Today, in collaboration with @Eigent_AI, we’re announcing AgenticRL-Bench, a benchmark that evaluates whether coding agents can diagnose and fix real incidents across the agentic RL stack (Miles, slime, verl, SGLang, Megatron, Harbor).
AgenticRL-Bench is built from 22 post-training RL tasks, and the Kaggle leaderboard scores agents on the 19 verifiable ones. These cover real agentic-RL engineering work, from debugging to training, in tasks such as Verl Upgrade Gate, Reward Determinism under Load, Trial Fault Attribution, and SQL Partial Credit Exploits. A task passes only if every hidden check passes, and each agent gets three attempts per task, run by Kaggle on Modal.
Reproducing AgenticRL-Bench surfaced an unusual pattern in capability benchmarking today: results often reflect frontier labs' risk tolerance as much as model intelligence. Seven of Eigent's 22 post-training RL tasks trip lab safety classifiers, a testament to the realism of the benchmark, and how often those classifiers trigger varies widely between labs, making it harder to measure true capability.
Current Leaderboard:
🥇 GPT-6.1 Sol — 22.8%
🥈 Gemini 4 Carbon — 19.3%
🥉 GPT-6 Astra and Claude Opus 5.5 — 14.0%
Check out the live leaderboard here: kaggle.com/benchmarks/eigent…

ALT A leaderboard graphic from Kaggle titled "Eigent AI AgenticRL-Bench". The ranking chart lists six AI models and their corresponding scores. The models are ranked as follows: 1. GPT 6.1 Sol with a score of 22.8% 2. Gemini 4 Carbon with a score of 19.3% 3. GPT-6 Astra with a score of 14% 3. Claude Opus 5.5 with a score of 14% 5. Gemini 4 Neon with a score of 10.5% 6. Gemini 4 Argon with a score of 7% The background features a subtle gradient with warm yellow and beige tones, and the source is cited at the bottom.
اولِ اَللهُ retweeted
We’ve closed our $5 Billion funding round co-led by Greenoaks and Sequoia Capital to build the future of medicine and bioengineering with AI. isomorphiclabs.com/articles/…
اولِ اَللهُ retweeted
From bakeries to vests, Pakistan's army now has its eye on the mines | After making and selling everything from bakery items to vests, Pakistan's military establishment has found a new source of economic and political leverage, the country's mineral wealth. Military chief Field Marshal Asim Munir has courted US President Donald Trump. He has also promoted American investment in Pakistan's mining sector. Together with the military's growing involvement in mineral deals, this points to a strategy that goes beyond attracting foreign capital. Last month's imposition of governor's rule in Khyber Pakhtunkhwa (KP), a province with considerable mineral resources, has added a contentious political dimension to this push.
📑 FULL STORY ➠ ow.ly/tYRf510882c
#Pakistan #World #Politics #Mining #Economy
اولِ اَللهُ retweeted
Keep your children STRICTLY away from this GTA-6 game.
This level of violence and brutality is not a joke. It will only create more psychopaths in our society, normalizing brutal violence and objectification of women. Unfortunate that most people playing this game and watching streams of this game are children.
Tem coisas que eu não concordo em jogos.
E essa cena de GTA, é uma delas.
Readers added context they thought people might want to know
GTA é um sandbox: liberdade não é incentivo. O jogo apenas reage às suas ações. Agir como um psicopata gera punição imediata, atrai a polícia e destrói sua reputação social. O mundo pune o crime e se adapta dinamicamente ao seu comportamento. É pura consequência.
pt.wikipedia.org/wiki/Grand_The…
اولِ اَللهُ retweeted
GPT-Live-1.1 is now available in the API.
It speaks with more natural and consistent accents across languages, supports lower idle costs with a new standby option, and completes tasks even more reliably.
اولِ اَللهُ retweeted
🚨یہ شخص پولیس سے بھاگ رہا تھا کہ اچانک اسے یاد آیا کہ یہ دنیا بھی ایک اوپن ورلڈ گیم ہے، جہاں کہیں بھی جایا جا سکتا ہے
Canada’s religious landscape set for major shift by 2070: Pew ctvnews.ca/world/article/can…
اولِ اَللهُ retweeted
I sat down with @OfficialLoganK from Google and he opened AI Studio live.
In one shot, he built a Minecraft-style game. A video studio. A native Android app. An ocean simulation. A Counter-Strike shooter.
Then he made an oil painting. 700,000 brushstrokes. All generated.
Then a live map tracking 16,000 objects orbiting Earth. In real time.
All with Gemini 4.
Then we got into the real stuff.
⏱️ CHAPTERS
00:00 Inside Google with Logan Kilpatrick
01:20 Intro
01:49 Logan's go-to test: "Build me a Minecraft game"
07:22 Why AI games matter
10:00 How Logan uses Google AI Studio every day
12:31 A native Android app in 23 minutes
14:02 Just-in-time software
15:37 An ocean simulation that's all code
18:00 An oil painting, 700,000 brushstrokes
20:14 Every satellite, live, from two prompts
22:25 Why you should scope the task tightly
23:56 A Counter-Strike clone in one shot
25:39 The big questions: what's your p(doom)?
26:32 The risk that actually keeps Logan up at night
29:00 What Google should be doing more of
31:00 Alignment, security and the Hugging Face incident
32:30 Why Google's Flash models kept winning
33:43 What changed with Gemini 4: the most ambitious
pre-training run
35:34 What really sets AI labs apart
38:21 Gemini 4 inside Antigravity and Google's products
39:04 "The model eats the harness"
43:04 From AI research labs to AI engineering labs
45:34 Research and engineering are blurring
46:37 Text diffusion and open research with Gemma
48:23 Is recursive self-improvement (RSI) already here?
51:02 How Logan actually uses AI
52:17 Why Logan still writes his own thoughts
56:24 Rapid fire with Logan Kilpatrick
1:02:15 "30 million AI engineering jobs"
1:04:31 Design taste and the comeback of generalists
1:06:30 One thing Gemini still can't do
1:07:17 We tested Gemini 4 Carbon on real projects
1:08:07 An animated mascot film from one picture
1:09:17 A 3D-printable bust from one photo
1:10:06 A Call of Duty clone from one line
1:11:17 The Feastables website from one link
1:12:30 A try-on app for clothes from one photo
1:13:30 My life story as a retro music video
1:14:35 A room redesign from five photos
1:15:47 An interactive textbook
1:16:34 Planning our next YouTube video
1:17:33 Verdict on Carbon
Bookmark this!
اولِ اَللهُ retweeted
Someone on Reddit built their own Netflix using Claude.
It’s called HolyArrs.
Most people who download and manage movies at home run six separate apps. One to find movies. One for TV shows. One for music. One for subtitles. One to search. One to request.
This guy replaced all six with just one 17mb app.
It connects to qBittorrent to download and Plex to watch. Everything in one dashboard.
In 8 days. 1,445 GitHub stars. 23,763 installs.
It also has an MCP server built in. That means you can literally ask Claude “download The Dark Knight in 4K” and it handles it.
Want something in 4K? It doesn’t just download it. It asks you first. So your hard drive doesn’t fill up overnight.
Still a beta.
اولِ اَللهُ retweeted
اسلام میں صوفیزم کچھ اور ھے اور پاک وھند میں اسکا کوئی نیا ھی ورژن پریکٹس کیا جاتا ھے۔"""""
#sahiladeem #sufism #islam #indo_pak #sacrifice
اولِ اَللهُ retweeted
Demis Hassabis: AI solved a 50-year biology challenge in just one year, saving a billion years of PhD research time.
“I think we'll be able to get down drug discovery from years to maybe months.”
“We managed to actually fold, using AlphaFold, in one year, all those 200 million proteins known to science.”
“So that's a billion years of PhD time saved.”
“We can expect to see some pretty dramatic health medicine breakthroughs in the coming few years.”
اولِ اَللهُ retweeted
“For those who desire the life of this world and its adornments, We will fully reward them for their deeds therein, and they will not be wronged in the slightest.”
Qur’an 11:15 — Surah Hud
اولِ اَللهُ retweeted
There is a town in Quebec where you can live your entire life without ever going outside. Your home, your school, your grocery store, your doctor, and your swimming pool are all in the same building. It is called Fermont, and it exists because the winter there will kill you.
Would you live in a town under one roof? Tell me in the comments?
اولِ اَللهُ retweeted
A China realizou o primeiro voo tripulado de um avião 100% elétrico. A aeronave CE-25A da COMAC decolou levando dois pilotos e um mecânico. O avião tem capacidade para 19 passageiros, autonomia de 400 km e velocidade de até 300 km/h. Antes disso, o avião fez vários voos autônomos.
Readers added context they thought people might want to know
O voo é o primeiro tripulado do demonstrador CE-25A da COMAC (após testes autônomos), não o primeiro voo tripulado de avião 100% elétrico do mundo (que foi em 1973 com o Militky MB-E1).
en.wikipedia.org/wiki/Electric_…
globaltimes.cn/page/202611/13…
aeroin.net/chineses-da-co…





















