sometimes just sounding confident gets the job done

i don't know if a lot of people realised this but anthropic models are so heavily stressed on the data the moment you work with those models outside their training data , like project management and upper levels, they break horrendously, but astra adapts a lot to new problems

alright naming my codex dot as sweetie

expensiveness downtime > cloudwatch > literally every observability solution startup launched in last 10 years you could say cloudwatch single handedly is responsible for 100s of thriving companies

every agent has its own shortcomings and you have to fight with it daily it is a bliss to see claude fighting other agents on behalf of me

before opus 5.5, the few higher points of its performance isn't enough to switch from codex, so never regretted not using claude, but now opus is genuinely impressive but their tooling is so frustrating :(

claude has the best models but they have biggest PITA tooling ever you can run out of disk, you can run out of memory, working with its finicky mcp, like 64GB RAM isn't enough for it, just frustrating

somethings you just feel "how awesome and amazing it is" you have to be super biased to see that as a marketing stunt

awesome problem to be working on right now

tell codex to not delete the ec2 instance and you delete it manually and see it go crazy and question everything in the world

maybe writing tests makes one to question their own internal mental models and help them to iterate on the mental models but at the code level, i still feel like, writing tests is just a means than the end goal end goal is fitting the code to match the mental model one would have about how modules are all getting together, and this might be achieved in a different way, like for example how easy it is now to visualise modules, their flows via AI generated html/mermaid etc,.. so you may not need to write tests to achieve modular codebase, you might be able to achieve that by just visualising the code in different ways and realising that certain modules are just connected in an unnatural way

AI generated code is the perfect example, all the modules AI writes are perfectly testable but the modules AI produces are complete slop, that makes every feature touch literally every module, no concept of reuse at all

xitter.cf/eightnoteight/status/2… having said that, i feel like testing & code modularity being related is a complete myth the more time you spend in code, just means you try to make the code modularity fit your mental model of how building blocks are coming together

#2 is more on the side of writing the kind of code that is only for tests but not for production like injecting time . Now(), sleep() etc,...

i might have cracked the testing problem

in terms of coding, we might have reached a time, where the models got cheaper faster than their accuracy, atleast at subsidised rates

this is like religious leaders seeing a printing press invention and saying only they can interpret the holy texts printing press should not print the holy texts commercial printing indicates total disregard for the religious norms

this has to be the most illogical thing ever

are visualisations for humans to understand ai or are they to make sure ai understood humans correctly