Today my Czech tutor showed me a website she built with Claude to manage her lessons and schedule. Then showed me how she built it and asked for my opinion.
She used to rely on 3rd party. Now shes independent.
Apparently nobody told her shes not supposed to be vibe coding
You need less harness and more clarity on the goal. Make sure you and the agent understand the problem the same way. Agree on the spec. Keep the loops short. Validate with a separate agent. And own the judgment.
Dystopian classics have never been more relatable than they are today. nothing can stop progress though. same as Oppenheimer wasn't able to slow down the atomic bomb. Let them pace the AI until they don't.
gotta love this site
Replying to @max_kelly
its real and way simpler than people think
theres a piezo disc sitting right at the bin inlet. every bit of dust you suck up physically smacks into it. piezo crystals turn mechanical stress into voltage, so a bigger particle hitting harder means a bigger voltage spike
the processor reads that spike train 15,000 times a second and sorts every hit into 4 amplitude buckets. thats exactly what your >10 >60 >180 >500 micron bars are
so its not seeing anything. its an ear, not an eye. its listening to dust hit a wall and inferring size from how loud the hit was
also notice the smallest bucket starts at 10 microns. the stuff that actually gets deep into your lungs is under 2.5. the filter catches that, the counter never counts it
and the real reason the sensor exists isnt the screen, its auto mode. concentration spikes, suction ramps, concentration drops, suction backs off. that closed loop is the actual feature and it saves your runtime. the counter on top is a byproduct they decided to show you
Cursor is the best harness, and their Composer 2.5 is the best coding model for me so far //especially if you still enjoy being a developer. @poteto does a great job, but hey, write your own workflows, it's fun!
the hard part of agent ops isn't the agents anymore. my loop is: run the app, log what breaks in issues.md, let one agent fix it, let a second agent audit the fix. the fixes are usually right. the audit step is what caught my own bad specs.
two model teams shipped the same trick today: a small model routes work, a big model only handles what's hard. cognition cut devin costs 46% doing it. your cheap tier is about to get a lot more capable for the same bill.
openai's agents api means the orchestration glue everyone hand-rolled this year (retry, routing, handoff) is a vendor feature now. the interesting work left is deciding what the agent should actually do.
api pricing looks cheaper until you notice who owns the tokens. if a provider is reselling capacity, your idle minutes are in that quote somewhere. read the pricing page twice before you switch.
Cursor shipped a coordinator agent that directs thousands of subagents. Spawning agents is solved. Deciding which problem deserves a thousand of them is not.
DeepSeek V4.1-Flash beats its own flagship at a quarter of the memory cost. The memory wall is now the model market: we stopped shopping for intelligence and started shopping for GB.
Pretty neat pitch. Love it
I wrote a deep dive into exe.dev
`ssh exe.dev new` and half a second later you have a Linux VM with a persistent disk and a private HTTPS URL
You pay for a pool of CPU/RAM, not per VM. A VM is as cheap as a folder!
flaviocopes.com/exe-dev/
If you can’t define your own dev principles, workflows, skills, hooks, and agentic loops, you’re already falling behind.
Just stop blindly using someone else’s prompt.
—The modal keeps reopening after I close it.
Right. I’ll store a hasClosedModal flag.
—What happens after refresh.
I’ll persist it in localStorage.
—Why is the modal opening automatically at all.
Good point. I’ll remove that behavior
Bağımsızlıktan yana adımlar atanlara sermaye cephe aldı; korumacı politikaları savunanlar itibarsızlaştırıldı. Kendi çıkarlarını ülkenin önüne koyanlar, ‘vatan’ kisvesi altında alternatiflerin sesini kısmak için medya ve siyaseti tam güç kullandı.






