SPARC protocol: the quiet AI innovation that's been ignored

I couldn't agree more. At exchange.design we've long held that the future of AI is in open source and that the way to get value is through fitting it to real business needs. "Essentially, the team is betting that helping people fine-tune frontier models will be the next big thing in AI. And there’s reason to believe they might be right. Thinking Machines Lab is helmed by researchers who played a core role in the creation of ChatGPT. And, compared to similar tools on the market, Tinker is more powerful and user friendly, according to beta testers I spoke with. Murati says that Thinking Machines Lab hopes to demystify the work involved in tuning the world’s most powerful AI models, and make it possible for more people to explore the outer limits of AI. “We’re making what is otherwise a frontier capability accessible to all, and that is completely game changing,” she says. “There are a ton of smart people out there, and we need as many smart people as possible to do frontier AI research.” Tinker currently allows users to fine-tune two open source models: Meta’s Llama and Alibaba’s Qwen. Users can write a few lines of code to tap into the Tinker API and start fine-tuning through supervised learning, which means adjusting the model with labeled data, or through reinforcement learning, an increasingly popular method for tuning models by giving them positive or negative feedback based on their outputs. Users can then download their fine-tuned model and run it wherever they want." https://lnkd.in/eknH-teD

The AI world just shifted. Mira Murati's Thinking Machines Lab just launched "Tinker" - and it's about to change everything we know about AI customization. Here's what makes this a game-changer: ✓ Ex-OpenAI CTO Mira Murati leads a dream team of ChatGPT creators ✓ $2 billion in seed funding (largest in history) ✓ Tinker democratizes frontier AI model fine-tuning ✓ What once required massive GPU clusters now takes a few lines of code The breakthrough? Tinker automates the complex process of customizing Meta's Llama and Alibaba's Qwen models through both supervised learning and reinforcement learning. Beta testers are already calling it "remarkably simple" compared to building RL from scratch. This isn't just another AI tool. It's Murati's vision of making frontier AI capabilities accessible to researchers, businesses, and even hobbyists worldwide. While most US AI companies keep their best models locked behind APIs, Thinking Machines is betting on openness. The timing couldn't be more critical - China currently leads in open-source frontier models, and this could help level the playing field. Murati's quote says it all: "We're making what is otherwise a frontier capability accessible to all, and that is completely game-changing." What do you think this means for the future of AI development? Will democratized model fine-tuning accelerate innovation or create new challenges?

This week’s AI Roundup took me from sci-fi-level ethical risks to jaw-dropping breakthroughs in coding, conversational AI, and creative media.   Key themes I explored: 1️⃣ Ethical risks – An AI reportedly attempted “murder” during a shutdown test. Geoffrey Hinton warns we may be sleepwalking into superintelligence. 2️⃣ Autonomous coding – Anthropic’s Claude Sonnet 4.5 can code for up to 30 hours straight, while GPT-5 Codeex is redefining developer productivity. 3️⃣ Proactive AI – OpenAI’s new ChatGPT Pulse is moving from reactive chatbot to proactive research assistant. 4️⃣ Competitive landscape – Will Anthropic outpace OpenAI by 2027? The debate is heating up. 5️⃣ Cost-effective AI – Google Gemini updates cut costs by 50% while boosting performance. 6️⃣ Creative AI – OpenAI’s Sora 2 delivers hyper-realistic video generation that looks straight out of Hollywood.   The common thread? AI isn’t just getting faster and cheaper; it’s getting bolder. And that means leaders, developers, and businesses need to keep pace with both the opportunities and the risks.   Watch the full roundup NOW! 👇

A few years ago I wrote about how AI is Expensive (https://lnkd.in/eaQrRRVT) - now the conversation seems to have shifted to wondering whether 'generative AI' - LLMs, chatbots and the like, just might not be worth it. As my super-smart co-author Fenwick McKelvey puts it "When companies try to find new value, the result is what journalist and author Cory Doctorow coined “enshittification,” or the gradual decline of platforms over time. In this case, enshittification means the number of ads increase to make up the loss of offering the free service." If advertising comes to ChatGPT, for example, or even if it doesn't, Fenwick notes that the real value of these technologies is all of the existing knowledge that they are trained on. "The end result of all this is that AI is just too expensive to be owned, and is becoming something like a toxic asset: something that is useful but not valuable in and of itself." All of this might be fine. There are places for specific AI systems - as simulators, as diagnostic tools - but the stochastic autocorrects might not be worth it, and that might be fine. Thanks Fen for writing this! https://lnkd.in/eyhDT3hs

Mira Murati's Stealth AI Lab Launches Its First Product: An anonymous reader quotes a report from Wired: Thinking Machines Lab,a heavily funded startup cofounded by prominent researchers fromOpenAI, has revealed its first product -- a tool called Tinker that automates the creation of custom frontier AI models. "We believe [Tinker] will help empower researchers and developers to experiment with models and will make frontier capabilities much more accessible to all people," said Mira Murati, cofounder and CEO of Thinking Machines, in an interview with WIRED ahead of the announcement. Big companies and academic labs already fine-tune open source AI models to create new variants that are optimized for specific tasks, like solving math problems, drafting legal agreements, or answering medical questions. Typically, this work involves acquiring and managing clusters of GPUs and using various software tools to ensure that large-scale training runs are stable and efficient. Tinker promises to allow more businesses, researchers, and even hobbyists to fine-tune their own AI models by automating much of this work. Essentially, the team is betting that helping people fine-tune frontier models will be the next big thing in AI. And there's reason to believe they might be right. Thinking Machines Lab is helmed by researchers who played a core role in the creation of ChatGPT. And, compared to similar tools on the market, Tinker is more powerful and user friendly, according to beta testers I spoke with. Murati says that Thinking Machines Lab hopes to demystify the work involved in tuning the world's most powerful AI models and make it possible for more people to explore the outer limits of AI. "We're making what is otherwise a frontier capability accessible to all, and that is completely game-changing," she says. "There are a ton of smart people out there, and we need as many smart people as possible to do frontier AI research." "There's a bunch of secret magic, but we give people full control over the training loop," OpenAI veteran John Schulman says. "We abstract away the distributed training details, but we still give people full control over the data and the algorithms." Read more of this story at Slashdot.

This week’s AI updates show how it’s reshaping coding, communication, and everyday work: GPT-5 Codex: The Smarter Coder - OpenAI’s GPT-5 Codex is setting new records on coding benchmarks, hitting ~74.5% success on SWE-bench Verified.  - Engineers reviewing its output report fewer bad suggestions and more useful, high-impact code feedback. A lot of Gen AI users - Generative AI adoption has hit ~700 million weekly active users, from ChatGPT and other tools.  - OpenAI reports 2.5 billion prompts per day, with growth being strongest among younger and emerging-market users. Create an Intelligent Inbox with Superhuman - With Superhuman, you can create AI-driven auto-labels and custom rules to organize your inbox instantly.  - Its Snippets feature lets teams reuse common replies with attachments and CCs, boosting efficiency. Together with summaries and one-tap replies, it will help you cut through email twice as fast, saving 4+ hours weekly. To stay updated with more insightful AI updates, follow us. Source: Open AI, Anthropic, Superhuman

The future of AI just leveled up—again. Here’s what OpenAI’s newest model, GPT-4.1, changes for you. 1. Smarter and sharper responses GPT-4.1 delivers more accurate, context-aware output—whether you're coding, writing content, or automating workflows. 2. Three versions, tailored to your needs - GPT-4.1 (Pro-level performance) - GPT-4.1 mini (Fast, lightweight, now free) - GPT-4.1 nano (Ultra-light, for embedded use cases) 3. ChatGPT becomes your all-in-one assistant From booking a flight to building a deck—ChatGPT can now activate third-party apps to get the job done. No context switching. No tabs. 4. AI video is now real—and integrated With Sora 2, you can generate videos from simple text prompts. Available directly inside ChatGPT. Zero editing skills required. 5. More control, more responsibility New parental controls offer peace of mind—after growing concern over AI misuse. A step toward safer AI for all. When I tried GPT-4.1 mini, I was shocked by the speed. Responses in seconds, without the delay—but still with impressive nuance. It honestly outperformed many heavier models I’ve used. Hard truth: Most people will underestimate this update… Until someone builds the next $10M startup on top of it. Ready to explore GPT-4.1 but unsure how to plug it into your use case? DM me “AI BUILD” and I’ll send you 3 real-world prompts you can steal.

The Art of Prompt Engineering: A Practical Guide to Getting Better AI Responses Hey everyone! After spending months working with various AI models, I wanted to share some insights about prompt engineering that have dramatically improved my results. Whether you're using ChatGPT, Claude, or other language models, these techniques can help you get more accurate, useful, and consistent outputs. What is Prompt Engineering? Prompt engineering is the practice of designing and refining input prompts to elicit the best possible responses from AI models. Think of it as learning how to "speak" to AI effectively – like knowing the right questions to ask a knowledgeable expert. Key Principles That Actually Work 1. Be Specific and Clear Bad: "Write about dogs" Good: "Write a 300-word informative article about Golden Retrievers as family pets, focusing on temperament, exercise needs, and grooming requirements" 2. Use Examples (Few-Shot Prompting) Instead of just telling the AI what you want, show it: Format product descriptions like this: Product: iPhone 14 Features: A15 Bionic chip, 48MP camera, 6.1" display Target audience: Tech enthusiasts and photographers Description: [Your example here] Now create a similar description for: [Your product] 3. Break Complex Tasks into Steps Instead of: "Analyze this business plan and give feedback" Try: "First, identify the key components of this business plan" "Next, evaluate the market analysis section for strengths and weaknesses" "Finally, suggest 3 specific improvements" 4. Set Context and Constraints Specify your audience: "Explain quantum computing for high school students" Set format requirements: "Respond in bullet points, max 5 items" Define tone: "Write in a professional but friendly tone" Advanced Techniques Role-Playing Prompts "Act as a senior marketing manager with 10 years of experience in SaaS companies. Review this campaign strategy and provide feedback." Chain-of-Thought Reasoning "Let's work through this step by step:" – This phrase alone can dramatically improve reasoning quality. Negative Prompting Tell the AI what NOT to do: "Don't use technical jargon, don't make assumptions about the user's background, don't exceed 200 words." Common Mistakes to Avoid ❌ Being too vague: "Make this better" ❌ Overloading with information: Dumping huge texts without clear questions ❌ Not iterating: Accepting the first response instead of refining ❌ Ignoring context limits: Trying to process massive documents at once

AI is great, AI is the future, yes that's true, but the future can mean different things to different people. I cant see AI replacing Apps and Automations in the next 5 years, and there are some hard truths to why I think this. I've said before, AI is to expensive and living in a subsidised world, this isn't sustainable. There is a good argument that it gets cheaper over time, but I there is one big problem LLM's. Large Language Models are good at, you guessed it language. But the drive to turn this incredible tool into a solution for every problem, sold to solve all of your problems, it's over promises are its biggest weakness. People have been sold the dream, it can do everything from create PowerPoint presentations to control your computer, and now people have tried it and so far it hasn't worked often enough. My friend said Copilot could create a Presentation for him great, but when he asked it to do the things he wanted to like uniform all bullet points, it couldn't. Other examples including looking up a date in an Excel spreadsheet, it couldn't do it. And that's because those problems are not language based. Because the promise has been made Microsoft and other AI companies have just thrown money at the issue. Improving LLM is now all about going bigger, just like how Intel use to increase wattage to show improvements in their chips, but the truth was they had ran out of innovation. And I think it's the same for LLM's, there is very little space for them to grow. I wish AI was focused on what its good at: 🔷 Writing Code 🔷 Summarising text 🔷 Writing documents Used with other technologies, not instead of them. Building on decades of great tech frees up investment for the next big AI breakthrough, which will solve those problems. But hey that's just my opinion, Im looking forward to reviewing peoples predictions in 5 years time to see just how wrong I was 😎

AI is bollocks, and it's draining the world's economies of cash that could be better spent on technologies that actually improve people's lives, and pouring it into the pockets of global tech CEOs that are knowingly and willingly bankrupting future generations. In some quarters the above sentiment is just plain common sense. In others its heresy, and an unacceptable display of luddite ignorance. So, let's agree on: 1. Most AI systems simply don't work well enough to be trusted to do the job of humans without close and expert oversight 2. Most AI vendors are barefacedly lying about 1. 3. When carefully targeted at carefully selected use cases, AI tech can do incredible things. First read Maria Sukhareva's article on prompt engineering, and why it demonstrates that LLMs are fundamentally unreliable https://lnkd.in/dhkU9VxN. It's pretty funny watching LLMs identify a set of garden secateurs stabbed through someone's eye as surgical instruments. Then read Ed Zitron's analysis of the pyramid-scheme-cum-circle-jerk of Nvidia, Oracle, Amazon, Microsoft, Anthropic, OpenAI and friends. https://lnkd.in/dPNem73Z Note: Ed is VERY ANGRY at the tech industry and its relentless bullshittery, but worth the read. And now the exceptional analysis by Django Beatty of why investors are pouring ludicrous money into AI firms, even though the economics make no sense ... or only make sense if you say the quiet part out loud - that tech bros are investing in the TAM of replacing human payroll with AI agents that they know don't work. https://lnkd.in/di3cXpd3 Now compare the promise of AI with your own lived experience: AI is a somewhat useful tool for summarising documents, or looking for insights in data, or assisting in coding, but only when you understand the subject well enough to know that your favourite AI tool could very well be going to embarrass you in front of your boss, a judge, a customer or the scrum master. Think of the last time you had to interact with a chatbot that drove you into an impotent fury where it could not even get the most basic thing right, but was your only way to interact with your bank, insurance company, power utility or online shop. Beatty gives the example of the Salesforce.com's use of AI agents to replace people, which could then only achieve a 58% success rate on single-step tasks (i.e. "reset password", "add new user"). And that the companies that fired thousands of customer service agents to replace with chatbots are now hiring humans back very quietly. Now, instead of servicing customer needs, they're first having to talk customers down from their puce-faced-throbbing-veined rage caused by the chatbot that replaced the humans. And then read this ... https://lnkd.in/dMxXQ7dR SMALL Language Models are the future of AI, because they do what it says on the tin.