Binx @BinxNet

I reverse-engineered a jev-like architecture given its type. You can find the repo here to train your own jevlikes: github.com/vinnylarouge/j…

Diogo Almeida @CompleteSkeptic

a day ago

After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI? I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev • 20-200x faster • 40-400x

3K 6K 60K 23.3M 50K

wow insane, they are literally livestreaming the RL training run of Mimo V2.6 Pro (1T, 42B active) and Flash (309B, 15B active) with per batch data/harness composition and a ton of internal training metrics mimo.xiaomi.com/rl/

Fuli Luo @_LuoFuli

5 hours ago

Nearly half a year of silence. We spent it studying one problem: how far RL can scale. MiMo-V2.6 is in the middle of its RL run right now. Three things we scaled: compute (~2B tokens per step, 1568 prompts × 16 rollouts, fully async), environments and harnesses (multi-task

168 369 4K 529K 2K

I've attempted to map everything in oncology in a public website and open source repo for all. The site is trying to get all cancers, products, technologies, bottlenecks, people, startup opportunities with 1000+ ideas for upgrading the field. If you have a loved one with cancer and you are technical or can engineer, go take a look, file improvement requests or bugs or help with the open repo and make this the best open and free info resource for individuals, researchers and educational use. This should save people time, aid AI oncology projects and generate positive action. If you are not technical just complain in this thread about broken or annoying or things you want and I'll fix them live. Some of the interesting pages: Treatments: onco.cc/drugs A gallery of the molecules being used onco.cc/molecules/ And targets: onco.cc/targets/ The technologies in oncology: onco.cc/technologies/ 1100 ideas for helping oncology: onco.cc/ideas/ Bottlenecks on oncology: onco.cc/bottlenecks/ Open questions (LETS GO RESEARCH PEOPLE) onco.cc/open-questions/ Startup requests (LETS GO STARTUP PEOPLE) onco.cc/startup-reques… Mechanics of cancer: onco.cc/mechanics Battlefronts: onco.cc/fronts/ Isotope supply: onco.cc/isotopes Key papers: onco.cc/key-papers Pipeline funnels: onco.cc/pipeline Cancer by type: onco.cc/cancers/ Institutional rankings: onco.cc/institutions/ The startups: onco.cc/startups/ Heros and heroines : onco.cc/heroes/ Key medical people: onco.cc/people/ Here is the project roadmap: onco.cc/roadmap/ Models and data sets: onco.cc/models/ There are other views as well, take a browse. Try making a PR if you have an upgrade to this on the repo here: github.com/judegomila/OnCo If you are biologically/medically minded and something is wrong, file a bug as well or say on the thread and we will get it fixed live. If this is a useful project star the repo and help get it calibrated. I've tried to add some other languages but I cannot speak them so tell me if that doesnt work well. I believe we will crack oncology and having total information dominance is key to the problem. Let the feedback flow!

keyboard_arrow_left Previous keyboard_arrow_right Next

Jev is cool not because it re-invented classification, but because it makes ARBITRARY classification into a type-safe programmable primitive. A general purpose zero shot decision model whose native interface is RUNTIME-DEFINED typed decisions, optimized for that exact interface

cocktail peanut @cocktailpeanut

6 hours ago

If you called Yann LeCun an idiot for saying we need to move beyond LLMs and build something new, you are banned from using Jev.

3 4 76 18K 4

i built computer use using @typesafeai ! it is 155x cheaper than opus 5, ~20x faster, and generalizes across OS's more on how it works in the vid & thread below:

Diogo Almeida @CompleteSkeptic

a day ago

After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI? I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev • 20-200x faster • 40-400x

3K 6K 60K 23.3M 50K

Astra ✅ Fast ✅ Frontier ✅ Efficient ✅ For everyone

Jasper Dekoninck @j_dekoninck

13 hours ago

We are releasing the latest version of BrokenArXiv and ArXivMath! These benchmarks now focus on conjectures that were refuted in the last month on ArXiv, and models are executed within a harness instead of directly via API. Performance remains impressive, with GPT-6 Astra on top

52 100 1K 1.1M 467

Here's the DeepSWE result for Union Alpha, a stealth model we just launched. Try it now! Works in every harness. $ ori [code | your-fav-harness] --model=stealth/union-alpha

OpenRouter @OpenRouter

9 hours ago

🥷 New stealth model: Union Alpha (@unionalphaai) A multimodal model for research, coding, and agentic workflows. - Free to use - 256K context - Tool calling - Frontier-level general-purpose performance Try it now and share your feedback: openrouter.ai/stealth/union-…

181 139 2K 688K 936

the bitter lesson is that most AI researcher spin-outs are basically just acquihire opportunities and their products and research are almost all completely worthless

Harsha Gundala @harshagundal

22 hours ago

They were building in stealth for 2 years, I was building in stealth for 2 hours… Happy to open source Qwen-2.5-1B-RLCD, 5x faster on-device inference for JSON workloads that need to be type-safe. ⚡️Demo below on a M4 MacBook⚡️ every LLM has the ability to efficiently batch

141 168 3K 464K 3K

Letting Astra add to the ASCII art for botslop.net

Interesting use of Jev , think worth more attention! Excited to find out exactly what this type of model enables folks to do (:

vogel @ryanvogel

20 hours ago

i built a codebase classifier with Jev and this might be the solution to overengineered code that agents create what should I test Jev on next?

43 38 984 73K 575

We can grow a tree from Number Theory🤩😍 Every branch begins with a Collatz trajectory, then symmetry does the rest.

the future will be built by observing things we’ve never observed before to do that, you need new observational tools (aka scientific instruments)

Markus J. Buehler @ProfBuehlerMIT

a day ago

We built recursive meta-intelligence, an AI that creates its own scientific instruments, turns them into persistent worlds that an agent ecology with hundreds of AIs inhabits, and uses those worlds to discover mechanistic principles in one of the hardest classes of physical

145 386 2K 349K 3K

Today we’re unveiling Odyssey-3, a big step forward for foundation world models. It can control robots, power humanoids, drive cars (on the roads of India!), train AIs, pilot drones, and even play video games. We can’t wait to see what intelligent systems it enables.

We midtrain + RL’d a trillion param LLM to analyze experimental data from our superconductor lab It’s better at it than Astra & Fable, and our scientists love it Read our first research report below

Liam Fedus @LiamFedus

a day ago

We built high-throughput materials labs in Menlo Park to create a loop between experiments and models. The labs generate fresh data, the models learn from it, and then help us decide what to try next. Using only 1,300 H200s, plus months of our experimental data, we mid-trained

266 496 5K 1.4M 2K

In addition to YuE2, we have 2 other new AI music models, DiffSynth Music and MuLaCover Both can take in reference audio and generate covers. IT NEVER ENDS modelscope.ai/models/DiffSyn… huggingface.co/HeartMuLa/MuLa…

Jev plays doom in realtime and it looks like an actual player. their numbers: 20 to 200x faster, 40 to 400x cheaper, output tokens free. Two things we might see now: - the opponent we're up against in a game might be a realtime AI. - a TA, or anything else we'd hire a real person for, can run on this instead. 10 calls a second is 36,000 calls an hour, and that costs around 7 dollars. anything that has to keep up with a person was off the table before this. can't wait to play against one of these. x.com/CompleteSkepti…

Diogo Almeida @CompleteSkeptic

a day ago

After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI? I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev • 20-200x faster • 40-400x

3K 6K 60K 23.3M 50K

在机器学习和信息论中，有一个著名的直觉叫流形假说。 现实世界产生的数据，在名义上往往拥有极高维度：一张高清图片可能对应数百万个像素维度，自然语言、声音、蛋白质序列、行为轨迹也都存在于极其庞大的状态空间中。 但有意义的数据并不会均匀散落在这些空间里，它会被压缩在一个极小、极其特殊的结构区域中。 一张随机生成的百万像素图片几乎不可能恰好构成一张真实人脸，因为自然图像同时受到身份、姿态、光照、透视、物体结构和物理规律等大量约束。 流形，只是这种低复杂度结构的一种几何表达，更一般地说，机器学习真正能够成立，是因为现实世界的数据根本不是随机噪声。 这也是理解智能最重要的起点。 智能存在的第一条件，在于这个世界必须具有可预测结构。 假如一个宇宙中的未来和过去完全独立，那么无论是人脑、Transformer还是无限大的超级计算机，都无法通过观察过去提高对未来的预测，因为这个世界根本没有留下任何可以学习的规律。 白噪声无法被理解，也无法被预测。 因此，智能能够存在，本身就意味着宇宙不是最大随机性的混沌，而包含大量跨尺度稳定的统计依赖、几何约束、组合结构、动力学规律和因果关系。 从这个角度看，机器学习做的事情是在巨量观测中寻找可以压缩的规律。 模型看过十亿张人脸，并不需要在参数里存储十亿份像素副本，它需要形成的是关于身份、轮廓、姿态、光照、空间关系等潜在变量的生成结构，最终用一个远短于全部训练数据本身的模型去近似。 因此，Learning在很深的意义上就是Compression。 如果一个庞大的数据集合能够被一个更短的生成程序描述，那么模型获得的就是规律。这与Kolmogorov Complexity的直觉高度一致。 生成或描述这些数据所需的最短程序究竟有多长，是一个重要的问题。 但是，数据不均匀本身仍然不足以产生智能。一个变量即使99%的概率取0、1%的概率取1，它的分布极度不均匀，却未必具有任何复杂结构。 重要的是变量之间存在可以利用的依赖。知道 (X) 之后，是否能够减少对 (Y) 的不确定性，这才是信息。 Mutual Information (I(X;Y)) 所度量的，本质上正是这种关系。 所以，智能依赖的是世界内部存在稳定的dependency； 统计规律、几何结构、层级结构、因果关系以及时间上的连续性共同构成了学习能够发生的基础。 这也给出了理解大模型的一种更准确方式。 一个随机初始化的神经网络最初只是一个巨大的、没有意义的函数族，而梯度下降不断利用数据排除那些与现实分布不一致的函数。 经过海量训练之后，原本随机的参数逐渐形成高度结构化的内部表征。 语言结构、空间关系、人物与概念关系、程序结构和部分物理规律也进入其中。 这里，我们不能简单说神经网络与现实世界同构。因为同构是一个过于严格的数学概念。 其实，模型是在自己的计算介质中形成了一套能够保存现实世界部分稳定关系的内部表示。 它是世界结构在另一种介质中的压缩投影。 于是，所谓智能涌现也没有那么神秘。 如果模型只能捕捉局部统计相关性时，它表现为模式匹配； 当它能够形成更长程、更抽象、更稳定的表示时，就出现泛化； 当不同领域的内部结构能够被重新组合，就表现为类比和联想； 当模型能够利用已有结构，对训练数据中没有直接出现过的状态进行推断，就表现为推理。 记忆、泛化、类比和推理，只是同一种机制在不同尺度上的表现：利用一个内部世界模型，在未观测区域重建结构。 模型规模之所以重要，是因为复杂世界需要足够大的函数空间才能被表示。 而某些能力的突然涌现，有时只是评价指标造成的阈值效应，有时也可能对应grokking或内部表示重组这样的真实非线性变化。 我想，我们需要更多地去关注参数空间在学习过程中发生了什么结构重组，使新的计算能力成为可能。 如果继续把这个问题推到信息论和统计物理层面，会得到更加统一的图景。 生命，依靠持续消耗自由能维持自身远离热平衡，而大脑和机器同样需要消耗能量进行信息处理。 由于Shannon entropy和热力学熵并不是同一个概念，我们可以更准确地定义，智能系统不断丢弃与任务无关的变化，同时保存对于预测和行动有价值的信息。 这与Information Bottleneck极其接近。 内部表示并不需要保存输入的一切，而只需要保留足够多能够预测未来、指导行为的结构。高级智能的标志，是知道什么必须留下，什么可以被遗忘。 由此，智能可以粗略地被写成Compression、Prediction、Control。 Compression意味着从海量观察中提取规律，Prediction意味着利用这些规律推断尚未看到的世界，Control意味着利用预测去改变未来。 一个有限的系统面对一个远比自身复杂的宇宙，通过观察世界、压缩世界、建立内部模型，再依靠这个模型预测并干预世界，这就是智能最一般的形式。

keyboard_arrow_left Previous keyboard_arrow_right Next

I told DeepSeek V4.1Flash to find a vuln in Samsungs sgpu GPU Driver 2 days later it found a Race UAF vuln that gives any unprivileged app arbitrary physical write.

Audio to Audio LoRa's able to be trained now. We basically have the ability to make edit audio models! @sin_ceriously Thanks for the initial training code github.com/filliptm/Comfy…

Some interesting results on algorithmic tasks.

keyboard_arrow_left Previous keyboard_arrow_right Next

Yifan Zhang @yifanzhang_

4 days ago

We are at the dawn of Superintelligence. Introducing the Recurrent Looped Transformer (RLT), We now have Transformers with Infinite Reasoning depth. From now on, we should pace progress at the Open Frontier of Superintelligence, Until Safe Superintelligence is achieved.

267 754 7K 1.3M 7K