AI Limitations Overview

Google DeepMind just exposed AI's limits - with Math. The same vector embeddings that power most modern AI search systems have mathematical limits we can't engineer around. No amount of training data or model scaling will fix this, according to them. Here's what's happening: When we ask AI to find relevant documents, we're essentially asking it to map meaning into geometric space - turning words into coordinates. But the researchers proved that for any given embedding dimension, there are combinations of documents that simply cannot be retrieved correctly. What sounds like a bug might be a fundamental limitation of these systems. To demonstrate this, they created LIMIT - a dataset so simple a child could solve it (matching "who likes apples?" with "Jon likes apples"). Yet even the best models, including those powering enterprise search systems, achieve less than 20% accuracy. GPT-class models with 4,096-dimensional embeddings still fail spectacularly. As we push AI to handle more complex retrieval tasks - think multi-criteria search, reasoning-based queries, or the instruction-following systems many companies are betting on - we're guaranteed to hit these walls. The paper shows that web-scale search would need embedding dimensions in the millions to handle all possible document combinations. So, what does this mean? Every company building RAG systems, every startup promising "ChatGPT for your documents," every enterprise search deployment - they're all constrained by this fundamental limit. The researchers found that alternative architectures like sparse models (think old-school keyword search) actually outperform modern neural approaches on these tasks. We've been treating retrieval as a solved problem, a building block we can rely on. But their research suggests we need to fundamentally rethink how we architect AI systems that need to find and reason over information. The good news? Once we understand the limits, we can design around them. Hybrid approaches, multi-stage retrieval, and careful system design can mitigate these issues. But it requires acknowledging that bigger models and more compute won't solve everything. For those of us working with AI, this is a reminder that understanding the fundamentals matters. The next breakthrough might not come from scaling up, but from stepping back and questioning our basic assumptions. What retrieval challenges has your organization faced that might be explained by these fundamental limits? ↓ 𝐖𝐚𝐧𝐭 𝐭𝐨 𝐤𝐞𝐞𝐩 𝐮𝐩? Join my newsletter with 50k+ readers and be the first to learn about the latest AI research: llmwatch.com 💡

The biggest shortcoming of AI resumes is simple. They fail to sell the human behind the document. Most AI-generated resumes follow the same pattern. They stack buzzwords. They inflate language. They sound polished. But they don’t communicate anything real. When you read enough resumes, you start to notice it fast. You see phrases like “results-driven,” “strategic thinker,” “cross-functional leader.” But you don’t see proof. You don’t see context and you don’t see the person. A resume is not supposed to sound impressive. It is supposed to be clear. When I review a resume, I’m looking for three things: 1. Why are you valuable? 2. What do you actually do? 3. How have you delivered results? If I can’t answer those in under 30 seconds, the resume failed. Strong resumes don’t hide behind language. They show specifics. Instead of saying you “led a team,” tell me how many people and what the outcome was. Instead of saying you “drove growth,” show me the numbers and the timeframe. Instead of saying you are “strategic,” show me a decision you made and what it changed. Clarity beats cleverness every time. AI can help you format a resume. It can help you clean up grammar. It can even help you organize your experience. But it cannot replace your thinking. If you rely on it to tell your story, you will end up with something that sounds good and says nothing. The best resumes feel human. They are specific. They are direct. They make it obvious why you matter. If someone finishes reading your resume and still has to guess what you’re good at, you didn’t do your job.

Some "AI" vocabulary: - LLM: Large Language Model. A statistical pattern machine that does NOT "know" things. They're just next-token predictors trained on existing texts. Ask a question and they do NOT check a database. They spin a statistical wheel of fortune. - Prompt: Your input. For example: "Marie Curie is remembered for…" Most of the time you'll get "radioactivity" or "Nobel Prize". You might also get "discovered penicillin". Wrong. But it sounds right, because in the vector space of the LLM, "scientist" lights up near "medical breakthrough" and "discovery" exist near "penicillin". The model MAY stitch it together into a fake fact. Confidently delivered, however. - Hallucination: from the model's perspective, this so-called "hallucination" is IDENTICAL to the truth. Same patterns. Same confidence. There is no model of confidence level, no warning light to the user. Because the model cannot know the distinction. It just predicts the next token. - RLHF: Reinforcement Learning from Human Feedback. This is what we sometimes call lipstick on the pig. Humans do NOT teach the model "truth" (because, remember, it cannot model truth), we're teaching polish. The model learns to hedge: "While I cannot verify…" It looks cautious and aware, but it's still guessing. Just smoother guessing. Will it be good enough for your particular service? Perhaps. But you (and your poor client that ordered "AI" from you) need to be aware of the limitations. From a truth perspective, RLHF might actually worsen the user experience because of the hand-wringing hedging. - RAG: Retreival-Augmented Generation. This is the industry's favorite patch: bolt on a traditional search engine and some internal documents; give the model a cheat sheet. It helps, a little. But the generator is still the same. It can misquote, remix, or confidently hallucinate on top of the retrieved facts. Garbage in, polished garbage out. And while researchers have warned us about this for YEARS, it has finally dawned on the biggest vendors (like OpenAI) that hallucinations aren't going away. Ever. They’re a feature of the architecture, not a glitch (see link in comments). LLMs are pattern machines. They're great at style, fluent in form, masters of mimicry. But they are NOT reliable knowledge engines or databases. If you build products on top of them, you are betting your future on a storyteller that never knows when it's lying. I feel like Cato the Elder since I always end with stuff like this: "Furthermore, I think we need to remember that 'AI' is not only referring to GPT-based LLMs."

The strange thing about AI is that you cannot think your way to knowing what it is good for. You have to use it, badly, for a while, before the map starts to appear. Most organizations get this backwards. They treat AI as a deployment problem: pick the use case, clean the data, get the governance sign off, then roll it out. But AI capability is jagged. It is brilliant at things you assumed were hard and too often useless at things you assumed were trivial, and the pattern shifts with every model and every workflow. You cannot predict the edges of that frontier from a conference room. You find them by walking into them. Look at SpaceX's first Raptor engine. Exposed wiring, plumbing everywhere, held together by what looks like improvisation. Then look at today's version: clean, dark, and beautifully engineered. The first version was never meant to be pretty. It existed so they could learn what to fix. You do not get the clean engine by designing it on paper. You get it by flying the ugly one and watching where it fails. AI rewards the same posture. The companies and teams pulling ahead are not the ones with the best AI strategy. They are the ones running the most cheap, fast, slightly embarrassing experiments, because that is the only way the knowledge actually transfers. Most of what you need to know about using AI well is tacit. It does not live in a slide. It lives in the hands of someone who has tried it on real work a hundred times. Here is the part worth sitting with. The cost of a bad AI experiment has collapsed to almost nothing. You can stand one up in an afternoon and kill it by Friday. The cost of waiting went the other way, because while you wait, your competitors are not building tools. They are building judgment. And judgment compounds. So the bias should be obvious. Stop hunting for the perfect first use case. There isn't one, and the hunt is just standing still with extra steps. Put something rough in front of real work, learn what only contact can teach, and fix the next thing. The frontier is jagged, and the only map is the one you draw with your own hands. ↓ 𝗜𝗳 𝗔𝗜 𝗶𝘀 𝗽𝗮𝗿𝘁 𝗼𝗳 𝘆𝗼𝘂𝗿 𝗷𝗼𝗯, 𝗺𝘆 𝗳𝗿𝗲𝗲 𝗻𝗲𝘄𝘀𝗹𝗲𝘁𝘁𝗲𝗿 𝗶𝘀 𝗳𝗼𝗿 𝘆𝗼𝘂. 𝗜 𝘀𝗵𝗮𝗿𝗲 𝗽𝗿𝗮𝗰𝘁𝗶𝗰𝗮𝗹 𝘄𝗼𝗿𝗸𝗳𝗹𝗼𝘄𝘀, 𝘁𝗼𝗼𝗹𝘀, 𝗮𝗻𝗱 𝘄𝗵𝗮𝘁’𝘀 𝗮𝗰𝘁𝘂𝗮𝗹𝗹𝘆 𝗰𝗵𝗮𝗻𝗴𝗶𝗻𝗴 𝗶𝗻 𝘁𝗵𝗲 𝗳𝗶𝗲𝗹𝗱: https://lnkd.in/dbf74Y9E

If your resume is too perfect, you might not get hired. Yes, you heard that right. I’ve reviewed over 100 resumes this month alone. And honestly, most of them looked great, with clean formatting, strong action words, and everything spelled correctly. Everything you’d expect from a top candidate. But guess what? > Almost all of them used the same templates. > Most achievements sounded… almost identical. > Very few actually made me pause. You could tell AI had cleaned things up, making the resumes neater… but also flatter. Less human. For example, a product analyst wrote, “Improved dashboard performance by 18%.” Not bad, right? But we changed it to: “Redesigned product dashboard after hearing repeated user complaints, which led to 18% faster decision-making for three internal teams.” Same number. A whole new story. Remember, recruiters don’t stop reading when something is perfect. They stop when something is forgettable. What sticks is the story, the context, the decisions behind those numbers, and the person behind the bullet point. So if you’re using AI to write your resume, that’s totally fine. But don’t stop there. Go back. Add your voice. Add the “why.” Bring in your personality, even if it’s just in small ways. Because at the end of the day, hiring managers aren’t just choosing a resume; they’re choosing a human. P.S. Have you looked at your resume recently and thought, “Wait… this doesn’t even sound like me”? If yes, it might be time for a rewrite. #ResumeTips #AIinHiring #JobSearch #InterviewCoach #cvtips

OpenAI recently posted a role that would have sounded strange even a few years ago: Head of Preparedness. The title alone sparked debate. Some read it as prudent. Others as performative. I see it as an admission that the field is entering uncharted territory. The role forces an obvious but uncomfortable question: what, exactly, are we preparing for? At first glance, the answers seem familiar - misuse, model failures, disinformation, cyber risk. These matter, but they are incomplete. If preparedness were only about preventing known harms, it would sit comfortably inside existing safety or policy teams. This role exists because the core problem is uncertainty that cannot be specified ahead of time. Modern AI systems increasingly exhibit emergent behavior. Not just better execution of known tasks, but capabilities that appear only after deployment, through interaction with users, tools, and incentives. These behaviors are hard to predict beforehand, to test exhaustively, or to explain cleanly after the fact. As models become more capable and more agentic, risk stops looking like a checklist of edge cases and starts looking like system dynamics. Feedback loops form between humans and models. Capabilities surface through prolonged use. Deployment pressures outpace institutional understanding. The failure mode is no longer a single obvious flaw, but patterns that compound over time. Preparedness also reflects a shift from invention to diffusion. The most consequential effects of AI are unlikely to come from a single breakthrough moment, but rather from countless small integrations into workflows, markets, and decision systems - individually benign, potentially destabilizing in aggregate. Preparedness is about watching the second derivative. There is an institutional dimension as well. Frontier labs now operate under sustained scrutiny from regulators, customers, governments, and internal stakeholders. A dedicated preparedness function creates a locus of accountability when something unexpected happens. At a deeper level, preparedness is an acknowledgment of limits. We are building general-purpose systems inside tightly coupled social and economic structures. The most honest posture is not confidence, but readiness. The polarized response to the role signals a field that is uncomfortable with its own maturity. Early-stage technologies tend to celebrate speed and dismiss caution as fear. Mature technologies institutionalize caution because the cost of being wrong compounds. AI is in the awkward middle: powerful enough to matter, yet young enough to still mythologize recklessness. Expect more labs to formalize similar roles. Expect preparedness to sit closer to product and deployment, not just research ethics. And expect the conversation to shift from whether such roles are necessary to how much authority they actually have. AI development is moving from optimism about capability to responsibility for consequences.

For many companies, proving the ROI of AI is hard enough. But in customer experience? It's often a struggle because the benefits can be complex and difficult to measure. While AI can clearly improve efficiency, its most significant impacts, like increasing customer lifetime value, are harder to connect directly to a financial return. This is especially true for customer-facing applications like chatbots or personalized recommendation engines. The problem typically starts with how companies define success. They often focus on what's easiest to measure rather than what's most valuable. For example, companies might measure a chatbot's resolution rate but not whether that resolution drove additional spending or reduced churn. The real ROI in CX isn't just about saving money on call center agents; it's about increasing customer lifetime value. Let's take AI-driven personalization as an example. It can make a customer feel understood and valued, but how do you put a dollar amount on that feeling? The benefits are often intangible, like a stronger brand reputation or higher loyalty, which are important for long-term growth but don't show up on a quarterly balance sheet. Many organizations deploy an AI chatbot or a new recommendation engine just because the technology is available, not because they've identified a specific customer pain point to solve. This leads to disconnected, siloed projects that don't align with a clear business strategy, making it impossible to calculate a meaningful return. And when the "AI Strategy" isn't integrated into the "Business Strategy," the negative impact is higher given the scale. But even with a clear vision, bringing an AI-powered CX solution to life is riddled with practical challenges. What are those, you might ask? For starters, AI models for CX, like chatbots or sentiment analysis tools, rely heavily on high-quality, clean data. If your customer interaction data is fragmented across different systems, incomplete, or biased, the AI will produce flawed results. The initial work of integrating, cleaning, and structuring this data is a massive, time-consuming effort that often gets underestimated. Integration with legacy systems, like your CRM or support systems, is not designed to seamlessly integrate with new AI technology. Connecting an AI engine to these older systems can be a complex and expensive technical nightmare that drains budgets and delays projects. Finally, we have employees. Customer service agents may resist using AI tools for fear of being replaced. Without a clear plan for change management and a focus on how AI can augment their abilities, like providing real-time information or summarizing a customer's history, adoption will be low and the project will fail to deliver value. Find a problem. Get your data ducks in a row. Connect systems. Solve the problem with AI. And help your people along the journey. #customerexperience #ai #technology #innovation #changemanagement

From Code Generation to System Integration: Why AI Coding Tools and Agentic IDEs Must Evolve to Solve Real Software Development Challenges Since GPT-3 went mainstream, AI coding tools have sprinted through three waves. 1. First came smart autocomplete. 2. Then came cloud companions tuned to specific stacks. 3. Now we’re in the agent wave – tools that read whole repos, open terminals, run tests and raise pull requests on their own. Every cycle starts the same way: Wow. Impressive. Look at how much this can do for me. But the uncomfortable truth is this: most of what these tools automate is commodity knowledge. Framework boilerplate, CRUD patterns, standard integration glue, typical test shapes – once a pattern exists in public code, a model can learn it and repeat it very well. That used to feel like expertise. Now it’s autocomplete on steroids. The real problems have barely moved: • Design and architecture. Not just file-by-file edits, but coherent system design: boundaries, contracts, data flows, failure modes, performance budgets – a holistic solution, not local patchwork. •  End-to-end SDLC integration. How change actually flows from idea to production: design, review, CI, approvals, environments, rollout strategies and on-call ownership. • Change management and legacy transformation. How to evolve decade-old systems, untangle hidden dependencies, migrate behaviour safely and avoid breaking everything that still quietly depends on “that old module”. • Traceability. Knowing who or what changed what, why, and what else was impacted – across code, configs, data pipelines and policies. • How strongly workflows enforce the top 10 principles like reliability, security, cost and maintainability that were outlined in the earlier post – not as posters on a wall, but as gates every change must pass through. This is where vibe-coding tools become dangerous. The model writes the feature, generates the tests, explains the diff. Everything looks green. It feels safe enough to ship on vibe. Without deep expertise and a solid workflow around it, that is not productivity. It is an efficient way to inject new risk into a live system. If code patterns are now cheap, differentiation shifts somewhere else: • To how clearly an organisation defines how systems should be built and evolved • To how tightly AI tools are integrated with that SDLC, not just with the editor • To how well workflows embody design principles, change discipline and traceability by default Writing code is becoming a commodity. However, writing holistic, thoughtful systems, and continuously evolving and governing them safely, is where the true value lies AI coding copilots and agentic IDEs now need to evolve from “look what I can generate” to “look how I help you integrate, operate and transform”. That’s when it stops being “wow, impressive demo” and becomes “yes – this is finally solving the real problem.”

🧑💻🐢 𝗔𝗜 𝗧𝗼𝗼𝗹𝘀 𝗦𝗹𝗼𝘄𝗲𝗱 𝗗𝗼𝘄𝗻 𝗧𝗼𝗽 𝗢𝗽𝗲𝗻-𝗦𝗼𝘂𝗿𝗰𝗲 𝗗𝗲𝘃𝘀 𝗯𝘆 𝟭𝟵% We all 𝗲𝘅𝗽𝗲𝗰𝘁 AI to 𝗯𝗼𝗼𝘀𝘁 productivity. But what happens when you rigorously test that assumption in the wild, with real code and experienced devs? A recent RCT (randomized controlled trial) study from METR (Feb–June 2025) tested exactly that. 𝗦𝘁𝘂𝗱𝘆 𝗮𝘁 𝗮 𝗚𝗹𝗮𝗻𝗰𝗲: 🧑💻 Participants: 16 experienced OSS developers (5+ years on their projects) 🗂️ Tasks: 246 real GitHub issues from large, mature repos 🛠️ Tools: Cursor Pro, Claude 3.5/3.7 Sonnet 🎲 Conditions: Randomized to AI-allowed vs. AI-disallowed 𝗥𝗲𝘀𝘂𝗹𝘁𝘀: 📉 Developers forecasted AI would speed up work by 24% 🧠 Post-task, they still believed it helped by 20% 🤯 Reality check: 𝗧𝗮𝘀𝗸𝘀 𝘁𝗼𝗼𝗸 𝟭𝟵% 𝗹𝗼𝗻𝗴𝗲𝗿 𝘄𝗶𝘁𝗵 𝗔𝗜 Even expert economists & ML researchers predicted a ~39% speedup. Instead: 𝗔𝗜 𝘀𝗹𝗼𝘄𝗲𝗱 𝘁𝗵𝗲𝗺 𝗱𝗼𝘄𝗻. 𝗪𝗵𝘆? With AI, devs: 🔍 Spent 𝗹𝗲𝘀𝘀 time 𝗰𝗼𝗱𝗶𝗻𝗴/searching ⏳ Spent 𝗺𝗼𝗿𝗲 time 𝗽𝗿𝗼𝗺𝗽𝘁𝗶𝗻𝗴, 𝘄𝗮𝗶𝘁𝗶𝗻𝗴, and 𝗿𝗲𝘃𝗶𝗲𝘄𝗶𝗻𝗴 AI outputs 💤 Faced more 𝗶𝗱𝗹𝗲 𝘁𝗶𝗺𝗲 and 𝗺𝗲𝗻𝘁𝗮𝗹 𝘀𝘄𝗶𝘁𝗰𝗵𝗶𝗻𝗴 𝗥𝗼𝗼𝘁 𝗖𝗮𝘂𝘀𝗲𝘀: 🙃 𝗢𝘃𝗲𝗿-𝗼𝗽𝘁𝗶𝗺𝗶𝘀𝗺 about AI's value ❌ 𝗟𝗼𝘄 𝗮𝗰𝗰𝗲𝗽𝘁𝗮𝗻𝗰𝗲 rate of AI suggestions (~44%) 🧱 Large, 𝗰𝗼𝗺𝗽𝗹𝗲𝘅 codebases too tricky for current AI 🧠 AI couldn't match developers 𝘂𝗻𝘄𝗿𝗶𝘁𝘁𝗲𝗻 understanding of the codebase 𝗜𝗺𝗽𝗼𝗿𝘁𝗮𝗻𝘁 𝗖𝗮𝘃𝗲𝗮𝘁𝘀: This doesn’t mean AI isn’t helpful, just that: 🧓 For very experienced devs on familiar repos, current tools may fall short 🆕 But for new projects, junior devs, or greenfield code, the story could be different 𝗟𝗼𝗼𝗸𝗶𝗻𝗴 𝗔𝗵𝗲𝗮𝗱: ✍️ Better prompting, lower latency, and domain-specific tuning might flip the results 🤖 Claude 3.7 already shows promise for partial task automation 𝗧𝗵𝗶𝘀 𝘀𝘁𝘂𝗱𝘆 𝘀𝘁𝗮𝗻𝗱𝘀 𝗼𝘂𝘁 𝗯𝘆 𝗯𝗲𝗰𝗮𝘂𝘀𝗲: 🌍 Using real-world tasks (not synthetic) 🧑🔬 Engaging expert developers ⏱️ Measuring fixed, real productivity (not just output volume) 🔗 Read the full study: Link in the comments 🤔 Have you noticed AI helping or hurting your coding workflow? #AI #SoftwareDevelopment #Productivity #MachineLearning

Most AI-Generated Resumes Are Garbage. But If You Use AI the Right Way… Let’s be honest—if you let AI blindly generate your resume, it’s going to be a generic, keyword-stuffed mess that recruiters toss aside. But if you use AI strategically, it can turn your resume into a powerful, ATS-friendly, and recruiter-approved document. Here’s how to make AI work for you (not against you): ✅ Use AI to Optimize, Not Write AI tools like Jobscan & ResyMatch can: Identify missing keywords so your resume matches job descriptions. Highlight skills & experience that ATS systems prioritize. Suggest better formatting for readability. 🚫 What AI shouldn’t do: Write your entire resume from scratch. It lacks personality, depth, and real impact. ✅ Tailor Your Resume with AI AI can help customize your resume for each job application by: Prioritizing relevant experience & skills for the role. Restructuring bullet points to match employer expectations. Refining wording to improve clarity & engagement. 🚫 What AI shouldn’t do: Copy-paste the same buzzwords for every job—recruiters see right through it. ✅ Let AI Help Quantify Your Achievements Strong resumes show impact, not just responsibilities. AI can: Suggest metrics & numbers to make accomplishments stand out. Turn “Managed social media” into “Increased engagement by 40% in 3 months.” Refine vague phrases into action-driven statements. 🚫 What AI shouldn’t do: Make up data—integrity matters! Why AI-Generated Resumes Get Rejected Overstuffed with buzzwords that make you sound robotic. Generic & impersonal—hiring managers want you, not an AI clone. Exaggerated skills that don’t hold up in interviews. The Smartest Way to Use AI? Balance AI with a human touch. Let AI suggest, but you decide what truly represents your experience. 💡 AI is your assistant, not your replacement. Use it to refine, enhance, and optimize—not to replace effort and authenticity. Have you used AI to improve your resume? What’s been your experience? Drop a comment! Follow Samichi for smarter job search strategies, AI-powered career hacks, and resume tips that actually get results. #JobSearch #ResumeTips #AIForJobSeekers #CareerGrowth #ATSOptimization #AIResumes