They are good at it when they are explicitly asked. The challenge is that this exercise is recursive at every level of the implementation and often the big picture needs to change based on findings during implementation. Once the AI has committed to a agreed solution, it tends to force it through at all cost and will do truly horrible things to the codebase to make it "work as designed". This is especially true when the AI spins out a subagent to do the work. The subagent has explicit instructions and will almost never push back if it hits a blocker. Instead it will force a solution to the detriment to the codebase.
The problem isn't its coding or its ability to review code... the problem is its ability to apply the right solutions. AI tends to do really well at high level architecture, but then starts to fall apart when there are 1000 mini decisions needed to implement that architecture. Further more, it will absolutely brute force whatever requirements it's been handed from an "authoritative" source. IOW, it's much more likely to be critical of design decision when they originated from its own context window, than those that are handed down from a ticket, design document, or another agent.
it feels like data centers in space only make sense if you're trying to protected them from the general population (mobs with pitchforks when we all lose our jobs)
you can easily run qwen locally in a pi harness (or similar) and just tell claude code to delegate all implementation work to it. claude will even setup the harness with the local model
I like what artcraft is doing with adobe, but a 1-to-1 clone of Salesforce is not as compelling. There are so many bad product and technical decisions baked in... Ideally you'd be building something better from the ground up (linear vs jira).
Also, why target multi-tenant? presumably the audience is a single org.
in theory, a company running custom software should be able to invest many more hours into running those same AI with "white-box" pen-testing (assuming hyperscalers actually start allowing this). An attacker would have to be willing to outspend the company (or get very lucky). If an attacker is resourced well enough to put that investment into owning a single company, you probably don't stand much chance reguardless.
Also, the attacker would have to run black box against the live environment which is much more difficult and likely to be noticed than offline white box.
i did this with opus5.5 this year and it caught an 8000$ mistake. Obviously i should find a new CPA, but i have found it incredibly helpful. One downside is that it will likely also flag a bunch of "by the book" libilities to how your filing that I'm pretty sure most CPA would let slide.
My first experience in antigravity was extremely disappointing.
I asked for a QUICK healthcheck on the sync status of a bigquery integration. It literally spun for 25m before I killed it.
The task, which is CLEARLY outlined in the agent readmes just involved using the gcloud cli to initiate auth, then just run a few queries to check logs and timestamps.
Even 3.6 flash can get this right (most the time).
I wanted to do this with a google home hub so i can run my own fully local "home assisant", but claude rejected it... instead i ordered a used lenovo thinkview device which appetently has a community around custom bootloading...
But it does make you think there might be a real market for unlocked devices like this now that it is trivial to build your own software to do exactly what you want.
I think that sould be obvious without even looking at the code... Like what artcraft did with adobe, its now viable to offer apps completely free by leveraging 100% ai. You could argue the code quality doesnt matter much as long as it works (for local apps security is less a concern).
Also, when building similar style apps, I often don't even start git until the app is in a stable state. The earliest code is often worthless and the overall direction not stable.
reply