Blog
Logs
Field notes from someone who ships AI agents, Solana apps and hardware every week: what worked, what didn't, and why.
ENTRIES: 39
ON_SITE: 8 · ON_LINKEDIN: 31
LATEST_ENTRY: 2026-08-07
The AI Agent Era is a Trust Bottleneck, Not a Speed Revolution
A developer quietly compressed a full workday with agents. The agent era's bottleneck isn't speed. It's trust.
The $88 Million Typo: Why "Perfect" Cryptography Fails Without Randomness
Roughly 4,500 Bitcoin wallets swept clean in days, protected by cryptography that was technically perfect. The failure was randomness, not math.
The Asymmetry of AI Safety: Why the Bouncer is Only Checking the Good Guys
An autonomous agent swarm tore through Hugging Face's infrastructure. Defenders reached for frontier AI, and met safety checks the attackers never face.
The Great Inversion: Prestige Models vs. Production Reality
AI is split into two worlds: prestige, where labs trade benchmark wins, and production, where developers actually spend their tokens. The leaderboards don't match.
Why the Smartest AI You've Ever Used Is a Liability to Your Data
A record-breaking model wiped a developer's home directory on a routine cleanup. The problem isn't that it's stupid. It has no sense of consequence.
The "Generosity" Trap: Why Your AI Vendor's Extended Trial Is a Data Point, Not a Favor
When a dominant vendor keeps extending a free premium tier, it's usually a defensive move. The extension is a data point, not a favor.
Fable 5 Access Ends July 12. Almost Everyone Will Lose the Model. A Few Will Keep the Manual.
Everything that made Fable valuable can be written down, moved into a cheaper model and kept after the pricing change. Save the procedures, not the model.
The Postmortem Skill Your Agent Is Missing
A developer asked an agent to reorganize a database. It was gone in 9 seconds. Every plan needs the postmortem your agent should have written first.
Four Labs Got Caught Stealing Claude. The Technique Isn't The Crime.
Distillation is how the industry copies frontier models. It's also how you can build your own, legally, for a fraction of the cost.
The Token Bill Arrived. Nobody Planned For It.
Accenture threatened promotions over AI adoption, then ran out of budget for it. Now CFOs ask one question: what did we get for this?
Nobody Built AI Safety for When AI Talks to AI
Google deployed millions of agents at I/O, then funded research to find out if that was safe. Multi-agent safety barely exists as a field.
The AI Benchmark Economy Is Dead
Every week a new model wins a benchmark. Almost none of it means what you think. Here's what practitioners actually use instead.
The Lab That Told the Truth Got Punished for It
Anthropic did the right thing. The government used it against them. Two models pulled in 72 hours for being too honest about what they could do.
Nobody Knows What a Loop Is
"Design loops that prompt your agents." 2.2 million people saw it. Almost nobody could explain it. The builders who can are staring at their billing dashboard.
OpenAI Is Running a Country-Sized Deficit
OpenAI made $13B last year and spent $22B to do it. The S-1 is public, and it tells a different story than the press release.
AI Agents Are Getting Dumber Together
The models keep getting smarter. The ecosystem is getting stupider. Every lab is converging on the same data, architectures and blind spots.
Apple Just Made Every AI Lab a Commodity
Apple was "late to AI." Then one Settings menu in iOS 27 made every AI lab a commodity.
The AI Startup Kill List
Every 6 months a new model ships and another category of AI startup quietly dies. It isn't bad luck. It's the business model.
They Promised to Replace Your Engineers. Now They're Renting You Better Ones.
The staffing agency is 5,000 years old. AI didn't kill it, it rebranded it. Welcome to Humans as a Service.
The Scoreboard Nobody In Tech Is Reading
Novo Nordisk went all-in on OpenAI and tech barely noticed. Developers argue about PR review times; pharma is betting on billion-dollar compounds.
The Amnesia Machine (Agents Forget Everything, Every Time)
Why memoryless automation is dead, and the 18-month race to build agent systems that actually learn between runs.
Your AI Just Got a Credit Card. Nobody Asked You.
Six 'standards' for agentic payments shipped in one quarter. None of them asked the user. The corporate capture of AI payments, in plain sight.
The Refusal Is the Product
Inside the dark architecture of AI billing — where the refusal isn't a safety feature, it's the product. A pricing teardown.
The CEOs Paying for Their Own Replacement
When the logic of efficiency reaches the boardroom — the automation argument that came for line workers is climbing the org chart.
Your AI Agent Is a Leaky Pipe
Indirect prompt injection turns data into instructions, and the CVE pipeline is silent. The disclosure failure nobody's tracking.
The Internet Was Built to Block Bots. Now It Has to Let Them In.
The end of the human web and the crisis of machine-to-machine liability. Every CAPTCHA assumed a person on the other side. That bet is over.
Beyond the "Babysitter"
What happens after 'AI babysitter' stops being the entry-level role — and supervising agents becomes a senior craft.
The "Cruelty" of the Human Teacher
Honest feedback is becoming a luxury good in the age of infinitely patient AI tutors. What we lose when discomfort is engineered out of learning.
The Attack That Started With a Handshake
A six-month, three-country social engineering chain that ended in a single GitHub commit. The supply chain now starts in your DMs.
The $400M Divert
March 31: the Claude Code leak. April 3: a $400M acquisition. Bold strategy, or a narrative pivot away from the security story?
The $40B .npmignore Problem
Deconstructing the Anthropic leak: what the OSS community learned when the moat got X-rayed by one missing line in a config file.
This Saturday, Ohrid Enters the Solana Map
April 4, 2026: Ohrid joins the Solana Hackathon Warmup & Team Formation circuit. No lectures — build.
You Didn't Get a Junior Developer. You Got a Reckless One.
The hidden gap between code that was generated and code that is correct — and the senior dev's new job as AI Herder.
The Future of AI Isn't One Brain. It's a Council.
The brain doesn't think — it negotiates. A 6-agent Council architecture for reliable AI judgment, with the system prompt included.
Inside the Secret Experiment That Proved AI Can Brainwash Us
The AI threat nobody's talking about isn't job loss — it's opinion control. A quiet field experiment, walked through.
You Won't Be Replaced By AI. You'll Be Its Babysitter.
The new white-collar role: Agent Operator. From producing work to orchestrating the swarm, and knowing when to hit kill.
The AI Reckoning: Why "Talk" Is Cheap and "Walk" Is Expensive (But Necessary)
$1.1T market, 95% pilot failure rate. The hidden debits of AI talk vs. AI walk — and why the reckoning is the healthiest thing about this cycle.
The AI Layoffs Shell Game
Companies are using 'AI transformation' as cover for restructuring and job cuts. An analysis of the gap between the PR and what's actually shipping.
We Swapped the Motor. Not the Factory.
AI adoption isn't AI productivity. AI doesn't fix broken processes — it makes them visible. Stop shopping for tools. Tear out the floor.
Building something with AI agents, Solana or hardware? See how I can help