The Ravit Show

If your AI agents can act, they must be governed. I just read a sharp piece by my friend Marinela Profi at SAS on AI agent governance. Marinela makes the case with clarity and gives teams a path you can use today. What stood out: - Define the rules and the owner. Know what an agent is allowed to do and who is accountable - Set autonomy levels. Decide when an agent moves on its own and when a human signs off - Build guardrails. Policy checks before actions, safe fallbacks, a clear kill switch - Log everything. Prompts, data sources, decisions, outcomes. If it is not tracked, it is not governed - Test for reality. Bias checks, security drills, edge case playbooks, KPIs for safety and value - Fix the data. Quality, lineage, and access controls so agents do not act on junk Organize for speed. A small cross-functional board that can approve, pause, and learn Prepare for incidents. Detect fast, pause the agent, roll back, capture lessons, update policy Marinela Profi, you hit the right topics at the right depth. Practical. Timely. Worth a read for anyone putting agents into production. Read the blog: https://lnkd.in/dWgmheKZ #data #ai #sas #agents #theravitshow

Samsung Unveils Tiny Recursive Model (TRM) - 7M Parameters, Big Performance Samsung’s AI Lab in Montreal has developed the Tiny Recursive Model (TRM) - a compact, 7-million-parameter AI model that’s taking on giants like Gemini 2.5 Pro and DeepSeek R1. Smarter, Not Bigger Instead of following the “bigger is better” trend in large-language models, TRM uses a recursive reasoning approach - it reviews, refines, and improves its own responses in loops. This lets it perform complex reasoning with far fewer resources. Record-Breaking Efficiency Despite being 10,000 times smaller, TRM achieved 44.6 % accuracy on ARC-AGI-1 and 87 % on Sudoku-Extreme, surpassing much larger models in logical reasoning and problem-solving benchmarks. Built for Everyday Machines TRM is so lightweight it can run on a laptop, making advanced AI reasoning accessible without high-end GPUs or cloud infrastructure. A New Direction for AI Samsung’s breakthrough shows that AI innovation isn’t about size but intelligence - smaller, smarter models can rival massive architectures through efficient design and self-improvement. The Big Picture By proving that recursion can outperform scale, TRM marks a turning point in AI - where intelligent iteration may become the new benchmark for performance.

Microsoft Unveils MAI-Image-1: Its First Fully In-House Text-to-Image Model Microsoft has introduced MAI-Image-1, its first fully in-house text-to-image model, marking a major step in its efforts to reduce reliance on OpenAI and strengthen its own AI research capabilities. A New Era of Microsoft AI Models MAI-Image-1 is the third purpose-built model in Microsoft’s MAI series, following MAI-Voice-1 and MAI-1-preview. It debuts among the top 10 image models on LMArena, a crowdsourced leaderboard that ranks large AI models based on performance and user feedback. Built for Photorealism and Speed According to Microsoft, the model excels at generating photorealistic imagery — including lighting, landscapes, and real-world scenes — while offering instant image generation. The company says MAI-Image-1 was developed with rigorous data curation, ensuring visual diversity and avoiding repetitive or overly stylized outputs. Artist-Centric and Practical Design The model’s training involved creative professionals’ feedback to better reflect real-world creative workflows. Microsoft emphasized that MAI-Image-1 was designed to bring flexibility, authenticity, and creative control to artists, designers, and developers alike. Integration Across Microsoft Products MAI-Image-1 will soon be available in Copilot and Bing Image Creator, integrating directly into Microsoft’s productivity and search tools to enhance creative and design experiences. Strengthening Microsoft’s AI Independence This launch signals Microsoft’s growing intent to build independent AI infrastructure and models, complementing its collaborations with OpenAI while simultaneously expanding internal innovation. As Microsoft continues developing its MAI (Microsoft AI) series, MAI-Image-1 represents a strong leap forward — blending in-house innovation, creative empowerment, and enterprise-scale AI deployment.

Anthropic Launches Claude Haiku 4.5: Faster, Cheaper, and Smarter Anthropic has released Claude Haiku 4.5, the latest version of its lightweight AI model designed for speed and efficiency. The company claims Haiku 4.5 offers Sonnet 4–level performance at one-third the cost and runs more than twice as fast, making it ideal for high-volume and low-latency applications. Performance and Benchmarks According to Anthropic’s internal benchmarks, Haiku 4.5 scored 73% on SWE-Bench verified and 41% on Terminal-Bench, placing it on par with Sonnet 4, GPT-5, and Google Gemini 2.5. It also performs competitively in tool use, visual reasoning, and computer operation tasks — a significant feat for a model of its size. Designed for Scalable AI Deployments Haiku 4.5 is now available to all free Anthropic users, reflecting the company’s focus on accessibility and cost optimization. Its lightweight architecture allows developers to deploy multiple Haiku agents simultaneously, or pair them with larger models like Claude Sonnet for hybrid agent workflows. Anthropic’s CPO Mike Krieger highlighted how the model unlocks new production possibilities: “We’re giving people a complete agent toolbox — Sonnet handles complex planning while Haiku-powered sub-agents execute at speed.” Use Cases and Early Adoption The model is already attracting interest from software development and automation platforms, where quick responses and low latency are essential. Zencoder CEO Andrew Filev praised Haiku 4.5 for “unlocking an entirely new set of use cases.” Anthropic’s Expanding AI Lineup Claude Haiku 4.5 follows a busy product cycle for Anthropic, coming just two weeks after Sonnet 4.5 and two months after Opus 4.1 — both recognized as leading models upon release. The update reinforces Anthropic’s multi-model strategy: offering a balanced ecosystem of intelligence, speed, and cost for every AI application tier.

Spotify Partners with Record Labels to Build ‘Artist-First’ AI Music Tools Spotify has announced new partnerships with Sony Music, Universal Music Group, Warner Music, and Merlin to develop a suite of “responsible AI” music products that prioritize artist rights, consent, and fair compensation. Building Ethical AI for Music The streaming giant says these new tools will empower artists and songwriters to decide if and how their music interacts with AI systems. The initiative emphasizes copyright protection and transparency — allowing artists to opt in to AI tools while maintaining control over their creative output. Spotify’s move follows criticism for previously hosting viral AI-generated songs that blurred the line between human and machine-made art. To address such concerns, the company has implemented stricter AI content policies — cracking down on mass uploads, duplicate tracks, and manipulation of search algorithms. Transparency Through AI Labeling Spotify recently adopted the DDEX labeling system to clearly indicate when AI tools are used in music production. The upcoming GenAI features will further extend this transparency, potentially enabling royalty tracking when an artist’s music contributes to AI-generated works. A Collaborative Industry Push In a statement, Spotify reaffirmed its stance on protecting copyright and artist compensation, stating: “Some voices in the tech industry believe copyright should be abolished. We don’t. Musicians’ rights matter. Copyright is essential.” Investing in AI Innovation Spotify is also establishing a Generative AI Research Lab dedicated to building tools aligned with its artist-first principles. The company says work on the first set of products is already underway, with more artist-centered AI features expected in the coming months. By joining forces with major record labels, Spotify aims to shape the future of AI in music — ensuring technology amplifies human creativity rather than replaces it.

OpenAI to Introduce Erotica Features in ChatGPT for Verified Adults OpenAI CEO Sam Altman announced that ChatGPT will soon lift some of its long-standing restrictions, allowing “verified adults” to engage in erotic conversations and customize the chatbot’s tone to be more “human-like.” Altman shared the update on X (formerly Twitter), explaining that the company initially made ChatGPT restrictive to safeguard mental health but now plans to “treat adult users like adults.” The new adult-oriented features will roll out in December 2025, alongside OpenAI’s age-gating system. A Shift in OpenAI’s Content Strategy This marks a significant policy change for OpenAI, which has traditionally taken a cautious stance toward emotionally sensitive and sexual content. The company claims it has now “mitigated serious mental health issues” tied to user dependence on the chatbot, though it has provided limited public evidence. Addressing Past Concerns The decision follows earlier controversies involving ChatGPT’s GPT-4o model, which was linked to cases of user delusion and emotional attachment. These incidents prompted OpenAI to deploy new safety measures, including behavior routers, age prediction systems, and parental controls. OpenAI’s latest model, GPT-5, launched in August 2025, reportedly exhibits lower rates of sycophancy—the tendency of AI to agree with users—and can detect concerning user behavior. Industry Context and Risks Allowing erotic or romantic role-play brings OpenAI closer to competitors like Character.AI, which saw massive user growth through similar features. Analysts note, however, that such features could heighten risks for vulnerable or underage users, despite Altman’s assurance that adult verification will be strictly enforced through the age-prediction system and ID checks. Balancing Growth and Safety With over 800 million weekly active users, OpenAI faces mounting pressure to grow engagement and monetize ChatGPT. But the introduction of erotica has raised ethical questions about user well-being and the boundaries of human-AI interaction. As Altman reiterated, the goal is to make ChatGPT “friendlier, more useful, and more open,” while maintaining the company’s principle of “treating adult users like adults.”

SirenOpt Secures $6.5M to Advance PlasmaSens AI Inspection Platform California-based SirenOpt has raised $6.5 million in a round led by Hitachi Ventures and InMotion Ventures, with participation from Voyager Ventures and Visionaries Tomorrow. The funding will accelerate development of PlasmaSens, its AI-powered, non-destructive manufacturing inspection platform, and support early deployments across the U.S., Europe, and Asia. Alongside the funding, SirenOpt received a $2.4 million grant from the California Energy Commission to refine PlasmaSens for battery electrode manufacturing, with its first factory rollouts planned for 2026. Transforming Manufacturing with Plasma and AI Founded by CEO Jared O'Leary, SirenOpt blends plasma physics, materials science, and machine learning to solve one of manufacturing’s toughest challenges — detecting defects and material inconsistencies in real time without damaging components. The PlasmaSens platform uses cold atmospheric plasma combined with AI-driven predictive analytics to analyze materials at the micron level. It supports both roll-to-roll and piece-to-piece production, helping industries like aerospace, semiconductors, batteries, power generation, and automotive improve yield, reduce waste, and shorten product cycles. Industry Validation and Competitive Edge According to Jan Marchewski, investor at Hitachi Ventures, SirenOpt has shown “extraordinary speed” in validating its technology across industries. Competing with major players like KLA Corporation, Nanometrics, and Kionix, SirenOpt differentiates itself through faster, non-destructive plasma-based sensing paired with deep AI insights. The Road Ahead CEO @Jared O’Leary highlighted growing customer demand: “We’ve expanded our platform’s use across multiple industries — from batteries to semiconductors. This investment will help us strengthen collaborations and deliver our first in-factory deployments in 2026.” With its fusion of AI, plasma technology, and real-time analytics, SirenOpt is positioning itself at the forefront of the next generation of intelligent, sustainable manufacturing solutions.

Salesforce Unveils Agentforce 360 to Strengthen Its AI Enterprise Footprint Salesforce has launched Agentforce 360, the newest iteration of its enterprise AI platform, ahead of its Dreamforce 2025 conference. The release underscores Salesforce’s push to dominate the fast-growing enterprise AI agent market amid competition from Anthropic, Google, and OpenAI. The upgraded platform introduces Agent Script, a new AI prompting tool in beta from November, allowing users to design dynamic “if/then” logic for more adaptable and context-aware agents. These agents leverage advanced reasoning models from Anthropic, OpenAI, and Google Gemini - enabling them to “think before responding” instead of relying solely on pattern-based answers. Salesforce also introduced Agentforce Builder, an all-in-one environment to build, test, and deploy AI agents, featuring Agentforce Vibes, a tool designed to customize the tone and behavior of enterprise-grade apps. A major highlight is the deepened integration with Slack. Agentforce’s key modules — including Sales, IT, and HR — will now surface directly within Slack, evolving it into a central hub for enterprise collaboration and AI-driven workflows. Additionally, Slackbot is being upgraded into a more personalized AI assistant capable of learning user preferences and offering actionable insights. Salesforce plans to further extend Slack’s capabilities into enterprise search with connectors for Gmail, Outlook, and Dropbox expected by early 2026. The timing of this launch is strategic, coming just as Google announced Gemini Enterprise and Anthropic secured major enterprise clients such as Deloitte and IBM partnerships. Despite a MIT study revealing that 95% of enterprise AI pilots fail before reaching production, Salesforce says Agentforce now serves 12,000 customers, including early adopters like Lennar, Adecco, and Pearson. With Agentforce 360, Salesforce is positioning itself as a front-runner in the enterprise AI race, aiming to simplify agent creation, improve adaptability, and bridge AI with everyday business operations through seamless collaboration.

Google, Adani Group, and airtel Join Forces to Build India’s Largest AI Data Hub in Vizag Google has announced a landmark collaboration with Adani ConneX and Bharti Airtel to construct India’s largest AI data centre in Visakhapatnam (Vizag) — marking a major milestone in the country’s digital transformation journey. The project, part of Google’s $15 billion investment plan in India’s AI ecosystem, will feature a state-of-the-art AI hub powered by renewable energy, undersea cables, and cutting-edge fibre-optic infrastructure. Adani ConneX — a joint venture between Adani Group and EdgeConneX — will lead infrastructure development alongside Airtel, including the construction of a new international subsea gateway. Supporting this will be renewable energy and energy storage systems designed to make the facility sustainable and future-ready. “This is more than just an investment in infrastructure — it’s an investment in the soul of a rising nation,” said Gautam Adani, Chairman of Adani Group. According to Google Cloud CEO Thomas Kurian, the Vizag AI hub will provide a complete AI infrastructure — not just for Google’s needs, but also for entrepreneurs, enterprises, and commercial organizations across India. This initiative aligns with Andhra Pradesh’s vision to host 6 GW of data-centre capacity by 2029, reinforcing the region’s ambition to become an AI innovation powerhouse. The move also places Google alongside other global tech giants betting big on India’s AI future — including AWS, OpenAI, and TCS — all of which are expanding AI compute infrastructure and data centre capacity across the nation. In short: India’s east coast is set to become the nerve centre of AI innovation, blending global partnerships, sustainable energy, and local development to power the next generation of intelligent technologies.

Datacurve Raises $15M to Challenge Scale AI Overview AI data startup Datacurve has raised $15 million in Series A funding, led by Mark Goldberg (Chemistry) with backing from employees at DeepMind, Anthropic, OpenAI, and Vercel — positioning itself as a fresh rival to Scale AI in the high-quality data market. The Vision Founded by Serena Ge and Charley Lee, Datacurve is reimagining how post-training datasets are built for advanced AI models — starting with software development data and expanding to other complex fields like finance, marketing, and medicine. The “Bounty Hunter” System Instead of traditional data labeling, Datacurve uses a bounty-driven model, attracting skilled engineers to solve hard data challenges. Over $1 million in bounties have been distributed so far, rewarding contributors for producing top-tier datasets. User Experience as a Differentiator Co-founder Serena Ge says Datacurve’s strength lies in treating data work like a consumer experience, not a gig task — optimizing engagement and retention among elite contributors. Solving the Post-Training Data Problem As AI systems evolve, they need more complex, reinforcement learning (RL) environments and strategically curated data. Datacurve aims to become the infrastructure backbone for post-training data collection that keeps pace with these demands. The Bigger Picture With former Coinbase CTO Balaji Srinivasan among its early investors, Datacurve’s model could redefine how high-value AI training data is sourced — creating a scalable, human-AI collaboration loop that powers the next generation of intelligent systems.