AI Model Development

Of course you need to use open-source models if you’re an enterprise leader. Close model providers, that are now forcing data retention, are gaining immense leverage on your business if you don’t. As you connect models to your business context, they see it and learn from it, and have a track record of going after their most successful customers thanks to this information. But that’s not enough, you also need to store your data and records in open systems, or your software vendors might block you from building AI systems outside of the walled garden they have set up for you. If you can’t convince them to give you complete access to the data they manage for you, AI fortunately allows you to migrate quite fast. Once you’ve got hold of your data, you’ll need to manage how AI systems can access this data on behalf of human users, because you don’t always want Bob to see what Alice is doing in your company. That’s hard and merciless, since AI models are great at finding need-to-know errors. It takes systems that check hard access rules and models that check soft access rules. Now comes the most important part. You need to set up your own continuous training flywheel, so that you can improve your AI systems based on their interaction with your employees and your users. This is how you turn the edges of your business into AI systems your vendors and competitors cannot replicate. It’s also how you reduce deployment cost as well, as you can shrink models according to model input distribution. Those bills are getting substantial, we need to collectively become efficient if we want AI development to continue, so that matters. All of these efforts might seem daunting – they are. This is both a complete replatforming of your IT, and a complete change in the way you’re developing software, and operating your business. AI lifecycle management requires understanding human behavior and gradient descent, that’s a stretch. At Mistral, we facilitate that work by providing all primitives that you need in a single control plane, Studio, and a training platform, Forge. With our applied AI engineers and scientists working hand-in-hand with our customers, we ensure that we transfer knowledge, and that we can disappear once the systems are up and running. We deploy on our customers' infrastructure, or through our zero-data-retention hosted services, so that your edges remain your edges, and the switch button can be fully in your hand. Frontier AI can accelerate the growth of your business, but if it’s not in your hands, it’s not going to be your growth.

Over the last two weeks, both the U.S. Government and Anthropic took significant actions that demonstrated their power to control access to AI by restricting what others can do with frontier models. This has been one of those moments that, once seen, will be hard to unsee, and it is significantly accelerating many businesses’ and nation states’ efforts to ensure reliable access to AI that no one else can terminate. Anthropic first released Claude Fable 5, a version of its Mythos model with additional guardrails, including some restrictions that seem justified on safety grounds (such as limitations on applying it to hacking, bioweapons, and so forth). However, it also restricted developers’ ability to use it to build competing LLM technology. This move was concerning, given that the whole AI community, including Anthropic, has benefitted tremendously from open research — indeed, the AI revolution was kicked off by my former team (Google Brain) freely publishing the Transformers paper! Imagine if Microsoft’s terms of use barred anyone from using their tools to build competitive software, or if Google barred using it to search for information to work on competing search engines. Anthropic’s argument that it was unsafe for others to be able to make advances in AI also rang hollow. Initially, Anthropic silently degraded Fable 5’s performance for users detected to be working on LLM research through invisible interventions that weakened the model’s outputs without notifying the user. After significant backlash, it walked back this decision and decided to be transparent when it did this, but it still refuses to use its latest capabilities to help AI researchers. This move represents a raw demonstration of power by Anthropic. It has used “safety” arguments to hinder potential competitors. Platforms succeed when they are viewed as stable, reliable partners that one can build on. The sudden rule changes by Anthropic (including a mandatory 30 day data retention policy for Fable usage) have made developers wonder about the stability of building on any one proprietary LLM provider, not just Anthropic. The U.S. Government then shortly followed with an even greater demonstration of power. It used the Commerce Department’s authority to regulate technologies that may be national security threats to restrict exports of Mythos and Fable, requiring a license for use by any foreign national, whether inside or outside of the U.S. This led Anthropic to disable access to Fable to all users worldwide. Sam Altman pointed out, referring to Anthropic, “It is clearly incredible marketing to say, ‘We have built a bomb, we are about to drop it on your head. We will sell you a bomb shelter for $100 million.’” But when one uses this type of fear-based marketing, it increases the odds that the U.S. Government will agree with you and slap export controls on the bomb you say you have built. [Truncated for length. Full text: https://lnkd.in/gph-rGtw ]

Microsoft just released a 35-page report on medical AI - and it’s a reality check for healthcare. The paper, “The Illusion of Readiness”, tested six of the most popular models (OpenAI, Gemini, etc)… across six multimodal medical benchmarks. And the verdict? The models scored high on medical exams. But they’re not even close to being real-world ready. Here’s what the stress tests revealed: ▶ 1. Shortcut learning Models often answered correctly even when key information, like medical images, was removed. They weren’t reasoning - they were exploiting statistical shortcuts. That means benchmark wins may hide shallow understanding. ▶ 2. Fragile under small changes Making small tweaks caused big swings in predictions. This fragility shows how unreliable model reasoning becomes under stress. In visual substitution tests, accuracy dropped from 83% to 52% when images were swapped - exposing shallow visual–answer pairings. ▶ 3. Fabricated reasoning Models produced confident, step-by-step medical explanations - but many were medically unsound… or entirely fabricated. Convincing to the eye, dangerous in practice. And more importantly, healthcare isn’t a multiple-choice exam. It’s uncertainty, incomplete data, and high stakes. So Microsoft’s team calls for new standards: - Stress tests that expose fragility - Clinician-guided guidelines that profile benchmarks - Evaluation of robustness and trustworthiness - not just leaderboard scores The takeaway is simple: Medical AI may ace tests today. But until it proves reliable under stress, it’s not ready for the clinic. When do you think popular LLMs will be clinic-ready? #entrepreneurship #healthtech #AI

Had to share the one prompt that has transformed how I approach AI research. 📌 Save this post. Don’t just ask for point-in-time data like a junior PM. Instead, build in more temporal context through systematic data collection over time. Use this prompt to become a superforecaster with the help of AI. Great for product ideation, competitive research, finance, investing, etc. ⏰⏰⏰⏰⏰⏰⏰⏰⏰⏰⏰⏰ TIME MACHINE PROMPT: Execute longitudinal analysis on [TOPIC]. First, establish baseline parameters: define the standard refresh interval for this domain based on market dynamics (enterprise adoption cycles, regulatory changes, technology maturity curves). For example, AI refresh cycle may be two weeks, clothing may be 3 months, construction may be 2 years. Calculate n=3 data points spanning 2 full cycles. For each time period, collect: (1) quantitative metrics (adoption rates, market share, pricing models), (2) qualitative factors (user sentiment, competitive positioning, external catalysts), (3) ecosystem dependencies (infrastructure requirements, complementary products, capital climate, regulatory environment). Structure output as: Current State Analysis → T-1 Comparative Analysis → T-2 Historical Baseline → Delta Analysis with statistical significance → Trajectory Modeling with confidence intervals across each prediction. Include data sources. ⏰⏰⏰⏰⏰⏰⏰⏰⏰⏰⏰⏰

AI is not failing because of bad ideas; it’s "failing" at enterprise scale because of two big gaps: 👉 Workforce Preparation 👉 Data Security for AI While I speak globally on both topics in depth, today I want to educate us on what it takes to secure data for AI—because 70–82% of AI projects pause or get cancelled at POC/MVP stage (source: #Gartner, #MIT). Why? One of the biggest reasons is a lack of readiness at the data layer. So let’s make it simple - there are 7 phases to securing data for AI—and each phase has direct business risk if ignored. 🔹 Phase 1: Data Sourcing Security - Validating the origin, ownership, and licensing rights of all ingested data. Why It Matters: You can’t build scalable AI with data you don’t own or can’t trace. 🔹 Phase 2: Data Infrastructure Security - Ensuring data warehouses, lakes, and pipelines that support your AI models are hardened and access-controlled. Why It Matters: Unsecured data environments are easy targets for bad actors making you exposed to data breaches, IP theft, and model poisoning. 🔹 Phase 3: Data In-Transit Security - Protecting data as it moves across internal or external systems, especially between cloud, APIs, and vendors. Why It Matters: Intercepted training data = compromised models. Think of it as shipping cash across town in an armored truck—or on a bicycle—your choice. 🔹 Phase 4: API Security for Foundational Models - Safeguarding the APIs you use to connect with LLMs and third-party GenAI platforms (OpenAI, Anthropic, etc.). Why It Matters: Unmonitored API calls can leak sensitive data into public models or expose internal IP. This isn’t just tech debt. It’s reputational and regulatory risk. 🔹 Phase 5: Foundational Model Protection - Defending your proprietary models and fine-tunes from external inference, theft, or malicious querying. Why It Matters: Prompt injection attacks are real. And your enterprise-trained model? It’s a business asset. You lock your office at night—do the same with your models. 🔹 Phase 6: Incident Response for AI Data Breaches - Having predefined protocols for breaches, hallucinations, or AI-generated harm—who’s notified, who investigates, how damage is mitigated. Why It Matters: AI-related incidents are happening. Legal needs response plans. Cyber needs escalation tiers. 🔹 Phase 7: CI/CD for Models (with Security Hooks) - Continuous integration and delivery pipelines for models, embedded with testing, governance, and version-control protocols. Why It Matter: Shipping models like software means risk comes faster—and so must detection. Governance must be baked into every deployment sprint. Want your AI strategy to succeed past MVP? Focus and lock down the data. #AI #DataSecurity #AILeadership #Cybersecurity #FutureOfWork #ResponsibleAI #SolRashidi #Data #Leadership

AI is rapidly moving from passive text generators to active decision-makers. To understand where things are headed, it’s important to trace the stages of this evolution. 1. 𝗟𝗟𝗠𝘀: 𝗧𝗵𝗲 𝗘𝗿𝗮 𝗼𝗳 𝗟𝗮𝗻𝗴𝘂𝗮𝗴𝗲 𝗙𝗹𝘂𝗲𝗻𝗰𝘆 Large Language Models (LLMs) like GPT-3 and GPT-4 excel at generating human-like text by predicting the next word in a sequence. They can produce coherent and contextually appropriate responses—but their capabilities end there. They don’t retain memory, they don’t take actions, and they don’t understand goals. They are reactive, not proactive. 2. 𝗥𝗔𝗚: 𝗧𝗵𝗲 𝗔𝗴𝗲 𝗼𝗳 𝗖𝗼𝗻𝘁𝗲𝘅𝘁-𝗔𝘄𝗮𝗿𝗲 𝗚𝗲𝗻𝗲𝗿𝗮𝘁𝗶𝗼𝗻 Retrieval-Augmented Generation (RAG) brought a major upgrade by integrating LLMs with external knowledge sources like vector databases or document stores. Now the model could retrieve relevant context and generate more accurate and personalized responses based on that information. This stage introduced the idea of 𝗱𝘆𝗻𝗮𝗺𝗶𝗰 𝗸𝗻𝗼𝘄𝗹𝗲𝗱𝗴𝗲 𝗮𝗰𝗰𝗲𝘀𝘀, but still required orchestration. The system didn’t plan or act—it responded with more relevance. 3. 𝗔𝗴𝗲𝗻𝘁𝗶𝗰 𝗔𝗜: 𝗧𝗼𝘄𝗮𝗿𝗱 𝗔𝘂𝘁𝗼𝗻𝗼𝗺𝗼𝘂𝘀 𝗜𝗻𝘁𝗲𝗹𝗹𝗶𝗴𝗲𝗻𝗰𝗲 Agentic AI is a fundamentally different paradigm. Here, systems are built to perceive, reason, and act toward goals—often without constant human prompting. An Agentic system includes: • 𝗠𝗲𝗺𝗼𝗿𝘆: to retain and recall information over time. • 𝗣𝗹𝗮𝗻𝗻𝗶𝗻𝗴: to decide what actions to take and in what order. • 𝗧𝗼𝗼𝗹 𝗨𝘀𝗲: to interact with APIs, databases, code, or software systems. • 𝗔𝘂𝘁𝗼𝗻𝗼𝗺𝘆: to loop through perception, decision, and action—iteratively improving performance.    Instead of a single model generating content, we now orchestrate 𝗺𝘂𝗹𝘁𝗶𝗽𝗹𝗲 𝗮𝗴𝗲𝗻𝘁𝘀, each responsible for specific tasks, coordinated by a central controller or planner. This is the architecture behind emerging use cases like autonomous coding assistants, intelligent workflow bots, and AI co-pilots that can operate entire systems. 𝗧𝗵𝗲 𝗦𝗵𝗶𝗳𝘁 𝗶𝗻 𝗧𝗵𝗶𝗻𝗸𝗶𝗻𝗴 We’re no longer designing prompts. We’re designing 𝗺𝗼𝗱𝘂𝗹𝗮𝗿, 𝗴𝗼𝗮𝗹-𝗱𝗿𝗶𝘃𝗲𝗻 𝘀𝘆𝘀𝘁𝗲𝗺𝘀 capable of interacting with the real world. This evolution—LLM → RAG → Agentic AI—marks the transition from 𝗹𝗮𝗻𝗴𝘂𝗮𝗴𝗲 𝘂𝗻𝗱𝗲𝗿𝘀𝘁𝗮𝗻𝗱𝗶𝗻𝗴 to 𝗴𝗼𝗮𝗹-𝗱𝗿𝗶𝘃𝗲𝗻 𝗶𝗻𝘁𝗲𝗹𝗹𝗶𝗴𝗲𝗻𝗰𝗲.

𝗢𝗻𝗲 𝗼𝗳 𝘁𝗵𝗲 𝗠𝗢𝗦𝗧 𝗱𝗶𝘀𝗰𝘂𝘀𝘀𝗲𝗱 𝗾𝘂𝗲𝘀𝘁𝗶𝗼𝗻: 𝗛𝗼𝘄 𝘁𝗼 𝗽𝗶𝗰𝗸 𝘁𝗵𝗲 𝗿𝗶𝗴𝗵𝘁 𝗟𝗟𝗠 𝗳𝗼𝗿 𝘆𝗼𝘂𝗿 𝘂𝘀𝗲 𝗰𝗮𝘀𝗲? The LLM landscape is booming and choosing the right LLM is now a business decision, not just a tech choice. One-size-fits-all? Forget it. Nearly all enterprises today rely on different models for different use cases and/or industry-specific fine-tuned models. There’s no universal “best” model — only the best fit for a given task. The latest LLM landscape (see below) shows how models stack up in capability (MMLU score), parameter size and accessibility — and the differences REALLY matter.  𝗟𝗲𝘁'𝘀 𝗯𝗿𝗲𝗮𝗸 𝗶𝘁 𝗱𝗼𝘄𝗻: ⬇️ 1️⃣ 𝗚𝗲𝗻𝗲𝗿𝗮𝗹𝗶𝘀𝘁 𝘃𝘀. 𝗦𝗽𝗲𝗰𝗶𝗮𝗹𝗶𝘀𝘁: - Need a broad, powerful AI? GPT-4, Claude Opus, Gemini 1.5 Pro — great for general reasoning and diverse applications.   - Need domain expertise? E.g. IBM Granite or Mistral models (Lightweight & Fast) can be an excellent choice — tailored for specific industries.  2️⃣ 𝗕𝗶𝗴 𝘃𝘀. 𝗦𝗹𝗶𝗺:  - Powerful, large models (GPT-4, Claude Opus, Gemini 1.5 Pro) = great reasoning, but expensive and slow. - Slim, efficient models (Mistral 7B, LLaMA 3, RWWK models) = faster, cheaper, easier to fine-tune. Perfect for on-device, edge AI, or latency-sensitive applications.  3️⃣ 𝗢𝗽𝗲𝗻 𝘃𝘀. 𝗖𝗹𝗼𝘀𝗲𝗱   - Need full control? Open-source models (LLaMA 3, Mistral, Llama) give you transparency and customization.   - Want cutting-edge performance? Closed models (GPT-4, Gemini, Claude) still lead in general intelligence.  𝗧𝗵𝗲 𝗞𝗲𝘆 𝗧𝗮𝗸𝗲𝗮𝘄𝗮𝘆? There is no "best" model — only the best one for your use case, but it's key to understand the differences to make an informed decision: - Running AI in production? Go slim, go fast. - Need state-of-the-art reasoning? Go big, go deep. - Building industry-specific AI? Go specialized and save some money with SLMs.  I love seeing how the AI and LLM stack is evolving, offering multiple directions depending on your specific use case. Source of the picture: informationisbeautiful.net

Not every foundation model needs to be gigantic. We trained a 1.5M-parameter neural network to control the body of a humanoid robot. It takes a lot of subconscious processing for us humans to walk, maintain balance, and maneuver our arms and legs into desired positions. We capture this “subconsciousness” in HOVER, a single model that learns how to coordinate the motors of a humanoid robot to support locomotion and manipulation. We trained HOVER in NVIDIA Isaac, a GPU-powered simulation suite that accelerates physics by 10,000x faster than real time. To put the number in perspective, the robots undergo 1 year of intense training in a virtual “dojo”, but take only ~50 minutes of wall clock time on one GPU card. The neural net then transfers zero-shot to the real world without finetuning. HOVER can be *prompted* for various types of high-level motion instructions that we call “control modes”. To name a few: - Head and hand poses - can be captured by XR devices like Apple Vision Pro. - Whole-body poses - via MoCap or RGB camera. - Whole-body joint angles - Exoskeleton. - Root velocity command - Joysticks. What HOVER enables: - A unified interface for us to control the robot using whichever input devices are convenient at hand. - An easier way to collect whole-body teleoperation data for training.  - An upstream Vision-Language-Action model to provide motion instructions, which HOVER translates to low-level motor signals at high frequency. HOVER supports any humanoid that can be simulated in Isaac. Bring your own robot, and watch it come to life! It's a big teamwork from NVIDIA GEAR Lab and collaborators: Tairan He, Wenli Xiao, Toru Lin, Zhengyi Luo, Zhenjia Xu, Zhenyu Jiang, Jan Kautz, Changliu Liu, Guanya Shi, Xiaolong Wang Team leads: Jim Fan, Yuke Zhu Website: https://lnkd.in/g6WrJyRC Paper: https://lnkd.in/g99yWTPa

🪂 How To Make Your Design System AI-Ready (https://lnkd.in/dtnpy7CM), a practical guide on how to reduce drifts, minimize mistakes, maintain context and improve the quality of AI-generated prototypes — with structured spec files, automated auditing and token layers. Put together by Hardik Pandya from Atlassian. --- 🔹 1. Design Decisions Are Infrastructure AI-generated prototypes often don't deliver consistently decent results because of tiny inconsistencies scattered all across a design system. Often it's decisions made but not documented, hard-coded values never cleaned up, or relying too much on AI making sense of mock-ups or design flows on its own. Unsurprisingly, better AI prototypes come from better data — but also from better human guidance. We shouldn’t assume that AI knows how to choose the right component, and how to design with accessibility in mind. It needs priorities, a clear path on how we make decisions, design principles, examples, do's and don'ts. In fact, we should treat design decisions as infrastructure. That means that every time we make a decision — not just a design decision, but even decision on how actually prioritize our work and how we make decisions around here — it must find a path into the spec file that is then consumed by AI. --- 🔶 2. Three Layers: Spec Files + Token Layer + Audit To ensure quality, we establish design principles, guidelines, rules in a form of “spec files”). It's structured Markdown files that include spacing rules, color choices, component usage guidelines, priorities etc. AI is going to read and reuse that spec file every time it's going to generate a prototype. Because the spec files are text files, it's much more cost-effective, but also much more accurate just because we don't rely on AI recognizing or decoding patterns from mock-ups, but gets specific guidelines instead. In fact, extending code is often a more effective way than generating code from mock-ups. Token layer lists and keeps updated all tokens used throughout the design system. AI always chooses from a closed set of named variables instead of inventing plausible values ad-hoc. An audit script catches what AI gets wrong. It scans the prototype and flags every hard-coded value and flags it if necessary. It can be a regular software doing that, with AI waiting for its feedback to come back. Finally, when a design system ships updates, a sync routine flags which spec files need updating. The goal is to make sure that AI always reads up-to-date, current specs, not the ones written against an outdated version. --- 🔺 3. Examples of AI-Ready Design Systems ⌾ Atlassian: https://lnkd.in/dVsGc3Cp ⌾ Carbon: https://lnkd.in/d4zq4WWb ⌾ CMS Design System: https://lnkd.in/dHHzV3en ⌾ Nordhealth: https://lnkd.in/d8C4j2ZA Yet again, AI can’t magically resolve technical debt or design debt — it needs guidance, decisions, priorities and principles.

The newly released 1-trillion-parameter Kimi K2 LLM takes the crown as the best available open-weight model. According to the first benchmarks, it rivals the best proprietary LLMs on coding tasks (Claude). How does it compare to the 673-billion-parameter DeepSeek V3 (the architecture underlying DeepSeek R1)? It's basically the same, except that it uses more experts in the Mixture-of-Expert (MoE) modules and fewer heads in the Multi-head Latent Attention (MLA) module. I compared the two architectures side by side in the figure below. I'd say the key novelty is not well reflected in the architecture but in the way it was trained. I.e., they replaced the popular AdamW optimizer with a modified version of Muon. According to the Kimi 2 developers, their MuonClip optimizer "substantially outperforms the widely used AdamW optimizer for LLM training". Given how expensive LLM training is, this probably multi-million dollar bet resulted in the smoothest LLM training loss curve I've ever seen! I am looking forward to discussing more when the official technical report is released!