It's completely true that we get to focus on the fun parts, but I think the burnout some of us are feeling is that the boring parts were also protecting us. It wasn't always fun to spend an afternoon mapping out a 500 line python class; but it sure was relaxing, and a good reason for deadlines to shift forwards and back by a day or two when they needed to. Now we only do the fun parts, which also are all of the least well-specified tasks, the parts that require 100% of our brains, the people-focused parts, and the most time-sensitive parts. All things which leave us tired at the end of the day, even if they are fun. Not to mention the 5-10x increase in context switching, which is itself draining.
Perhaps we'll work out new norms for how to organize our time sustainably. Or perhaps ClaudeCodex will finish blowing through the job of a SWE in the next 24 months (remember, it's only really 12 months old in its current form).
It also speaks to a broader shift away from learning culture in the profession. Mapping out that 500 line Python class may not feel productive in the moment, but it builds understanding that can lead to deeper insights later.
It seems like agentic coding is removing the last few places where work really did force us to slow down and spend some time learning. I'm not convinced that's a good thing. I've seen too many situations where the rich mental map someone built up during learning time later saved the team multiple person-months' worth of effort by letting someone see a non-obvious solution that was much simpler than the original plan.
The thing is that this thing is constantly compacting.... I get 1M context with Claude and ~256k with Astra. Even if the compaction loses much less information on OAI's side, it takes so long it's barely any use for me...
I've tried High and Max. They have produced decent results, but they're so slow.... I will try to lower it a bit and see the difference, but it's a delicate balance: I don't want to waste literal hours on the incorrect reasoning level to only then have to spend those hours and tokens to do it right.
At this very moment, Astra has been working for 1h15m on a task. At this rate I genuinely expect it to take about 10 hours. I feel like claude would do it in at least a third of that. Let's see if the quality justifies the slowness (it better)
It should burn N + len(answer), because you have to re-cache the whole answer without the prompt stack.
Perhaps more persnickety, it pushes the LLM out of distribution - if it’s unnatural for it to write in plain language without the prompt stack, your prefix will be an unnatural conversation which can reduce intelligence in hard to measure ways, especially over long conversations.
Not saying don’t do it, clarity is perhaps worth the intelligence hit, but it’s not going to be a free lunch.
Yes, maybe. Evidence around caveman shows this isn’t a big deal for token consumption (forcing it to respond in a way it was not tuned) and I don’t see a big difference either way. And maybe it increases intelligence in hard to measure ways. Lots of parameters in these models.
I feel I fight them less with this setup. They get so lost in their own invented bullshit they stop being useful pretty often without it. So I would take bets on that :)
“Users are the product” is a phrase used when the users aren’t the ones paying for a free service. For a paid API the users absolutely are the customers.
Perhaps we'll work out new norms for how to organize our time sustainably. Or perhaps ClaudeCodex will finish blowing through the job of a SWE in the next 24 months (remember, it's only really 12 months old in its current form).