Even if it's a black box (I agree with people talking and sharing but to make a point), the minute you do something that gives you an edge publically, it will be studied and replicated to near perfection. You can see this in a rush defense.
It also happens outside of sport, what's not mentioned is the teams and culture that makes innovations is where the magic is at. By the time you figure out what they've done their already onto the next since they already have all the ingredients to innovate. I'd be willing to bed all coaching staff know what it takes to have that culture but nurturing and creating it is so damn hard.
By the time Tony is on board with AB's all of his knowledge will likely be stale and could potentially be harmful
To me there is something to locality of behaviour with types that improve expressiveness. When there are no types, you have to switch to a caller which could also be dynamic, but with types you know what is expected without additional hops. Imo that makes the token debate even more up in the air, yes it's fewer tokens for input and output but is it really to know all the code paths and use cases?
I had a similar experience, I have a loop that ensures the output stays the same, and a benchmark target, it did improve performance 8x, but when I looked at the code I could easily see more than 20x improvements, however that wasn't what I was focused on, the 8x was enough to not be disruptive to regenerating but I'm also under no illusion that the AI optimised the code well
@thdxr I feel like, before AI it was the classic, the first 90% takes 1 month, the last 10% takes 3 months.
Now it's the first 80% takes 5 minutes, and the last 20% takes 6 months
My tinfoil hat theory is that anthropic knows models aren't a moat, and slop cannoning every vertical to get vendor lockin, and that's also why they don't allow other providers. Because once enterprise buys in, you can't leave without pain. Look at Microsoft and teams. All the features are hype to suck you in, and then when other people build an actual good product, you can't leave.
I can deal with the feature slop cannon, it's all the things that no longer work, and behaviours they change. Dumbest example go to /insights, it tells you to compact more, but I have auto compact on but 4.7 is 1 million tokens, so it won't compact before context poisoning. I go okay cool, I remember plan mode let me reset context and execute plan. Nope plan mode no longer does that.
If I have multiple agents running, I try to have no overlap, I have a bit of a workflow that tells them to wait if a test or compiler error happens from unrelated work and to not fix that. Then I yank completed work to a branch, review and merge that work while other agents are going.
I felt all other options are either too isolated or too chaotic. I do feel like I'm in a bit of a niche workflow, but I haven't found another option where I feel I can see all changes happening and have them work together
@theo Imo this is such an obvious distraction tactic.
"You're going to forget we screwed you over in a week or so, so here's a small something to keep you around until you forget".
It sucks, and can get out of control easily. I've also been working on a greenfield project where I needed to iterate and figure out my foundation.
I managed to clean the mess, by creating stable and unstable internal apis, with ai not allowed to touch stable unless told. I iterate on unstable, and have a slop fest, then meticulously move over parts I am ready to lock in.
If I did that from the start, I think I would have been much further ahead, but untangling the mess took two weeks, and it would have been a similar time to write by hand, however the iterations have saved me months of figuring out what to build
@ZachSDaniel1 Definitely feel this, I've also been a victim of world class gas lighting, where I actually have to step away and think "Am I crazy?", only to come back to find Claude forgot to take its meds again.
@mikehostetler To be fair, we can tell.
I am a bit vendor locked, but Claude code, desktop, design, their platform are all bloated with almost good ideas almost working nicely
Love this experiment. The more I'm using AI, the more I find myself in the driving seat, even when I'm not writing code.
I feel that it's game changing at discovering things, and suggesting possible directions, but falls flat at understanding a design and slotting into that. That's led me to often feeling like I've had a net zero productivity boost.
But working on the same problem, using what it's good at and also putting forward my solutions and being very explicit, I'm starting to see a net productivity gain. The ironic part is I'm more involved than ever, and I'm also not doing more than 2-3 sessions at a time. Where both sessions are likely on the same problem.
It feels like stead and intentional is where the wins are at
All the people I've spoken too that feel really productive have minimal setups and approaches like that.
And agree on the instincts part, the most subtle lesson for me has been what's a clear description to humans might not be one for ai (eg. Intuitive domain assumptions). And also when to commit a lot of focus and oversight. Both come from just practicing and using it, and there's no magic or trick behind it
As a metric, Prs can be useful to show signals of whether there is a good level of productivity but once you go beyond that number e.g. 1 change a day it becomes a vanity metric.
There's too many variables for it to be useful, trunk based dev vs normal.
But also super high pr counts make start asking *are you really thinking about what you are building", "are you doing busy work? I would bet those 100 prs should have been 3 if you understood the problem and fixed it properly
It didn't click until this tweet, I have been using opus and codex daily for rate limit reasons. Most of the time I thought they were the same but recently I had a much better time with codex.
Now I realise it's not codex getting better but opus getting worse in ways I didn't notice
@zeddotdev The three points that would change my life.
- able to copy paste preview
- prettier rendering of preview
- this one might be less straight forward but a nice way to jump between preview and rendering. Kind of frustrating right now especially without copy and paste
@thdxr When my interface depends on the outcome of experiments and ideas, I use agents, when an interface is close to formed, I can the whole project and start over by hand.
I think this model is a bit of a fallacy. It encourages the pirate to not think about the problem, and the architect becomes a glorified garbage man, praying for the day he cleans faster than the mess that's created. In my experience, the real power is understanding your problem well enough to build rigid boundaries where it matters and creating space for experiments and bad code because it might not matter but have no side effects. Embracing that model imo is the way to ship and iterate fast sustainably. Martin fowler did an excellent article on cruft a few years back that's sort of related