The issue I have with loops is that for truly complex work, where I care about building a generalized solution for a complex problem, the agents frequently reward hack and end up burning indefinitely without finishing until I step in.
Even with Claude Code and Opus I genuinely don't understand how people are actually doing productive things with these loops. Unmanaged by humans, these things go deep, deep into their own pits of internal domain language, depth-first development, and tail-chasing. The stuff that I've seen at work produced by people exploring these kinds of "meta loop" approaches typically have a high ratio of slop and fluff to useful code and documentation.
(I haven't tried Fable yet, but people have been writing about this stuff long before Fable came out so that shouldn't count.)
It makes sense for something like autoresearch because you're trying to exhaustively explore a space that would be impossible for a human to explore by hand. But for something like building the software to run a business, there's absolutely no way I would trust anything like this. It's ultimately just superpowered vibecoding, giving you the ability to churn out more slop faster, not build useful things faster.
Show me one example of a successful application developed using this kind of meta-meta-loop architecture. Supposedly Claude Code itself is entirely AI generated code, but it's not a `while` loop, it still has a team of designers and developers.
(Also this meta-loop stuff is different from the tactical-level Ralph Wiggum loop stuff. That's a strategy to make smaller/weaker LLMs perform like better LLMs, using say 5x more tokens but at 1/10 the cost per token so that the cost-benefit math works out.)
Telemetry tooling for local Claude/Codex usage so I can analyze old sessions and fill tooling gaps, make sure I'm using the right models for various tasks, update my processes, etc.
I've also replaced Linear with a local sqlite-backed tool, added tooling to speed up code nav, and am building "no-slop", a tool for enforcing architectural guidelines on vibe-coded projects.
It’s not easy to buy such a large tranche of shares at a fixed and fair price in a single transaction!
Both parties get something they want this transaction. Alphabet gets the Berkshire halo effect and a guaranteed buyer of $10 billion worth of equities, Berkshire gets a large tranche of equity at a price they believe is fair.
I think they view Alphabet as their next Apple, and a relatively safe place to ride out whatever happens with AI: Alphabet is fairly well positioned for the upturn or the downturn, especially now with this expanded warchest of cash.
They are buying 10B$ worth of shares for 10% discount from current valuation, and if their goal is to hold for 10-20 years, then it could be a good hedged buy in favour of AI.
Even if AI crashes 90% SpaceX, OpenAI & Anthropic are worth say 200B each post IPO. In 10-20 years with similar effects to Internet they might be the next Meta. Apple, Microsoft of the world.
But Google will likely still be the leader if it can make good on it's advantages.
The responses to this are wild. I have worked with and built smaller systems like this and it is an incredible speed up.
So much reflexive hate against a genuinely transformative tech. Yes AI has annoying people and grifters, but it is genuinely incredible at some things and finding out how to use it effectively within a company is the most fun I’ve had in my career.
The issue of it burning through tokens grepping around should be fixed with language server integration, but that’s broken in Claude Code and the MCP code nav tools seem to use more tokens than just a home-built code map in markdown files.
They got so many things right in the beginning but now seem to lose touch with their core fan base, the developers. It's the typical corporate grind, a million competing interests arise where it's not anymore about the user but about politics and whoops, that's when you know you're not anymore a startup.
That is not a good idea. To deal with LLMs one need to have knowledge about the topic of the query, case contrary one will not be able to detect the errors of their output-prompts. The test is easy, if after one or three queries one do not detect the errors, one is done, the person is reading the output-prompts in passive mode.
The self-learn path require also to cultivate a intuition that comes from searching and reading technical doc that a LLM will not give you, among other things.
Anyway, I observe how the warning of the other user about this got downvoted and critiqued. I expect the same, and leave this thread with peace of mind subscribing to such warning, as a message to the OP.
>OP wants to brush up on their skills, not have AI do it for them.
the two things aren't mutually exclusive.
if an AI tells you "solder A to B" you're going to learn some technique whether you want to or not. Extrapolated entirely into a robotics project.. there's a lot to gain just through sheer osmosis of instruction.
the barrier to entry for a lot of playing around is getting a working scaffold to be able to run all your testing from
id expect it could pick you out a breadboard, a micro, some actuators and sensors, along with get a code deploy and run harness going for you, so you can focus on doing the robotics, rather than anything else.
Some people learn by doing, or learn by example, and the faster they can get into doing an example, the faster they will learn.
I am one of those people, and I can't count how many textbooks I own, of which I've read the first few chapters, and lost interest because I wasn't doing anything, only reading.
When I can instead start doing something, as GP emphasized, I can learn the applicable concepts as they're applied, which works well for me. AI helps me do that, because it is like a textbook that follows along with me, rather than asking me to follow it. Also I ask a lot of questions.
Very true, and doing it this way lets me learn at midnight for no extra cost while I go from zero to one and keep my day job that lets me pursue such passions.
They're very varied, so not a clear path to a job there, and I'm not sure I would want to make a job out of all of them.
1. I was prepared my to roll my eyes, but I actually think the framing is correct. AI hasn't replaced legacy vendors yet, but companies are now in a position to at least assess whether "Cheap External Tool + AI" beats "Expensive Tool", which starts to compress margins for existing tooling.
2. A suspicious number of "It's not X, it's Y" in this piece.
Jonathan Haidt recently made the point on Ezra Klein's podcast that while adults can take a break from phones and reset their attention/hormones in a couple of weeks, we don't know what impact similar addiction has on a developing mind. It's possible the addiction sets in much deeper.
I'm sympathetic to folks who grew up shaped by this. Not for nothing, but The Conversation also has a compelling start/end, but has a long, arguably slow, boring middle. So it's like being forced into withdrawal on hard mode.
I think that's true. Alcohol addiction modifies the brain and it can take over a year to recover dopamine sensitivity, focus, mood, cognitive ability. It's called PAWS (post acute withdrawal syndrome). Given TikTok etc. has a similar profile of long usage over a long period, I'd also expect it to take a long time to fully recover from.
Curious how you're addressing this