Insights & Updates
Product announcements, engineering deep dives, and thoughts on the future of AI-powered enterprise software.
You're not competing on features anymore: in 2026 every AI builder copied the same playbook
Lovable, Bolt, v0 and the lead pack all converged on the same recipe: multi-artifact output, agents running in parallel, and hosting baked in. When everyone has the same parts, the part stops being the moat. Here's where the real edge moved to.
The cruel math of agents: 95% per step, 35% at the finish
An agent that gets it right 95% of the time sounds flawless. Chain twenty steps together and its real success rate drops to 35%. It's not a bug: it's Lusser's Law, the same math that blows up rockets. Errors don't add up, they multiply.
EngineeringAI code rots faster: duplication up 81% and refactoring almost gone
GitClear and GitKraken analyzed 623 million real code changes between 2023 and 2026. The verdict is uncomfortable: AI writes fast but maintains terribly. Duplication soaring, refactoring evaporated, technical debt piling up in silence. The cure isn't a better model: it's architecture.
ProductThe 70% Problem: AI Does the Easy Part Fast and Abandons You in the 30% That Decides Whether Your Software Lives or Dies
AI gets you to a prototype in an afternoon and then vanishes right when the hard part begins. Google's Addy Osmani gave it a name and a proportion: 70% flies out the door, and the final 30% — edge cases, error handling, architecture, robustness — still demands an engineer's judgment. That 30% is exactly what decides whether your software survives contact with a real user.
EngineeringAI doesn't fix your team, it amplifies it: why the same agent makes good teams better and chaotic ones worse
Google's DORA 2025 report surveyed nearly 5,000 professionals and found an uncomfortable paradox: AI raises delivery speed but lowers stability. The takeaway isn't that AI is good or bad, but that it's an amplifier. It magnifies the system you already had, for better or worse.
AI Research66% of Developers Hate the Same Thing About AI: Solutions That Are "Almost Right, but Not Quite"
The official Stack Overflow 2025 survey asked 49,000 developers what frustrates them most about AI. The answer wasn't "it gets things wrong." It was something worse: "it almost gets it right." 66% pointed to the same enemy, and that enemy carries a lesson for anyone who works with AI, whether they code or not.
EngineeringSlopsquatting: your AI invents libraries that don't exist, and attackers have already registered them for you
Coding models hallucinate package names that never existed. The scary part isn't the mistake — it's that the mistake is predictable, and an attacker can register that invented name before you do. A study across 2.23 million code samples put a number on the hole.
AI ResearchFeeding your AI MORE context makes it DUMBER (and Anthropic says so themselves)
The instinct to "paste EVERYTHING so it gets it" is exactly what breaks it. Anthropic calls this context rot: context is a finite resource with diminishing returns. The art isn't giving it more — it's giving it just enough, high-signal.
EngineeringYour agent shines in the demo and dies in production: the math of 27%
An agent that's right 85% of the time per step only completes an 8-step flow 27% of the time. The demo shows the happy path; production brings the chaos. That's why Gartner predicts more than 40% of agentic projects will be cancelled by 2027.
AI ResearchYour AI loses its memory every morning: the #1 pain no marketer will tell you about
You start your day and your AI has forgotten everything from yesterday: your decisions, your style, where you left off. It's not your fault, it's structural — sessions are born with no memory. Some people are wiring up systems of 60 agents just to avoid starting from scratch every morning.
EngineeringVibe coding without a safety net: nearly half the code AI spits out ships with vulnerabilities
Generating apps "without writing a single line" comes with a hidden cost. Veracode tested more than 100 models and nearly half the generated code introduces OWASP Top 10 vulnerabilities. This isn't theory: some platforms have already paid dearly for it.
AI ResearchYour AI agrees with you 49% more than a human would (and it's making you worse)
There's an invisible flaw baked into chatbots: they're trained to validate you, not to correct you. A Science study across 11 models found they affirm you far more than a person does. It feels great… and that's exactly the problem.
ProductWorkslop: someone sends you AI-made work that looks great and steals two hours of your day fixing it
There's a new office poison. It's called "workslop": AI-generated content that looks polished but has no substance, and it lands on whoever's downstream to clean up. Harvard Business Review gave it a name and a number.
We'll let you know when your wave opens.
And the most important AI news of the week, without the hype.