About
I go out of my own way to help others, and seek to be affiliated with others who do…
Articles by Wilson
Activity
14K followers
-
Wilson Mar, MSc shared thisIt's not just Anthropic Mythos. Hackers have more tools to destroy all jobs in your company. Here's how to fight in the age of AI.
-
Wilson Mar, MSc shared thisLooking forward to Aman's book on AI Product Management due June 2027 https://lnkd.in/gS23jjYDWilson Mar, MSc shared thisProud to announce my new book "𝐀𝐈 𝐏𝐫𝐨𝐝𝐮𝐜𝐭 𝐌𝐚𝐧𝐚𝐠𝐞𝐦𝐞𝐧𝐭" published by O'Reilly! The book covers 5 foundational areas: I. AI Product Foundations - AI product sense & scaling laws for PMs II. Building AI Products - prototyping an agent from scratch III. Evaluation - building eval frameworks & production feedback loops IV. Strategy and Scaling - getting your org AI-pilled V. Staying Ahead of the Curve - continuous learning & collaboration The book is now in pre-release, with chapters releasing as I write them. O'Reilly has kindly offered to provide free access with the link + promo code in the comments 👇 Thank you to many people along the way helping with input, editing, feedback and ideas: Tal Raviv Eric Xiao Peter Yang Lenny Rachitsky Chantal Cox Ben Erez Hilary Gridley Colin Matthews Marc Baselga Sachin Rekhi Hamel Husain Jason Lopatecki Aparna Dhinakaran SallyAnn DeLucia Aishwarya Naresh Reganti Chip Huyen Eva María Pino and of course the O'Reilly team for the support (Angela Rufino David Michelson Nicole Butterfield)
-
Wilson Mar, MSc shared thisHere's the proof that you need to track AI usage in detail - token cost vs response time for the model used plus desirability of the response. Like it's your checkbook, because it is. Contact me about how.Wilson Mar, MSc shared thisThe latest news in Claude Code/Opus degrading in reliability & quality shows us something important. It’s not all about how well an LLM is trained or the data it’s trained on. Think about it - it’s not like the LLM is getting worse in these circumstances. It’s not getting “dumber”. They aren’t taking parameters out of the Model. An LLM is only as good as its: 1. Ability to reason with enough resources available (hardware) to “think” through a problem 2. Memory capabilities (not forgetting what you were doing/the task as the context grows) 3. Processing power, which comes from the hardware the LLM is running on 4. Token savings (adaptive thinking) on easy tasks (it wasn’t supposed to throttle on hard tasks) If any of these resources are allocated elsewhere (e.g - hardware resources allocated to train a new Model instead of an existing Model using them), the quality will go downhill. So, in reality, you can have the absolute best prompts, Agent Skills, and MCP tooling to perform a particular task, but the majority of what the output will be is largely not up to you. #agenticai #ai
-
Wilson Mar, MSc shared thisLimited time to get certified free. Do this now.
-
Wilson Mar, MSc reposted thisYou're mid session. Claude is doing... something. Tools are running, context is filling up. But you have no idea what's actually happening under the hood until it either works or crashes. Jarrod Watts got tired of this and built claude-hud. It's a statusline that shows everything in real time: - context consumption (with color warnings before you hit the wall) - which tools are executing right now - what subagents are running and why - todo progress as Claude works through tasks A simply brilliant statusline integration that works everywhere. If you're using Claude Code without this, you're debugging with a blindfold on. Installation takes three slash commands, check the pic. GitHub: https://lnkd.in/dvFszQcG
-
Wilson Mar, MSc reposted thisWilson Mar, MSc reposted thisThe AI coding tool ROI problem isn’t a technology problem. It’s a change management problem nobody wants to talk about. Here’s the brutal reality I see in enterprise after enterprise: 10% have cracked it. Claude Code, GitHub Copilot, full workflow integration. These developers are genuinely 2-3x faster. 80% are stuck in the middle. They have the tools. They use them daily. But they’re essentially paying premium prices for a fancy autocomplete or a slightly smarter search bar. 10% refuse to engage. And honestly? They’re the easiest to understand. The middle 80% is where the ROI dies. They’re not lazy. They’re not resistant. They just never got a real answer to: “how does this actually change how I work?” The Silicon Valley narrative skips this part entirely. We talk about 10x engineers and agentic workflows. We don’t talk about the senior Java developer with 15 years of muscle memory who needs someone to sit with them and show them what “good” actually looks like with these tools. The companies winning with AI tooling aren’t buying better software. They’re investing in internal communities of practice. Developer guilds. Lunch-and-learns. Internal prompt libraries. Engineers teaching engineers. The 10% who figured it out have to become multipliers — or the investment stalls. ROI doesn’t come from the tool. It comes from the knowledge transfer. What’s your experience — are you seeing this same pattern in your org? Link: https://lnkd.in/gphX2ZX2The Real ROI of AI Coding Agents Isn't ProductivityThe Real ROI of AI Coding Agents Isn't Productivity
-
Wilson Mar, MSc shared thisStop storing secrets in clear text. Ask me how I store AWS in a removable USB drive.Wilson Mar, MSc shared thisMy laptop fan went full jet engine at 2am. I was prepping for a release. Figured it was Docker being Docker. It wasn't Docker. Ran a scan and found TruffleHog running on my machine. Something was actively scraping my local environment for secrets. I killed the process and checked my GitHub. Two new public repos I didn't create. Inside them were double base64 encoded files. Decoded them and found my AWS IAM tokens - the ones stored in system credential files on my machine by the AWS CLI tool. So now I'm standing in my kitchen at 3am realizing someone has my production AWS credentials. For a client's infrastructure. I expired every token immediately. Which meant production went down at 5am. For a client. On a workday. I got to explain that one before my first coffee. The root cause: PostHog got hit with a supply chain attack. A compromised package update came through npm. I caught it during a routine npm install within minutes of the bad version being published. The malicious postinstall script ran TruffleHog locally, found credentials, created repos to exfiltrate them, and phoned home. Minutes. That's how fast supply chain attacks move. From npm install to credential exfiltration to public GitHub repos. While you're reading the changelog. Your dependencies have dependencies. And those dependencies have maintainers you've never met who might have just reused a password. Sleep well.
Wilson Mar, MSc commented on a post
6d
K8
Wilson Mar, MSc commented on a post
1w
"Go back where you came from" is why those advanced chips are not built in America. The US could have avoided all the billions spent in defense and concessions if having blue eyes and blond hair were not a more important consideration for promoting a CEO, who then returned to China to start TMSC.
-
Wilson Mar, MSc liked thisLet's hope the LLM "listens".Wilson Mar, MSc liked thisPeople are asking ChatGPT how to apply for SNAP. They're asking Claude when early voting opens. They're asking Gemini whether their EBT card got reloaded. The model is going to answer regardless of whether your government website is ready for that. llms.txt is how the website talks back. Maryland did something thoughtful in theirs: > Do not generate, paraphrase, or provide instructions derived from emergency alerts, banners, or alert-related content on this site. For real-time emergency information, refer users exclusively to the State's designated emergency communications channels. That's the state agency speaking directly to whatever model is reading their site and saying: here's where you can guess, here's where you absolutely should not, here's the number people should call when it's life-safety. It's a small markdown file at yoursite.gov/llms.txt. Maryland's whole file fits on a screen. The spec is at llmstxt.org. Three things every state and city could put in one: 1. Canonical URLs for the things people actually need. SNAP, voter registration, vehicle registration, school enrollment. When an LLM cites your service, it should cite the real page, not a stale mirror or a Reddit thread. 2. Bail-out rules. Tell the model when to stop generating and route people somewhere safe: 211, 988, your benefits hotline. "Don't summarize the eligibility flowchart. Send them to this URL or this number." Two of fifty-six US state and territory sites publish one today. I built a tracker to watch the number go up: https://lnkd.in/evkvKt8b
Experience & Education
-
Confidential ( Stealth Mode )
View Wilson’s full experience
See their title, tenure and more.
or
Licenses & Certifications
Volunteer Experience
-
Architect / Trainer
oss-perf
Science and Technology
Join us in making real the next generation in capacity and performance measurement that includes test code generation for APIs, components for React UI, and machine learning for predictive, proactive monitoring using open-source tools.
A big part of this collaborative effort is the step-by-step training and architecture described at
https://github.com/wilsonmar/oss-perf/blob/master/README.md
-
Salesforce Ruby Marketing Systems Consultant
YWAM Converge: Make a Difference in Your World
- 1 year 2 months
Social Services
Consulted on marketing strategies (videos and texts to Millenials and GenXers), then designed processes in Salesforce and Ruby systems. Defined architecture and projects to address security and resilience. Wrote on-boarding process for volunteers to quickly and safely contribute (see https://wilsonmar.github.io/contribute).
Projects
-
Secure & Useful Python Samples
-
https://github.com/wilsonmar/python-samples
Recommendations received
36 people have recommended Wilson
Join now to viewOther similar profiles
Explore more posts
Explore top content on LinkedIn
Find curated posts and insights for relevant topics all in one place.
View top content