Manu Manu

GPT5 tomorrow ?

What ?
I'm from Paris and didn't know all that

I've read complicated explanations about how gerrymandering works—with terms like 'stacking + packing' & so on.

But here's a simple & effective graphic that tells you most of what you need to know.

Another try at LLM based on diffusion.

I really think there are applications for this kind of models.
Like for instance auto completion in a IDE.

seed.bytedance.com/en/seed_diff...

Since the Aider Leaderboard is no longer being updated, here’s the Roo Code Leaderboard.

roocode.com/evals

I'm always surprised when I see gemini costing this much.
Probably it's more verbose ?

I didn't know any of these tools, thanks, will give them a try

I believe one of the new mRNA vaccines was expected to treat colon cancer, and another had the potential to keep pancreatic cancer from recurring. That’s what they’re taking away.

🚀 Introducing Gemini CLI GitHub Actions: Automatically triage issues, review code, write custom prompts or tag @gemini-cli to have Gemini CLI act on your behalf. Time to level up!

Blog: blog.google/technology/d...
Repo: github.com/google-githu...

#GitHubActions #GeminiCLI #OpenSource

I was surprised by the cutoff in June 2024.
Is this usual that the training data is more than 1 year old ?

And also I discovered the developer role in the messages !

GitHub - openai/harmony: Renderer for the harmony response format to be used with gpt-oss Renderer for the harmony response format to be used with gpt-oss - openai/harmony

Reading this :
github.com/openai/harmo...

I'm surprised :
* The system prompt is actually the developer prompt, the 'real' system prompt usage is much more limited
* the cutoff is June 2024, more than one year ago. It sounds quite old.

Create personal illustrated storybooks in the Gemini app. Simply describe any story, and Gemini instantly generates a unique 10-page book with custom art and audio.

I also tested today to create a storybook with Gemini.
it's incredibly simple ( just ask for it in Gemini ), and well executed.

it's a small thing that gives Google more points with Gemini !

blog.google/products/gem...

OpenAI's Harmony

OpenAI's response format for its open-weight model series gpt-oss

github.com/openai/harmony

Today is incredible, at least for me :

- openAi open source models
- deepmind genie 3 model
- eleven music, the music model from elevenlabs.

And there may be more in this week !

Genie 3: A New Frontier for World Models Today we are announcing Genie 3, a general purpose world model that can generate an unprecedented diversity of interactive environments. Given a text prompt, Genie 3 can generate dynamic worlds...

Google released today Genie 3, an amazing 'world' model.
The quality looks very good.
It stays coherent.
It has memory.

Really incredible,

deepmind.google/discover/blo...

Open models by OpenAI Advanced open-weight reasoning models to customize for any use case and run anywhere.

OpenAI's gpt-oss-120b and gpt-oss-20b

They perform at the level of o4-mini and run on a high-end laptop.

openai.com/open-models/

pfff...I came here to try to beat you ( ie to post before you ), but failed :D

AI Music Generator | Free Song Maker & Music Creator Generate original, royalty-free songs with our AI music generator. Turn simple text prompts into custom music in seconds. Free song maker for musicians, producers, and content creators.

ElevenLabs released a music model.

Training data has been cleared.
An API is coming "soon".
Very good quality.

a new concurrent to Suno / Udio ...

elevenlabs.io/music

Incredible ratings !
And there is a reason for that!

I agree, I'm "just" on twitter and bluesky - and 2 are already a lot for me.
But even if I'd love to say bluesky is enough for me, way too much news is missed on bluesky.
And there are more vocal, anti AI people.

Pie chart. Title reads "2025. Game engines used to create 9724 GMTK Game Jam Games" Unity has 41%. Godot has 39%. Other has 11%. Game Maker has 5%. Unreal has 4%.

These are the engines used to make 9,724 games for the GMTK Game Jam in 2025!

1️⃣ Unity - 41%
2️⃣ Godot - 39%
3️⃣ Other / No Engine - 11%
4️⃣ Game Maker - 5%
5️⃣ Unreal Engine - 4%

Alibaba's Qwen-Image — a 20B MMDiT model for next-gen text-to-image generation. Especially strong at creating stunning graphic posters with native text.

🔹 SOTA text rendering — rivals GPT-4o in English, best-in-class for Chinese
🔹 In-pixel text generation — no overlays, fully integrated
🔹 Bilingual

Tencent Hunyuan releases open-weight of four compact models (0.5B, 1.8B, 4B, 7B)! Designed for low-power scenarios like consumer-grade GPUs, smart vehicles, smart home devices, mobile phones, and PCs, these models support cost-effective fine-tuning for vertical applications.

YouTube video by Game Industry Conference Physically Based Rendering: Putting Pieces Together - Paweł Stolecki || CD Projekt Red ||

Nice introduction (or refresher) to PBR: "Physically Based Rendering: Putting Pieces Together", Game Industry Conference 2024 presentation recording www.youtube.com/watch?v=OwyM...

Trump And Miller Compel Colleges Not To Enroll International Students Trump officials are using rules, policies and agreements to compel and discourage U.S. universities from enrolling international students.

Wow. Had not realized that Columbia’s settlement with Trump admin included a commitment to decrease international student enrollment. In what way does this help “combat antisemitism,” which is supposedly the administration’s objective?
www.forbes.com/sites/stuart...

GitHub - MetaStone-AI/XBai-o4 Contribute to MetaStone-AI/XBai-o4 development by creating an account on GitHub.

A new chinese open source model with very , very good performances :
XBai-o4

It's a 32.8B parameter model, and with the performance of o3 mini / Claude 4 Opus.

It's crazy that we can reduce the weight this much !

github.com/MetaStone-AI...
huggingface.co/MetaStoneTec...

A photo of an Apple keyboard with the alphabet keys changed depending on the frequency of letters in English in this case. the E, A, D, T, I, O, L and N keys are quite large, while Q, X, Z, K, and V are quite small.

Studies have found the most logical and fastest keyboard layout is not between key orders like qwerty and dvorak, but one that prioritises the size of keys depending on the frequency of those letters in the language being typed, regardless of the ordering.

So Mastercard is playing a role in steam and itch.io , but doesn't want people to know about it ?

Horizon Alpha has been upgraded to Horizon Beta.

openrouter.ai/openrouter/h...

Waoo , beautiful!