Ryan Mitchell

"Graphs make provenance first-class. Instead of burying "where did this come from?" inside a dense vector, a graph records sources, activities, and agents as nodes and edges you can query..." https://lnkd.in/efAdV9dx

26

I don't think AI has killed SaaS (yet) but I do wonder if it's killed "value-based pricing" I've been doing a lot of software shopping the last month, particularly in the conversational AI space, where there's a Cambrian explosion of vendors figuring out business models in real time. You can find it all: conversation-based, outcome-based, platform + usage-based, and tier-based tied to some specific platform feature. Overall, "value-based" is the framing I'm most allergic to these days. Time was, a vendor could win with a high sticker price by anchoring to the outcome or business impact the solution would drive. This is generally a sales best practice. But at a time where the line between build and buy grows ever thinner, when customers have built their own agents, understand inference costs, and have seen how thin some product wrappers actually are -- value-based feels untethered. The age of AI doesn't mean I'm going to build every tool in Claude Code. But higher buyer literacy DOES mean that the relationship between your sticker price and your underlying cost-of-goods-sold is going to get more scrutiny. I'm happy to rent your application layer (and the expertise it encodes) and pay a reasonable cost for maintenance and support. But charging a disproportionate premium that's not moored in variable costs like inference and compute seems increasingly hard to justify. If your pricing can't survive a conversation with a buyer who's been tinkering in Claude Code for six months, you have a problem.

59

17 Comments

This looks useful for folks in research or scanning what's coming out on the frontier: Lumi from Google's PAIR research team, a tool for better understanding papers on arxiv. Reading smarter, not harder -- that tagline brings me back to Mortimer Adler's classic 'How to Read a Book'! From their GH repo: Lumi uses AI to help you quickly read and understand arXiv papers. Features include: ✏️ AI-augmented annotations - read summaries at multiple granularities 🔖 Smart highlights - highlight text + ask questions 🖼️ Figure explanations - ask Lumi about images in the paper Demo: https://lnkd.in/ex7erTBb GH repo: https://lnkd.in/emz_Ghcs Medium article: https://lnkd.in/enky8dKC GitHub discussions: https://lnkd.in/efAmiCcp

19

Putting a knowledge graph into an LLM is usually not feasible. However, you can put an ontology (structure + taxonomy + rules) that then allow you to query (and update) a knowledge graph that in turn can drive LLM content. In effect, the SHACL becomes your contract between the language model and the knowledge or context graph. This is a long article, but I cover a lot of ground here, including why I believe that such an approach can both dramatically improve accuracy and provide a working environment for dynamic data that can better feed the LLM in its role not as database but as transformer. Thoughts on Queens and cabbages and sailing ships, or at least the role of Steampunk as a programming aesthetic, all of course in … The Ontologist!

148

19 Comments

Another day, another LLM vulnerability: The team at Anthropic (the folks behind Claude) showed that a small number of samples is all it takes to poison an LLM of any size. > "As few as 250 malicious documents can produce a 'backdoor' vulnerability in a large language model—regardless of model size or training data volume. […] Even though our larger models are trained on significantly more clean data, the attack success rate remains constant across model sizes." What this means in practical terms is that large language models can be fairly easily backdoored; all it takes is a small stash of malicious documents in the training set. As AI companies are gobbling up data left, right, and center, it is close to impossible to ensure training data isn't tainted. ↗ https://lnkd.in/ggiD6VHn

12

3 Comments

AI agents are dominating the 2025 tech landscape with so many options available. This post breaks down the strengths and weaknesses of the some of the widely used agent frameworks so that others can make informed decisions Key Takeaways: ► OpenAI Agents SDK – Best for simplicity and production-readiness, with a low learning curve and built-in primitives like Agents, Handoffs, and Guardrails for structured workflows. ► LangGraph – Ideal for complex, graph-based workflows but has a steeper learning curve. ► CrewAI – Most intuitive for role-based multi-agent systems, but lacks scalability and is still evolving. ► AutoGen – Supports diverse conversation patterns and agent topologies, but documentation is unclear. Link: https://lnkd.in/gr6pVhJ2

13

AI & Market Research - Can AI take surveys for us? I haven't posted about too many uses of AI with Market Research to date, however I think this case study should get a mention. A recent study by PyMC Labs and Colgate‑Palmolive finds that large language models (LLMs) can closely simulate human responses in consumer product surveys—hitting ~90% of human test–retest reliability across 57 surveys (9.3k human responses). How it worked: Instead of asking a bot for a number, the team asked for a short written answer with a why. That text was then compared to five Likert “anchor” statements using embedding similarity (their Semantic Similarity Rating / SSR method). Result: 90% of human test-retest reliability. They mirrored distributions of responses across demographic segments such as age and income. Researchers emphasise that this doesn’t mean LLMs think like people, but their outputs can align with human behavioural patterns under defined conditions Want to try this? 1) Start with a low-risk, high-volume concept test you already run. 2) Write clear Likert anchors (1–5) as plain sentences. 3) Prompt an LLM for a short rationale per concept and then map that text to anchors using embedding similarity. 4) Create a small human holdout to benchmark; only proceed if your synthetic vs human gap is tight. 5) Be transparent: label synthetic vs human, document prompts, and keep humans in the loop for final calls. Imagine being able to test new tourism or retail products or ad concepts instantly across synthetic “audiences”, before spending on full panels. Huge potential for insight and speed — as long as we keep the human layer where it matters most: judgement, context, and creativity. This isn’t about replacing people; it’s more about "simulate → iterate → validate" so teams can test more ideas, faster, and spend human budget where judgement matters most. Full story below 👇 #AI #MarketResearch #Insights #DigitalTransformation #SyntheticData #Tourism #Innovation https://lnkd.in/eRBWtRcs

6

1 Comment

Here's a few things I found after playing with commercial and transactional prompts on Perplexity Comet for ~3 hours: > Agents prefer websites with streamlined access to data, and that preference seems to be hardcoded to some extent. In doing 'research + add to my cart' agent shopping tasks, Comet consistently started out by navigating to one retailer's website, even when I stated I preferred it's direct competitor! > One reason for that might be the quality and consistency of the website's data structure. I am thinking back to some of the monster ecommerce retailer sites I have seen. The ones built on a homegrown CMS in 1996, with another CMS bolted on in 2002. New product categories and experiences are added to the site in a way similar to the Weasley's house in Harry Potter. If agentic commerce takes off and requires consistency in data structure, those sites are going to be cooked. > There are at least two clear agentic user flows that brands need to consider. One is a research flow where the agent will go off, do research, and come back with a comprehensive recommendation that mentions many brands and cites many webpages. This is just what SEO will become. > The other is a true virtual assistant flow where, if you know what you want, the agent will find it for you and present you with a filled cart. For these user flows, you may have a customer who's at the finish line ready to convert and gets snatched away at the last second because the virtual assistant found a better price, better stock, or (worst of all) just had an easier time navigating a competitor's site. > Shopping feeds are going to be critically important. If you're an ecommerce brand and your shopping feeds are jacked up, you are not going to have a good time acquiring customers on search. It's easy to say "who cares about Perplexity Comet, they have no market share" but I would caution against that mindset. Google is laying the foundation to enable the same experience in their ecosystem. Comet should be the sandbox you test in to prepare for Google's agentic browsing capabilities.

17

5 Comments

📝 Blogged: "You Gotta Push If You Wanna Pull" Wrote up some thoughts on push vs. pull queries, materialized views, and why data duplication isn't something to fear. (Link to the post in the first comment below 👇)

46

2 Comments

Long Live Spark, Down with Spark 🙃 Spark has been the de facto processing engine for the last 10 years in the "Big Data" community. It still has its place in data processing. Some have branched off their own versions, spun off custom engines, and developed new approaches in an attempt to overcome the shortcomings of spark in diverse landscapes. Our focus is to enable the right compute for the problem with a focus on interoperability with other systems and to have open source compute engines as first-class citizens in platform. Enterprises are highly complex, with high heterogeneity of solutions, tech debt, data gravity, and modalities. I would argue that most pipelines in enterprise are not web scale; they are small and medium sized. A pipeline with a few million or hundreds of millions of rows, sometimes a few billion. Spark is overkill for 90%+ of these workloads and carries a lot of overhead that impacts speed and cost. With the Palantir Multi-Modal Data Plane (MMDP), we provide the framework to store data anywhere, compute data anywhere, and use any model. If we break down this a bit more, this means I can have data spread across different platforms, file formats, structures and use any compute (inside or outside of the platform) to process my data. The permutations of storage locations, formats, and compute engines are massive. It's the right tool for the problem at hand, not one tool you need to shove everything into. Re-platforming should never be the goal. With new frameworks like Polars, DataFusion, DuckDB and the availability of larger node sizes that this means you can run increasingly bigger and more complex workloads in a fraction of the time and cost. You can mix and match these as needed in Foundry and AIP, even having one step run one engine and another step use a different engine. (You can even bring your own engine.) Spark is always there on the table, but it should no longer be the default for most enterprises to start with.

232

24 Comments

Stop calling everything an AI "Agent." Most "Agents" people talk about are just chatbots. You give it a prompt, it gives you text (you've seen it before). That’s a linear path, not an agentic one. Real agents operate in a loop. I’ve broken down the architecture below based on the latest framework logic. If you want to build Agent systems that actually move the needle, we need to be damn sure we understand what an Agent is. Here is the 4-step breakdown of how a real AI Agent actually functions: 1. 𝗧𝗵𝗲 𝗜𝗻𝗶𝘁𝗶𝗮𝗹𝗶𝘇𝗮𝘁𝗶𝗼𝗻 (𝗧𝗵𝗲 𝗣𝗲𝗿𝘀𝗼𝗻𝗮) - 𝗧𝗼𝗽 𝗹𝗲𝗳𝘁 𝗼𝗳 𝗱𝗶𝗮𝗴𝗿𝗮𝗺 It starts with a user message. But the Agent doesn't just reply. It initializes with specific instructions. This is the "Identity." It defines the boundaries of what the agent can and cannot do. 2. 𝗧𝗵𝗲 𝗕𝗿𝗮𝗶𝗻 (𝗧𝗵𝗲 𝗟𝗟𝗠) - 𝗧𝗼𝗽 𝗿𝗶𝗴𝗵𝘁 𝗼𝗳 𝘁𝗵𝗲 𝗱𝗶𝗮𝗴𝗿𝗮𝗺 The Agent sends the request + context to the LLM. The LLM is acting as a Reasoning Engine. It decides: "Do I have enough info to answer, or do I need a tool?" 3. 𝗧𝗵𝗲 𝗛𝗮𝗻𝗱𝘀 (𝗧𝗼𝗼𝗹𝘀 & 𝗠𝗖𝗣) - 𝗜𝗻𝘀𝗶𝗱𝗲 𝘁𝗵𝗲 𝗯𝗹𝘂𝗲 𝗯𝗼𝘅 This is where the ROI happens. If the LLM decides it needs data, it triggers a Tool Call (via the Model Context Protocol or similar). The Agent executes the function (e.g., searching a database, checking a calendar). The Tool returns the result. The Agent feeds that result back to the LLM. 4. 𝗧𝗵𝗲 𝗟𝗼𝗼𝗽 (𝗜𝘁𝗲𝗿𝗮𝘁𝗶𝗼𝗻) - 𝗜𝗻𝘀𝗶𝗱𝗲 𝗼𝗳 𝘁𝗵𝗲 𝗯𝗹𝘂𝗲 𝗯𝗼𝘅 This is the "Agentic" part. The process repeats until the task is complete. The system doesn't stop until the logic is satisfied. It’s a self-correcting cycle of: Reason → Act → Observe → Repeat. So, whats the difference between what most are doing now, and what an Agent actually is?: 𝗦𝘁𝗮𝗻𝗱𝗮𝗿𝗱 𝗣𝗿𝗼𝗺𝗽𝘁𝗶𝗻𝗴: 1 Human + 1 Prompt = 1 Output. (Efficiency: 1x) 𝗔𝗴𝗲𝗻𝘁𝗶𝗰 𝗟𝗼𝗼𝗽: 1 Human + 1 Goal = 10+ Autonomous Iterations. (Efficiency: 10x+) Charles Lamanna gives an interesting take that in 2026, more people aren't just paying for an LLM to 𝘬𝘯𝘰𝘸 things anymore, but to build systems that 𝘥𝘰 things instead. P.S. We break down concepts like this in the Power Hour Newsletter, subscribe for free here: https://lnkd.in/gsKBUaM7 #AIAgents #LLM #SoftwareArchitecture #Productivity #AI

74

8 Comments

Most people know Claude Chat. But very few know that Claude also has Claude Code and Cowork. If you only use Claude Chat only, you are using just a small part of what Claude can do. Here is the clear difference: 1. Claude Chat Claude Chat is the standard conversational AI interface used for thinking, writing, and analysis. You give a prompt and Claude responds in the chat window. use cases: • A SaaS marketer generates 10 LinkedIn post variations from one idea and picks the best one • A product manager pastes a 30 page PRD and asks Claude to extract features, risks, and dependencies • A founder converts meeting notes into a clear action plan with tasks and owners • A content team turns a blog article into a LinkedIn thread, email, and short summary When to use: • Research • Writing • Analysis • Quick problem solving 2. Claude Code Claude Code runs inside your terminal as an AI coding agent. Instead of answering questions, it modifies files, writes code, and executes tasks across a repository. use cases: • A developer asks Claude to refactor a legacy authentication module across 40+ files • A startup team asks Claude to generate unit tests for an entire API layer • A security engineer runs Claude to scan the repository and flag vulnerable packages • A backend developer asks Claude to convert synchronous code to async across the project When to use: • Large codebase changes • Automated development tasks • Code reviews and testing • Engineering automation 3. Cowork Cowork is a desktop AI agent for non developers. You describe a task and it performs actions across files and apps. It can access folders, documents, and connected tools. use cases: • An operations manager asks Cowork to sort 500 downloads into folders by file type and project • A finance team converts hundreds of receipt PDFs into a structured expense spreadsheet • A content team asks Cowork to rename and organize media files for a marketing campaign • A research team compiles multiple reports into one structured summary document When to use • File organization • Document processing • Repetitive office workflows • Multi app tasks Quick summary • Claude Chat → thinking and writing • Claude Code → coding and engineering tasks • Cowork → automating office work Most people only use the first one. But the real value comes when you use the right mode for the right job. If this post helped you: • Save it for later • Repost to help others • Follow Shubham Vora to learn how to use AI in your day to day life and business.

109

31 Comments