About
I found the highest-leverage AI work in a regulated business is not the cleverest agent…
Services
Articles by David
-
A data center isn't an AI thing:
A data center isn't an AI thing:
A data center isn't an AI thing: That's the first move in the whole campaign against them, and it's the one that does…
1 Comment -
The Platform Beats the Agent: Founding an AI Delivery Function at a Regulated Lender
The Platform Beats the Agent: Founding an AI Delivery Function at a Regulated Lender
I spent two months this spring on a contract founding the AI delivery function at a regulated commercial real estate…
3 Comments -
How AGI Reprices Everything Else
How AGI Reprices Everything Else
If AGI arrives within a decade, AI becomes the gravitational center of the global economy — not by replacing every…
3 Comments -
The Israel-Iran Conflict Is Not Ancient. It Was Built
The Israel-Iran Conflict Is Not Ancient. It Was Built
If there is one thing people should understand about the Israel-Iran conflict, it is this: Israel and Iran are not…
-
What AI Knows About Movies (and What It Can't)
What AI Knows About Movies (and What It Can't)
AI can talk about movies like a critic. But it has never actually seen one.
2 Comments -
When does AI have rights?
When does AI have rights?
The Challenges of Testing Whether AI Has Rights Large language models, such as GPT-4/ChatGPT, have evolved to possess…
2 Comments -
In Defense of Generative AI ("AI Art")
In Defense of Generative AI ("AI Art")
Introduction For simplicity, I refer to generative AI images as "AI art." Whether these qualify as "art" in the…
-
The Technosphere War
The Technosphere War
Every day, you are assaulted by billions of agents trying to eat you from the inside out. Over billions of years…
1 Comment
Activity
5K followers
-
David Veksler shared thisThis entire website -- hundreds of pages of original research (!!!) was just two prompts. 1: define concept in Claude Fable 5 2: build in ChaGPT 5.6 Sol (Extra High) Zero characters typed! I brainstormed the whole thing while walking my baby. GPT built everything, configured the DNS and deployed the site to production all on its own. whopaysforai.org
-
David Veksler shared thisOnce upon a time, I listed some of my photos as stock photography. My #2 image is a court room photo. This means that if you search for my name, there will be dozens of articles about people convicted of many, many different sordid crimes.
-
David Veksler shared thisright: the amplifier metaphor, and especially "below 90 it makes you dumber." Amplifiers don't care about the sign of what they scale, so bad judgment gets amplified just as fast as good. wrong: the x-axis. It's not IQ, it's domain competence plus the judgment to catch the model when it's confidently wrong. A 110-IQ senior engineer beats a 140 who's never shipped code, because he can write the spec and smell a bad implementation. The buff lands on relevant expertise, not raw g.
-
David Veksler shared thisI rebuilt my hedge fund's website this morning. Total typing time: under 2 minutes. Four prompts: plan it, build it, deploy it, switch it over. Vellum Capital has run on WordPress since 2018 — plugins, a database, a PHP stack absorbing 500+ comment-spam hits and constant bot probes a week. The maintenance tax nobody budgets for. Now it's a static site behind a CDN. No database, no PHP, no plugin updates — nothing dynamic left to exploit. Eight years of writing, 46 posts, all at their original URLs. RSS feed intact for subscribers. Claude Fable drew up the migration plan. GPT Sol 5.6 built and shipped it. My contribution was knowing what to ask for, in what order, and how to verify the result. Typing the prompt was the easy part. The agents ran for about an hour, and then a human still had to check the work — redirects, feed, URL mapping. That review step is the part you can't hand off yet. My real contribution was the spec and the verification, not the keystrokes.
-
David Veksler shared thisAnyone else loading up the overnight hours with Claude automations? The usage window you sleep through is quota you'd otherwise waste. I run a stack of routines between 1 and 3 AM. What I had to learn the hard way: Routines don't remember your defaults. Every new one, you set model, permissions, and branch from scratch. Pin the model. Otherwise it reruns whatever you used last — for me that's often Fable, at 2x the token rate. Fastest way to torch your quota on a job that didn't need the horsepower. Set permissions to Auto, or the run stalls waiting on an approval that never comes. Branch selection is broken right now. It'll happily run on the wrong branch — a real problem if the job commits its output. Timing: I start no later than 3 AM. The usage window is 5 hours, so 3 + 5 lands a fresh block at 8 AM when I actually sit down. The overnight work never bleeds into the workday. Everything runs through git — code, documents, all of it — and I have Claude commit each job's output. So I always know which routine produced what, and a bad run is one revert away.
-
David Veksler shared thisI told an LLM to remove my personal info from all public databases and now I can't sign legal documents because Lexis-Nexis can't run identify verification on me 😭
-
David Veksler posted thisI've run WordPress sites since it was called b2/cafelog circa ~2001. But the last three sites I shipped have no CMS at all, and adding one would have made every part of them worse: A CMS earns its keep by doing three jobs: an editing workflow, a database, and a theme. With AI, all three now have better answers. The editorial process is git. Every content change is a commit or a pull request. I get review, diffs, rollback, and a complete audit trail of who changed what and why. No revisions table, no admin screen. Merge or don't. The database is JSON and markdown in the repo. coloradofirearmswatch.org tracks every version of the state's firearms list as versioned JSON files. Diffable, greppable, portable. When the data is wrong, the fix is a pull request, not a phpMyAdmin session. The theme is whatever the model designs. This was the last blocker to fall. I used to pick a theme and then fight it. Now AI designs the layout from scratch, iterates on it, and runs QA in a real browser. No theme marketplace, no license key, no CSS archaeology. And the part that actually forced the decision: page builders like Elementor are hostile to agents. They store layout as opaque data in a database, behind a UI built to keep humans away from the code. An agent cannot diff that, review it, or reason about it. Plain files in a repo are the substrate agents are good at. The moment your best co-worker is a model, the CMS is the thing standing between it and the work. So the stack behind the three sites from my last post is deliberately boring: Astro generates coloradofirearmswatch.org, Eleventy generates walletrecovery.info, and https://lnkd.in/gEmtnFRK is static HTML written directly by an agent pipeline from a comnmon spec. Static files on a CDN. Fast and secure -- no PHP, no MySQL, no Redis cache, no plugin updates, no admin login to defend. This doesn't work for every site of course, and databases are far from dead. But this is about what I build now that my collaborators are LLM agents. A CMS is an editor, a database, and a theme engine. Git is a better editor, files are a better database, and the model is a better designer. Links in the first comment.
-
David Veksler posted thisAt 10pm Wednesday I handed Claude an idea I had sketched on a long drive. By Thursday afternoon the site was in production and going viral on Reddit. coloradofirearmswatch.org is a sourced reference for Colorado's SB25-003: every version of the state's firearms list, where each of the 64 counties stands, the litigation tracker. Concept to production in under 24 hours, and I did not write the code. I gave Claude a credit card plus Cloudflare and SSH keys, and it did the rest: scaffolded the Astro site, wrote the scraper agents, wired up CI, deployed to Cloudflare Pages. It is the third site I have shipped this way. walletrecovery.info and https://lnkd.in/gEmtnFRK were also built end to end by agentic AI. It's over 340,000 lines and more than 780 commits. I reused the same workflow every time: 1. Concept. A paragraph of intent, dictated on a drive. "A sourced reference for SB25-003. Field manual tone. No advocacy." 2. Spec. Claude Fable 5 does the brainstorming and turns intent into contracts: data schemas, page inventory, provenance rules, exactly what agents may and may not touch. 3. Implementation. Claude Opus builds the site. OpenAI's Codex handles design iteration and QA in a real browser, on cheaper tokens. This is the part that changed recently: current models can critique their own designs, iterate on layout, and run QA autonomously. Agentic design skill is now real. The models swap in and out because the repo carries shared agent instructions, a CLAUDE.md and an AGENTS.md that any model reads on entry. Frontier tokens for judgment, commodity tokens for construction, and seamless context switching between vendors. What makes the result safe to publish is not the model. It is the guardrails, because hallucinated specifics are the failure mode of AI-built reference content: - Git is the database. Every data change arrives as a pull request. A human reviews merges. Agents never ship without review. - Every fact on the site carries its source URL, the date it was checked, and who verified it. - Agents fail closed. On any anomaly they open an issue and stop. No PR, no silent guess. Anyone can get a model to write a website. The engineering is deciding what the model has to prove before its work ships. Links in the first comment, including the how-it's-built writeup and a governance guide for running agents like this inside your org.
-
David Veksler shared thisMy weekend project publishes more than I do: https://lnkd.in/gEmtnFRK is 151 interactive reference pages pulling about 3,500 page views a day, with 680+ commits since April 2025. No CMS, no writers, no build step. The entire editorial operation is one skill file that an AI agent executes end to end: brainstorm topics against real traffic data, write a spec, build the page, QA it in a real browser, commit, deploy to production. The part worth stealing is not the content generation. It is the harness around it. Model delegation by cost. Claude Fable 5, the expensive model, does the judgment work: it picks topics, writes the spec, and reviews the result. Implementation is delegated to cheaper models, each owning exactly one page, with shared files reserved for the supervising agent. Frontier tokens for decisions, commodity tokens for construction. Guardrails as hard gates. Every number in a spec is an anchor, not a fact: versions, prices, and limits must be re-verified against a primary source before publishing. Nothing ships without a browser-level QA pass, a preview image at exact dimensions, and a commit staged by explicit path only. A closed feedback loop. Cloudflare edge analytics feed a popularity score back into the repo nightly, and the ideation stage ranks new topics against what people actually read. The traffic dashboard is public. The top page alone gets around 1,500 views a day. I still review every diff. Each guardrail on that list exists because an early run earned it. Hallucinated specifics are the failure mode of AI-generated reference content, so the design goal is making fabrication structurally hard rather than hoping the model behaves. Anyone can get a model to write a page. The engineering is the pipeline: spec contracts, verification gates, delegation boundaries, and telemetry feeding back into what gets built next. Same discipline I use for production AI systems, applied to a content site to see how far it goes. Links in the first comment, including the live traffic dashboard and the how-it's-built writeup.
-
David Veksler liked thisDavid Veksler liked thisI’m focused on building mio.money because I believe businesses should be able to own their money instead of renting access to it through unnecessary intermediaries. While I bootstrap the company, I have capacity for one, maybe two, strategic consulting engagements (5–20 hours per week) focused on payments, embedded finance, Bitcoin, and product strategy. Over the past 30+ years, I’ve built fintech platforms and products in highly regulated markets, including the launch of the world’s first Bitcoin-backed Visa® consumer card programs and bringing entirely new payment categories to market. The rest of my time is dedicated to building self-sovereign financial infrastructure. I believe the future of payments is verification, not custody. Businesses should control their own assets, and infrastructure should empower them rather than stand between them and charge perpetual rent for access to property they already own. If you’re building something ambitious and need someone who knows how to turn complex payment ideas into products that survive compliance, banking, interchange, settlement, and scale, my DMs are open. These are the problems I enjoy solving. #FinTech #Bitcoin #EmbeddedFinance #Payments
-
David Veksler liked thisDavid Veksler liked thisTwo new flagship models, Claude on every major desktop OS, and a handful of features designed to make Claude easier to work with across teams and devices. June, in review: Claude Fable 5 Claude Sonnet 5 Claude Tag Claude Design updates Claude Code artifacts Claude Desktop for Linux Cowork on web and mobile Claude Science Chat and Cowork, together A new way to reflect with Claude
Experience
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
Education
-
Texas A&M University - Mays Business School
-
-
Activities and Societies: President of the Texas A&M Objectivism Club
• Focus on information architecture. Studied C, VB.Net, ASP.Net, and Java.
• Built and maintained a successful roommate/apartment matching service for the Texas A&M Department of Student Life -
-
-
-
-
-
Licenses & Certifications
Publications
-
Information Security Fundamentals
See publicationPart 1: Secure Web Browsing
Part 2: Secure Networks
Part 3: Secure Email and IM
Part 4: Secure Operating Systems & mobile devices
Part 5: Secure Organizations
Courses
-
Data Science at General Assembly (9 week course)
2015
Projects
-
MacOS Utilities With ChatGPT
A collection of utilities for MacOS, designed to simplify and automate various tasks. These utilities have been created with the help of ChatGPT, an advanced AI language model.
https://github.com/DavidVeksler/MacOS_Utilities_With_ChatGPT -
CryptAByte
See project"CryptAByte.com is a free online drop box that enables secure (encrypted) message and file sharing over the web using a public-key infrastructure."
Website, API, & desktop apps. -
ObjectivismOnline
ObjectivismOnline is an online community dedicated to bringing together the vast marketplace of Objectivist ideas, commentary on current events, essays written by intellectuals, discussion with Objectivists, and the vast number of other Objectivist websites that have spread across the Internet.
Other creators -
-
CryptAByte
-
CryptAByte.com is a free online drop box that enables secure (encrypted) message and file sharing over the web using a public-key infrastructure.
Messages and files are encrypted using a public key and can only be decrypted using the passphrase entered when your key is created. Your data is never stored in plaintext, and is impossible to decrypt without your passphrase.
https://github.com/DavidVeksler/CryptAByte -
Mises.org
-
The world center of the Austrian School of economics and libertarian political and social theory. The #1 economics website in the world.
Other creators -
-
Crypto-Portfolio-Tracker
-
Languages
-
English
Native or bilingual proficiency
-
Russian
Professional working proficiency
-
Spanish
Elementary proficiency
-
Chinese
Elementary proficiency
Recommendations received
7 people have recommended David
Join now to viewExplore more posts
-
Selina Li
We spent the last 4 months rebuilding the way developers automate. Today, we’re launching Bubble Lab. 🚀 If you’ve ever wrestled with n8n or similar tools: debugging silently breaking workflows, staring at black-box logs, or trying to ship fragile JSON configs into production… you’re not alone. We felt that pain firsthand while building internal automations for our previous startup, gymii.ai. So we fixed it. Bubble Lab is the first agentic workflow builder where every step is real, inspectable TypeScript. Prompt your workflow → get a clean visual editor → inspect the underlying code → export it into your stack. Say goodbye to black-box automation and trial and error debugging, and hello to clean, code-based, agentic workflows you can trust!! Oh yeah and it’s open source. 😄 🔗 Build your first workflow for free: https://app.bubblelab.ai 🔗 GitHub: https://lnkd.in/eyFsgJkd And with Google Gemini 3.0 dropping just hours ago, Bubble Lab already supports it!! #developer #automations #AI #agentic #workflows #bubblelab
138 Comments
Explore top content on LinkedIn
Find curated posts and insights for relevant topics all in one place.
View top content