Building software in the AI revolution and sharing what I learn about staying creative, product and valuable.
Previously Atlassian, Trello, Ubiquiti Networks.
lol...
Ah yes! mandated restriction on bottoms up demand for a good or service due to spend that far out paces what was budgeted for / can be absorbed, the first horsemen of every bubble pop. 🤦♂️
While I do think this is the default path and is one of the contributors to why we aren't seeing real productivity scale at the same time code generation output explodes...
It also, doesn't have to be the case. It is fully possible to code with an agent and come along the for the ride on understanding and skill uplevel. It requires intentional effort and care. It's the new engineer craft to be plugged in at the right level of abstraction when the agent will happily let you tune out if you so please.
When your engineer finishes a session with a coding agent, they have more code than they started with and exactly the same amount of skill.
The agent produced output. It did not produce understanding.
You paid frontier prices to make your codebase bigger and your engineer no smarter. That is not an accident. It is a design choice, and it is worth remebering who it serves.
@FBallAnalysisYT The thing most people are missing is that you can sell positions on these platforms, so this is just a bet on further downward movement in the odds. Not a 9 month 0.8% investment.
@sflorimm ~3-5 avg. That's not counting the sub agents that get spun up automatically. Those are the ones my brain tends to. I can and have ratcheted up to 10 for a day or two at a time as an experiment but my real leverage/impact goes down at about 5.
Dear @ClaudeDevs / @bcherny / @_catwu , I love your product and am an hourly active user. I love that your product velocity is blistering and you throw things at the wall to see what sticks.
Respectively, the decision to overload the word "workflow" as a reserved word is a WILD product decision. At least in my work, "workflow" is a very frequently used word and is NOT a word I would naturally use to invoke this new fancy feature.
Thankfully the model/harness has been good about understanding my intent and not summoning the agent army. However, it's really annoying to see these bits of reasoning show up every few turns.
All of you out here still saying "[X] SOTA foundation model is So mUCh bEtTEr than [Y] SOTA foundation model" clearly do not understand the ecosystem.
I'll give you that you might be more comfortable with it. You might now how to prompt it better. You might have skill files you tuned against it, etc...
But the models themselves are WILDLY similar and increasingly so. Every well-formed benchmark proves this out.
I'm not saying there aren't differences... Subtle flavor nuances that impact workflows. I'm just saying 95% of X > Y the takes I'm seeing in the last month are massive hot air.
@zachtratar 💯 I've found there are some new traps when the models write meaningless performative tests that can further intrench the wrong patterns BUT if you can resist the temptation to get lazy about the middle D, it feels like the true fulfillment of the TDD promise.
Angie is 100% correct
(Yes, encouraging usage did make sense early on, aka last year, when there was both resistance for usage and costs were manageable)
This is the same as saying "gasoline is a temporary phenomenon and will eventually be replaced"
Like... yes, but not for the reasons you state.
Inefficient + expensive directly conflicts with the idea that they create the market incentive of tokenmaxxing. both of those things can't be true. Those are opposite claims.
My take is that tokenmaxxing is not a market incentivized behavior, it's a temporary and expected experimental phase. It's finding the limitations and utility boundaries of a new technology.
Inefficient and expensive are wildly relative and subjective labels. LLMs are profoundly efficient in terms of transforming some raw energy resource into applied intelligence in many use cases but that doesn't mean we can't do better.
It just so happens that human intelligence has largely been encoded in language, but clearly languages are arbitrary evolved human centric artifacts that are fundamentally inefficient from a computational perspective.
Sources: Amazon has shut down an internal leaderboard that tracked employees' use of AI tools after workers tried to boost their scores with needless tasks (@rafeuddin_ / Financial Times)
(Visit Techmeme dot com for the link and full context!)
> if something is important it will come up again in a better form
This is that famous line from someone at basecamp. I think it's valid for prioritization. Backlogs are classically, just a list of stuff that isn't actually important.
BUT there are projects / features that deserve thinking + observation over time, and keeping notes to share perspective between humans (and now agents) can prove super useful. Often times reading how someone was thinking about a feature at it's inception time can round out or inform how I'm thinking about it now.
All that said, I agree backlogs are almost aways dumpsters and there is probably a better place to store those "let this idea simmer and cook" a little bit things.
@zachtratar What is the "right amount" of startups from your vantage point? What / whose metrics are we optimizing for in terms of startup volume?
Failure rate going up isn't inherently bad, though it does certainly necessitate evolution of the composition of the ecosystem.