{"id":1558,"date":"2025-05-27T14:02:06","date_gmt":"2025-05-27T14:02:06","guid":{"rendered":"https:\/\/musictechohio.online\/site\/why-this-ai-veteran-left-google-to-make-models-unconditionally-open\/"},"modified":"2025-05-27T14:02:06","modified_gmt":"2025-05-27T14:02:06","slug":"why-this-ai-veteran-left-google-to-make-models-unconditionally-open","status":"publish","type":"post","link":"https:\/\/musictechohio.online\/site\/why-this-ai-veteran-left-google-to-make-models-unconditionally-open\/","title":{"rendered":"Why this AI veteran left Google to make models \u2018unconditionally open\u2019"},"content":{"rendered":"<div>\n<p><b><a href=\"https:\/\/gradientflow.substack.com\/subscribe\">Subscribe<\/a>\u00a0\u2022<\/b><a href=\"https:\/\/gradientflow.com\/newsletter\/\">\u00a0<b>Previous Issues<\/b><\/a><\/p>\n<h3>Beyond Open Weights: The Path to Unconditionally Open AI<\/h3>\n<p><span style=\"font-weight: 400;\">While I routinely work with both proprietary LLMs and open-weights models, my heart lies with models that are open in the fullest sense. Very early on, I <\/span><a href=\"https:\/\/gradientflow.com\/open-source-principles-in-foundation-models\/\"><span style=\"font-weight: 400;\">noted<\/span><\/a><span style=\"font-weight: 400;\"> that for foundation models, \u2018open\u2019 must comprehensively cover not just weights but also data, code, and the detailed recipes crucial for genuine reproducibility. Measured against this standard, few initiatives have captured my attention like <\/span><a href=\"https:\/\/oumi.ai\/?utm_source=gradientflow&amp;utm_medium=newsletter\"><b>Oumi Labs<\/b><\/a><span style=\"font-weight: 400;\">. This public-benefit corporation champions \u2018unconditionally open\u2019 foundation models and the collaborative tooling to build them, recently demonstrating their prowess with <\/span><a href=\"https:\/\/oumi.ai\/blog\/posts\/introducing-halloumi?utm_source=gradientflow&amp;utm_medium=newsletter\"><span style=\"font-weight: 400;\">HallOumi<\/span><\/a><span style=\"font-weight: 400;\">, a claim-verification model that punches above its weight. Guiding this mission is CEO <\/span><a href=\"https:\/\/www.linkedin.com\/in\/koukoumidis\/\"><b>Manos Koukoumidis<\/b><\/a><span style=\"font-weight: 400;\">, whose journey through AI includes foundational research in edge intelligence at Princeton and MIT, pioneering NLP work at Microsoft \u2013 where he developed an LSTM-based ChatGPT precursor and an early RAG prototype back in 2016 \u2013 and scaling PaLM at Google Cloud with a 300-person team. What follows is a carefully edited excerpt from our conversation, delving into Oumi\u2019s vision for making open models the standard for serious, production-grade AI.<\/span><\/p>\n<hr>\n<p style=\"text-align: center;\"><strong>Enjoy these insights? Consider supporting our work. <img decoding=\"async\" src=\"https:\/\/s.w.org\/images\/core\/emoji\/15.1.0\/72x72\/1f64f.png\" alt=\"\ud83d\ude4f\" class=\"wp-smiley\" style=\"height: 1em; max-height: 1em;\"><\/strong><\/p>\n<\/p>\n<p><center><iframe loading=\"lazy\" style=\"border: 1px solid #EEE; background: white;\" src=\"https:\/\/gradientflow.substack.com\/embed\" width=\"480\" height=\"320\" frameborder=\"0\" scrolling=\"no\"><\/iframe><\/center><\/p>\n<hr>\n<h5><b>Company Structure and Team<\/b><\/h5>\n<p><b>What is Oumi and what does the name stand for?<\/b><\/p>\n<p><span style=\"font-weight: 400;\">Oumi is an open source AI lab. The name stands for Open Universal Machine Intelligence, with the tagline \u201cLet\u2019s build better AI and open is the path forward.\u201d It\u2019s structured as a public benefit corporation (PBC), which means it\u2019s for-profit but has a strong, legally binding mission to benefit the public.<\/span><\/p>\n<p><b>What are \u201cfounding scholars\u201d at Oumi?<\/b><\/p>\n<p><span style=\"font-weight: 400;\">Founding scholars are academics who were involved with Oumi in the very early days, some even before the company was officially incorporated. They have equity stakes and are more deeply involved than typical advisors. There are around 15 founding scholars with more collaborators joining as the project grows.<\/span><\/p>\n<h5><b>Open Source Vision for AI<\/b><\/h5>\n<p><b>What do you mean by AI having a \u201cLinux moment,\u201d and what does \u201cunconditionally open\u201d mean?<\/b><\/p>\n<p><span style=\"font-weight: 400;\">When we say \u201ctruly open,\u201d we adhere to the <\/span><a href=\"https:\/\/opensource.org\/osd\"><span style=\"font-weight: 400;\">OSI standard<\/span><\/a><span style=\"font-weight: 400;\"> which requires open data, open code, and open weights. But we go beyond this to \u201copen collaboration\u201d \u2013 making it easy for others to reproduce, extend, and contribute to making models better. If something is open but people can\u2019t push it forward, it doesn\u2019t help much.<\/span><\/p>\n<p><span style=\"font-weight: 400;\">Just as Linux became the foundation for operating systems, AI models should become a common utility that anyone can build upon. The community needs all the necessary pieces to replicate and improve upon the work without barriers.<\/span><\/p>\n<p><b>Why is this openness important for AI development?<\/b><\/p>\n<p><span style=\"font-weight: 400;\">AI has become the foundation not just for the tech industry, but for healthcare and all sciences. It would be a disservice not to make it a public utility that\u2019s easy for anyone to leverage and contribute to. The foundation models should be a common utility that benefits everyone.<\/span><\/p>\n<p><b>How does the current state of \u201copen\u201d models compare to your vision?<\/b><\/p>\n<p><span style=\"font-weight: 400;\">Currently, even the most open models like Llama, DeepSeek, and Alibaba\u2019s models only provide open weights. While this is a great start and we\u2019re grateful for these efforts, it\u2019s not the full picture of what \u201copen\u201d should mean.<\/span><\/p>\n<p><span style=\"font-weight: 400;\">For the near term, we\u2019ll likely see openness primarily in <\/span><a href=\"https:\/\/gradientflow.com\/post-training-rft-sft-rlhf\/\"><span style=\"font-weight: 400;\">post-training<\/span><\/a><span style=\"font-weight: 400;\"> rather than pre-training (which requires enormous resources). Pre-training massive models from scratch is currently prohibitive for smaller organizations, but there\u2019s a huge opportunity for the open community to take existing open models and make them better through post-training collaboration.<\/span><\/p>\n<figure id=\"attachment_45798\" aria-describedby=\"caption-attachment-45798\" style=\"width: 2041px\" class=\"wp-caption aligncenter\"><img data-recalc-dims=\"1\" fetchpriority=\"high\" decoding=\"async\" data-attachment-id=\"45798\" data-permalink=\"https:\/\/gradientflow.com\/why-this-ai-veteran-left-google-to-make-models-unconditionally-open\/oumi-gradientflow-components-characteristics-of-an-open-source-llm\/\" data-orig-file=\"https:\/\/i0.wp.com\/gradientflow.com\/wp-content\/uploads\/2025\/05\/oumi-GradientFlow-Components-Characteristics-of-an-Open-Source-LLM.png?fit=2041%2C577&amp;ssl=1\" data-orig-size=\"2041,577\" data-comments-opened=\"0\" data-image-meta='{\"aperture\":\"0\",\"credit\":\"\",\"camera\":\"\",\"caption\":\"\",\"created_timestamp\":\"0\",\"copyright\":\"\",\"focal_length\":\"0\",\"iso\":\"0\",\"shutter_speed\":\"0\",\"title\":\"\",\"orientation\":\"0\"}' data-image-title=\"oumi \u2013 GradientFlow -Components &amp; Characteristics of an Open Source LLM\" data-image-description=\"\" data-image-caption=\"&lt;p&gt;(click to enlarge)&lt;\/p&gt;\n\" data-medium-file=\"https:\/\/i0.wp.com\/gradientflow.com\/wp-content\/uploads\/2025\/05\/oumi-GradientFlow-Components-Characteristics-of-an-Open-Source-LLM.png?fit=300%2C85&amp;ssl=1\" data-large-file=\"https:\/\/i0.wp.com\/gradientflow.com\/wp-content\/uploads\/2025\/05\/oumi-GradientFlow-Components-Characteristics-of-an-Open-Source-LLM.png?fit=750%2C212&amp;ssl=1\" class=\"size-full wp-image-45798\" src=\"https:\/\/i0.wp.com\/gradientflow.com\/wp-content\/uploads\/2025\/05\/oumi-GradientFlow-Components-Characteristics-of-an-Open-Source-LLM.png?resize=750%2C212&amp;ssl=1\" alt=\"\" width=\"750\" height=\"212\" srcset=\"https:\/\/i0.wp.com\/gradientflow.com\/wp-content\/uploads\/2025\/05\/oumi-GradientFlow-Components-Characteristics-of-an-Open-Source-LLM.png?w=2041&amp;ssl=1 2041w, https:\/\/i0.wp.com\/gradientflow.com\/wp-content\/uploads\/2025\/05\/oumi-GradientFlow-Components-Characteristics-of-an-Open-Source-LLM.png?resize=300%2C85&amp;ssl=1 300w, https:\/\/i0.wp.com\/gradientflow.com\/wp-content\/uploads\/2025\/05\/oumi-GradientFlow-Components-Characteristics-of-an-Open-Source-LLM.png?resize=1024%2C289&amp;ssl=1 1024w, https:\/\/i0.wp.com\/gradientflow.com\/wp-content\/uploads\/2025\/05\/oumi-GradientFlow-Components-Characteristics-of-an-Open-Source-LLM.png?resize=768%2C217&amp;ssl=1 768w, https:\/\/i0.wp.com\/gradientflow.com\/wp-content\/uploads\/2025\/05\/oumi-GradientFlow-Components-Characteristics-of-an-Open-Source-LLM.png?resize=1536%2C434&amp;ssl=1 1536w, https:\/\/i0.wp.com\/gradientflow.com\/wp-content\/uploads\/2025\/05\/oumi-GradientFlow-Components-Characteristics-of-an-Open-Source-LLM.png?resize=1568%2C443&amp;ssl=1 1568w\" sizes=\"(max-width: 750px) 100vw, 750px\"><figcaption id=\"caption-attachment-45798\" class=\"wp-caption-text\">(<a href=\"https:\/\/gradientflow.com\/wp-content\/uploads\/2025\/05\/oumi-GradientFlow-Components-Characteristics-of-an-Open-Source-LLM.png\"><strong>click to enlarge<\/strong><\/a>)<\/figcaption><\/figure>\n<h5><b>Collaboration and Governance Models<\/b><\/h5>\n<p><b>How do you envision collaboration in open AI development to work effectively?<\/b><\/p>\n<p><span style=\"font-weight: 400;\">We need a standardized platform with standardized benchmarks where contributions can be validated and combined. When someone contributes an innovation in data curation, training methods, or other areas, it should be done on an end-to-end platform that captures all aspects of the work. Contributions that demonstrably improve performance can be combined to create better models.<\/span><\/p>\n<p><span style=\"font-weight: 400;\">AI is an incredibly complex field with diverse use cases and modalities; it truly requires \u201call hands on deck\u201d \u2013 not just model builders but data engineers working on pipelines and data cleaning can make valuable contributions.<\/span><\/p>\n<p><b>How do you address the signal-to-noise problem in open collaboration?<\/b><\/p>\n<p><span style=\"font-weight: 400;\">We focus on contributions that move the needle on benchmarks without regressions in other areas. When a contribution shows promise, many eyes look at it. As Linus says, \u201cgiven enough eyeballs, all bugs are shallow\u201d \u2013 or for AI, all issues are shallow.<\/span><\/p>\n<p><span style=\"font-weight: 400;\">Rather than reviewing every single fork, we focus on results. If a recipe moves a trusted benchmark positively with no regressions (including safety), it bubbles up for consideration. You don\u2019t need to review every contribution in depth if you focus on those that show clear, positive impact.<\/span><\/p>\n<p><b>How do you handle potentially problematic contributions, like data with IP violations or unsafe model behaviors?<\/b><\/p>\n<p><span style=\"font-weight: 400;\">I believe in Linus\u2019s Law: \u201cGiven enough eyeballs, all bugs are shallow.\u201d When development is done in the open, with many people scrutinizing the process and the data, it\u2019s arguably safer. For AI, given enough eyeballs, all AI issues (safety, bias, IP) are more transparent and addressable.<\/span><\/p>\n<p><span style=\"font-weight: 400;\">A glass-box approach to how models are built and what data is used is crucial. If a contribution moves the needle positively, it will attract more scrutiny, which helps in vetting it for potential issues.<\/span><\/p>\n<h5><b>The Oumi Platform and Technical Capabilities<\/b><\/h5>\n<p><b>What does the Oumi platform provide to developers?<\/b><\/p>\n<p><span style=\"font-weight: 400;\">The <a href=\"https:\/\/github.com\/oumi-ai\/oumi\">Oumi platform<\/a> enables experimentation with foundation models across different training types and model families. Think of it as the \u201cDevOps layer\u201d for foundation-model R&amp;D. It includes:<\/span><\/p>\n<ul>\n<li style=\"font-weight: 400;\" aria-level=\"1\"><span style=\"font-weight: 400;\">A unified API covering data curation, data synthesis, all types of training (pre-training and various post-training techniques like LoRA), and evaluation with academic or custom benchmarks.<\/span><\/li>\n<li style=\"font-weight: 400;\" aria-level=\"1\"><span style=\"font-weight: 400;\">Flexible deployment options \u2013 run on your laptop or scale to the cloud with just a configuration change.<\/span><\/li>\n<li style=\"font-weight: 400;\" aria-level=\"1\"><span style=\"font-weight: 400;\">Built-in pipelines to synthesize data with any LLM, score quality, and clean datasets.<\/span><\/li>\n<li style=\"font-weight: 400;\" aria-level=\"1\"><span style=\"font-weight: 400;\">Extensible trainer and benchmark harness.<\/span><\/li>\n<\/ul>\n<p><span style=\"font-weight: 400;\">These capabilities address common needs in both academia and enterprise.<\/span><\/p>\n<p><b>How does the platform handle compute resources?<\/b><\/p>\n<p><span style=\"font-weight: 400;\">Users bring their own compute. You can run the platform on your laptop or deploy to AWS, GCP, Azure, Lambda, Together, RunPod, or HPCs by changing the deployment configuration. We\u2019ve tested it on national lab HPCs with over 1,000 GPUs. For some collaborators, we provide compute, and our enterprise offering will include compute resources.<\/span><\/p>\n<p><b>What data processing capabilities does Oumi provide?<\/b><\/p>\n<p><span style=\"font-weight: 400;\">We make it easy to synthesize new data or curate data using existing models. You can use any open or closed model through Oumi to handle batch inference for data synthesis. You can also use LLMs to rate data quality and clean it up.<\/span><\/p>\n<p><span style=\"font-weight: 400;\">While we optimize data loading and streaming for training, our current focus for data tooling is more on synthesizing new data or creating\/augmenting datasets using LMs. These jobs can be scheduled on various cloud providers or on-premise clusters.<\/span><\/p>\n<p><b>Are you tied to a specific orchestration stack like Ray?<\/b><\/p>\n<p><span style=\"font-weight: 400;\">No. We don\u2019t yet use Ray internally, but the platform is designed such that someone could integrate <\/span><a href=\"https:\/\/www.ray.io\/#why-ray\"><span style=\"font-weight: 400;\">Ray for distributed training<\/span><\/a><span style=\"font-weight: 400;\"> if they wish. You can schedule Oumi jobs on Ray or Slurm if that fits your infrastructure.<\/span><\/p>\n<p><img loading=\"lazy\" data-recalc-dims=\"1\" decoding=\"async\" data-attachment-id=\"45800\" data-permalink=\"https:\/\/gradientflow.com\/why-this-ai-veteran-left-google-to-make-models-unconditionally-open\/oumi-halloumi_graph\/\" data-orig-file=\"https:\/\/i0.wp.com\/gradientflow.com\/wp-content\/uploads\/2025\/05\/oumi-halloumi_graph.png?fit=570%2C403&amp;ssl=1\" data-orig-size=\"570,403\" data-comments-opened=\"0\" data-image-meta='{\"aperture\":\"0\",\"credit\":\"\",\"camera\":\"\",\"caption\":\"\",\"created_timestamp\":\"0\",\"copyright\":\"\",\"focal_length\":\"0\",\"iso\":\"0\",\"shutter_speed\":\"0\",\"title\":\"\",\"orientation\":\"0\"}' data-image-title=\"oumi-halloumi_graph\" data-image-description=\"\" data-image-caption=\"\" data-medium-file=\"https:\/\/i0.wp.com\/gradientflow.com\/wp-content\/uploads\/2025\/05\/oumi-halloumi_graph.png?fit=300%2C212&amp;ssl=1\" data-large-file=\"https:\/\/i0.wp.com\/gradientflow.com\/wp-content\/uploads\/2025\/05\/oumi-halloumi_graph.png?fit=570%2C403&amp;ssl=1\" class=\"aligncenter wp-image-45800\" src=\"https:\/\/i0.wp.com\/gradientflow.com\/wp-content\/uploads\/2025\/05\/oumi-halloumi_graph.png?resize=436%2C308&amp;ssl=1\" alt=\"\" width=\"436\" height=\"308\" srcset=\"https:\/\/i0.wp.com\/gradientflow.com\/wp-content\/uploads\/2025\/05\/oumi-halloumi_graph.png?w=570&amp;ssl=1 570w, https:\/\/i0.wp.com\/gradientflow.com\/wp-content\/uploads\/2025\/05\/oumi-halloumi_graph.png?resize=300%2C212&amp;ssl=1 300w\" sizes=\"auto, (max-width: 436px) 100vw, 436px\"><\/p>\n<h5><b>Halloumi: AI Claim Verification<\/b><\/h5>\n<p><b>What is Halloumi and why did you start with this project?<\/b><\/p>\n<p><a href=\"https:\/\/oumi.ai\/blog\/posts\/introducing-halloumi\"><span style=\"font-weight: 400;\">Halloumi<\/span><\/a><span style=\"font-weight: 400;\"> is an \u201cAI lie detector\u201d or more precisely, AI claim verification. It checks that every part of an LLM\u2019s answer is grounded in the provided context and not hallucinated. It works for summaries, question answering, or any context-based LLM output.<\/span><\/p>\n<p><span style=\"font-weight: 400;\">It provides per-sentence confidence scores along with citations and explanations \u2013 pointing to the specific lines in the original document to verify against. It offers state-of-the-art quality, significantly better than general-purpose models like GPT-4 or Gemini for this specific task.<\/span><\/p>\n<p><span style=\"font-weight: 400;\">We started with Halloumi because hallucinations are a major blocker for enterprises adopting AI in production. It also served as an excellent test case for developing and validating the Oumi platform itself.<\/span><\/p>\n<p><b>How is Halloumi being used?<\/b><\/p>\n<p><span style=\"font-weight: 400;\">We\u2019ve seen interest from fintech firms using it for RAG scenarios and others wanting to cross-check if articles they\u2019ve written align with their notes. Some use it as a final QA pass when drafting articles to flag ungrounded statements before publication. The inputs to Halloumi are simple: the context provided to the LLM, the original prompt, and the LLM\u2019s generated response.<\/span><\/p>\n<h5><b>Open AI Safety Considerations<\/b><\/h5>\n<p><b>How do you respond to concerns that open models could lead to safety issues?<\/b><\/p>\n<p><span style=\"font-weight: 400;\">While I deeply respect figures like Hinton who have higher \u201cP(doom)\u201d estimates, I align more with Yann LeCun\u2019s perspective. The current development approach at big labs focuses on racing each other rather than investing enough in safety.<\/span><\/p>\n<p><span style=\"font-weight: 400;\">The best way to address safety is to do it openly, with the community collectively working on it before AI becomes even more powerful. When development is done in the open, with many people scrutinizing the process, issues are more likely to be caught early.<\/span><\/p>\n<p><b>What\u2019s your approach to AI safety?<\/b><\/p>\n<p><span style=\"font-weight: 400;\">Even if the probability of catastrophic outcomes is small, the potential consequences mean we need to address it. Beyond concerns about AI going wrong independently, there\u2019s the risk of it learning harmful behaviors from human data or being used by bad actors.<\/span><\/p>\n<p><span style=\"font-weight: 400;\">We need to build \u201cprotector AI\u201d that can detect and prohibit misuse, and the best way is through open development since closed labs aren\u2019t doing enough. Safety research and development must happen openly and collaboratively, now, not later.<\/span><\/p>\n<h5><b>Future Vision<\/b><\/h5>\n<p><b>What\u2019s your vision for the future of open AI by 2030?<\/b><\/p>\n<p><span style=\"font-weight: 400;\">I\u2019d like to see the majority of the enterprise AI market powered by open source AI that\u2019s developed collectively and transparently. Ideally, this would be fully open source AI \u2013 with open data, open code, open weights, and open collaboration \u2013 enabling both enterprise and scientific applications.<\/span><\/p>\n<p><b>Some argue that as AI models become proficient at coding and AI development (a \u201cflywheel\u201d effect), the need for \u201call hands on deck\u201d might diminish. What\u2019s your take?<\/b><\/p>\n<p><span style=\"font-weight: 400;\">It\u2019s a reasonable point. My hope is that if such a <\/span><a href=\"https:\/\/gradientflow.substack.com\/p\/deconstructing-openais-path-to-125\"><span style=\"font-weight: 400;\">flywheel<\/span><\/a><span style=\"font-weight: 400;\"> develops, it\u2019s built in the open. The open community is well-positioned to compete, especially in post-training, where innovation and novel ideas are paramount, often more so than just raw GPU power. The ability to rapidly iterate and prove concepts at a smaller scale, leveraging diverse global talent, is a strength of the open approach.<\/span><\/p>\n<hr>\n<figure id=\"attachment_45846\" aria-describedby=\"caption-attachment-45846\" style=\"width: 756px\" class=\"wp-caption aligncenter\"><img data-recalc-dims=\"1\" loading=\"lazy\" decoding=\"async\" data-attachment-id=\"45846\" data-permalink=\"https:\/\/gradientflow.com\/why-this-ai-veteran-left-google-to-make-models-unconditionally-open\/computing-power-for-ai\/\" data-orig-file=\"https:\/\/i0.wp.com\/gradientflow.com\/wp-content\/uploads\/2025\/05\/Computing-Power-for-AI.jpg?fit=5005%2C2608&amp;ssl=1\" data-orig-size=\"5005,2608\" data-comments-opened=\"0\" data-image-meta='{\"aperture\":\"0\",\"credit\":\"\",\"camera\":\"\",\"caption\":\"\",\"created_timestamp\":\"0\",\"copyright\":\"\",\"focal_length\":\"0\",\"iso\":\"0\",\"shutter_speed\":\"0\",\"title\":\"\",\"orientation\":\"1\"}' data-image-title=\"Computing Power for AI\" data-image-description=\"\" data-image-caption=\"&lt;p&gt;Source: We did the math on AI\u2019s energy footprint. Here\u2019s the story you haven\u2019t heard ; click HERE to enlarge&lt;\/p&gt;\n\" data-medium-file=\"https:\/\/i0.wp.com\/gradientflow.com\/wp-content\/uploads\/2025\/05\/Computing-Power-for-AI.jpg?fit=300%2C156&amp;ssl=1\" data-large-file=\"https:\/\/i0.wp.com\/gradientflow.com\/wp-content\/uploads\/2025\/05\/Computing-Power-for-AI.jpg?fit=750%2C391&amp;ssl=1\" class=\" wp-image-45846\" src=\"https:\/\/i0.wp.com\/gradientflow.com\/wp-content\/uploads\/2025\/05\/Computing-Power-for-AI.jpg?resize=750%2C391&amp;ssl=1\" alt=\"\" width=\"750\" height=\"391\" srcset=\"https:\/\/i0.wp.com\/gradientflow.com\/wp-content\/uploads\/2025\/05\/Computing-Power-for-AI.jpg?w=5005&amp;ssl=1 5005w, https:\/\/i0.wp.com\/gradientflow.com\/wp-content\/uploads\/2025\/05\/Computing-Power-for-AI.jpg?resize=300%2C156&amp;ssl=1 300w, https:\/\/i0.wp.com\/gradientflow.com\/wp-content\/uploads\/2025\/05\/Computing-Power-for-AI.jpg?resize=1024%2C534&amp;ssl=1 1024w, https:\/\/i0.wp.com\/gradientflow.com\/wp-content\/uploads\/2025\/05\/Computing-Power-for-AI.jpg?resize=768%2C400&amp;ssl=1 768w, https:\/\/i0.wp.com\/gradientflow.com\/wp-content\/uploads\/2025\/05\/Computing-Power-for-AI.jpg?resize=1536%2C800&amp;ssl=1 1536w, https:\/\/i0.wp.com\/gradientflow.com\/wp-content\/uploads\/2025\/05\/Computing-Power-for-AI.jpg?resize=2048%2C1067&amp;ssl=1 2048w, https:\/\/i0.wp.com\/gradientflow.com\/wp-content\/uploads\/2025\/05\/Computing-Power-for-AI.jpg?resize=1568%2C817&amp;ssl=1 1568w, https:\/\/i0.wp.com\/gradientflow.com\/wp-content\/uploads\/2025\/05\/Computing-Power-for-AI.jpg?w=2250&amp;ssl=1 2250w\" sizes=\"auto, (max-width: 750px) 100vw, 750px\"><figcaption id=\"caption-attachment-45846\" class=\"wp-caption-text\">Source: <a href=\"https:\/\/www.technologyreview.com\/2025\/05\/20\/1116327\/ai-energy-usage-climate-footprint-big-tech\/\"><strong>We did the math on AI\u2019s energy footprint. Here\u2019s the story you haven\u2019t heard<\/strong><\/a> ; click <a href=\"https:\/\/gradientflow.com\/wp-content\/uploads\/2025\/05\/Computing-Power-for-AI.jpg\"><strong>HERE<\/strong><\/a> to enlarge<\/figcaption><\/figure>\n<hr>\n<h3>Claude Opus 4 &amp; Claude Sonnet 4: Cheat Sheet<\/h3>\n<p><b>Claude Opus 4 and Claude Sonnet 4<\/b><span style=\"font-weight: 400;\"> are Anthropic\u2019s latest hybrid-reasoning language models. The <\/span><a href=\"https:\/\/www-cdn.anthropic.com\/4263b940cabb546aa0e3283f35b686f4f3b2ff47.pdf\"><b>system card<\/b><\/a><span style=\"font-weight: 400;\"> explains how the models were trained on a blend of public web data, opted-in user content and proprietary sources, and how they can switch between a quick default mode and a slower \u201cextended thinking\u201d mode for harder problems. The card outlines a battery of safety checks, ranging from single-turn content filters and bias benchmarks to red-team exercises and agentic coding trials. These evaluations informed Anthropic\u2019s decision to release Opus 4 under the stricter AI Safety <\/span><a href=\"https:\/\/www.anthropic.com\/news\/anthropics-responsible-scaling-policy\"><span style=\"font-weight: 400;\">Level 3 standard <\/span><\/a><span style=\"font-weight: 400;\">and Sonnet 4 under <\/span><a href=\"https:\/\/www.anthropic.com\/news\/anthropics-responsible-scaling-policy\"><span style=\"font-weight: 400;\">Level 2<\/span><\/a><span style=\"font-weight: 400;\">.<\/span><\/p>\n<figure id=\"attachment_45860\" aria-describedby=\"caption-attachment-45860\" style=\"width: 660px\" class=\"wp-caption aligncenter\"><img data-recalc-dims=\"1\" loading=\"lazy\" decoding=\"async\" data-attachment-id=\"45860\" data-permalink=\"https:\/\/gradientflow.com\/why-this-ai-veteran-left-google-to-make-models-unconditionally-open\/claude-opus-4-claude-sonnet-4-system-card-2\/\" data-orig-file=\"https:\/\/i0.wp.com\/gradientflow.com\/wp-content\/uploads\/2025\/05\/Claude-Opus-4-Claude-Sonnet-4-system-card.png?fit=1920%2C1080&amp;ssl=1\" data-orig-size=\"1920,1080\" data-comments-opened=\"0\" data-image-meta='{\"aperture\":\"0\",\"credit\":\"\",\"camera\":\"\",\"caption\":\"\",\"created_timestamp\":\"0\",\"copyright\":\"\",\"focal_length\":\"0\",\"iso\":\"0\",\"shutter_speed\":\"0\",\"title\":\"\",\"orientation\":\"0\"}' data-image-title=\"Claude Opus 4 &amp; Claude Sonnet 4 \u2013 system card\" data-image-description=\"\" data-image-caption=\"&lt;p&gt;(click to enlarge)&lt;\/p&gt;\n\" data-medium-file=\"https:\/\/i0.wp.com\/gradientflow.com\/wp-content\/uploads\/2025\/05\/Claude-Opus-4-Claude-Sonnet-4-system-card.png?fit=300%2C169&amp;ssl=1\" data-large-file=\"https:\/\/i0.wp.com\/gradientflow.com\/wp-content\/uploads\/2025\/05\/Claude-Opus-4-Claude-Sonnet-4-system-card.png?fit=750%2C422&amp;ssl=1\" class=\" wp-image-45860\" src=\"https:\/\/i0.wp.com\/gradientflow.com\/wp-content\/uploads\/2025\/05\/Claude-Opus-4-Claude-Sonnet-4-system-card.png?resize=660%2C371&amp;ssl=1\" alt=\"\" width=\"660\" height=\"371\" srcset=\"https:\/\/i0.wp.com\/gradientflow.com\/wp-content\/uploads\/2025\/05\/Claude-Opus-4-Claude-Sonnet-4-system-card.png?w=1920&amp;ssl=1 1920w, https:\/\/i0.wp.com\/gradientflow.com\/wp-content\/uploads\/2025\/05\/Claude-Opus-4-Claude-Sonnet-4-system-card.png?resize=300%2C169&amp;ssl=1 300w, https:\/\/i0.wp.com\/gradientflow.com\/wp-content\/uploads\/2025\/05\/Claude-Opus-4-Claude-Sonnet-4-system-card.png?resize=1024%2C576&amp;ssl=1 1024w, https:\/\/i0.wp.com\/gradientflow.com\/wp-content\/uploads\/2025\/05\/Claude-Opus-4-Claude-Sonnet-4-system-card.png?resize=768%2C432&amp;ssl=1 768w, https:\/\/i0.wp.com\/gradientflow.com\/wp-content\/uploads\/2025\/05\/Claude-Opus-4-Claude-Sonnet-4-system-card.png?resize=1536%2C864&amp;ssl=1 1536w, https:\/\/i0.wp.com\/gradientflow.com\/wp-content\/uploads\/2025\/05\/Claude-Opus-4-Claude-Sonnet-4-system-card.png?resize=1568%2C882&amp;ssl=1 1568w\" sizes=\"auto, (max-width: 660px) 100vw, 660px\"><figcaption id=\"caption-attachment-45860\" class=\"wp-caption-text\">(<strong><a href=\"https:\/\/gradientflow.com\/wp-content\/uploads\/2025\/05\/Claude-Opus-4-Claude-Sonnet-4-system-card.jpeg\">click to enlarge<\/a><\/strong>)<\/figcaption><\/figure>\n<p><span style=\"font-weight: 400;\">For practitioners <\/span><a href=\"https:\/\/www-cdn.anthropic.com\/4263b940cabb546aa0e3283f35b686f4f3b2ff47.pdf\"><span style=\"font-weight: 400;\">the document<\/span><\/a><span style=\"font-weight: 400;\"> is more than corporate disclosure: it is a technical manual that signals where the models excel, where guard-rails bite and what deployment obligations follow. Understanding these details helps product teams gauge reliability, tailor prompts and comply with risk policies, while giving end-users a clearer view of why the models may refuse, summarise or slow down in certain scenarios. In short, the card turns abstract assurances of \u201chelpfulness, honesty and harmlessness\u201d into concrete, auditable claims that can be tested in the real world. <\/span><\/p>\n<p><a class=\"a2a_button_bluesky\" href=\"https:\/\/www.addtoany.com\/add_to\/bluesky?linkurl=https%3A%2F%2Fgradientflow.com%2Fwhy-this-ai-veteran-left-google-to-make-models-unconditionally-open%2F&amp;linkname=Why%20this%20AI%20veteran%20left%20Google%20to%20make%20models%20%E2%80%98unconditionally%20open%E2%80%99\" title=\"Bluesky\" rel=\"nofollow noopener\" target=\"_blank\"><\/a><a class=\"a2a_button_linkedin\" href=\"https:\/\/www.addtoany.com\/add_to\/linkedin?linkurl=https%3A%2F%2Fgradientflow.com%2Fwhy-this-ai-veteran-left-google-to-make-models-unconditionally-open%2F&amp;linkname=Why%20this%20AI%20veteran%20left%20Google%20to%20make%20models%20%E2%80%98unconditionally%20open%E2%80%99\" title=\"LinkedIn\" rel=\"nofollow noopener\" target=\"_blank\"><\/a><a class=\"a2a_button_facebook\" href=\"https:\/\/www.addtoany.com\/add_to\/facebook?linkurl=https%3A%2F%2Fgradientflow.com%2Fwhy-this-ai-veteran-left-google-to-make-models-unconditionally-open%2F&amp;linkname=Why%20this%20AI%20veteran%20left%20Google%20to%20make%20models%20%E2%80%98unconditionally%20open%E2%80%99\" title=\"Facebook\" rel=\"nofollow noopener\" target=\"_blank\"><\/a><a class=\"a2a_button_reddit\" href=\"https:\/\/www.addtoany.com\/add_to\/reddit?linkurl=https%3A%2F%2Fgradientflow.com%2Fwhy-this-ai-veteran-left-google-to-make-models-unconditionally-open%2F&amp;linkname=Why%20this%20AI%20veteran%20left%20Google%20to%20make%20models%20%E2%80%98unconditionally%20open%E2%80%99\" title=\"Reddit\" rel=\"nofollow noopener\" target=\"_blank\"><\/a><a class=\"a2a_button_email\" href=\"https:\/\/www.addtoany.com\/add_to\/email?linkurl=https%3A%2F%2Fgradientflow.com%2Fwhy-this-ai-veteran-left-google-to-make-models-unconditionally-open%2F&amp;linkname=Why%20this%20AI%20veteran%20left%20Google%20to%20make%20models%20%E2%80%98unconditionally%20open%E2%80%99\" title=\"Email\" rel=\"nofollow noopener\" target=\"_blank\"><\/a><a class=\"a2a_button_mastodon\" href=\"https:\/\/www.addtoany.com\/add_to\/mastodon?linkurl=https%3A%2F%2Fgradientflow.com%2Fwhy-this-ai-veteran-left-google-to-make-models-unconditionally-open%2F&amp;linkname=Why%20this%20AI%20veteran%20left%20Google%20to%20make%20models%20%E2%80%98unconditionally%20open%E2%80%99\" title=\"Mastodon\" rel=\"nofollow noopener\" target=\"_blank\"><\/a><a class=\"a2a_button_copy_link\" href=\"https:\/\/www.addtoany.com\/add_to\/copy_link?linkurl=https%3A%2F%2Fgradientflow.com%2Fwhy-this-ai-veteran-left-google-to-make-models-unconditionally-open%2F&amp;linkname=Why%20this%20AI%20veteran%20left%20Google%20to%20make%20models%20%E2%80%98unconditionally%20open%E2%80%99\" title=\"Copy Link\" rel=\"nofollow noopener\" target=\"_blank\"><\/a><\/p>\n<p>The post <a href=\"https:\/\/gradientflow.com\/why-this-ai-veteran-left-google-to-make-models-unconditionally-open\/\">Why this AI veteran left Google to make models \u2018unconditionally open\u2019<\/a> appeared first on <a href=\"https:\/\/gradientflow.com\/\">Gradient Flow<\/a>.<\/p>\n<\/div>\n<div style=\"margin-top: 0px; margin-bottom: 0px;\" class=\"sharethis-inline-share-buttons\" ><\/div>","protected":false},"excerpt":{"rendered":"<p>Subscribe\u00a0\u2022\u00a0Previous Issues Beyond Open Weights: The Path to Unconditionally Open AI While I routinely work with both proprietary LLMs and open-weights models, my heart lies with models that are open&hellip;<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[176,1],"tags":[],"class_list":["post-1558","post","type-post","status-publish","format-standard","hentry","category-newsletter","category-uncategorized"],"_links":{"self":[{"href":"https:\/\/musictechohio.online\/site\/wp-json\/wp\/v2\/posts\/1558","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/musictechohio.online\/site\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/musictechohio.online\/site\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/musictechohio.online\/site\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/musictechohio.online\/site\/wp-json\/wp\/v2\/comments?post=1558"}],"version-history":[{"count":0,"href":"https:\/\/musictechohio.online\/site\/wp-json\/wp\/v2\/posts\/1558\/revisions"}],"wp:attachment":[{"href":"https:\/\/musictechohio.online\/site\/wp-json\/wp\/v2\/media?parent=1558"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/musictechohio.online\/site\/wp-json\/wp\/v2\/categories?post=1558"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/musictechohio.online\/site\/wp-json\/wp\/v2\/tags?post=1558"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}