Yuqi Wang

member of technical staff @RekaAILabs

[OpenAI @OpenAI]

The result is a faster, more natural conversation with ChatGPT Voice from the moment a session starts.

How we built it: https://t.co/TvTXbaMQ0M

[Reka @RekaAILabs]

Reka has signed the Open Weights and American AI Leadership letter.

We believe an open AI ecosystem is essential to advancing research, expanding access, and building a more innovative and resilient field. That's why we've openly released datasets like CS2-10k and Research-Eval, models like Reka Flash 3 and Reka Edge, and frameworks like Reka Quant.

https://t.co/7WDzElEvXZ

[Reka @RekaAILabs]

Today we're opening the WorldModelGym leaderboard 📊

Most world model evals ask one question: does the generated video look real? WMGym asks a different one. Give a model a menu of possible actions, can it correctly predict which one actually leads to the best outcome?🎯

We ran our own Dreamer-v3 across all four benchmark families as the first entry. It currently leads three of them — Meta-World, DeepMind Control, and Classical Control (where it hits 84% decision fidelity), and sits second on Atari.

We invite you to also enter your own model: you host a /score endpoint, we call it and compute the score ourselves.

Read our full breakdown and methodology: https://t.co/oqVMqJcHUT

[Ti Morse @ti_morse]

My first interview with @sama, Co-Founder of @OpenAI.

0:04 How to start a startup
3:30 Trusting exponentials
4:57 Operating in chaotic environments
6:12 Learning to enjoy painful experiences
8:03 Creating abundant intelligence
11:15 Keeping core suppliers on OpenAI’s timelines
12:10 Invention of the joint-stock company
15:30 The best CEOs aren’t sociopaths
16:46 We are in the singularity
18:09 AI authoritarianism vs liberty
19:24 Texting 300-400 people a day
20:32 Having a small number of deep beliefs about the future
21:41 Critical path
22:25 Thinking about what’s next
23:38 Getting on planes in marginal situations
28:16 Buying lots of compute
30:51 Ambition
34:00 Google shouldn’t have let OpenAI survive
37:54 Having his life shot through a cannon after the launch of ChatGPT
41:20 The growth of Codex
42:02 The Death Star tweet
44:46 Status games and desire to be useful
47:57 Not being ambitious enough on compute investments
50:26 Execution
51:45 Ask for what you want
54:11 First few weeks of OpenAI
55:15 Shutting down Sora to focus on Codex
57:44 Designing beautiful products
59:01 Getting addicted to TikTok
1:01:03 Inventing a new device
1:02:53 Try to get better at your strengths
1:05:57 Masa is an n of 1
1:07:11 Real trends vs fake trends

[Reka @RekaAILabs]

Before a single model weight is updated, someone has to answer a deceptively hard question: What's actually in your data? 🤔

Julian Geoffrey López and Fedor Zhdanov share how they think about composition, quality, and annotation, and why getting this right determines what your world model knows, and what it doesn't.

🎥 Watch the full version: https://t.co/h31Uclx08X
📄 Read more: https://t.co/BBily52xyx

#PhysicalAI #WorldModels #DataPipeline #AI #RekaLabs

[熊师傅 weight decay 了吗 @bigeagle_xd]

https://t.co/MWUVsw1Ypx

K3 发布前两天内部 vibe 出来的 win xp 模拟器，功能相当完整，有“千千静听”可以听歌，有“QQ2005”可以聊天，朋友们会在你的QQ空间里聊超女，新浪首页在报道神州6号…��

在内部传开的那个下午，站起身来，发现一半的同���都在用 XP，瞬间有一些恍惚，此情此景，仿佛回到了学校的微机课，又像是当年的网吧。我回想起 20 年前中考后的夏天，每天在家听着《宁夏》和《突然很安静》，在不舍得开空调的炎热的下午，在家里的老电脑上刷冒险岛……

20 年弹指一挥间，感谢 K3，能在这个剧变的时代，让我怀念一下年轻的自己。

[Reka @RekaAILabs]

@KonradJam leads the data platform team at Reka. His job: prepare hundreds of thousands of hours of video data for world model training, keeping pace with a research team that moves fast.

In this episode, he and Julian Lopez share how the data pipeline for world model training actually works.

Less than 100 people. Petabytes of data. No two days are the same.

📄Learn more about the work we do → https://t.co/9SE1HnO0mD

#PhysicalAI #WorldModels #DataPipeline #AI #RekaLabs

[Reka @RekaAILabs]

🤔What does it take to train an omni world model from scratch?

📽️Petabytes of video. 6⃣pipeline stages. And every improvement in data quality pays off twice when you're training a model that both generates and understands video.

Our Reka Labs' data team on how it works → https://t.co/BBily52xyx

#PhysicalAI #WorldModels #AI #RekaLabs

[Shilong Liu @Shilong_Liu_AI]

https://t.co/lkoQPTFVnq

[Reka @RekaAILabs]

Today Roberto and Marianna walk through PhysicalRealismBench-U — how the benchmark works, how they built it, and what VLMs still get wrong about basic physics.

🔗 https://t.co/nxvMIcfnwy

#PhysicalAI #AI #RekaLabs #ComputerVision #VLM #PhysicalRealismBench

[Reka @RekaAILabs]

In this video, @MateuszOnAI and @zaheri_h_r sat down to talk through two questions driving recent Reka Labs research.

1️⃣ Can VLMs understand physics the way we do?
2️⃣ And if an agent uses a world model to choose between actions, does it pick the right one?

Two benchmarks, two complementary questions about what it actually means for a model to understand the physical world.

Read more on the benchmarks here:
📄 PhysicalRealismBench-U → https://t.co/z6pPl2ly8S
📄 WorldModelGym → https://t.co/9n2ei3LuyY

[Lilian Weng @lilianweng]

new post on harness engineering for AI self-improvement: https://t.co/ZYvGfVs61k

It is hard to forecast how much the future of RSI will rely on harnesses. Likely harness engineering will evolve in the direction of self-improvement and enable auto-research, and, in turn, smarter models keeps harnesses simple.

Even when many harness improvement get eventually internalized into core model, the need to specify goals and context will not disappear.

[Reka @RekaAILabs]

Most evaluations of world models ask, "Does this look right?" 🌎

We built WorldModelGym to ask a different question: if an agent actually uses this model to make decisions, does it still choose well? 🏋️🏃‍♂️🏁🦾

A simulation can look physically plausible and still lead an agent completely astray. 📉⚠️

Stay tuned. More tomorrow at https://t.co/gaf6iEgEiw. 👀

[Reka @RekaAILabs]

🎮🕹️🖥️ CS2-10k is now available on @huggingface 🚀
600,000+ egocentric gameplay videos. 10,000+ hours.

Every frame paired with the exact keyboard, mouse, and 3D position data that produced it.
If you're working on world models, action-conditioned video generation, or egocentric navigation, this is ready to download and use today.

[Lilian Weng @lilianweng]

A super long overdue (3+ years?) post on scaling laws.

Compute is expensive. Scaling laws are a way to help us reason about the optimal compute allocation between data and model size before committing to a large run.

The post covers what scaling laws predict, how compute-optimal allocation works, why Kaplan et al. and Chinchilla disagree, and how data limits + fitting details make extrapolation tricky.

https://t.co/HP26eJvjHB

[Reka @RekaAILabs]

Training world models needs egocentric video and dense action signals, synchronized. That data is genuinely hard to find.

We built it from Counter-Strike 2 demos.

CS2-10k: 600K+ player-round videos, 10K+ hours, per-frame annotations — keyboard state, mouse delta, 3D position, camera yaw/pitch. All paired to the visual stream.

Why CS2 demos? Matches are deterministic replays. We can reconstruct first-person video and extract the exact control inputs that caused every visual change. No labeling, no estimation.

We're also releasing cs2-dem-renderer — the open-source pipeline we used to build it. Give it a .dem file, it outputs .mp4 + .parquet. Run your own dataset at whatever scale you need.

[Tianqi Chen @tqchenml]

We taught a brand-new mini-series this year at @SCSatCMU on Modern GPU Programming for ML Systems, as part of the ML Systems course, touching on fun questions like what data layout swizzling is, how to use 3D TMA, and state-of-the-art Blackwell programming. We released a curated online book based on the materials: https://t.co/5ZJg2lySNO check it out

[OpenAI @OpenAI]

Introducing LifeSciBench, a benchmark for measuring and improving how well AI supports real-world life science research.

Developed with 173 scientists from biotechnology and pharmaceutical research, LifeSciBench includes 750 expert-authored tasks across seven biological research workflows.

https://t.co/JTk0wXHFrT

[OpenAI @OpenAI]

Let’s talk about evals.

We’re always looking for better ways to measure and forecast model progress, especially as benchmarks get saturated or gamed.

@tejalpatwardhan, who leads our frontier evals team, spoke to @andrewmayne about why evals matter and what models need to be judged on next.

[Reka @RekaAILabs]

We just released PhysicalRealismBench-U — a benchmark for testing whether VLMs actually understand physics in programmatically generated videos, fully attributable.

This is an important step toward models that understand and generate physically realistic outputs.

Best result across 9 frontier models: 57.7% realism F1.

Read our blog post here: https://t.co/yOEyZ2CASA

Visit the benchmark: https://t.co/Z9gPr4CiYt