kyegomez / OpenMythos↗

过去24小时AI圈挺热闹，几条值得说说的： 1. 一个22岁的辍学生Kye Gomez把Anthropic Claude的架构猜了个七七八八，搞了个叫OpenMythos的开源项目。不管猜得准不准，这事本身说明——大厂的模型护城河，可能比想象的薄得多。 2. OpenAI CFO想压一压支出，私下建议IPO从2026推到2027。另一边，Altman倒是挺放松，张罗办GPT-5.5发布派对，还在X上说Elon"想来就来"。搞得好像那场官司是别人家的。 3. Snap这周确认裁了大约1000人，CEO Evan Spiegel那句"AI能减少重复工作、提高效率"成了众矢之的。省了5亿美元的账怎么算都漂亮，但被裁的人不会这么觉得。这一波之后，大厂谁还敢把"AI提效"写进裁员公开信？ 4. 奥斯卡出了新规：用了生成式AI的作品不能参选表演和编剧奖。你可以用AI搞特效、做音乐，但台前的荣誉还是留给真人。界线划得挺清楚。 现在AI发展的速度，已经不是在跑，是在窜。技术快半步、监管慢半拍、就业掉一层——这三者怎么对齐，今年内必须给答案。

Matched through a repo-specific project phrase query.

【海外で話題】 Claudeの「考える仕組み」を 再現したOpenMythosが面白い👀 普通のAI👇 100層のユニークな層を積む →パラメータが爆増 OpenMythos👇 同じ層を何度もループ →パラメータ増やさずに思考を深化 これが「Recurrent-Depth Transformer」 の核心。 AIの仕組みに興味ある人は コード読むだけで論文1本分の学びがある。 https://t.co/TUAehz6wCC

Matched through a repo-specific project phrase query.

OpenMythos is getting attention: the creator of Swarms has released an open-source implementation of a possible Claude Mythos-style architecture, and the repo is already gaining thousands of stars on GitHub. https://t.co/erFBgFJFOZ Important disclaimer: this is not a leak. There is no confirmed public architecture for Claude Mythos. OpenMythos is a hypothesis, built from public information, papers, architectural trends, and educated speculation from people following the field closely. But the hypothesis itself is interesting. The author proposes a model based on a Recurrent-Depth Transformer architecture with MoE routing and adaptive computation. At a high level, the model has three major parts: 1. Prelude These are standard transformer layers. They process the input once and initialize the hidden state. 2. Recurrent Block This is the main idea. Instead of stacking many different transformer layers, the model repeatedly applies the same block N times. So the model’s effective depth comes not only from having more layers, but from cycling through a shared reasoning block multiple times. Each recurrent step can also use depth-specific LoRA adapters, meaning the base weights are shared, but each pass through the loop can still behave differently. 3. Coda These are the final layers. They run once after the recurrent loop and produce the final logits. The deeper idea is latent-space reasoning. Traditional chain-of-thought reasoning spends more tokens to “think longer.” A recurrent-depth model can instead spend more internal computation before producing tokens. That is a very different geometry of reasoning. Instead of making the model explain its thinking in text, you let it refine hidden states internally, step by step, before committing to an answer. If this works at scale, it points toward a future where “thinking time” becomes a controllable compute budget inside the model, not just a longer visible reasoning trace. Of course, this is still speculative. OpenMythos is not Claude Mythos. It is an open implementation of one possible architectural guess. But as a research direction, it is worth watching. Recurrent computation, adaptive depth, MoE routing, and latent reasoning are all plausible ingredients for the next generation of agentic and cybersecurity-capable models. The current implementation is around 770M parameters, but other developers are already starting to scale the idea and test whether the architecture holds up at larger sizes. Code: https://t.co/erFBgFJFOZ Beautiful hypothesis. Unproven, but technically fascinating. #AI #OpenMythos #ClaudeMythos #Anthropic #LLM #MachineLearning #Transformers #MoE #RecurrentTransformer #LatentReasoning #AIAgents #DeepLearning #OpenSourceAI #Swarms #ArtificialIntelligence

Matched through a repo-specific project phrase query.

Lots of new stars on these today 11.4k ⭐ OpenMythos > A theoretical reconstruction of the Claude Mythos architecture. 9.6k ⭐ obscura > The headless browser for AI agents and web scraping 5.5k ⭐ llm_wiki > LLM Wiki is a cross-platform desktop application that turns your documents into an organized. Links below

Matched through a repo-specific project phrase query.

我讲一下自己对当前和今后的AI产业发展的判断。 首先，绝大部分积极使用AI的用户，都只是把AI当成是一种玩具，而不是生产力——我不否认一些企业已经把AI投入生产，但这种投入是相当有限的，而且更多人的确是把AI当成玩具。这可以从Github上各种无脑跟风的热门AI项目可以看出——尤其是OpenMythos。(1/11)

Matched through a repo-specific project phrase query.

🚨 Dev solo de 22 anos quebrou o Claude Mythos em 2 semanas e publicou o OpenMythos no GitHub. Moat de IA saiu dos pesos. Agora vive em quem tem compute pra rodar inferência em escala. Governar capacidade que vaza por inferência virou jogo perdido. #IA #InteligenciaArtificial #AI #AgenticAI #GenAI

Matched through a repo-specific project phrase query.

@BrianRoemmele @KyeGomez @openmythos https://t.co/VVBxd0b9LM

Returned by a high-confidence repo query and contains a visible project phrase, but the exact URL, slug, or package name was not visible.