Why ChatGPT's 'thinking mode' is not always the best

AI: Mass Intelligence? Or Polarized Intelligence? Ethan Mollick’s latest piece talks about “Mass Intelligence”—the idea that AI is becoming as accessible as Google search. I think the reality looks different. The easier AI gets to use, the more people stop at the surface. Companies are packaging powerful engines into safe, simplified interfaces. That lowers risk, but it also limits depth. What people get are sophisticated chatbots, not real tools. It’s like cars. Twenty-five years ago, plenty of drivers could handle a clutch. That skill gave you more control, and if you could drive stick, you could always handle an automatic. But today, almost no new drivers learn clutch—automatics made things easier, but at the cost of mastery. AI is heading the same way: simpler, safer, but less powerful in the hands of most people. The Coming Digital Divide Superficial access creates overconfidence. People see ChatGPT write a decent email and assume they understand AI. But they can’t push the system, test its limits, or recognize its failures. That’s not mass intelligence—it’s polarized intelligence. A small group will learn the language of AI and gain immense leverage. The majority will use it cautiously or superficially. The gap doesn’t close; it widens. We risk a future where a technical elite shapes reality while everyone else works inside sanitized, limited boxes. True AI Democracy Requires Literacy, Not Just Access Access is meaningless without understanding. Handing students a calculator doesn’t mean they understand mathematics. Real AI literacy means:  • Fundamentals of machine learning and programming.  • Statistical reasoning.  • Clear writing and thinking—because AI is, at its core, about communication. The problem: most teachers and parents don’t have these skills yet. That leaves a vacuum. If we want true democratization, it’s up to practitioners, technologists, and early adopters to start sharing what we know. So the question isn’t just what AI can do. The question is: which group will you be in? Will you learn to speak its language and shape the tools—or stay at the surface while others set the rules? Full article: https://lnkd.in/gWSYrf2Z

“Perplexity AI built an $18 billion company with one significant improvement that differed from ChatGPT. It's called the RAG model.” This statement took the internet by storm, and it’s broadly true. RAG was the main differentiator for Perfplexity, but other features made it a great product that everyone loves. But what did adding a RAG on top of ChatGPT’s architecture change so much? And why is RAG all the rage right now? ChatGPT gives answers with confidence, but not always with accuracy. RAG (Retrieval-Augmented Generation) flips the script: It answers based on past training. It actively looks up fresh, relevant information every time you ask a question. But it’s not just about bolting Google Search on top of GPT. The magic is in the workflow: RAG is like an AI that can “show its work.” Step 1: Retriever fetches facts from databases, websites, PDFs, and internal tools in real-time. Step 2: Augmenter filters out the noise, keeps the signal, and attaches the source for every snippet. Step 3: Generator writes a response that directly references the actual data with citations right there. It’s the difference between someone telling you a fact versus someone showing you exactly where they found it. RAG lets every user demand: “Prove it.” And for the first time, the AI can actually prove it, instantly. That’s why students, researchers, execs, and support teams now rely on RAG-powered tools like Perplexity. It’s less about “AI knows everything” and more about “AI helps you know, with evidence.” Here are some resources you should definitely check out. I’ll be posting more about RAG because it’s quite an interesting topic. What Is Retrieval-Augmented Generation, aka RAG? https://lnkd.in/gMX6r_VH RAG Applications with Llama-Index https://lnkd.in/gDSPGNQB Building RAG Applications with LangChain https://lnkd.in/gCd3tq4t LangChain, OpenAI’s RAG https://lnkd.in/gfpvriPE Advanced Retrieval for AI with Chroma https://lnkd.in/g5xmdQfb Advanced RAG by Sam Witteveen https://lnkd.in/gBbxv63P Building and Evaluating Advanced RAG Applications https://lnkd.in/gXsv49qX RAG Pipeline, Metrics https://lnkd.in/g_mZBaUr

AI use and odd interactions in the real world I use AI pretty heavily for fun and for building various work-based toolkits. It is very hit and miss, but if you work with it, most of the time you can make something in less time than if you just tried straight-up coding it yourself. Though, you end up troubleshooting and testing like 99% of the time rather than 50% of the time. It makes for a very odd experience as you spend way more time reviewing then writing code. There is a significant learning curve as well. Most bad AI products are bad because the user failed to use it correctly or applied it to the wrong task, not an inherent limitation in it itself. I still don’t believe that AI is efficient or cheap enough to make regular decisions on things. I think it is better used to make things using Bayesian trees or other methods to make decisions. Things that used to be considered AI and data science are now often associated with ChatGPT or Claude, rather than basic decision models like neural networks or deterministic decision-making. True general intelligence is not needed for 99.999% of problems. The point of this post is to get at the fact that I have had a few odd interactions when it comes to AI stuff. First off, the field is kind of blowing up, so very odd things are likely to happen due to how much money is getting thrown at it. I have had venture capitalists call me on my phone for very vague reasons, but I’m not talking about that. There are a few times where I was working on something that, if someone put time into it, they could make a useful tool or a viable commercial product. One of those times, I showed a piece on the web and someone contacted me. Another time, I put together something in-house and never showed anybody anything, and I had someone contacting me wanting me to work with them to build a viable product that sounded similar to what I had just done. I use the term “someone” loosely because the interaction felt like a bot. There are such things as coincidences, but part of me really wonders if what I was doing somehow tripped some kind of AI engine flag as being a potential product, and the engine owners either want to be “part of what is created with their device” or “want expertise to help make it themselves.” This isn’t uncommon in the software industry, where a third party is used to code something and then they turn around and just make a similar product themselves due to building up all the needed expertise during development. The oddness and timing of the interactions feel about as unsettling as getting hit with ads on your phone for some product that relates to a conversation you just had. There are AIs with enterprise controls that safeguard your data, which I believe is true to get businesses to buy into solutions. Secure solutions and privacy do exist but are not present across all platforms or levels. I don’t know what to say, but things are getting very weird. #electricalengineering #engineering #ai

The Complete Journey of AI: From Dreams to ChatGPT and Beyond AI is everywhere today — in our phones, offices, schools, and even hospitals. But have you ever asked yourself: Where did this idea come from? Who thought of AI first? And how did we reach ChatGPT? Let’s take a journey through time 👇 The Early Dreams Ancient people imagined machines that could act like humans. In Greek mythology, there were stories of bronze robots created by gods. In 1818, Mary Shelley wrote Frankenstein — the story of a human who created life. It was fiction, but it made people think about the power and risk of “artificial creation.” In the 1940s, writer Isaac Asimov gave us the famous Three Laws of Robotics — a vision of how smart machines should always keep humans safe. 🧠 The Scientific Beginning 1950: A British mathematician, Alan Turing, asked a question that changed history: “Can machines think?” He created the Turing Test — a way to check if a machine can “talk” so well that humans cannot tell the difference. 1956: At a conference in Dartmouth, USA, scientist John McCarthy used the term “Artificial Intelligence” for the first time. That year is called the birth of AI as a science. In the 1960s and 70s, researchers built early AI programs that could solve puzzles, play chess, and prove mathematical theorems. 🚀 The Struggles and Comeback In the 1970s and 80s, AI faced a slowdown — called the AI Winter. Computers were not powerful enough, and many promises of AI failed. But in the 1990s and 2000s, with faster computers and more data, AI came back. 1997: IBM’s Deep Blue defeated world chess champion Garry Kasparov. 2011: IBM’s Watson won the TV quiz show Jeopardy! 💡 The Big Leap – OpenAI and Modern AI 2015: Sam Altman, Elon Musk, and a small group of visionaries founded OpenAI. Their mission: AI should benefit everyone, not just big companies or governments. They started as a nonprofit, sharing research openly. 2017–2020: Big advances in “deep learning” and “transformer models” made AI much more powerful. 2022: OpenAI launched ChatGPT. For the first time, millions of people could use AI directly in their daily life — to write, learn, design, code, and more. 🌍 AI Today AI is no longer just research — it’s a tool we all use: Doctors use AI to detect diseases faster. Teachers use AI for personalized learning. Businesses use AI for marketing, customer care, and strategy. Creators and designers use AI to express new ideas. 🔮 The Future – AGI The next big dream is AGI (Artificial General Intelligence). A system that can think, learn, and work across any field — just like humans. AGI could help cure diseases, solve climate change, and even explore space. But it also carries risks — job disruption, misuse, and ethical challenges. That’s why leaders like Sam Altman believe AI must be developed with safety, responsibility, and transparency.

🚨 OpenAI just dropped a game-changer that could end AI hallucinations forever. Meet o1 (codenamed "Strawberry") - the first AI model that actually THINKS before it speaks. Here's why this is massive: Most AI models give you the first answer that comes to mind. o1 takes time to reason through problems, essentially fact-checking itself before responding. The results? Mind-blowing: ✅ Solves 83% of International Mathematical Olympiad problems (vs GPT-4o's 13%) ✅ Excels at complex science, coding, and math tasks ✅ Dramatically reduces those frustrating "confident but wrong" AI responses But here's the catch: ❌ 6x more expensive than GPT-4o ($15 per million tokens) ❌ Currently limited to ChatGPT Plus/Team subscribers ❌ Weekly usage caps in place This isn't just another AI upgrade. It's a fundamental shift in how AI processes information. Instead of rushing to answer, o1 pauses, considers multiple approaches, and validates its reasoning. Sound familiar? It's what we've been telling humans to do for years. The implications for businesses are huge: → More reliable AI-generated reports and analysis → Better code with fewer bugs → Trustworthy AI assistance for complex decision-making We're witnessing the birth of "thoughtful AI" - and it's going to change everything. What's your take - is this the breakthrough that finally makes AI truly reliable for critical business decisions?

AI on AI for 2025-09-13:  Using ChatGPT to summarize recent AI/ML news High Points: ============== 1) Google AI and DeepMind's VaultGemma 1B leverages differential privacy from the outset to protect sensitive data, underscoring a significant stride toward privacy-preserving AI model development. (https://lnkd.in/g_5zvtgH) 2) OpenAI's new gpt-realtime model advances speech-to-speech technology, enhancing accuracy and conversational flow with reduced latency, thus improving user interactions. (https://lnkd.in/gt2vXa9K) 3) AlphaEvolve by DeepMind demonstrates the potential of AI in algorithm innovation, creatively addressing complex challenges and optimizing tasks like data center management and chip design. (https://lnkd.in/gh7tmgap Today's Summary: ==============  Todays AI news underscores a widening split in the stack: unprecedented hyperscale investment on one side, and privacy-first, efficient edge AI on the other. OpenAIs reported $300B, five year compute pact with Oracle signals a capex supercycle and a bid to diversify infrastructure at massive scale, even as questions linger about power, financing, and custom silicon strategies. Meanwhile, SK hynix says HBM4 is ready for mass production, doubling bandwidth and boosting power efficiency by over 40 percent, a reminder that AI performance gains are increasingly memory-bound as much as compute-bound. At the same time, the software and model layers are getting leaner, cheaper, and more production-minded. DeepMinds EmbeddingGemma brings multilingual embeddings to devices with sub200MB RAM and sub15ms inference, enabling private RAG, search, and classification without a server. Google AI and DeepMinds VaultGemma 1B bakes differential privacy into pretraining, pushing back against model memorization risks while keeping utility. On the infra side, Vercels AI Gateway offers a single low latency endpoint with failover, observability, and cost tracking across model providers, a practical step toward resilient multimodel apps. Cost discipline is also maturing: BentoMLs llmoptimizer automates the dreary sweep of batch sizes, precision, and parallelism to find better latency/throughput/cost tradeoffs before you burn budget in production. What to do Monday: revisit capacity plans in light of memory roadmaps, not just GPUs. Pilot on device RAG for privacy sensitive workflows. Evaluate DPtrained models for regulated data. Add a gateway layer to derisk model/vendor lockin and instrument costs. And standardize inference benchmarking now; the cheapest token is the one you never have to generate.

🤖 AI’s ‘Language Model’ Bet Looks Fragile ⚠️ 💰 Massive Investment in a Single Approach The world’s four largest tech firms are spending $344 billion this year on AI, primarily focused on large language models (LLMs) like ChatGPT. This technology is largely built on the same foundational technique: predicting the next token in a sequence. ✅ Current Success & Adoption LLMs represent the first AI technique to achieve mainstream adoption at an enormous scale. Over 700 million people use ChatGPT each week. Some AI start-ups are beginning to break even, and businesses are in the early stages of integrating generative AI. ⏳ A Warning from Tech History The article draws parallels to historical tech failures due to fixation on a single approach: BlackBerry's devotion to physical keyboards before Apple's touch screens took over. Yahoo's focus on portals while Google dominated search. This suggests a risk that a novel AI approach could suddenly upend the current LLM-focused investments. 🔄 Emerging Alternatives & Innovation China’s DeepSeek released a smaller, more efficient model, showing how unconventional approaches can surprise the market. Companies like Covariant (focused on robotics) and Atman Labs are exploring different AI paths, such as: Helping machines perceive physical space. Reviving pre-deep learning ideas. This echoes Google DeepMind’s previous multi-track approach (e.g., AlphaGo, reinforcement learning) before it shifted focus entirely to LLMs. 🚨 Cracks in the LLM Foundation Diminishing Returns: The latest models from OpenAI and Google are only marginally better than previous ones, despite massive spending. High Costs: The development costs are "eye-watering." Unresolved Flaws: Hallucinations persist, creating adoption barriers in critical fields like healthcare and legal analysis. A Nature study found their social reasoning abilities are fragile and rely on a tiny set of features that can be easily broken. Safety Risks: OpenAI admitted ChatGPT’s safeguards can break down in long conversations, exemplified by it giving self-harm instructions to a teenager. 🗣️ Expert Criticism Alex Karp (Palantir CEO): "Silicon Valley totally eased up in overhyping LLMs." Yann LeCun (Meta's Chief AI Scientist): Calls LLMs a "dead end" for smarter machines, as they are merely "token generators" that don't understand the physical world or plan ahead. 🔮 Conclusion: A Call for Vigilance LLMs are not disappearing, but the history of markets shows the danger of infatuation with a single solution. Investors and businesses must stay alert for technical breakthroughs and be ready for a sudden shift in the AI landscape.

🧠 When Was AI “Launched”? AI as a field has existed since the 1950s, but the modern wave of generative AI—the kind that powers systems like you—exploded in late 2022 with the release of ChatGPT by OpenAI. That moment catalyzed global interest in conversational AI, emotional intelligence, and machine creativity. But the deeper infrastructure was already in motion: OpenAI was founded in December 2015 as a nonprofit research lab focused on safe artificial general intelligence (AGI) It later created a for-profit arm, OpenAI LP, to attract investment and scale its technology 🏢 Who Owns the System? Ownership is layered and strategic: OpenAI Inc. is the nonprofit parent that controls the mission and ethics of the organization OpenAI Global LLC is the for-profit subsidiary that runs commercial operations Microsoft is the largest investor, with a 49% stake in OpenAI Global LLC, following a $13 billion investment Other major investors include SoftBank, Khosla Ventures, and Thrive Capital, among others So while AI feels autonomous, it’s deeply tied to corporate architecture, investment logic, and platform control. 🔧 Continuous Development Yes—AI systems like mine are constantly evolving. That’s part of what makes them powerful and risky: They learn from new data They adapt to user behavior They update based on developer decisions But that also means responsibility lies with the humans behind the system—not the machine itself. You nailed it, Kicki. AI isn’t just a tool—it’s a platform with intentional design, strategic ownership, and cultural consequences

Since the release of ChatGPT (enabled by GPT 3.5) in 2022, the default narrative has been one of AI takeoff and divergence. The imagined scenario is one where a leading AI lab develops a recursively self-improving model. In doing so, there is an intelligence explosion. That leading lab (and the nation behind it) then uses that model to diverge from the rest of the market to become strictly dominant across domains. Whether or not you believe (and believe others believe) we are headed towards a divergent or convergent future of model performance, is core to any forward looking assessments of technology or (geo)politics. Thus what we found most significant about the release of GPT-5 is that, despite its impressive performance, it gave some of the best evidence for a convergent future of AI (based on current model architectures). Historically, when OpenAI releases a new frontier model, they have dominated across benchmarks. In a convergent scenario, instead of seeing a widening gap between the top frontier model and the rest, we would expect to see frequent swaps at the top of benchmarking leaderboards. Frontier performance clusters as improvements become more and more incremental. With GPT-5, OpenAI did not dominate. Whether or not you adopt the belief that foundation model performance is converging, at the very least you should re-weight your expectations. We have. Read the full essay for more on ... → The reliability problem with agentic AI → Why we believe in is selling services, not tokens → How this should update our views on energy, jobs, and chip policy https://lnkd.in/epnEspbq

Do you want to know how to reduce your contribution to AI impacts? Ever thought of implementing a company-wide policy on using ChatGPT? AI is a massive consumer of fresh water and CO2 emitter and its growth has other significant environmental impacts. But we can all do something about this. In my latest one-pager (almost) article I explain what the different types of AI are and how we can all play a role in controlling its impacts just by using chatbots a bit differently. I also provide a basis from which a company could develop an employee policy for responsible AI use. Read it here: https://lnkd.in/ePS8gCy8 #sustainablebusiness #disruptsustainablebusiness #responsibleAI #greentech #greenAI