AI News Archive: June 24, 2026 — Part 9
Sourced from 500+ daily AI sources, scored by relevance.
- Nous Research Adds /learn to Hermes Agent's Skills System, Capturing Workflows as Slash Commands Without Hand-Writing SKILL.md
Nous Research Adds /learn to Hermes Agent's Skills System, Capturing Workflows as Slash Commands Without Hand-Writing SKILL.md MarkTechPost
- RAG Evaluation 101: What to Measure (and What Not to)
Five questions, five papers, five things your RAG eval is probably getting wrong. Continue reading on Towards AI »
- BMS opens Mumbai tech hub to scale AI and data operations
The Mumbai GCC is intended to provide BMS specialist capacity as it deploys AI tools across its broking and operations footwork
- Why South African AI investments are stalling
While 83% of local CEOs say they are embedding AI across multiple workflows, only 18% of their workforce actively use AI regularly, according to IBM's 2026 CEO study.
Score: 37🌐 MovesJun 24, 2026https://www.itweb.co.za/article/why-south-african-ai-investments-are-stalling/mYZRX79gjmRqOgA8 - The Payer AI Readiness Gap: Why Better Data Will Define the Next Era of Health Plan Performance
The Payer AI Readiness Gap: Why Better Data Will Define the Next Era of Health Plan Performance MedCity News
- This AI Subscription Deal Gives You GPT, Claude, and Gemini for $70
This AI Subscription Deal Gives You GPT, Claude, and Gemini for $70 PCMag
Score: 36🌐 MovesJun 24, 2026https://www.pcmag.com/deals/this-ai-subscription-deal-gives-you-gpt-claude-and-gemini-for-70 - Ora Computing raises €3.5M to build the efficiency layer of the AI stack
Ora Computing, a startup developing softwareto optimise and compress AI foundation models, has closed a €3.5 million seedfunding round led by Constructor Capital and Greencode Ventures, with continued...
Score: 36💰 MoneyJun 24, 2026https://tech.eu/2026/06/24/ora-computing-raises-eur35m-to-build-the-efficiency-layer-of-the-ai-stack/ - Pangram CEO says language models give themselves away by making the same arguments
Language models may write cleaner prose than most humans, but ask one for 100 arguments on a topic and they'll all cluster together. Human reasoning is far more diverse, says Pangram CEO Max Spero, and that's what might give AI away. The article Pangram CEO says language models give themselves away by making the same arguments appeared first on The Decoder .
Score: 35🌐 MovesJun 24, 2026https://the-decoder.com/pangram-ceo-says-language-models-give-themselves-away-by-making-the-same-arguments/ - The Designer’s AI Glossary: Familiar Words, New Meanings
These terms help designers navigate the shift from interfaces to intelligent systems.
- Arokia IT Expands AI-Powered B2B Marketing Services to Help Technology, Manufacturing, and Healthcare Companies
Arokia IT Expands AI-Powered B2B Marketing Services to Help Technology, Manufacturing, and Healthcare Companies azcentral.com and The Arizona Republic
- Cisco says Saudi firms need secure AI ready networks by 2030
Cisco says Saudi firms need secure AI ready networks by 2030 Arabian Business
- Bubble vs. Replit Comparison: Which AI App Builder Is Right for You?
Bubble provides all-in-one AI app generation alongside a visual editor that gives anyone full control and customizability of their app. Replit provides an IDE alongside AI support for code generation, iteration, and debugging. Which is right for you? We analyzed 11 key factors to help you decide.
- Firefox Now Lets You Shake Your Android Phone to Summarise Websites
Mozilla rolled out a new AI summarisation feature for Firefox on Android. The new feature known as Page Summaries can generate a brief overview of a webpage, helping users quickly understand its key points without having to read through the entire content. The new feature works on Firefox AI cloud-based solution powered by Mistral Small 3.1. Mozilla confirmed that the...
- Zelara lands €3M to bring continuous learning to customer engagement
Zelara,a Berlin-based startup developing an AI-native learning system for lifecyclemarketing, has raised €3 million in a pre-seed funding round led by NAP, withparticipation from Heartfelt and Angel I...
Score: 35💰 MoneyJun 24, 2026https://tech.eu/2026/06/24/zelara-lands-eur3m-to-bring-continuous-learning-to-customer-engagement/ - SuperPlane secures $2.6M to turn production operations into an AI-native workflow layer
SuperPlane has raised $2.6 million in a Pre-Seed round to bring AI to the engineers managing production infrastructure. The investment was led by Credo Ventures, with participation from First Momentum...
- Wakeline lands €2.1M to bring continuous learning to AI
Düsseldorf-basedWakeline, a deeptech startup developing continuously adaptive AI systems, hasraised €2.1 million in pre-seed funding. The round was led by Aachen-basedTechVision Fonds (TVF), with part...
Score: 35💰 MoneyJun 24, 2026https://tech.eu/2026/06/24/wakeline-lands-eur21m-to-bring-continuous-learning-to-ai/ - Compri secures €3.2M to build AI-powered procurement teams
Compri, a Milan-based startup developingAI-powered procurement software for industrial companies, has raised €3.2million in a seed funding round led by Picus Capital, bringing its totalfunding to more...
Score: 35💰 MoneyJun 24, 2026https://tech.eu/2026/06/24/compri-secures-eur32m-to-build-ai-powered-procurement-teams/ - Meta Looks Good With Glasses On
Meta Looks Good With Glasses On The Information
Score: 35🌐 MovesJun 24, 2026https://www.theinformation.com/newsletters/the-briefing/meta-looks-good-glasses - A24 Knows You’re Mad About the Google AI Collab
Indie movie fans are upset about Google DeepMind’s $75 million investment in the studio, which comes as AI companies are deepening their influence in Hollywood.
Score: 35🌐 MovesJun 24, 2026https://www.wired.com/story/a24-knows-youre-mad-about-the-google-ai-collab/ - 8 generative AI certifications to grow your skills
Generative AI continues to disrupt nearly every industry as companies seek to harness the technology for digital and operational initiatives. As a result, IT professionals with highly sought gen AI skills are in high demand. Deloitte’s 2026 State of Generative AI in the Enterprise report found that 75% of organizations expect gen AI technology to impact talent strategies within the next two years, with 32% of organizations reporting very high levels of gen AI expertise already on course to make those changes. With so much at stake, gen AI skills are earning significant premiums today. To learn more about gen AI processes and tools while demonstrating to employers you have the skills to tackle gen AI projects, there are several certifications available to show your capabilities using AI in the workplace. How to choose a gen AI certification When deciding which gen AI certification is best for your career, you’ll want to reflect first on your career path and how you plan to use AI moving forward. Identify what technologies you’ll need to use, and if there are specific AI tools or services you’ll want to be trained in for your desired career path. This will help you identify the right certifications to either start or grow your career alongside the changing job market. Whether you want to learn the foundational aspects of AI or you’re an IT pro looking to get ahead of the next trend in skills, there’s an AI certification best suited for your needs. These are all designed to help you grow and validate your AI skills, and are a great place to start if you want to gain an edge in the AI hiring market. AWS Certified AI Practitioner Amazon’s AWS Certified AI Practitioner certification validates your knowledge of AI, ML, and gen AI concepts and use cases , as well as your ability to apply those in an enterprise setting. The exam covers the fundamentals of AI and ML, gen AI, applications of foundation models, and guidelines for responsible use of AI, in addition to security, compliance, and governance for AI solutions. This cert is best suited for business analysts, IT support and marketing professionals, product and project managers, LOB or IT managers, and sales professionals who want to gain an edge with AI skills. Format: Online Prerequisites: Entry-level certification with no prerequisites. Cost: $100 Certified Generative AI Specialist (CGAI) Offered through the Chartered Institute of Professional Certifications, the Certified Generative AI Specialist (CGAI) certification is designed to teach you the in-depth knowledge and skills required to be successful with gen AI. The course covers principles of gen AI, data acquisition and preprocessing, neural network architectures, natural language processing (NLP) , image and video generation, audio synthesis, and creative AI applications. On completing the learning modules, you’ll need to pass a chartered exam to earn the CGAI designation. This certification is best suited for IT professionals, software developers, data scientists and IT business leaders who want to better understand how to apply gen AI in an enterprise setting. Format: Online chartered exam with 50 multiple choice questions. Prerequisites: No prerequisites, but a foundational understanding of programming and basic AI concepts will be helpful. Cost: $550 Certified Generative AI Expert The Certified Generative AI Expert certification offered by Blockchain Council is aimed at anyone interested in gaining more expertise of gen AI. You’ll be tested on your knowledge of generative models, neural networks, and advanced ML techniques. Modules include introduction to gen AI, and gen AI for text, images, enterprises, public services, as well as data privacy in AI, prompt engineering for text analysis, and upcoming trends in gen AI. Through the self-paced certification program, you’ll have the opportunity to gain hands-on experience with gen AI tools, and you’ll earn a lifetime certification upon passing the final exam. This cert is best suited for AI professionals, data scientists, and software developers interested in understanding prompt engineering and AI as an enterprise tool. Format: Online, self-paced Prerequisites: Entry-level certification with no prerequisites. Cost: $249 Certified Prompt Engineer The Certified Prompt Engineer certification offered by Blockchain Council is designed to validate your knowledge of foundational prompt engineering topics . The self-paced course covers prompt engineering in real-world case studies and gives you the opportunity to gain hands-on experience with the OpenAI API. Modules include introduction to prompt engineering, understanding prompts, principles of effective prompt engineering, creating effective prompts, advanced prompt engineering, future of prompt engineering and AI conversations, and working with popular AI tools. This certification is best suited for AI researchers, software engineers, data scientists, and any tech professionals working with natural language processing. Format: Online, self-paced with 40 comprehensive lessons and 10 learning modules. Prerequisites: Entry-level certification with no prerequisites. Cost: $349 Databricks Certified Generative AI Engineer Associate Databricks offers the Certified Generative AI Engineer Associate certification , which evaluates your ability to use Databricks to design and implement LLM -enabled solutions. Passing this exam shows you’re capable when it comes to building and deploying performant RAG applications and LLM chains that run on the Databricks platform. The 90-minute exam covers topics such as design applications, data preparation, application development, assembling and deploying apps, governance, and evaluation and monitoring. There are no prerequisites to take the exam but it’s recommended you have six or more months of hands-on experience performing the gen AI solutions tasks outlined in the exam guide. This certification is best suited for data professionals, developers, AI solutions architects, and ML engineers. Format: Online or in-person 90-minute multiple choice exam. Prerequisites: None, but related training is highly recommended. Cost: $200 Google Generative AI Leader Google offers the Generative AI Leader certification , to validate your knowledge of how gen AI can transform businesses. This certification is for leaders who have a business-level knowledge of Google Cloud’s gen AI lineup and how to use those tools to move an enterprise forward. The exam covers the fundamentals of gen AI, Google Cloud’s gen AI offerings, techniques to improve gen AI output, and business strategies for a successful gen AI solution. It’s an entry-level certification, so there are no prerequisites to take the exam and it’s a good fit for anyone interested in learning more about gen AI, even without hands-on experience. This cert is best suited for IT business leaders, managers who must grasp the strategic value of AI without writing code, and other non-technical professionals. Format: Online 90-minute exam consisting of 50-60 multiple choice questions. Prerequisites: Entry-level certification with no prerequisites. Cost: $99 GSDC Certified Generative AI Professional The Global Skills Development Council offers the Certified Generative AI Professional certification covering the latest AI-powered agentic technologies, gen AI skills, and ethical best practices around producing and evaluating AI generated content. The courses cover topics including foundations of AI and ML, gen AI tools, governance, ethical and responsible AI, how to implement gen AI into organizations, and more. The certification consists of two exams you’ll need to pass after completing the video-led courses, which will give you hands-on experience through real-world use cases, templates, and expert resources. This cert is best suited for IT professionals, managers, consultants, developers, entrepreneurs, and beginners who want to gain foundational knowledge of AI. Format: Hybrid of online and in-person with a 48-question multiple choice exam. Prerequisites: Entry-level exam with no prerequisites. Cost: $200 Nvidia Generative AI LLMs Certified Associate The Nvidia Generative AI LLMs certification is an associate-level credential that covers the foundational concepts for developing, integrating, and maintaining AI-driven applications using gen AI and LLMs with Nvidia solutions. The exam covers topics such as fundamentals of ML and neural networks, prompt engineering, alignment, data analysis and visualization, experiment design, software development, Python libraries for LLMs, and LLM integration and deployment. Certification lasts two years, but you can get recertified by retaking the exam. Upon passing, you’ll receive a digital badge and optional certification that indicates your certification level and the topic. This certification is best suited for AI strategists, data scientists, cloud architects, gen AI specialists, ML engineers, software engineers, and software architects. Format: Online with a multiple choice exam consisting of 50 to 60 questions. Prerequisites: A basic understanding of gen AI and LLMs. Cost: $125
Score: 35🌐 MovesJun 24, 2026https://www.cio.com/article/2128415/generative-ai-certifications-and-certificate-programs.html - How Entrepreneurs Apply AI Speed Breakthroughs to Cut Costs (and Scale Smarter)
How Entrepreneurs Apply AI Speed Breakthroughs to Cut Costs (and Scale Smarter) entrepreneur.com
- Three insights you may have missed from theCUBE’s coverage of Pure Accelerate
As enterprises advance their artificial intelligence initiatives, they’re discovering that the real constraint isn’t model sophistication — It’s data. AI outcomes now depend on whether organizations can access, mobilize and operationalize data as an active system rather than a passive repository. This shift was a defining theme at Pure Accelerate 2026. The challenge is not simply […] The post Three insights you may have missed from theCUBE’s coverage of Pure Accelerate appeared first on SiliconANGLE .
Score: 34🌐 MovesJun 24, 2026https://siliconangle.com/2026/06/24/three-insights-ai-outcomes-pure-accelerate-pureaccelerate/ - Dubai Centre for Artificial Intelligence organises AI Economy for Media Professionals event
To enhance the media industry’s coverage of the technologies of the future
- C3.ai Inc. Research & Ratings | AI
C3.ai Inc. Research & Ratings | AI Barron's
Score: 33🌐 MovesJun 24, 2026https://www.barrons.com/market-data/stocks/ai/research-ratings?amp%2525253Bmod - How to Run DeepSeek Locally on Your Own Computer, and the Catch Most Guides Skip
DeepSeek became famous overnight as the open model that matched the best reasoning systems for a fraction of the cost. You can run it on your own machine, free and private, and watch it think through problems step by step. But there is a catch almost every excited tutorial glosses over, the version that made headlines is not the version you are actually running. Here is how to run it, which model truly fits your hardware, and the honest truth about what you get. DeepSeek arrived in January 2025 and did something open models were not supposed to be able to do. It matched the best closed reasoning models on hard math and coding, it was released under a permissive license anyone could use commercially, and it had reportedly been trained for a tiny fraction of what its rivals spent. It became a genuine phenomenon, the rare AI story that escaped the tech world and made the evening news. And because it is open, you can download it and run it on your own computer, with no subscription and nothing sent to anyone’s servers. Here’s the part the excited tutorials tend to skip, though, and it matters. The DeepSeek that made headlines, the full model that matched the frontier, is enormous, and you almost certainly can’t run it on your own machine. What you can run, and what nearly everyone actually runs, is a set of smaller versions that carry DeepSeek’s reasoning in a much lighter package. They are genuinely useful, and for most people they are the right answer, but they’re not the same thing as the model from the headlines. So this guide does two things, it shows you exactly how to run DeepSeek locally with real commands, and it’s honest about which version you are really getting and what that means. First, the distinction the whole thing hinges on. The catch, the full model versus the distills DeepSeek’s famous reasoning model comes in two very different forms, and understanding the difference is the key to a setup that does not disappoint you. The full model is a giant. It uses a mixture-of-experts design with hundreds of billions of parameters, and running it requires data-center hardware, multiple high-end GPUs and hundreds of gigabytes of memory. This is the version that matched the best closed models in the benchmarks you read about. It’s not something you run on a laptop or even a powerful gaming desktop, and any guide implying otherwise is setting you up for frustration. What you actually run at home are the distilled versions. DeepSeek took its big model’s reasoning behavior and trained it into a range of much smaller models, a process called distillation, where a small model learns to imitate the reasoning patterns of the large one. These distilled models come in sizes from tiny to fairly large, and they run comfortably on ordinary hardware. They keep a real portion of the reasoning ability, enough to be genuinely useful and to clearly outperform normal models of the same size on math and logic, but they’re not the full model and they fall short of it on the hardest problems. That’s the honest trade. You’re getting most of the reasoning behavior in a package you can actually run, not the headline model itself. Keep that framing and you’ll be happy with the result. Forget it, and you’ll wonder why your laptop isn’t matching the benchmarks. With that settled, here is how to run one. The setup, in a few minutes with Ollama Running DeepSeek locally is genuinely easy, because a free tool called Ollama handles all the hard parts, the download, the memory management, the GPU acceleration, behind a single command. The whole process is four steps, and the only real decision, which size to download, is covered in the next section, so glance at that first to find your model name, then come back here. First, install Ollama. Go to ollama.com, download the version for your system, and install it. On a Mac you drag it to your Applications folder, on Windows it installs like any normal program, and on Linux it is a single command in the terminal. # Linux install (Mac and Windows use the downloaded installer instead) curl -fsSL https://ollama.com/install.sh | sh Second, open a terminal and confirm it works. On a Mac, open the Terminal app, found in Applications under Utilities or by searching Spotlight for Terminal. On Windows, open PowerShell from the Start menu. On Linux, open your usual terminal. Then type the command below, and a version number means you are ready. ollama --version Third, pull and run your DeepSeek model. Use the size you picked from the next section. As an example, the command below downloads and starts the 8-billion version, which is a good default for most laptops. The first run downloads the model, which takes a few minutes, then drops you straight into a chat where you simply type your question at the prompt. ollama run deepseek-r1:8b Fourth, test it and exit. Ask it a reasoning question, a math word problem is a good one, and watch it think through the answer. When you are done, type /bye to leave the chat. To start it again later, you run the same ollama run command, and since the model is already downloaded, it starts instantly. To use the model from your own apps or scripts rather than the terminal, Ollama also runs a local server at http://localhost:11434 that anything on your machine can send requests to, which is how you would wire it into a code editor or a program. That’s the whole mechanical process. The only real decision is which size to pull, so here is how to choose. Picking the size that fits your machine DeepSeek’s distilled models come in a wide range, and matching the size to your hardware is what separates a smooth experience from a sluggish one. Here’s the honest breakdown, with the command for each. ollama run deepseek-r1:1.5b # ~2 GB, runs on almost anything, for testing ollama run deepseek-r1:8b # ~8 GB RAM, the everyday pick for most laptops ollama run deepseek-r1:14b # 12-16 GB, the value sweet spot ollama run deepseek-r1:32b # ~24 GB VRAM (RTX 3090/4090), best local quality ollama run deepseek-r1:70b # ~40 GB, needs dual GPUs or a high-memory Mac On a basic laptop or any modest machine, start with the small distills. The smallest version runs on almost anything, needing only a couple of gigabytes of memory, and it’s useful mostly for testing that your setup works. The 7-billion or 8-billion version is the realistic everyday pick, needing around 8 gigabytes of memory, and it produces reasoning that is meaningfully better than ordinary models its size. For most people on a normal laptop, the 8-billion model is the right starting point. On a machine with 16 gigabytes of memory or a midrange graphics card, the 14-billion version is the value sweet spot. It needs roughly 12 to 16 gigabytes, produces consistently strong reasoning on math, coding, and logic, and is the size where the reasoning starts to feel genuinely capable rather than just better-than-small. If your hardware can hold it, the 14-billion model is the choice that gives most people the best balance. On a desktop with a strong graphics card, around 24 gigabytes of video memory like an RTX 3090 or 4090, the 32-billion version is the local sweet spot for quality. It uses roughly 20 gigabytes, runs fast on that hardware, and delivers reasoning close to the full model on many tasks, which is about as good as local reasoning gets on a single consumer card. There are larger distills still, a 70-billion version that needs around 40 gigabytes and therefore dual GPUs or a high-memory Mac, but for almost everyone the 32-billion model is the practical ceiling and a genuinely powerful place to land. The way to check whether you chose well is to run ollama ps in another terminal while the model is loaded. If it shows a high CPU percentage, the model is spilling out of your graphics memory onto your processor, which works but is slow, and the fix is to drop to the next size down. A fast smaller model beats a crawling larger one for nearly everything. What makes DeepSeek different to use, the visible thinking One thing will stand out the moment you run it, and it’s worth understanding because it’s the whole point of this model. Before DeepSeek gives you an answer, it shows its thinking, a stream of reasoning wrapped in think tags where it works through the problem, considers approaches, catches its own mistakes, and only then arrives at a final answer. This isn’t a quirk, it’s the feature. Most models hand you a conclusion and hide how they got there. DeepSeek lets you watch the reasoning unfold, which is genuinely useful, because you can see where its logic is solid and where it went astray, and that transparency helps you trust or question the answer rather than taking it on faith. For debugging code or checking a math result, watching the step-by-step work is often more valuable than the answer itself. If you find the thinking output distracting for simple tasks, you can turn it off, in Ollama with the command /set parameter think false typed in the chat, but for anything that benefits from reasoning, leaving it on is the reason you are running this model in the first place. A few DeepSeek-specific gotchas worth knowing This model behaves a little differently from a standard chatbot, and a few non-obvious settings make a real difference. First, raise the context length. Ollama defaults to a fairly short context, and because DeepSeek’s reasoning chains can run long, that default will cut its thinking off mid-thought. The fix is one command typed inside the chat session, which raises the context for that session. # Run this inside the chat (at the >>> prompt) to give it room to reason /set parameter num_ctx 16384 Set it to at least 8192, and 16384 is a safer choice that gives the model room to reason fully. Second, the prompting advice is the opposite of what you might expect. Do not tell DeepSeek to think step by step, because it already reasons internally and the instruction interferes with its native process. Do not lean on system prompts or worked examples either, since this model tends to do better with a direct question and all your instructions placed in the main message. A temperature setting around 0.6 is the recommended sweet spot, low enough to stay coherent and high enough to avoid repetitive output, and you can set it the same way with /set parameter temperature 0.6 inside the chat. And finally, if you are wiring the model into an automated workflow that expects clean structured output, be aware that the visible thinking tags appear before the answer and can confuse a program parsing the response, so you strip them out before passing the result downstream. A note on privacy, since it comes up Because of where DeepSeek comes from, privacy questions come up often, and the honest answer depends entirely on how you use it. The privacy concerns you may have read about relate to DeepSeek’s hosted service, its website and its cloud API, where your data is processed on the company’s servers and is subject to the laws and policies that govern them. That’s a legitimate consideration for the hosted product. But running the open model locally through Ollama is a completely different situation. You are using only the downloaded model weights on your own hardware, with no connection to the company’s servers at all, so nothing you type ever leaves your machine. This is, in fact, one of the strongest reasons to run it locally rather than use the app, you get the model’s capability with none of the data leaving home. For sensitive work, local is the answer, and it sidesteps the concern entirely. What you end up with Put it together and you have a capable reasoning model running privately on your own computer, for free. You can hand it a hard math problem, a logic puzzle, or a buggy piece of code, watch it reason through the steps, and get an answer you can actually inspect, all without sending anything to anyone. For a lot of people, that is a genuinely powerful tool to own outright. Just hold on to the honest framing this guide opened with. You’re running a distilled version that carries most of DeepSeek’s reasoning, not the full headline model, and within that reality the experience is excellent, especially at the larger sizes your hardware can support. Pick the biggest distill your machine runs comfortably, give its reasoning room to breathe with a longer context, and resist the urge to over-instruct it. Do that, and you have one of the most capable open reasoning models thinking through your problems on your own machine, in about five minutes and at no cost. That is a striking thing to be able to say, and the only trick is being clear-eyed about which version you are actually getting. This is one in a set of hands-on guides to running the major open models yourself. If you run DeepSeek locally, drop a comment with your hardware, the distill size you settled on, and how the reasoning held up on your real tasks. The honest experience of people actually running these models is worth more than any benchmark table. Resources The DeepSeek R1 models on Ollama, with every size and tag: https://ollama.com/library/deepseek-r1 Ollama, the easiest way to run DeepSeek locally on your own computer: https://ollama.com The DeepSeek R1 paper, on the reasoning approach and the distillation results: https://arxiv.org/abs/2501.12948 A hardware-focused breakdown of which DeepSeek distill fits which GPU or Mac: https://llmhardware.io/guides/deepseek-hardware-requirements A practical Ollama setup walkthrough with VRAM requirements for each size: https://www.sitepoint.com/deepseek-r1-local-deployment-guide-2026/ How to Run DeepSeek Locally on Your Own Computer, and the Catch Most Guides Skip was originally published in Towards AI on Medium, where people are continuing the conversation by highlighting and responding to this story.
- Future Ready Now: Menlo College Expands AI & Analytics Education with Student Innovation and Silicon Valley Partnerships
Future Ready Now: Menlo College Expands AI & Analytics Education with Student Innovation and Silicon Valley Partnerships azcentral.com and The Arizona Republic
- Vornado and Other Office REITs Are Rebounding as AI Fears Fade
Vornado and Other Office REITs Are Rebounding as AI Fears Fade Barron's
- 'Intelligentisation' phase of hospital transformation
'Intelligentisation' phase of hospital transformation Healthcare IT News
Score: 31🌐 MovesJun 24, 2026https://www.healthcareitnews.com/news/asia/intelligentisation-phase-hospital-transformation - Figma's 2026 AI report: Can AI help us collaborate better?
AI used to be a solo act. According to our research, it’s not anymore.
- The 9 best AI voice generators
Recording a voiceover is challenging enough. You go through way too many takes to get what you want. You don't have enough time to rehearse and hit your tone and intention targets. You read endless audio editing software guides to make sure your voice sounds good. And even if you nail all of these things, if you don't have access to a studio, your perfect performance will be riddled with background noise. So should you give up and hire a voice actor? Not yet: AI voice generators can deliver impr
- 72% of Indian BFSI leaders report growing demand for instant resolution; only one-third are redesigning workflows to meet It: Genesys and Dun & Bradstreet research
Genesys released findings from a new study conducted in partnership with Dun & Bradstreet examining the trends in customer experience across India's Banking, Financial Services and Insurance (BFSI) sector. The report reveals a significant gap between the pace of rising customer expectations and the operational readiness of BFSI organisations to meet them. The post 72% of Indian BFSI leaders report growing demand for instant resolution; only one-third are redesigning workflows to meet It: Genesys and Dun & Bradstreet research appeared first on Express Computer .
- And what happens next?
In the game "The choice before us" by Nick Shapiro, [1] you are put in the shoes of an AI company leader. You grow your business. You unlock "wonders", such as curing cancer. All the while, you're attempting to avoid your product getting smart enough to escape and take over. You win by achieving 5 wonders without unleashing uncontrolled AI. I love this game, but it has the major flaw that when you win, you are normally very close to superintelligence. What happens afterwards? You turn the GPUs off? Go home? Get some sleep? The game seems to think so. This failure to ask "What happens next?" seems to be a broader phenomenon within the AI community. It was in fact the sole question I needed to ask a capabilities researcher for them to take the threat of superintelligence seriously. It's my main weapon against people claiming there are many possible worlds "where only 90% of people die" (if a rogue AI has gone off the rails and killed 90% of your population, you probably no longer have control of the planet, and I have little faith in the survival of everybody else). More broadly, I just wish people would ask this question more often. "But Sean!" you say. "I cannot keep asking what comes next forever. I'll end up wondering what happens after the pope becomes US president in the year 2124." And you would be correct! You cannot, in fact, keep on doing this forever. The tree of possibilities is infinite, and spending your life exploring them is reserved for the brave of heart, the noble of mind, and those who have nothing more productive to be doing. We would be better off finding a place to stop. To work out how, let's look at a simpler version. In particular, the game of chess. More specifically, this position: In case you're wondering, this position is a mess For the uninitiated, this is a complex position. The white king is threatened. The black queen is under attack. So is the white rook, a white bishop and one of the white knights. More broadly, the arrows in the image show the main pieces of tension that one would want to resolve before attempting any sort of evaluation; it's no use counting up the pieces if you lose your queen next turn! In chess engine parlance, we would like to perform a "quiescence search" . This search goes only through the moves which resolve tension, arriving at a position like the following: [2] With all the mess cleared, we can now see that black is up a bishop and a knight and will probably win the game. The point of this exercise? When evaluating the future, only evaluate futures where things are relatively stationary. The world will be changing, naturally, but there is a difference between worlds where Anthropic takes over the US government and worlds where capabilities plateau. (Oh, I'm sorry, did you think that a government takeover was a straightforward lock-in scenario? When there is no precedent for it in US history? When this is flipping the power structure of the entire globe?). Don't think your way halfway into the singularity and declare that we're in a good position. That's how you lose your queen. Get to a stable point. They're safer. ^ Which is great and you should try out by the way ^ After the moves 1. Kh1 Qxb4 2. Rb1 gxh6 3. d3 Bxa7 Discuss
- Config 2026: New materials, new tools and a more expressive canvas
Push past what you thought was possible with code layers, Figma Motion, shaders, generative plugins and Weave tools, all on the canvas.
- How to Securely Connect Your AI Agent to Telegram with Azure
Personal AI agents are starting to take a major leap forward. Projects like OpenClaw show that we are no longer talking only about assistants capable of answering questions, but about systems that can access tools, consult information on your computer, execute actions, and even act on our behalf. But the more power we give them, the more important one issue becomes an issue that is often pushed into the background: security . Imagine that we have built our local agent, connected tools to it, and now want to interact with it from anywhere using our mobile phone. Telegram seems like an ideal option: it is reliable, available on virtually any device, and offers a free, very mature API. However, there is a problem that is rarely addressed in enough depth. The most natural way to connect any agent to Telegram forces you, as a user, to make an uncomfortable decision: open a port on your computer or leave the bot token exposed on your local machine . Either option creates an attack surface that, in an agent with access to your tools, you simply cannot afford. In this article, we will see how we solved this problem in AzulClaw through an architecture based on Azure Bot Service, Azure Functions, and Azure Service Bus, completely eliminating the need to expose the local machine to the Internet . How the Problem Is Usually Solved Telegram offers two main mechanisms for receiving messages from a bot. Long Polling: With this approach, the application makes periodic requests to Telegram to check whether there are new messages. The advantage is that there is no need to expose any port or have a public URL; we are the ones contacting Telegram. However, it has several limitations when we work with AI agents. If the agent needs several seconds to reason through a response using a language model, the reading process can become blocked. Messages start to accumulate, and handling errors, restarts, or process crashes ends up depending entirely on our code . As the agent becomes more complex, this approach becomes harder to maintain. Webhooks The second option is to register a public URL so Telegram can send messages directly in real time. From an operational point of view, it is a more elegant and efficient solution. The problem appears when the agent runs on our own machine. To receive those messages, we must expose an endpoint accessible from the Internet . In many projects, this involves configurations similar to: { "channels": { "telegram": { "webhookUrl": "https://mi-dominio.com/telegram-webhook", "webhookHost": "0.0.0.0" } } } That value means the application will listen on all available network interfaces. It is not a bad practice in itself. In fact, it is the approach used by many open-source projects and is perfectly valid for many scenarios. But when the agent has access to personal files, corporate tools, private APIs, or execution capabilities, the situation changes considerably. How Can We Approach It Securely? Placing the Cloud Between Telegram and Your Agent The key is to question an assumption we usually take for granted. What if the agent did not have to communicate directly with Telegram? Instead, we can introduce an intermediate cloud layer responsible for receiving messages, validating them, and securely transporting them to our local agent. In this way, Telegram never interacts directly with the machine where the agent is running . This is how the architecture we implemented in AzulClaw works: Each component has its role: Azure Bot Service is the official Telegram channel. It registers the webhook, manages the bot token, and translates messages into the standard Bot Framework format. The Telegram token lives in Azure, not on your computer. Azure Function is the relay. It receives activities from Bot Service, authenticates them with a JWT signed by Microsoft, and puts them into the queue. It is also responsible for collecting the response and returning it. Azure Service Bus is the message bus. Two queues: `bot-inbound` for incoming messages and `bot-outbound` for responses. Your local agent only needs the connection string: no public IP, no open port . AzulClaw running locally reads from the queue, processes the message, and puts the response back. Everything is outbound traffic. Never inbound. Three Additional Security Layers Removing public exposure of the agent is a major step, but it is not enough. We must also protect the endpoint that receives messages in Azure . For that reason, the relay implements three complementary security mechanisms. Layer 1: Bot Framework JWT. Each incoming request includes a token signed by Microsoft. The Function validates it against the Bot Framework authentication services. If it fails, `401`. No exceptions. Layer 2: User and chat allowlist. Even if a request is authentic and comes from Azure Bot Service, you can restrict which Telegram users or groups are allowed to interact with your agent. Numeric IDs are configured as environment variables in the Function. Layer 3: Correlation ID for synchronous responses. The relay assigns a unique ID to each message. When AzulClaw processes and responds, it uses that same ID as the session ID in Service Bus. The Function waits for the response to that specific message , not just any message, using Service Bus Sessions. # The relay in the Function: receives, authenticates, enqueues, waits correlation_id = str(uuid.uuid4()) # Authenticate Bot Framework JWT is_authorized, _, _ = await _authenticate_request(req_body, auth_header) if not is_authorized: return func.HttpResponse(status_code=401) # Only allowed users pass if not evaluate_telegram_access(req_body, ALLOWED_USERS, ALLOWED_CHATS).authorized: return func.HttpResponse(status_code=200) # silent # Enqueue for AzulClaw await sender.send_messages(ServiceBusMessage( json.dumps(req_body), correlation_id=correlation_id )) # Wait for synchronous response by session ID reply = await _await_outbound_reply(client, correlation_id) The Local Experience: Zero Firewall Changes From the user’s point of view, the integration is extremely simple. The local runtime includes a worker that: 1. Reads messages from the input queue. 2. Processes them using the agent. 3. Publishes the response to the output queue. The only required dependency is a Service Bus connection string . There is no need to open ports . There are no TLS certificates to maintain . It does not matter whether the agent is behind a corporate VPN, a home network, or NAT. In addition, if the agent restarts or remains temporarily disconnected, messages continue to be stored in the queue until it can process them. The architecture naturally provides resilience . More Than a Telegram Integration Although the example uses Telegram, the real value of this pattern goes far beyond one specific platform. The core idea is to completely decouple external channels from the agent we run locally . Telegram is only one of those channels. The same approach can be applied to mobile applications, enterprise services, messaging platforms, or any system that needs to communicate with a private agent. In addition to improving security, we gain further advantages: Decoupling between components. Greater resilience against restarts or failures. Scalability. Message traceability. The ability to incorporate multiple agent instances. Security in the Enterprise Environment As AI agents move from experimental projects into corporate environments, the standards change. In an enterprise network, opening ports or relying on vulnerable local configurations is not acceptable . To solve this at the root, at AzulClaw we chose to fully isolate execution . When you delegate critical actions to a system that acts on your behalf, the architecture must be secure by default . If you want to implement this model, the full code — the Azure Function, the Terraform module, and the integration with the local runtime — is available in [AzulClaw] . The official Telegram skill includes everything needed to deploy this in any Azure subscription. How to Securely Connect Your AI Agent to Telegram with Azure was originally published in Towards AI on Medium, where people are continuing the conversation by highlighting and responding to this story.
- Opinion | Human professors are the best antidote to AI cheating
Opinion | Human professors are the best antidote to AI cheating The Boston Globe
Score: 30🌐 MovesJun 24, 2026https://www.bostonglobe.com/2026/06/24/opinion/letter-artificial-intelligence-college-classroom-cheating/ - Stocks Ignore Oil Slump As Tech, Fed, and AI Worries Weigh
Stocks Ignore Oil Slump As Tech, Fed, and AI Worries Weigh Barron's
Score: 30🌐 MovesJun 24, 2026https://www.barrons.com/articles/stocks-oil-slump-markets-worries-5c64f6f7?mod - When Doing More Work in Less Time Feels Wrong: The Rise of AI Productivity Guilt
As automation tools compress days of work into mere minutes, employees are facing a strange psychological side effect: the toxic pressure to constantly look busy.
- The 5 best AI app builders in 2026
Even using no-code, building a new app can take a big chunk of your time. Setting up data sources requires smart planning and foresight. Building an intuitive user interface takes multiple tries until you find the perfect layout. And tying it all together with bug-free app logic demands attention to detail and many rounds of testing. AI helps in two ways here. The first is by turning your prompt into a first-draft app, speeding up setup. The other is by building solutions with code and placing t
- The 8 best AI presentation makers in 2026
The days of spending hours dragging images to just the right place on a slide are far behind us. You can now create a presentation with AI, giving the robots the job of setting the structure, adding the initial content, and executing on the aesthetics of your deck. All you have to do is tweak it with your information, human insights, and flair—and rehearse it a bit before the big performance. I spent a month testing all the best presentation software, focusing on the ones that use AI in a way t
Score: 30🌐 MovesJun 24, 2026https://zapier.com/blog/best-ai-presentation-maker-ai-presentation-maker - OpenAI says ChatGPT Instant now better understands what users actually want
OpenAI is updating GPT-5.5 Instant, its most-used ChatGPT model. The update targets conversation quality, with better intent recognition, improved context across multiple turns, and more reliable handling of complex, multi-condition prompts. The article OpenAI says ChatGPT Instant now better understands what users actually want appeared first on The Decoder .
Score: 29🌐 MovesJun 24, 2026https://the-decoder.com/openai-says-chatgpt-instant-now-better-understands-what-users-actually-want/ - Why healthcare CIOs must fix their foundations before scaling AI
Why healthcare CIOs must fix their foundations before scaling AI Healthcare IT News
Score: 29🌐 MovesJun 24, 2026https://www.healthcareitnews.com/news/why-healthcare-cios-must-fix-their-foundations-scaling-ai - DoubleLine’s Sherman: AI ‘Early’ in Tapping Bond Market
Jeff Sherman, Deputy Chief Investment officer at DoubleLine Capital, joined the program to analyze recent developments in the U.S. Treasury market. The discussion focused on recent moves in Treasury yields, including a notable drop in the 10-year yield by 10 basis points to just above 4.3%. He speaks with Romaine Bostick & Katie Greifeld on "The Close." (Source: Bloomberg)
Score: 28🌐 MovesJun 24, 2026https://www.bloomberg.com/news/videos/2026-06-24/doubleline-s-sherman-ai-early-in-tapping-bond-market-video - Linear Trees: What If Every Decision-Tree Leaf Had Its Own Linear Model?
Bridging the gap between the threshold-finding power of trees and the predictive elegance of linear math. Link: https://unsplash.com/photos/a-computer-screen-with-a-bunch-of-code-on-it-ieic5Tq8YMk As data scientists, our machine learning toolkits often force us into a frustrating ultimatum, making us choose between two entirely different worlds of modeling. On one side of the ring, we have standard linear regression. It is the reliable workhorse of the industry — fast, incredibly simple, and highly interpretable. Every feature receives a clean coefficient, allowing us to walk into a meeting and clearly explain exactly how a prediction changes when a specific feature changes. But linear regression carries one massive, glaring weakness: it arrogantly assumes that a single, rigid equation can accurately describe your entire dataset. On the other side of the ring, we have decision trees. They are wonderfully flexible, capable of capturing nonlinear relationships, complex thresholds, and feature interactions without requiring us to manually engineer them. However, standard regression trees have their own fatal flaw: they make constant predictions inside each leaf. They treat continuous data like flat platforms, making their predictions look like a clunky staircase rather than a smooth, realistic curve. So, let’s ask the obvious question: what happens when we combine the hierarchical structure of a decision tree with the smooth predictive behaviour of linear regression? We get a Linear Tree . A Linear Tree divides your data into different mathematical regions using decision-tree rules, and then fits a separate, distinct linear model inside each of those regions. It is an algorithm that acknowledges a simple truth: one straight line cannot describe the entire world but several local straight lines absolutely can. Problem with Linear Regression To understand why this is so powerful, let’s look at real estate. Suppose we want to build a model to predict apartment rental prices in a bustling tech corridor like Mahadevapura, using only property size. A simple linear regression model might learn a global rule that looks like this: Rent = ₹15,000 + (₹30 × Square Footage) This model assumes that every single additional square foot contributes the exact same amount of value to the rent, no matter what. But anyone who has ever hunted for an apartment knows real markets do not behave that way. For a tiny studio apartment, an additional 100 square feet is life-changing and highly valuable. For a massive luxury Colive space, a few extra square feet barely register. The real relationship in the data is piecewise. It actually looks more like this: If Size < 500 sq ft: Rent = ₹10,000 + (₹40 × Size) If Size 500 to 1,500 sq ft: Rent = ₹20,000 + (₹25 × Size) If Size > 1,500 sq ft: Rent = ₹45,000 + (₹15 × Size) Instead of forcing one stubborn line through every single observation, we need a different line for each distinct market segment. A standard linear model completely chokes on this unless we spend hours manually creating data transformations, interaction variables, and arbitrary threshold indicators. A Linear Tree learns these specific regions automatically. Problem with Decision Trees A standard regression tree attempts to solve this nonlinearity problem in a completely different way. It repeatedly slices the data using rigid, binary rules. It might ask: Is the apartment size below 500 square feet? Yes: Predict a flat ₹25,000. No: Is the size below 1,500 square feet? Yes: Predict a flat ₹45,000. No: Predict a flat ₹70,000. This allows the model to capture those market thresholds easily. However, look at what happens inside the leaf. Every single apartment that reaches the same leaf receives the exact same prediction. A 510-square-foot flat and a 1,490-square-foot flat will receive the identical ₹45,000 prediction because they landed in the same bucket. If we plot its predictions on a graph, we get a chunky staircase. We could try to fix this by building a massively deep tree to create smaller steps, but that introduces a host of other headaches: massive complexity, zero interpretability, terrible behavior outside the training range, and a massive risk of overfitting. Instead of creating dozens of tiny, constant, flat regions, a Linear Tree creates a few meaningful, broad regions — and then fits a mathematical trendline inside each one. So what exactly does a linear tree do? A Linear Tree is simply a decision tree where the terminal leaves contain linear models instead of flat constants. A standard tree looks like this: Feature A < 10? Yes: Predict 25 No: Predict 68 A Linear Tree looks like this: Feature A < 10? Yes: y = 4 + (2 × Feature X) - (0.5 × Feature Z) No: y = 30 + (0.8 × Feature X) + (2 × Feature Z) The tree structure still acts as the traffic cop, deciding which specific region an observation belongs in. But once the observation reaches its final leaf, the final prediction is calculated dynamically using that leaf’s unique linear equation. How does a linear tree learn? Step 1: Fit a global linear model At the root of the tree, the algorithm looks at all available training data and fits a standard linear model. This is the baseline. Step 2: Evaluate candidate splits The algorithm then looks for ways to slice the data. It considers possible splits like age < 35 or income < ₹70,000. For every single candidate split, it divides the observations into two distinct groups, and fits a brand-new, separate linear model to each group. Step 3: Measure the mathematical improvement The algorithm calculates the error of the single parent model, and compares it to the combined error of the two new child models. If splitting the data into two linear regimes significantly reduces the overall error, the split is locked in. Step 4: Repeat recursively This process continues deeper into the child branches. The algorithm keeps splitting and fitting lines until it hits a stopping condition: reaching a maximum depth, running out of data points in a leaf, or failing to find a split that improves the error. Step 5: Predict When a new data point arrives, it trickles down the decision rules until it lands in a terminal leaf, where the local linear model calculates the final output. Implementing Linear Trees in Python The Python ecosystem has a package called linear-tree that provides scikit-learn-style estimators. Let’s look at how clean this is to implement. We can even pass regularized models (like Ridge or Lasso) into the leaves to prevent our coefficients from going crazy. import numpy as np import matplotlib.pyplot as plt from sklearn.linear_model import LinearRegression, Ridge from sklearn.tree import DecisionTreeRegressor from lineartree import LinearTreeRegressor # --------------------------------------------------------- # 1. Generate the Piecewise "Real Estate" Dataset # --------------------------------------------------------- np.random.seed(42) X = np.linspace(0, 20, 300).reshape(-1, 1) # Create a behavioral shift at X = 8 (The Mathematical Breaking Point) # If X < 8, steep slope. If X >= 8, shallower slope. y = np.where( X < 8, 10 + 4.5 * X, # Regime 1: Steep linear trend 46 + 1.2 * (X - 8) # Regime 2: Flatter linear trend ) # Add some real-world noise y += np.random.normal(0, 2.5, size=y.shape) # --------------------------------------------------------- # 2. Initialize and Train the Models # --------------------------------------------------------- # Linear Regression (The mediocre middle line) linear_model = LinearRegression() # Regression Tree (Max depth 2 creates a 4-step staircase) tree_model = DecisionTreeRegressor(max_depth=2, random_state=42) # Linear Tree (Max depth 1 creates a single split with two lines) # We pass Ridge to keep the coefficients stable inside the leaves linear_tree_model = LinearTreeRegressor(base_estimator=Ridge(), max_depth=1) # Fit all models linear_model.fit(X, y) tree_model.fit(X, y) linear_tree_model.fit(X, y) # --------------------------------------------------------- # 3. Generate Predictions for the Plot # --------------------------------------------------------- X_plot = np.linspace(0, 20, 500).reshape(-1, 1) y_lr = linear_model.predict(X_plot) y_tree = tree_model.predict(X_plot) y_lt = linear_tree_model.predict(X_plot) # --------------------------------------------------------- # 4. Create the Visualization # --------------------------------------------------------- plt.figure(figsize=(12, 7)) # Plot the raw training data plt.scatter(X, y, color='gray', alpha=0.4, label='Actual Data', edgecolors='k') # Plot the competing models plt.plot(X_plot, y_lr, color='#3498db', linestyle='--', linewidth=3, label='Linear Regression (Global)') plt.plot(X_plot, y_tree, color='#e74c3c', linestyle='-.', linewidth=3, label='Regression Tree (Staircase)') plt.plot(X_plot, y_lt, color='#2ecc71', linewidth=4, label='Linear Tree (Piecewise)') # Formatting for a clean, Medium-ready aesthetic plt.title('Algorithm Showdown: Forcing One Line vs. Finding Two', fontsize=16, fontweight='bold', pad=15) plt.xlabel('Feature (e.g., Square Footage)', fontsize=12, labelpad=10) plt.ylabel('Target (e.g., Rent Price)', fontsize=12, labelpad=10) # Add a vertical line to explicitly show the breaking point plt.axvline(x=8, color='black', linestyle=':', alpha=0.5, label='Hidden Threshold (X=8)') plt.legend(fontsize=11, loc='lower right', framealpha=0.9) plt.grid(True, linestyle='--', alpha=0.5) plt.tight_layout() # Show the plot plt.show() Output Limitations of Linear Trees 1. Discontinuous Boundaries: Because each leaf is an isolated mathematical island, predictions can jump abruptly at the boundaries. If the tree splits at an income of ₹50,000, two users making ₹49,999 and ₹50,001 might receive drastically different predictions because they are evaluated by entirely different leaf equations. 2. Coefficient Instability in Small Leaves: If you let your tree grow too deep, you might end up with a leaf containing only 15 data points, trying to fit a linear model with 10 features. The math will completely break down. This is why using regularized models (like Ridge or Lasso) inside the leaves is critical — it shrinks the coefficients and prevents the local models from overfitting to noise. Conclusion Linear Trees are built on a beautifully pragmatic observation: complex, messy, global data is usually just a collection of simpler, local trends. A single linear model is far too restrictive for the real world. A standard decision tree is far too crude. A Linear Tree strikes the perfect balance, giving you the threshold-finding power of tree splits, combined with the smooth, trend-capturing elegance of linear math. Linear Trees: What If Every Decision-Tree Leaf Had Its Own Linear Model? was originally published in Towards AI on Medium, where people are continuing the conversation by highlighting and responding to this story.
- In the Age of AI, Your Expertise Matters More Than Ever
A LinkedIn executive explains why solo entrepreneurs who pair deep expertise with AI tools are finding greater success.
Score: 28🌐 MovesJun 24, 2026https://www.inc.com/brian-honigman/age-of-ai-your-expertise-matters-more-than-ever-solo-entrepreneur/91364578 - What are the minimum skills for AI use?
What are the minimum skills for AI use? IT Pro
Score: 27🌐 MovesJun 24, 2026https://www.itpro.com/business/careers-and-training/what-are-the-minimum-skills-for-ai-use - FOD#157: What People Still Don’t Understand About AI Agents
The BioNeMo Confusion
Score: 26🌐 MovesJun 24, 2026https://www.turingpost.com/p/ai-agent-toolkits-what-people-still-don-t-understand - Your enterprise AI agents should automatically remember which model is right for which task. Mindstone built the capability with Rebel
AI agent orchestration platforms are popping up like weeds these days, but London-based AI transformation startup Mindstone's Rebel might be among the most promising I've come across. That's because the system, which officially launched this week, is a local-first, agentic AI operating system distributed under a " Fair Source " license, allowing teams of under 100 users to freely adopt and customize it to suit their needs, while those organizations with more users will require paying for an enterprise license. The marquee features are its simplicity and extensive customizability to fit any given team, no matter how unique or specific the workflows, all based around the common, open source standard file format markdown, and, as a result, an organizational memory layer that ensures agents reliably use the enterprise's preferred AI models for each given task or even subtasks — dynamically switching between local and cloud ones in a predictable, visible way to save costs and maintain data privacy and security as needed. "Shared memory is the most empowering thing you could possibly do with a knowledge-worker AI," said Greg Detre, chief technology officer (CTO) of Mindstone, in a recent video call interview with VentureBeat. "You get this feeling of being a super-organism as a company that just gets smarter and smarter." Rebel is available now for macOS on Intel and Apple Silicon machines, as well as Windows, with Linux support in development. Mindstone has raised $5 million from private investors including Pearson Ventures, Moonfire Ventures and Zanichelli Venture. A distinctive, local-first architecture based on markdown files What makes Rebel distinctive is its local-first architecture. Instead of the approach found in developer-heavy agent frameworks such as as LangGraph, CrewAI and AutoGPT, which require teams to wire together databases, cloud infrastructure and state-management logic, Rebel's core agent memory and instructions live across local markdown ( .md ) text files — arguably the simplest, easiest, and most popular way to steer AI agents, one that has been widely adopted by AI developers and power users around the globe. Mindstone says Rebel stores its state, prompts, task instructions and memory hierarchy in these files, allowing users and companies to easily inspect, move or modify them as needed. A primary configuration file, agents.md, acts as the agent’s core instruction layer and runtime boundary. That architectural choice is partly about cost. Mindstone argues that common office formats such as Word documents and PDFs often carry formatting and metadata overhead that consumes model token context and raises API costs. Markdown keeps the information closer to raw text, allowing more of the model’s context window to be spent on the actual task rather than document structure. The company also positions the approach as a hedge against vendor lock-in. If a company’s agent instructions, automations and memory are stored locally as text files, they are not trapped inside one SaaS provider’s interface or database. That matters more as enterprises begin giving AI systems broader access to email, calendars, documents and internal workflows. Rebel also lets users create repeatable AI workflows. “Skills” are saved multi-step procedures an agent can reuse. “Operators” adjust how the agent behaves for a given task, such as reviewing a pitch deck from an investor’s perspective or evaluating work through a security lens. “Automations” can run scheduled background tasks, such as scanning messages or files, finding relevant updates, drafting responses, or preparing work before an employee opens the app. Automatically selecting the best, enterprise-preferred AI model for every task (and subtask) Another important feature is multi-model orchestration. Rebel can break a task into parts and route different steps to different models, including splitting between local and cloud-based ones depending on the sensitivity of the information or as guided by enterprise policies. A more powerful model can handle planning or complex reasoning; a cheaper model can handle routine work; a local model can handle sensitive steps or approval checks. This matters for enterprises that want flexibility or are seeking cost controls: not every task need be sent to the same expensive cloud model, and some enterprise workflows prohibit sensitive corporate data leaving local infrastructure. “I want to be able to say, ‘Help me with this,’ and it knows what’s personal, what’s sensitive, and what can be shared with the whole company," Detre explained. That model-agnostic setup gives companies more control over cost and security. Data-heavy work can run on lower-cost models such as Llama or DeepSeek. Higher-level reasoning can be reserved for more expensive models. Sensitive work can be routed through a local model running on the user’s machine, keeping that information from leaving the device. This approach also gives enterprise teams a way to mix cloud and local inference without treating the choice as all-or-nothing. By shifting away from centralized, monolithic cloud interfaces toward a local file-driven architecture, Mindstone is introducing a model for how enterprise technical decision-makers orchestrate autonomous workflows without forfeiting data sovereignty or predictability How it works in practice Mindstone CTO Greg Detre designed Rebel’s memory system to avoid a common problem in enterprise AI: dumping large amounts of company information into a database and hoping search will retrieve the right context later. Instead, Rebel uses a tiered memory structure. When an interaction happens, the system estimates how likely that information is to be useful again. Information with a high expected value is written into a local readme.md file tied to a specific project space. Information with a moderate expected value becomes a reference link back to deeper historical records. Lower-priority material is stored in an indexed memory directory, where it remains available but dormant until a relevant task calls it back. An ROI dashboard for enterprise buyers For larger organizations, Mindstone Pro adds an Impact Dashboard designed to show where Rebel is saving time and money across business units. Mindstone says the dashboard uses a separate, closed LLM to evaluate telemetry and calculate business impact. The company says the system is calibrated conservatively, using the lower end of estimated performance gains to avoid inflated productivity claims. That feature speaks to a practical problem for enterprise AI buyers: proving value without over-surveilling employees. Mindstone says the dashboard is isolated from individual workspaces, allowing IT and business leaders to evaluate adoption and return on investment without reading employees’ private agent activity. Fair Source licensing aims to reduce platform risk Mindstone is releasing Rebel under a Fair Source license, a model meant to sit between fully closed SaaS and permissive open source. Under the license, Rebel’s code is viewable, auditable, modifiable and deployable. Individuals and organizations with up to 100 concurrent users can run it for free. Once an organization exceeds that threshold, it needs a commercial Mindstone Pro license. The license also includes a two-year sunset clause. Twenty-four months after a given version is released, that version automatically converts to the MIT open-source license. For enterprise buyers, the practical pitch is that Rebel reduces the risk of being trapped. If every automation, memory file and agent instruction is stored locally in markdown, a company can move its data and workflows elsewhere if needed. The product may be commercial, but the underlying work is designed to remain inspectable and portable. Security questions focus on local approvals and shared memory Rebel’s debut on the open access tech product sharing platform Product Hunt this week prompted technical questions about how a local-first agent should handle permissions, safety checks and shared memory. One developer, Nikita Pokryschko, asked whether approval checks for sensitive actions could run entirely on a local model, or whether the gating logic still required a cloud call. Detre responded by explaining Rebel’s separation between planning, execution and background safety logic. Wöhle added that companies can configure Rebel to rely entirely on a local model for gating decisions. That distinction matters for corporate security teams. Autonomous agents often need broad permissions to read files, draft emails or interact with internal systems. If the final approval layer depends on an external cloud model, some companies may see that as a compliance risk. Mindstone is arguing that Rebel can keep those approval boundaries local. A second discussion focused on how Rebel decides what memory can be shared. Product developer Clement Morel asked whether shareability is determined by content, user settings or learned behavior, and what happens if the system gets it wrong. Detre said Rebel uses the user’s local “Chief-of-staff README” and defined spaces to separate private, team and company-wide information. When the agent encounters ambiguous context, the system pauses and asks the user for approval before proceeding. That emphasis on visibility is part of Mindstone’s broader argument against opaque agent systems. As CEO Joshua Wöhle put it in a post on his LinkedIn account : “If an agent is going to sit inside your workspace, remember your context, and ask permission before changing the world, you should be able to see how it works. Not because everyone will read the code, but because someone can.” Mindstone points to customer rollout as early proof Mindstone says Rebel has already been deployed across the 250-person workforce of customer Epignosis, covering sales, engineering, product, finance and customer success teams. "The entire organization is operating on Rebel today," Wöhle told VentureBeat. Over a 12-week deployment, Mindstone says Epignosis recaptured the equivalent capacity of eight full-time roles. The company says adoption spread organically after employees saw colleagues automate time-consuming work, a pattern employees reportedly called the “potatoes effect.” The Epignosis case is central to Mindstone’s argument that enterprise AI should not be treated as a set of isolated personal tools. Rebel’s shared-memory design is meant to let workflows move across teams and improve as more employees use them. “The border between learning and doing is fading out - and that changes everything about how you scale,” Epignosis CEO Dimitris Tsingos said in a statement provided to VentureBeat by Mindstone. Background on Mindstone Mindstone Learning Limited, headquartered in London, launched in 2020 under the direction of CEO Joshua Wöhle, previously a co-founder of the digital child safety firm SuperAwesome. Originally positioned in the consumer education technology market, the company built a digital curation tool likened to a "Spotify for learning" that utilized compound learning methodologies. However, following the widespread commercialization of generative artificial intelligence platforms between 2022 and 2024, Mindstone moved into business-to-business enterprise enablement. Leadership identified a critical "last-mile" barrier: while AI tools promised substantial productivity gains, traditional corporate training failed to equip the workforce to practically integrate them into daily operations. Today, Mindstone functions as a comprehensive enterprise software and training ecosystem designed to maximize corporate return on investment for existing AI licenses. The product architecture systematically addresses different organizational tiers through highly contextualized, "live-fire" software applications rather than abstract slide presentations. Financially, Mindstone utilizes a hybrid capitalization strategy that interweaves institutional venture capital from entities like Moonfire Ventures and Pearson Ventures with community-based equity crowdfunding on platforms such as Seedrs and Crowdcube. Mindstone has successfully penetrated the enterprise market, securing commercial contracts with blue-chip corporations including The Home Depot, Hyatt Hotels Corporation, Pearson, and Ernst & Young. Ultimately, Mindstone positions itself as the crucial antidote to corporate inertia, ensuring organizations establish the internal competency required to execute successful AI transformations. Mindstone’s bet: enterprise AI needs shared memory, not more seats Rebel arrives as companies are trying to move from AI experimentation to AI operations. The first wave of enterprise adoption centered on access: giving employees chatbots, copilots and model subscriptions. Mindstone is betting the next wave will center on coordination. That means shared memory, reusable workflows, local control, flexible model routing and measurable business impact. It also means giving enterprises a way to inspect the systems they are being asked to trust. The company’s challenge now is execution. Local-first software can be harder to manage than cloud SaaS. Shared memory raises governance questions. Multi-model routing adds complexity. And enterprises will still need proof that agentic workflows can deliver reliable productivity gains without creating security or compliance headaches. But Mindstone is making a clear argument: buying AI seats is not the same as building AI infrastructure. Rebel is its attempt to turn scattered employee experiments into an operating layer for work.
- Focused on AI at the Enterprise Level: Ganguly
Raj Ganguly, co-founder and co-CEO of B Capital, joins Dani Burger on "Bloomberg Deals." They discuss how the use case for software is not going away and Ganguly's belief that enterprise services are the next wave AI growth. (Source: Bloomberg)
Score: 25🌐 MovesJun 24, 2026https://www.bloomberg.com/news/videos/2026-06-24/focused-on-ai-at-the-enterprise-level-ganguly-video - Amazon Prime Day 2026 India: Dates, deals, AI shopping tools, and more
Amazon Prime Day 2026 India: Dates, deals, AI shopping tools, and more
- 5 Prompts That Get Dramatically Better Answers Out of Claude
Most people use Claude like a search box, ask a question, take the first answer, move on. The people getting genuinely useful work out of it are doing something different. They are giving it a handful of short, specific instructions that change the kind of answer it returns. Here are five of them, why they work, and the honest limits of each. None are secret, and all of them work on any strong model, but most people never use them. There’s a large gap between the answers Claude gives most people and the answers it is capable of giving, and the gap is almost entirely about how you ask. A model trained to be helpful and agreeable will, by default, give you something pleasant, balanced, and a little shallow, because that’s what a vague prompt rewards. The work happens when you override that default with a specific instruction about what you actually want. What follows are five short prompts that reliably change the quality of what comes back. I’m using Claude as the example because it responds well to this kind of direction, but none of these are Claude-specific tricks, they work on any capable model, and none of them are magic. They are just ways of telling the model to drop its defaults and do the more useful thing. Think of them less as cheat codes and more as the instructions the model was waiting for you to give. /brutal The single most useful instruction, because it fixes the most common failure. By default, a model leans toward encouragement. Ask it to review your business plan, your code, or your writing, and it will lead with what is good, soften every criticism, and leave you feeling better than your work probably deserves. That’s pleasant and nearly useless when you’re trying to improve something. The fix is to explicitly ask for the opposite. Tell it to be brutal, to skip the praise, to act as a harsh critic whose only job is to find what is wrong. A prompt like critique this as harshly as you can, assume it is flawed and find every weakness, no encouragement changes the entire character of the response. Instead of a supportive note, you get the actual problems, the weak argument, the unhandled edge case, the paragraph that does not work. The honest caveat is that it’ll find problems whether or not they are serious, because you told it to, so you have to weigh the critique rather than take every point as gospel. But a harsh review you filter is far more useful than a gentle one that told you nothing. When you genuinely want to make something better, this is the instruction that gets you there. /eli5 Short for explain it like I am five, and it’s the fastest way to actually understand something rather than receive a wall of jargon. Ask a model to explain a complex topic and it often answers at the level of the source material, dense, technical, full of terms that assume you already know the thing you are asking about. You end up with a correct answer you can’t use. Asking for an explanation as if you were five, or for a plain version with no jargon, forces a different mode. The model has to find the simple structure underneath the complexity, reach for an everyday analogy, and drop the vocabulary that was hiding the idea. Explain how a mortgage-backed security works like I am five produces something you will actually remember, where the technical version produces something you will reread three times and still not grasp. The limit is that simplification loses precision, an analogy is never the whole truth, so for anything you need to get exactly right you follow up and ask for the rigorous version once the basic shape is clear. Used as a first step, though, it’s the difference between understanding a thing and merely reading about it. Start simple, then add the detail back. /steelman The instruction that makes a model genuinely useful for thinking, because it forces it to argue against you well. When you ask whether your idea is good, or whether some position is right, the model tends to agree with the framing you handed it, especially if you have signaled which answer you want. That feels validating and teaches you nothing. Steelmanning is the opposite of a straw man. Instead of attacking the weakest version of a position, you ask the model to build the strongest possible version of it. Tell it to steelman the case against your plan, or to make the best argument an intelligent person on the other side would make. Now you are not getting agreement, you are getting the strongest opposition to your own view, which is the only thing that actually tests it. It’s the fastest way to find out whether your idea survives contact with a smart adversary, before a real one finds the hole. The caveat is to use it in both directions, steelman the side you disagree with to understand it, and steelman the side you agree with to see if the strong version holds up. The value is in the discomfort, a good steelman should make you slightly less sure of yourself, and that is the point. /missing Five words, and the closest thing to a blind-spot detector you’ve got. Most prompts ask the model to answer the question you posed. This one asks it to look at the question itself, at the things you did not think to ask, the assumptions baked into your framing, the option you never considered. It works because a model has a broad view of how problems like yours usually go wrong, and your framing is necessarily narrow, shaped by what you already know. Asking what am I missing here, or what would someone with more experience catch that I would not, pulls in the considerations outside your current view. You ask it to plan a product launch, then you ask what you are missing, and it surfaces the legal step, the support load, the dependency you forgot, the thing that was going to bite you in three weeks. The honest limit is that it can produce generic checklists if your context is thin, so the more of your actual situation you give it, the sharper the blind spots it finds. But even the generic version catches things, and the specific version regularly catches the one thing that mattered. It costs five words, and it’s worth asking on every plan you make. /10x The instruction that refuses the first answer, because the first answer is almost never the best one. A model gives you a reasonable response to your prompt and stops, because reasonable is what you asked for. It won’t, on its own, reach for the dramatically better version unless you tell it to. So you tell it to. After the first answer, push it, ask for a version ten times better, or tell it that response was a starting point and you want it to push much harder. This forces the model past its first, safe attempt into more ambitious territory, a bolder structure, a sharper angle, a more creative solution it would not have offered to a prompt that seemed satisfied with ordinary. The same is true going the other way, when an answer is overcomplicated, asking for the version ten times simpler strips it to what actually matters. The caveat is that more ambitious isn’t always better, sometimes the first answer was right and the pushed version is just louder, so you keep what genuinely improved and discard what only got bigger. But the habit of not accepting the first draft, of treating it as the opening move rather than the final word, is most of what separates people who get extraordinary output from people who get ordinary output. The model can almost always do better. You just have to ask. The pattern underneath all five Step back and these five are the same move in different clothes. Each one overrides a default. Be brutal overrides agreeableness. ELI5 overrides jargon. Steelman overrides validation. What am I missing overrides your narrow framing. 10x it overrides the first acceptable answer. A model left to its defaults gives you pleasant, balanced, jargon-tinged, first-draft, on-the-rails answers, because that’s the safe response to an unspecified request. Every one of these prompts is just you specifying the more useful thing instead. That’s the real lesson, and it’s bigger than any single prompt. The quality of what you get out of a model is mostly a function of how clearly you tell it what kind of answer you want. These five are a starting set, not a complete one, and once you see the pattern you will start writing your own. The model is usually capable of far more than its default gives you. These are simply five ways of asking for it. If you have your own prompt that reliably changes the quality of what you get back, drop it in the comments. The genuinely useful ones tend to be short, specific, and a little uncomfortable, and they spread by word of mouth more than any guide. 5 Prompts That Get Dramatically Better Answers Out of Claude was originally published in Towards AI on Medium, where people are continuing the conversation by highlighting and responding to this story.