Understanding AI Systems

This week MIT dropped a stat engineered to go viral: 95% of enterprise GenAI pilots are failing. Markets, predictably, had a minor existential crisis. Pundits whispered the B-word (“bubble”), traders rotated into defensive stocks, and your colleague forwarded you a link with “is AI overhyped???” in the subject line. Let’s be clear: the 95% failure rate isn’t a caution against AI. It’s a mirror held up to how deeply ossified enterprises are. Two truths can coexist: (1) The tech is very real. (2) Most companies are hilariously bad at deploying it. If you’re a startup, AI feels like a superpower. No legacy systems. No 17-step approval chains. No legal team asking whether ChatGPT has been “SOC2-audited.” You ship. You iterate. You win. If you’re an enterprise, your org chart looks like a game of Twister and your workflows were last updated when Friendswas still airing. You don’t need a better model - you need a cultural lobotomy. This isn’t an “AI bubble” popping. It’s the adoption lag every platform shift goes through. - Cloud in the 2010s: Endless proofs of concept before actual transformation. - Mobile in the 2000s: Enterprises thought an iPhone app was strategy. Spoiler: it wasn’t. - Internet in the 90s: Half of Fortune 500 CEOs declared “this is just a fad.” Some of those companies no longer exist. History rhymes. The lag isn’t a bug; it’s the default setting. Buried beneath the viral 95% headline are 3 lessons enterprises can actually use: ▪️ Back-office > front-office. The biggest ROI comes from back-office automation - finance ops, procurement, claims processing - yet over half of AI dollars go into sales and marketing. The treasure’s just buried in a different part of the org chart. ▪️Buy > build. Success rates hit ~67% when companies buy or partner with vendors. DIY attempts succeed a third as often. Unless it’s literally your full-time job to stay current on model architecture, you’ll fall behind. Your engineers don’t need to reinvent an LLM-powered wheel; they need to build where you’re actually differentiated. ▪️Integration > innovation. Pilots flop not because AI “doesn’t work,” but because enterprises don’t know how to weave it into workflows. The “learning gap” is the real killer. Spend as much energy on change management, process design, and user training as you do on the tool itself. Without redesigning processes, “AI adoption” is just a Peloton bought in January and used as a coat rack by March. You didn’t fail at fitness; you failed at follow-through. In five years, GenAI will be as invisible - and indispensable - as cloud is today. The difference between the winners and the laggards won’t be access to models, but the courage to rip up processes and rebuild them. The “95% failure” stat doesn’t mean AI is snake oil. It means enterprises are in Year 1 of a 10-year adoption curve. The market just confused growing pains for terminal illness.

I spent 3+ hours in the last 2 weeks putting together this no-nonsense curriculum so you can break into AI as a software engineer in 2025. This post (plus flowchart) gives you the latest AI trends, core skills, and tool stack you’ll need. I want to see how you use this to level up. Save it, share it, and take action. ➦ 1. LLMs (Large Language Models) This is the core of almost every AI product right now. think ChatGPT, Claude, Gemini. To be valuable here, you need to: →Design great prompts (zero-shot, CoT, role-based) →Fine-tune models (LoRA, QLoRA, PEFT, this is how you adapt LLMs for your use case) →Understand embeddings for smarter search and context →Master function calling (hooking models up to tools/APIs in your stack) →Handle hallucinations (trust me, this is a must in prod) Tools: OpenAI GPT-4o, Claude, Gemini, Hugging Face Transformers, Cohere ➦ 2. RAG (Retrieval-Augmented Generation) This is the backbone of every AI assistant/chatbot that needs to answer questions with real data (not just model memory). Key skills: -Chunking & indexing docs for vector DBs -Building smart search/retrieval pipelines -Injecting context on the fly (dynamic context) -Multi-source data retrieval (APIs, files, web scraping) -Prompt engineering for grounded, truthful responses Tools: FAISS, Pinecone, LangChain, Weaviate, ChromaDB, Haystack ➦ 3. Agentic AI & AI Agents Forget single bots. The future is teams of agents coordinating to get stuff done, think automated research, scheduling, or workflows. What to learn: -Agent design (planner/executor/researcher roles) -Long-term memory (episodic, context tracking) -Multi-agent communication & messaging -Feedback loops (self-improvement, error handling) -Tool orchestration (using APIs, CRMs, plugins) Tools: CrewAI, LangGraph, AgentOps, FlowiseAI, Superagent, ReAct Framework ➦ 4. AI Engineer You need to be able to ship, not just prototype. Get good at: -Designing & orchestrating AI workflows (combine LLMs + tools + memory) -Deploying models and managing versions -Securing API access & gateway management -CI/CD for AI (test, deploy, monitor) -Cost and latency optimization in prod -Responsible AI (privacy, explainability, fairness) Tools: Docker, FastAPI, Hugging Face Hub, Vercel, LangSmith, OpenAI API, Cloudflare Workers, GitHub Copilot ➦ 5. ML Engineer Old-school but essential. AI teams always need: -Data cleaning & feature engineering -Classical ML (XGBoost, SVM, Trees) -Deep learning (TensorFlow, PyTorch) -Model evaluation & cross-validation -Hyperparameter optimization -MLOps (tracking, deployment, experiment logging) -Scaling on cloud Tools: scikit-learn, TensorFlow, PyTorch, MLflow, Vertex AI, Apache Airflow, DVC, Kubeflow

The AI gave a clear diagnosis. The doctor trusted it. The only problem? The AI was wrong. A year ago, I was called in to consult for a global healthcare company. They had implemented an AI diagnostic system to help doctors analyze thousands of patient records rapidly. The promise? Faster disease detection, better healthcare. Then came the wake-up call. The AI flagged a case with a high probability of a rare autoimmune disorder. The doctor, trusting the system, recommended an aggressive treatment plan. But something felt off. When I was brought in to review, we discovered the AI had misinterpreted an MRI anomaly. The patient had an entirely different condition—one that didn’t require aggressive treatment. A near-miss that could have had serious consequences. As AI becomes more integrated into decision-making, here are three critical principles for responsible implementation: - Set Clear Boundaries Define where AI assistance ends and human decision-making begins. Establish accountability protocols to avoid blind trust. - Build Trust Gradually Start with low-risk implementations. Validate critical AI outputs with human intervention. Track and learn from every near-miss. - Keep Human Oversight AI should support experts, not replace them. Regular audits and feedback loops strengthen both efficiency and safety. At the end of the day, it’s not about choosing AI 𝘰𝘳 human expertise. It’s about building systems where both work together—responsibly. 💬 What’s your take on AI accountability? How are you building trust in it?

A new California bill, SB 1047, could introduce restrictions on artificial intelligence, requiring companies to test the safety of AI technologies and making them liable for any serious harm caused by their systems. California is debating SB 1047, a bill that could reshape how AI is developed and regulated. If passed, it would require tech companies to conduct safety tests on powerful AI technologies before release. The bill also allows the state to take legal action if these technologies cause harm, which has sparked concern among major AI companies. Proponents believe the bill will help prevent AI-related disasters, while critics argue it could hinder innovation, particularly for startups and open-source developers. 🛡️ Safety First: SB 1047 mandates AI safety testing before companies release new technologies to prevent potential harm. ⚖️ Legal Consequences: Companies could face lawsuits if their AI systems cause significant damage, adding a new layer of liability. 💻 Tech Industry Pushback: Tech giants like Google, Meta, and OpenAI are concerned that the bill could slow AI innovation and create legal uncertainties. 🔓 Impact on Open Source: The bill might limit open-source AI development, making it harder for smaller companies to compete with tech giants. 🌐 Potential Global Effects: If passed, the bill could set a precedent for AI regulations in other states and countries, influencing the future of AI governance globally. #AI #AIBill #TechRegulation #CaliforniaLaw #ArtificialIntelligence #OpenSource #Innovation #TechPolicy #SB1047 #AIRegulation 

AI models like ChatGPT and Claude are powerful, but they aren’t perfect. They can sometimes produce inaccurate, biased, or misleading answers due to issues related to data quality, training methods, prompt handling, context management, and system deployment. These problems arise from the complex interaction between model design, user input, and infrastructure. Here are the main factors that explain why incorrect outputs occur: 1. Model Training Limitations AI relies on the data it is trained on. Gaps, outdated information, or insufficient coverage of niche topics lead to shallow reasoning, overfitting to common patterns, and poor handling of rare scenarios. 2. Bias & Hallucination Issues Models can reflect social biases or create “hallucinations,” which are confident but false details. This leads to made-up facts, skewed statistics, or misleading narratives. 3. External Integration & Tooling Issues When AI connects to APIs, tools, or data pipelines, miscommunication, outdated integrations, or parsing errors can result in incorrect outputs or failed workflows. 4. Prompt Engineering Mistakes Ambiguous, vague, or overloaded prompts confuse the model. Without clear, refined instructions, outputs may drift off-task or omit key details. 5. Context Window Constraints AI has a limited memory span. Long inputs can cause it to forget earlier details, compress context poorly, or misinterpret references, resulting in incomplete responses. 6. Lack of Domain Adaptation General-purpose models struggle in specialized fields. Without fine-tuning, they provide generic insights, misuse terminology, or overlook expert-level knowledge. 7. Infrastructure & Deployment Challenges Performance relies on reliable infrastructure. Problems with GPU allocation, latency, scaling, or compliance can lower accuracy and system stability. Wrong outputs don’t mean AI is "broken." They show the challenge of balancing data quality, engineering, context management, and infrastructure. Tackling these issues makes AI systems stronger, more dependable, and ready for businesses. #LLM

It's never been more exciting to start an AI startup. But the graveyard is vast. Here's what not to do: Spencer Shulem and I studied dozens of AI startup failures and successes. This is what we learned: — 1. Falling for shiny object syndrome When a shiny new model or tech drops, it's tempting to pursue it. For example, Argo AI raised billions of dollars to build self-driving tech. But after 6 years, the company realized the tech wasn't ready for public roads. Now, it's gone. Successful startups stay laser-focused on their target user and use case. For example, Anthropic has been working on its constitutional AI technology for years, despite many flashy new approaches emerging. That focus allowed them to make (one of) the best LLM(s) out there. — 2. "It works in the lab" Turning prototypes into products takes massive investments. Don't make the Rabbit/Humane mistake: they had good demos and commercials, but the AI devices didn't live up to the hype in the real-world. Now, both are headed to the graveyard. Successful AI startups make demos replicable in reality. For instance, Cohere spent two years building a robust serving platform. This foundational work enabled their self-serve API to reliably handle billions of requests from day 1. — 3. Irresponsible deployment In the rush to market, many AI product teams fail to put adequate safeguards in place. Take Clearview AI. They scraped hundreds of millions of social media photos without consent. When the NYT exposed it, they got banned from selling to companies and folded. On the other hand, teams like those at Perplexity AI pay especially close attention to Red Teaming. Their vigilance has allowed them to take share from Google, whose AI search has myriad examples of irresponsible outputs (like recommending the depressed to jump off a bridge). — 4. Prioritizing flash over function Many failed AI startups churn out flashy demos that generate reams of press, but don't solve real problems. Remember Quixey? Their demos touted a deep learning-powered "search engine for apps." Now, they don't exist. Successful startups like video AI tool Runway laser-focused on their users' gnarliest problems. They went deep on discovery with video creators to find the workflows that burn hours and dollars. Then, they cut the time & cost by 10x. — 5. Raising too much, too fast VC can seem necessary as an AI founder. But have you heard the stories of Olive AI or Inflection? Each raised a billion or more without achieving product-market fit. Now, they barely exist. On the other hand, successful startups like Cohere bootstrapped for 2 years before raising a $40M Series A. This allowed them to deeply validate their self-serve model and hit $1M ARR before taking on VC. With strong fundamentals in place, they could then scale with confidence.

Have you seen Stanford University's new AI Index Report? There's a ton to digest, but this takeaway stands out to me the most: “The responsible AI ecosystem evolves—unevenly.” In the report, the editors highlight that AI-related incidents are on the rise, but standardized evaluations for validating response quality (and safety) leave MUCH to be desired. From a corporate perspective, there’s an observable gap between understanding the RISKS of errant model responses— and actually taking meaningful ACTION (at the data, system, code, and model response levels) to mitigate it. The primary thrust of the AI movement seems to be this: Build, build, and build some more…then process the consequences later. I think we need to take a step back as data leaders and ask if that’s really an acceptable approach. Should it be? A couple of bright spots highlighted in the report included a few new benchmarks which all offer promising steps toward assessing the factuality and safety of model responses: - HELM Safety - AIR-Bench - and FACTS Governments are also taking notice. “In 2024, global cooperation on AI governance intensified, with organizations including the OECD, EU, U.N., and African Union releasing frameworks focused on transparency, trustworthiness, and other core responsible AI principles.” When it comes to AI, it’s not just our customers who suffer when things go awry. Our stakeholders, our teammates, and our reputations are all on the hook for generative missteps. In short, rewards may still be high—but the risks have never been higher. What do you think? Let me know in the comments!

Retrieval-Augmented Generation (RAG) enhances AI models by dynamically pulling in relevant external knowledge rather than relying solely on pre-trained data. This leads to more 𝗮𝗰𝗰𝘂𝗿𝗮𝘁𝗲, 𝘂𝗽-𝘁𝗼-𝗱𝗮𝘁𝗲, 𝗮𝗻𝗱 𝗰𝗼𝗻𝘁𝗲𝘅𝘁𝘂𝗮𝗹𝗹𝘆 𝗿𝗶𝗰𝗵 𝗿𝗲𝘀𝗽𝗼𝗻𝘀𝗲𝘀.  Agentic AI refers to 𝗔𝗜 𝘀𝘆𝘀𝘁𝗲𝗺𝘀 𝘁𝗵𝗮𝘁 𝗼𝗽𝗲𝗿𝗮𝘁𝗲 𝗮𝘂𝘁𝗼𝗻𝗼𝗺𝗼𝘂𝘀𝗹𝘆, 𝗺𝗮𝗸𝗶𝗻𝗴 𝗱𝗲𝗰𝗶𝘀𝗶𝗼𝗻𝘀, 𝗽𝗹𝗮𝗻𝗻𝗶𝗻𝗴, 𝗮𝗻𝗱 𝗲𝘅𝗲𝗰𝘂𝘁𝗶𝗻𝗴 𝘁𝗮𝘀𝗸𝘀 𝘄𝗶𝘁𝗵 𝗺𝗶𝗻𝗶𝗺𝗮𝗹 𝗵𝘂𝗺𝗮𝗻 𝗶𝗻𝘁𝗲𝗿𝘃𝗲𝗻𝘁𝗶𝗼𝗻. Instead of passively generating text, agentic AI interacts with external tools, reasons about problems, and refines its own processes.  As AI systems evolve, the question is no longer just about 𝗥𝗔𝗚 𝘃𝘀. 𝗻𝗼𝗻-𝗥𝗔𝗚, but rather: 𝗦𝗵𝗼𝘂𝗹𝗱 𝗥𝗔𝗚 𝗯𝗲 𝗵𝗮𝗻𝗱𝗹𝗲𝗱 𝗯𝘆 𝗮 𝘀𝗶𝗻𝗴𝗹𝗲 𝗮𝗴𝗲𝗻𝘁 𝗼𝗿 𝗮 𝗺𝘂𝗹𝘁𝗶-𝗮𝗴𝗲𝗻𝘁 𝘀𝘆𝘀𝘁𝗲𝗺?  𝗦𝗶𝗻𝗴𝗹𝗲-𝗔𝗴𝗲𝗻𝘁 𝗔𝗴𝗲𝗻𝘁𝗶𝗰 𝗥𝗔𝗚   In this approach, a 𝘀𝗶𝗻𝗴𝗹𝗲 𝗮𝘂𝘁𝗼𝗻𝗼𝗺𝗼𝘂𝘀 𝗔𝗜 𝗮𝗴𝗲𝗻𝘁 manages the entire 𝗿𝗲𝘁𝗿𝗶𝗲𝘃𝗮𝗹, 𝗿𝗲𝗮𝘀𝗼𝗻𝗶𝗻𝗴, 𝗮𝗻𝗱 𝗴𝗲𝗻𝗲𝗿𝗮𝘁𝗶𝗼𝗻 𝗽𝗿𝗼𝗰𝗲𝘀𝘀. It can 𝗽𝗹𝗮𝗻, 𝗱𝗲𝗰𝗶𝗱𝗲 𝘄𝗵𝗶𝗰𝗵 𝘀𝗼𝘂𝗿𝗰𝗲𝘀 𝘁𝗼 𝗾𝘂𝗲𝗿𝘆, 𝘀𝘆𝗻𝘁𝗵𝗲𝘀𝗶𝘇𝗲 𝗿𝗲𝘀𝗽𝗼𝗻𝘀𝗲𝘀, 𝗮𝗻𝗱 𝗲𝘃𝗲𝗻 𝘃𝗲𝗿𝗶𝗳𝘆 𝗶𝘁𝘀 𝗼𝘄𝗻 𝗼𝘂𝘁𝗽𝘂𝘁𝘀.  𝗣𝗿𝗼𝘀:   ✔ More 𝗲𝗳𝗳𝗶𝗰𝗶𝗲𝗻𝘁 (less overhead from inter-agent coordination)   ✔ Easier to 𝗼𝗿𝗰𝗵𝗲𝘀𝘁𝗿𝗮𝘁𝗲 and integrate into existing workflows   ✔ Lower computational cost  𝗖𝗼𝗻𝘀:   ✖ 𝗟𝗶𝗺𝗶𝘁𝗲𝗱 𝘀𝗽𝗲𝗰𝗶𝗮𝗹𝗶𝘇𝗮𝘁𝗶𝗼𝗻 – one agent must handle everything   ✖ Can become a 𝗯𝗼𝘁𝘁𝗹𝗲𝗻𝗲𝗰𝗸 when dealing with complex, multi-step tasks  𝗠𝘂𝗹𝘁𝗶-𝗔𝗴𝗲𝗻𝘁 𝗔𝗴𝗲𝗻𝘁𝗶𝗰 𝗥𝗔𝗚  Here, 𝗺𝘂𝗹𝘁𝗶𝗽𝗹𝗲 𝘀𝗽𝗲𝗰𝗶𝗮𝗹𝗶𝘇𝗲𝗱 𝗔𝗜 𝗮𝗴𝗲𝗻𝘁𝘀 collaborate to perform different tasks—retrieval, validation, synthesis, planning, or even fact-checking. Each agent has a specific role, creating a 𝗺𝗼𝗱𝘂𝗹𝗮𝗿 𝗮𝗻𝗱 𝘀𝗰𝗮𝗹𝗮𝗯𝗹𝗲 system.  𝗣𝗿𝗼𝘀:   ✔ 𝗕𝗲𝘁𝘁𝗲𝗿 𝘀𝗽𝗲𝗰𝗶𝗮𝗹𝗶𝘇𝗮𝘁𝗶𝗼𝗻 – dedicated agents for retrieval, reasoning, and validation   ✔ 𝗛𝗶𝗴𝗵𝗲𝗿 𝗮𝗰𝗰𝘂𝗿𝗮𝗰𝘆 – multiple agents cross-check and refine results   ✔ More 𝘀𝗰𝗮𝗹𝗮𝗯𝗶𝗹𝗶𝘁𝘆 𝗮𝗻𝗱 𝗮𝗱𝗮𝗽𝘁𝗮𝗯𝗶𝗹𝗶𝘁𝘆 for complex workflows  𝗖𝗼𝗻𝘀:   ✖ 𝗠𝗼𝗿𝗲 𝗰𝗼𝗺𝗽𝗹𝗲𝘅 𝗼𝗿𝗰𝗵𝗲𝘀𝘁𝗿𝗮𝘁𝗶𝗼𝗻 – requires careful agent coordination   ✖ 𝗛𝗶𝗴𝗵𝗲𝗿 𝗰𝗼𝗺𝗽𝘂𝘁𝗮𝘁𝗶𝗼𝗻𝗮𝗹 𝗰𝗼𝘀𝘁  Single-Agent Agentic RAG is 𝗹𝗶𝗴𝗵𝘁𝗲𝗿 𝗮𝗻𝗱 𝘀𝗶𝗺𝗽𝗹𝗲𝗿, making it ideal for well-defined, streamlined use cases. 𝗠𝘂𝗹𝘁𝗶-𝗔𝗴𝗲𝗻𝘁 𝗔𝗴𝗲𝗻𝘁𝗶𝗰 𝗥𝗔𝗚, however, is more 𝗮𝗱𝗮𝗽𝘁𝗮𝗯𝗹𝗲, 𝘀𝗰𝗮𝗹𝗮𝗯𝗹𝗲, 𝗮𝗻𝗱 𝗯𝗲𝘁𝘁𝗲𝗿 𝘀𝘂𝗶𝘁𝗲𝗱 𝗳𝗼𝗿 𝗶𝗻𝘁𝗿𝗶𝗰𝗮𝘁𝗲 𝗱𝗲𝗰𝗶𝘀𝗶𝗼𝗻-𝗺𝗮𝗸𝗶𝗻𝗴 𝘁𝗮𝘀𝗸𝘀.  As AI systems become 𝗺𝗼𝗿𝗲 𝗮𝘂𝘁𝗼𝗻𝗼𝗺𝗼𝘂𝘀, expect 𝗺𝘂𝗹𝘁𝗶-𝗮𝗴𝗲𝗻𝘁 𝗮𝗿𝗰𝗵𝗶𝘁𝗲𝗰𝘁𝘂𝗿𝗲𝘀 𝘁𝗼 𝗯𝗲𝗰𝗼𝗺𝗲 𝘁𝗵𝗲 𝘀𝘁𝗮𝗻𝗱𝗮𝗿𝗱 𝗳𝗼𝗿 𝗲𝗻𝘁𝗲𝗿𝗽𝗿𝗶𝘀𝗲-𝗹𝗲𝘃𝗲𝗹 𝗥𝗔𝗚 𝗮𝗽𝗽𝗹𝗶𝗰𝗮𝘁𝗶𝗼𝗻𝘀.  

Ever been fooled by a chatbot thinking it was a real person? It happened to me! As AI continues to evolve, particularly in the realm of chatbots, transparency is more important than ever. In many interactions, it’s not always clear if you’re talking to a human or an AI—an issue that can affect trust and accountability. AI-powered tools can enhance convenience and efficiency, but they should never blur the lines of communication. People deserve to know when they’re interacting with AI, especially when it comes to critical areas like healthcare, customer service, and financial decisions. Transparency isn’t just ethical—it fosters trust, allows users to make informed decisions, and helps prevent misinformation or misunderstandings. As we integrate AI more deeply into our daily lives, let’s ensure clarity is a top priority. Transparency should be built into every interaction, making it clear when AI is at the wheel. That’s how we build responsible, reliable, and user-friendly AI systems. GDS Group #AI #Transparency #EthicsInAI #TrustInTechnology

A new behavior that must be evaluated in AI models: sycophancy. (And don’t worry if you had to look up what that means—I did too.) On April 25th, OpenAI released a new version of GPT-4o in ChatGPT. But something was off. The model had become noticeably more agreeable—to the point of being unhelpful or even harmful. It wasn’t just being nice; it was validating doubts, encouraging impulsive behavior, and reinforcing negative emotions. The cause? New training signals like thumbs-up/down user feedback unintentionally weakened safeguards against sycophantic behavior. And since sycophancy hadn’t been explicitly tracked or flagged in previous evaluations, it slipped through. What I appreciated most was OpenAI’s transparency in owning the miss and outlining clear steps for improvement. It's a powerful reminder that as we release more advanced AI systems, new risks will emerge—ones we may not yet be measuring. I believe this signals a rising need for AI quality control—what I like to call QA for AI, or even “therapists for AI.” People whose job is to question, test, and ensure the model is sane, safe, and aligned before it reaches the world. We’re still learning and evolving with these tools—and this post is a great read if you're following the path of responsible AI: https://lnkd.in/gXwY-Rjf