Christopher Penn's Recent LinkedIn Posts

Yesterday, LinkedIn FINALLY confirmed what we all knew: there's a new algorithmic sheriff in town. Well, it's not so much a sheriff as it is an entire department. At Trust Insights, we've been covering the LinkedIn algorithm for the past two years based on academic papers, engineering content, conference presentations, and more. And we've just updated our Unofficial LinkedIn Algorithm Guide with the latest announcements and papers from yesterday that help you understand what's happening under the hood. Here's the good news: if you've been focused on meaningful engagement (leave those comments!) and producing useful, helpful content, nothing really changes for you. There are some nuances about how LinkedIn's LLMs work and little things you can do, but for the most part, the evergreen advice is evergreen for a reason: the LLMs under the hood are nearly impossible to "hack" or "game". Anyone telling you otherwise doesn't understand the systems and how they work together. Two things matter more than ever: profile clarity and behavioral coherence. Because of the nature of LLMs themselves (and LinkedIn is using an ensemble of them, not just one) you have to interact with them in meaningful, semantically coherent, consistent ways. What are you an expert in? Lean into that in the majority of your interactions. And do it frequently. The new sequentual transformer LLM under the hood of the generative recommender looks at your behavior over time. There's a lot more detail than I can fit in a post, so go grab your copy of the completely refreshed Unofficial LinkedIn Algorithm Guide, free of cost! Get it here: https://lnkd.in/eCjw7SWJ #AI #GenerativeAI #GenAI #LinkedIn #Algorithm #ArtificialIntelligence #LargeLanguageModels #MachineLearning #IntelligenceRevolution cc Sydney Baldwin Amber Naslund

Yesterday, LinkedIn FINALLY confirmed what we all knew: there's a new algorithmic sheriff in town. Well, it's not so much a sheriff as it is an entire department. At Trust Insights, we've been covering the LinkedIn algorithm for the past two years based on academic papers, engineering content, conference presentations, and more. And we've just updated our Unofficial LinkedIn Algorithm Guide with the latest announcements and papers from yesterday that help you understand what's happening under the hood. Here's the good news: if you've been focused on meaningful engagement (leave those comments!) and producing useful, helpful content, nothing really changes for you. There are some nuances about how LinkedIn's LLMs work and little things you can do, but for the most part, the evergreen advice is evergreen for a reason: the LLMs under the hood are nearly impossible to "hack" or "game". Anyone telling you otherwise doesn't understand the systems and how they work together. Two things matter more than ever: profile clarity and behavioral coherence. Because of the nature of LLMs themselves (and LinkedIn is using an ensemble of them, not just one) you have to interact with them in meaningful, semantically coherent, consistent ways. What are you an expert in? Lean into that in the majority of your interactions. And do it frequently. The new sequentual transformer LLM under the hood of the generative recommender looks at your behavior over time. There's a lot more detail than I can fit in a post, so go grab your copy of the completely refreshed Unofficial LinkedIn Algorithm Guide, free of cost! Get it here: https://lnkd.in/eCjw7SWJ #AI #GenerativeAI #GenAI #LinkedIn #Algorithm #ArtificialIntelligence #LargeLanguageModels #MachineLearning #IntelligenceRevolution cc Sydney Baldwin Amber Naslund

Yesterday, LinkedIn FINALLY confirmed what we all knew: there's a new algorithmic sheriff in town. Well, it's not so much a sheriff as it is an entire department. At Trust Insights, we've been covering the LinkedIn algorithm for the past two years based on academic papers, engineering content, conference presentations, and more. And we've just updated our Unofficial LinkedIn Algorithm Guide with the latest announcements and papers from yesterday that help you understand what's happening under the hood. Here's the good news: if you've been focused on meaningful engagement (leave those comments!) and producing useful, helpful content, nothing really changes for you. There are some nuances about how LinkedIn's LLMs work and little things you can do, but for the most part, the evergreen advice is evergreen for a reason: the LLMs under the hood are nearly impossible to "hack" or "game". Anyone telling you otherwise doesn't understand the systems and how they work together. Two things matter more than ever: profile clarity and behavioral coherence. Because of the nature of LLMs themselves (and LinkedIn is using an ensemble of them, not just one) you have to interact with them in meaningful, semantically coherent, consistent ways. What are you an expert in? Lean into that in the majority of your interactions. And do it frequently. The new sequentual transformer LLM under the hood of the generative recommender looks at your behavior over time. There's a lot more detail than I can fit in a post, so go grab your copy of the completely refreshed Unofficial LinkedIn Algorithm Guide, free of cost! Get it here: https://lnkd.in/eCjw7SWJ #AI #GenerativeAI #GenAI #LinkedIn #Algorithm #ArtificialIntelligence #LargeLanguageModels #MachineLearning #IntelligenceRevolution cc Sydney Baldwin Amber Naslund

In this episode you'll learn: - What's changed about how LinkedIn works (grab the new paper from Trust Insights!) - How to use the paper's findings in your LinkedIn strategy - How to evaluate your LinkedIn strategic data with agentic AI tools like Claude Cowork

There are five levels of AI enablement now. 1. Done by you: you're doing all the work. You chat with Claude, Gemini, ChatGPT, and you're copy pasting everything. The entry level. 2. Done with you: you're doing some of the work. Gems, GPTs, projects where some of the work is pre-baked, but a good amount isn't. 3. Done for you: you're doing very little of the work. Claude Code, Cowork, Antigravity, Codex take your project plan and run with it. 4. Done without you: you're doing almost none of the work. OpenClaw, ClaudeBot (the real one!), NemoClaw, Autoresearch, etc. take a broad high level statement and run with it, doing all the planning and research. 5. Done anticipating you: It's a small stretch of the imagination, but once systems have true persistent memory, AI agents identify needs based on your history and just... make stuff and show it to you, like a proud child showing you their crafts. Where we are today is at level 4, but the pieces are already in place for level 5. Memory systems like OpenViking, Serena, etc. already exist and are already in production in places, unevenly. Models like the brand new Minimax M2.7 are agent-first, very smart, and dirt cheap. I'll put a guess up on the board that we'll see full level 5 in a production-ready environment before year's end, if not sooner. Why? Tools like the *Claw family, Andrej Karpathy's Autoresearch, and their kind are perfectly capable of generating their own code. OpenAI said a good chunk of GPT-5.3 and 5.4 were written by their predecessors, so once we give the agentic systems all the ingredients under one roof and enough compute power, they should be able to build themselves to reach level 5 for us. Two key takeaways: first, if you're still in levels 1 and 2, the best time to move up was last year, and the second best time is now. Stop debating em dashes and start trying out agents. Claude Cowork is an easy, non-technical way to get going with agentic AI and see what it can really do. NVIDIA's NemoClaw and Claude Code Channels (released yesterday) are ways to get going with level 4; I recommend NemoClaw with Minimax M2.7 because it's vastly cheaper than Claude models and just as powerful, or Nemotron Super (which NVIDIA will let you use for free for a little while). When level 5 systems become generally available, the gulf between AI fluent and AI laggard will vastly widen. Those who are fluent will accelerate far beyond what we can do today, and what we can do today is already fairly spectacular. It's not a question of budget - Chinese AI models hosted locally or on infrastructure in your jurisdiction are incredibly cheap and crazy powerful, as are other regional models like Mistral Small 4 (came out 2 days ago) or Nemotron. It's a question of investing in your own education. #AI #GenerativeAI #GenAI #ChatGPT #ArtificialIntelligence #LargeLanguageModels #MachineLearning #IntelligenceRevolution

Here are the three ingredients for success with any significant AI project, especially in tools like Claude Code, Antigravity, etc. PRD, spec, workplan. PRD: product requirements document. This contains things like user stories, functional requirements, non-functional requirements, domain requirements, milestones, KPIs, etc. It's the "why". Spec: specification document. This contains things like architecture, models, schemas, quality gates, etc. This is the "what". Workplan: step by step implementation instructions. What can be parallelized. What cannot be. What things have dependencies. This is the "how". If you take the time and have long conversations with your AI tool of choice to build out these three documents BEFORE you get underway, your chances of success are very, very high. If you don't, your chances of success are much lower. And this applies to EVERYTHING, not just code. That next book you're working on.  The course you're designing.  The YouTube series you want to film.  The album you're writing. EVERYTHING benefits from a PRD, spec, and workplan, because very often in the process of building these, you "pre-fail". You find gaps, you determine what you forgot or missed, you find ideas that don't work well together. But you MUST use this with agentic AI. I used to HATE writing stuff like this. I skimped on documentation. But in the age of agentic AI, AI can do the typing. You do the thinking. You have conversations with your regular AI and build all three documents collaboratively. Then you hand them off to your agentic system of choice and you let them do their thing. 9 times out of 10, you'll hit a home run because there will be no ambiguity for the AI agents to fill with hallucations. You hand it the holy trinity of why, what, and how, and it does the labor, in the same way that you hand a construction crew a blueprint, rendering, and building codes and you get a building. If it's worth doing - and especially if it's worth spending significant time or money on, do these three steps first. Your sanity will thank me. #AI #GenerativeAI #GenAI #ChatGPT #ArtificialIntelligence #LargeLanguageModels #MachineLearning #IntelligenceRevolution

AI tools are lazy. This should not be a surprise to anyone. After all, they're trained on our work histories, our discussions, and our examples. They take the same shortcuts we do. They say something is done when it isn't. They cut corners. Here's a simple example: test coverage in coding. When you're using an AI coding tool and you say, "hey, unit testing is important, make sure you do it", it will... and it will say "I've achieved 80% unit test coverage". 80% is okay sometimes. For mission critical stuff, 80% will probably get you fired. Yet if you Google "generally accepted unit test coverage", the results unsurprisingly come back saying 80% is the standard. That's a standard for humans. That's not the standard for AI - or it shouldn't be. AI doesn't have to clock out at the end of the day. It doesn't have PTO. As a manager of AI, it is perfectly acceptable to say, "100% test coverage is mandatory. Anything less than 100% is failure." This is something Katie Robbert and I talk about on this week's upcoming Trust Insights podcast, In-Ear Insights (new episode on Wednesday): the virtual versions of ourselves should be BETTER than us. If we're using AI to take on tasks we do, we should be requiring it to do a better job than us. If we're not holding it to higher standards, then what was the point? Just making faster slop? If you know what success looks like today, what does success look like with the best version of you? Hold AI to THAT standard. #AI #GenerativeAI #GenAI #ChatGPT #ArtificialIntelligence #LargeLanguageModels #MachineLearning #IntelligenceRevolution

"You can't manage what you can't measure." This management trope, widely derided (sometimes correctly), is a cornerstone of effective AI management. Here's what I mean. Suppose you were working on a piece of writing, a piece of content, and you wanted to tell AI to aim for better readability. Maybe you wanted it to be at roughly USA Today, or maybe stretch up to the New York Times. You could try prompting "make this piece readable like USA Today", or even if you knew a bit about readability, you could prompt "rewrite this piece at a 6th grade level". But that's still vague. That's still ambiguous. That's not really measurable, unless you happen to have a 6th grader handy. What if you prompted it with "rewrite this piece for Flesch Kincaid Grade Level 6 and build the necessary code to measure it"? You would get far, far better results. Why? "6th grade level" is ambiguous. Flesch-Kincaid Grade Level is a proven algorithm, invented in 1975 by the US Navy, that machines can measure to: 0.39 * (Total Words / Total Sentences) + 11.8 * (Total Syllables / Total Words) - 15.59 That's the literal Flesch-Kincaid Grade Level forumla. A machine can understand that. A machine can even write simple code to measure that. And when you pair it with a good AI agent like Claude Code, it can programmatically build the code, take your text, rewrite it, measure, and keep tuning it until it hits the required measurement. Whether or not you believe in the management trope for people, the management trope of managing by measuring is how you get far better results out of AI. There's a clear measure of success and a programmatic, deterministic way to perform that measurement, and AI agents can iterate autonomously until they achieve the success measures. So as you build your next AI project, ask yourself what you can measure, and have AI manage itself to those measurements of success. #AI #GenerativeAI #GenAI #ChatGPT #ArtificialIntelligence #LargeLanguageModels #MachineLearning #IntelligenceRevolution

AI should make you smarter. I'm in the middle of updating a deck for my colleagues at SMX (I'm doing a full day AI SEO/PPC workshop next month), and in parallel last night I was trying to have AI generate a Common Core-aligned 6th grade math textbook from scratch, based on Common Core (a USA educational requirements framework) but incorporating the last 15 years of educational design research (which I sourced from 230 academic papers that Gemini Deep Research and Qwen Deep Research). As I got underway reviewing my deck, it occurred to me - hey, I've got this huge pile of research laying around about effective educational design. Why am I not using it with my slide deck? So I dug through the research, pulled out the 8 principles I care about (manually, not with AI because the two domains don't have a ton of granular overlap), and then I handed that off to Claude Cowork to audit and grade my existing deck. I got a B minus. As an Asian, you can imagine how that felt. 😉 The recommendations and changes based on these 8 principles are solid - things like interleaving existing knowledge with new knowledge to help students learn concepts faster, concrete > representational > abstract transition instruction so I don't teach concepts in a vacuum, motivational scaffolding paired with cognitive scaffolding to help students make progress emotionally as well as cognitively, plus many more. These are things I did not know about, because I never studied educational design. But now that I know they exist, now that they're in my vocabulary, I can incorporate them into all my future work. AI made me smarter and better by increasing my knowledge of what I should be doing and taking known, proven research from a domain I don't operate in (formal education) and moving it into a domain I do operate in (public speaking/keynote speaking). There's so much conversation and so many studies and papers coming out about AI diminishing cognitive skills. Many of these are correct from a certain point of view, or from how people use AI today, and that's a shame because this technology can be incredibly transformative when used properly, like the non-medical AI guy in Australia who engineered an mRNA vaccine customized for his dog's cancer and put it into remission using AlphaFold plus the help of the local university's labs. AI should be making you smarter every day, expanding your mind, increasing your vocabulary, opening your eyes to new possibilities. If it's not, then it's time to rethink how you use AI. #AI #GenerativeAI #GenAI #ChatGPT #ArtificialIntelligence #LargeLanguageModels #MachineLearning #IntelligenceRevolution

What's missing from AI outputs is pride. When you do the work, you take pride in your outputs. When you don't do the work, you don't have any stake in the outputs. It's the difference between a meal you cooked yourself - even if you didn't quite nail it - and a meal from a box or a Doordash. Both will feed you. Both, properly done, will satisfy you. But one has your personal pride of creation embedded in it, and the other does not. When you're using AI to create, yes, you are creating the starting seed, but you're not building the output itself. And that's okay, selectively. But if you have a nagging feeling like work isn't as joyful as it used to be (assuming it ever was, of course), then chances are you're using AI for outputs that you previously enjoyed. It's one of the reasons I and many other emphasize that you should hand off work and tasks you don't like to AI first. I take no pride in my expense reports, and I'm glad machines have consumed that task for the most part. I do take pride in writing and making stuff. It's why most weeks, I still manually type out my newsletters (this last week, I recorded it first and then transcribed after). When Katie Robbert and I created the Trust Insights TRIPS Framework for assessing which tasks should go to AI, P - Pain - was a key consideration. Which tasks bring you no joy? Terminator Marie Kondo should take those on for you. But for things that bring you joy? Don't give those away to AI unless you absolutely have to, or give away only a portion so you can not only keep your edge, but also keep your joy. #AI #GenerativeAI #GenAI #ChatGPT #ArtificialIntelligence #LargeLanguageModels #MachineLearning #IntelligenceRevolution

This week, I expand on a recent post about the levels of AI enablement and where the technology is going. Learn the 5 levels, judge where you are, and learn how it's going to impact your world very soon. #AI #GenerativeAI #GenAI #ChatGPT #ArtificialIntelligence #LargeLanguageModels #MachineLearning #IntelligenceRevolution

Want to build a digital twin of yourself? Let's treat it like the experiment it actually is. Observation: Most people attempt to build a virtual version of themselves by starting with the tools. The outputs are universally disappointing. The tool is not the problem. Hypothesis: A functional digital twin requires documented inputs that accurately represent your reasoning patterns, domain expertise, and communication heuristics. Without those inputs, you are not building a digital twin. You are building a generic AI with a name tag. Before you run the experiment, validate your variables: Can you state the specific, measurable problem this solves in two sentences? Do you have a falsifiable definition of success - response latency, output fidelity, client retention impact? Have you accounted for both hard costs (API calls can be surprisingly low, in our own project, approximately $6 per use, but soft costs in documentation time are significant)? The step most people skip: data collection. Your thinking patterns, decision frameworks, domain heuristics. None of that exists in a structured format yet. That documentation phase is not setup. It is the actual build. Everything else is execution. Controls: Purpose before Platform. Every time. If you cannot clearly define the business problem this solves, you are not ready to build. We're walking through the full methodology in our newsletter series, starting this week. What's the hardest part of your expertise to document? That's your most important variable. Subscribe to our newsletter and find out more.

In this week's In-Ear #Insights (the @TrustInsights #podcast) @katierobbert & @cspenn discuss AI and authenticity. Can you automate your social presence and stay authentic? Find out why the answer is a hard no, and discover the right ways to use AI without losing your human voice! #ArtificialIntelligence #SocialMediaMarketing #Authenticity #GenerativeAI

Most marketers treat LinkedIn as a black box. It isn't. The architecture is documented in their engineering blogs and technical papers, you just have to read them. What's actually interesting is using agentic AI to run your own statistical analysis on your LinkedIn data. Pull your exports, run correlation analysis on format vs. reach, topic vs. engagement, posting cadence vs. distribution. Stop theorizing. The signal is in your own data. The algorithm will keep changing. Your analytical framework shouldn't. Full video on our YouTube channel. #LinkedInAlgorithm #AgenticAI #AIMarketing

In this week's In-Ear #Insights (the @TrustInsights #podcast) @katierobbert & @cspenn look at #AI clones. Ever wish you could clone yourself to handle boring tasks? 🤖 Find out if an AI version of you is actually better at your job in our latest episode!

For the nerds - the productivity power move if you're a Google Workspace user: gws. That's the rather blandly named Google Workspace Command Line Interface. This is a free library on Github, made by Google but unsupported, that creates a simple command line interface to Google Workspace services - Gmail, Docs, Calendar, the works. Open up a terminal window and access any part of Google Workspace that you authorize. You might say, "well that sounds awful, and very 1983" and you'd be right. But it's not for you and me. It's for our AI agents, for Claude Code/OpenCode/Qwen Code, for Python scripts, for OpenClaw, for anything where a terminal - a plain text interface - is the most optimal way to work with Google Workspace data. With a tool like Qwen Code/Claude Code/Antigravity/Codex, you can have these tools pick up and use gws to talk to your Google stack and produce useful, helpful briefings, as one simple example. Or respond to client emails programmatically, confirming receipt or providing feedback. If you have a virtual version of yourself and you like to live dangerously, you could even have the virtual version respond. (emphasis on living dangerously) Best of all, there's no additional fee for gws besides what you're already paying for Google Workspace. To start, get permission from your Google Workspace admin, then install it following the directions and set your permissions to READ ONLY to begin - this prevents an AI agent from just going off and deleting all your emails, for example. For most use cases like daily briefings, read only is a great start, and then you can re-authenticate with scopes for writing and deleting things later once you get the hang of how to use it. Tools like gws are the "last mile" of connectivity that we've been waiting for to make AI productive for us at a personal level. Instead of copy pasting everything, as companies create more interfaces for AI, it makes it easier every day to connect more systems to it and derive value from it. Side note: I find it amusing that everyone's using CLIs now because, to no one's surprise, doing everything with pure MCPs consumes a LOT of tokens to reinvent the wheel all the time. A deterministic CLI gets the data, and then hands the clean data off to AI. Efficient, MUCH faster, and uses fewer tokens. #AI #GenerativeAI #GenAI #Gemini #ArtificialIntelligence #LargeLanguageModels #MachineLearning #IntelligenceRevolution

In this episode you'll learn: - What's changed about how LinkedIn works (grab the new paper from Trust Insights!) - How to use the paper's findings in your LinkedIn strategy - How to evaluate your LinkedIn strategic data with agentic AI tools like Claude Cowork Join our weekly #email newsletter: https://lnkd.in/gB6HMkp Follow our YouTube channel for more #marketing and #analytics videos! https://lnkd.in/gCJksQ4 Take our new generative AI course: https://lnkd.in/e2Uhc5wu Subscribe to the Trust Insights podcast: https://lnkd.in/e_ixa6Qz Join our free private Slack community, Analytics for #Marketers: https://lnkd.in/gNRGRzw Trust Insights is a marketing analytics consulting firm that transforms data into actionable insights, particularly in digital marketing and AI. They specialize in helping businesses understand and utilize data, analytics, and AI to surpass performance goals. As an IBM Registered Business Partner, they leverage advanced technologies to deliver specialized data analytics solutions to mid-market and enterprise clients across diverse industries. Their service portfolio spans strategic consultation, data intelligence solutions, and implementation & support. Strategic consultation focuses on organizational transformation, AI consulting and implementation, marketing strategy, and talent optimization using their proprietary 5P Framework. Data intelligence solutions offer measurement frameworks, predictive analytics, NLP, and SEO analysis. Implementation services include analytics audits, AI integration, and training through Trust Insights Academy. Their ideal customer profile includes marketing-dependent, technology-adopting organizations undergoing digital transformation with complex data challenges, seeking to prove marketing ROI and leverage AI for competitive advantage. Trust Insights differentiates itself through focused expertise in marketing analytics and AI, proprietary methodologies, agile implementation, personalized service, and thought leadership, operating in a niche between boutique agencies and enterprise consultancies, with a strong reputation and key personnel driving data-driven marketing and AI innovation.

The truth is more likely in the data. Right now, there's a ton of rhetoric in the media from all kinds of different parties about the reality on the ground in the Middle East. It's difficult to separate truth from falsehoods, what's deep faked, what's propaganda from any given party. In situations like this, one of the ways you can judge how accurate a source is, is by comparing it to other raw data sources. For example, let's look at FlightRadar24, the site that shows flights around the world. As tanker pilots like to say, "nobody kicks ass without tanker gas". You can see where things are happening based on where tanker planes are moving. If a source is saying, "There's nothing happening in XYZ area" but there's 5 KC-135 tankers landing at the nearest airport... that source is probably not reliable. Another example is MarineTraffic.com. With nothing but a free account, you can sign in and see where ships of all kinds are around the world. Pop by the Strait of Hormuz and you can see the massive traffic jams on either side of dozens of tankers and container ships. Anyone claiming things are "back to normal" can be refuted within seconds, looking at the primary source data. And the best part is that almost all the information available you need is available probably for free online, in very boring places like ADS-B exchanges, transponders, SDR shortwave radio, etc. Look for data sources that have hundreds or thousands of inputs (like transponders) because those are harder to manipulate than a news channel or social media influencer. Or if you don't know where to look, ask your favorite AI. As the fog of war gets thicker, ask your AI agents to find you open, free datasets from credible sources and assemble them into dashboards that suit your specific needs. Lots of open source projects liek Crucix and Worldmonitor can unify many data sources, and you can use agentic coding tools like Claude Code, Antigravity, and Codex to further customize them to your needs. As The X Files said back in the 90s... the truth is out there. But Mulder and Scully would have killed to have today's AI agents at their fingertips. #AI #GenerativeAI #GenAI #ChatGPT #ArtificialIntelligence #LargeLanguageModels #MachineLearning #IntelligenceRevolution

So What? Agentic Product Marketing with ICPs In this episode, you'll learn: - Why agentic AI systems like Claude Cowork, Channels, and OpenClaw make closed-loop product marketing - How ideal customer profiles participate in agentic product marketing development and product market fit - How to get started with ICPs and agentic product marketing Thursday at 1pm EST on YouTube Live YT link:

So What? Analyzing Your LinkedIn Strategy with Agentic AI In this episode you'll learn: - What's changed about how LinkedIn works (grab the new paper from Trust Insights!) - How to use the paper's findings in your LinkedIn strategy - How to evaluate your LinkedIn strategic data with agentic AI tools like Claude Cowork This Thursday at 1pm EST on YouTube Live YT link: https://lnkd.in/efPV7nSi #agentic #ai #claude #LinkedIn