I see a lot about adding skills to make Claude test things.
How about making Claude generate scripts that you can run?
Algorithms are better than agents. In fact, the whole point of AI is to reduce thinking to algorithmics. I didn't see a lot of evidence that the team understands this. If you need a magic box to test your other magic box, you have magic boxes all the way down.
LLMs will write a lot of tests by default, but anybody doing any kind of engineering knows that automated suites are just one layer of testing. There is always that manual layer, which LLMs can now help with. Even for hardware, they can control bench equipment and run through tests that can't be simulated (they are simulated, but you'd still want a final test on real hardware).
"skills" is just text files, you just have docs in your repo outlining the manual test procedures which you'd otherwise do yourself.
So I'd say you need both. Scripts plus manual testing, either done by you or the LLM.
Not everyone is using AI to make shittier software. Most big important software projects I know of are frantically fixing an avalanche of AI-found bugs. Important software is getting more robust and resistant to attacks, while still adding functionality.
There's a lot of slop, and it will almost certainly grow quicker than non-slop, but there is a big opportunity for those of us who still work at the bottom to make the pillars more solid, and take on re-engineering projects that would be infeasible otherwise.
Case in point: Before Claude a port of my compiler to a new architecture might take anywhere from months to a year. After AI, now that there are multiple backends as examples and an even larger--massive even--test suite, this can take days or less. It took Claude about 3 days to come up with an arm64 backend for Virgil (around April time frame). Then in August it took just 3 hours to build a complete RiscV backend. Both of them pass the entire test suite. And they shook out other bugs elsewhere. Overall, more targets, and more robust.
Things can have rocket boosters, you just gotta know where to place them.
Every TPU they keep for Google Cloud reduces the need to buy GPUs from Nvidia. Or really, in this market, means they have additional ability to support AI compute beyond what they can buy from Nvidia.
Also, it doesn't just come out and say what the peak throughput of these chips are, just 2.7x the prior generation. This is one of those things that Apple does every Mac release--it's always some big dumb bar that is bigger or some line up and to the right that has absolutely no explanation.
I worked for Google for 10 years. All of my personal email and all of my work email (back then, Google, now CMU) goes through Gmail. They know basically everything that's ever been sent as a receipt in an email. They know what I watch on YouTube, what I search for Google, and Android's little Google rootkit knows everywhere I go, and by extension, everyone I hang out with. They know I install an adblocker everywhere I can. I rail against advertising and tracking on the internet (this site) all the time.
Yet they don't freaking get it. Google has not discovered that I do not like ads. Advertising (mostly) doesn't work on me. I don't buy the shit that is advertised to me. Sometimes, it's the absolute opposite.
...and yet, on the one device I haven't got an ad blocker on, they insist on showing me ads for stuff I WILL NOT BUY. (one specific insurance ad I have been seen for 4+ years--no, no no).
I just don't get it. They expect me to pay for YouTube premium or something, I don't know. I won't pay them.
Google bought my 30s and now they just can't help but nickle-and-dime my mind from here to eternity. It's kind of infuriating. They don't need to know more about me. What they do know about me should be enough!
i worked at google for 18 years. they certainly "know" you don't like ads, and they definitely don't care. they may also "know" your consumer account is a safe place to dump all the non-converting impressions they need to burn somewhere, but even if you could prove that, they wouldn't care. tbh, whenever somebody comes at me with tin foil hat assertions i point to this exact hilarious scenario as why i do not believe they are actually surveillance-oriented business a la the palantirs of the world. i think their ideal state is 100 trillion anonymous bank accounts they can pull from, not 8 billion unique human profiles to care about.
Google's culture is to just burn ungodly amounts of disk space as a kind of flex for how important a team's project is. So every Google service is collecting as much as it can, all the time, with legal counsel and some "oh we value your privacy" sophistry going on.
If you still work for Google I recommend reading privacy design docs for random large scale services if you're able. It's very telling what the priorities are.
And I agree that they mostly don't want to be Palantir-like. They mostly don't care about people. They care about advertisers. Any new targeting mechanism they can offer advertisers they absolutely will. The problem is they sit on a gold mine of very, very sensitive personal data (most of it "anonymized"). Most of it can be de-anonymized with little work. Worse, they will absolutely cough it up to the government when required. They also have a lot of motivated engineers thinking of ways to try to use said data to "improve products". Like the whole saga of doing activity inference for Android users. You don't even want to know about how stupid and creepy that is/was.
I think you're missing the point of ads. They're not (only) there so you click on them and buy something, they're there because it implants the brand into your mind.
"one specific insurance ad I have been seen for 4+ years"
And the one time you need some random specific insurance, what are you going to think about?
I'm starting to realize that trusting some random people who wrote the gruddiest, boringest, most torturous low-level stuff might have been a bad plan.
At this point I feel embarrassed in front of the AIs.
Well, with fewer architecture astronauts overdoing things, all those extra abstractions and dimensions of extensibility that never get used never get written, so one might hope the code is simpler in the long run. But right now coding agents just love writing code, and they don't have a lot of taste for boundaries and architecture, so we have the loop-unrolled, big-pile-of-ifs kind of logic that isn't good.
How about making Claude generate scripts that you can run?
Algorithms are better than agents. In fact, the whole point of AI is to reduce thinking to algorithmics. I didn't see a lot of evidence that the team understands this. If you need a magic box to test your other magic box, you have magic boxes all the way down.
reply