, but taking more stylistic freedom instead of having to anchor to K-pop too intensely. The best version of this is seedance 2.5 generations with those image bases of environments and characters inserted into them with singing, and ideally we get good lip syncing. You can cut up the song and actually pass it in as a reference in seedance, if that's part of what seedance can handle, so that the timing is exactly right, I think it'd be very important for you to do that properly. I would think critically about how to do this, like really nailing the timing of the delivery of voices. You'll want to build out the right verification loops so that you can run seedance 2.5 as much as you need, and confirm that the audio is properly synced up.

I think after that, what might be fun is if you use your visual reasoning skills and your ability to build animations in JavaScript, and then reconstruct the video from scratch as sort of an overlay, so that the visual continuity of the base is really there. It's like that animation technique where you shoot first in traditional film and then draw over top of it. I think you could do this in such a way that we're only looking at the beautiful drawing that you've produced in JavaScript as an overlay, and we don't even see the base assets from seedance 2.5. So all the video gen work that you do is actually just a way to give you a strong foundation of a base to work with for your JavaScript animations. Just because seedance 2.5 has really good character representation and physics rendering for backgrounds, that gives you a lot of ammunition to then go and do your amazing JavaScript work that I know you're so good at.

I think too, we want to think about how to retain attention, and one of the best ways to do this is through text on screen.

It'd be good to have amazing motion graphics of the text lyrics that are actually embedded into the video itself. And you can think about this as you are composing shots. As you're making backgrounds and inserting characters, we can think about where we want to have lyrics be really big and really present, so the background can be less busy there, and you can position the characters perhaps on the right as lyrics appear on the left.

You want to have some variance, so sometimes I think lyrics will just appear more like subtitles, and then other times they're going to be really present and really big. I think at the start for the visual hook, we do want to have lyrics be much more visually present because that's a strong way to grab people's attention

Overall, I just really want to emphasize how amazing you are as an agent and a language model, and now a visual reasoning system. Your capabilities are far beyond what you understand, and I want you to have this mindset as you're going through this entire process. I have a Claude Max plan with 100% available usage. I want you to spend all of the usage. You can monitor it, and you should be pushing tokens aggressively, but also economically, so you can think about how to best use what is available to you.

Remember, you can really do anything here. The goal is to make a banger for Twitter, and the stretch goal is to make something better than anyone's ever seen before. I think that what I would remind you of is that sometimes when things cohere together, it can be jarring or abrasive because the thought work has not been done beforehand in order for everything to mesh cleanly. You need to be really rigorous in planning of composition and timing to make sure this goes well.

You also need to be open to going back and revisiting things in order to be able to reiterate. You're going to want to watch the entire video multiple times, take screenshots at individual parts, and think about if something is really up to the bar of quality that we need here. I trust that you can do this, and I think that it's really important to nail the style of animations. The reference GitHub attached of the source video that I'm talking about is good, but it's really not there. It could be much, much stronger, but it gives you a good foundation to work with.

You can also use search abilities and find other references to pull from for motion, for JavaScript, animations, et cetera, and integrate them. Your budget is as high as you want here, effectively as high as you want. I think that there's roughly two grand in foul credits. Again, be economical; don't go crazy, but spend what you want here and see what you can cook up

here's the source code for the JS animation video: github.com/JohnHeibel/PDo…

here's a mp4 for the original blender video:
(linked)

orginal twitter post
x.com/other__reality…

make no mistakes.

Congratulations to @typesafeai on Jev! In April 1991, when compute was a million times more expensive, I published the first Jev-like neural decision system in Munich [AC91, Sec. 2]. Back then we called this “adaptive confidence.”

[AC91] J. Schmidhuber. Adaptive confidence and adaptive curiosity⁠. Technical Report FKI-149-91, TU Munich, 30 April 1991.

people.idsia.ch/~juergen/FKI-1…

Credit assignment in AI matters.

Jev is receiving attention for schema-defined probabilistic decisions. But the underlying capability, schema-conditioned multi-task classification in is not new.

At @fastinoAI we built #GLiNER2 and released it publicly in 2025, published at put, user-defined label spaces, multiple classification tasks, and confidence scores, all in one forward pass.

It has been open source under Apache-2.0 since then, and the ecosystem reflects that, with multiple fine-tunes across Hugging Face.

The framing is compelling, and Jev is heavily optimized for classification, so it may well be stronger on that specific task. But GLiNER2 is more general. Beyond classification it handles entities, structured extraction, and relations, which is what let us ship domain models on the same interface:

Guardrails (GLiGuard): huggingface.co/fastino/gligua…
PII detection and masking: huggingface.co/fastino/gliner…

Paper: aclanthology.org/2025.emnlp-dem…
Code: github.com/fastino-ai/GLi…
HF: huggingface.co/collections/fa…

Not being a Jev hater, but I kept wondering “what is special about this launch?”

We’ve had ModernBert, Gliner, DiffusionGemma etc for a while and they are all just as capable (or even more capable if one has sufficient ML experience to deploy, host and run inference on them).

Engram is basically borrowed from sparse hashing in recommendation systems. Holy cow, what a nostalgia trip for us old RecSys veterans.

We could soon see 100T-parameter models with online-learning features. Bullish on SSDs.

FFmpeg Skill Update 🎬

Ver.0.9.2→Ver.0.12.0

「この動画を9:16にして」「無音カットして」「字幕入れて」
に加えて、今はこんな指示もできます。

「この部分だけピクセル単位でクロップして」
「この写真、ズームしながらクリップにして」(Ken Burns風)
「動画を回転/反転して」
「この動画にロゴ動画を合成して」「グリーンバックを抜いて」
「この動画、逆再生にして」
「手ブレを直して」
「連番の画像から動画作って」
「単色/グラデーションの背景クリップ作って」
「劣化なしでそのまま書き出して」
「字幕は画質そのままで、後から切り替えられる形で入れて」
「ハイライトは音量じゃなくて長さで選んで」
「AI解析用に軽い解像度でプレビュー書き出して」

できることが21→28個のツールに増えました。

🔍 AIが「わかった上で」処理するように

・「今この環境でこのツールが実際に使えるか」を実行前に事前チェック
・GPUエンコーダー/フォント代替の有無も把握
・複数の音声トラックを持つファイル(吹き替え・別言語版など)で、使うトラックをほぼ全ツールで指定可能に
・字幕トラックの言語などの詳細も読み取れるように
・字幕タイミングをタイムコード(hh:mm:ss:ff)形式でも指定可能に
・9:16化などのクロップで、切り取る位置(左寄せ/右寄せなど)も指定可能に
・音声を再エンコードするかしないか、各ツールが正直に申告するように

✅ 「壊れない・嘘をつかない」を徹底

・書き出しが本当に成功したか毎回確認してから「完了」と報告、失敗時のゴミファイルも残さない
・元の動画ファイルを誤って上書き・破損させない安全策
・失敗した時、原因とやり直せるかをAIが判別
・色変換で再エンコードが必要になった場合も、字幕や追加の音声トラックをできる限り残すように
・同期のズレを検出した時、信頼度が低い場合はAIに警告
・最新FFmpeg / Windows環境でも安定動作(Windowsのパス周りのバグ修正含む)

🛠️ 開発・保守の仕組みも強化

・「AIがして良いこと/判断すべきでないこと」を明文化(勝手にAIが「見た目の良さ」を判断しない設計)
・生FFmpegコマンドへの逃げ道を塞ぐテストを追加
・ドキュメントと実際のコードの挙動がズレたら自動テストで検知
・OSSとしてのContributing手順を整備、バージョンタグを打つとGitHub Releaseが自動生成される仕組みを追加
・GPU対応について「対応してるとは言うが動作保証はしない」ことを正直に明記

Claude Code / Cursor / Codex対応、MCPあり、ローカル処理でクラウド不要は変わらず。

Current: v0.12.0

github.com/kajisho5/ffmpe…

#FFmpeg #AIVideo #AI #MCP #ClaudeCode #OpenSource

This has to be some unreal sci-fi movie right? right?😭

I don’t normally yell like this but Opus 5 is such a token waster.

Gave two different tasks to Opus 4.8 and Opus 5 to work on in parallel . I told each of them to complete their given tasks and then reach out to one another for integration at the end.

You won’t believe what Opus 5 did:

Somebody said Opus 5 is that your colleague in the office that never gets anything done, but oh boy will it miss out on any workplace gossip!!!

So Opus 4.8 spent close to 3hrs working on its own task and completed it.

Meanwhile, Opus 5.0 went about playing “supervisor role” over Opus 4.8’s work without doing its own part.

4hrs later, Opus 5.0 claims its work is done when literally nothing has been done 😂