fhd2

In my experience non-technical folks quite dig the memory feature. For me that's kinda context poisoning as a service, but I know people that get value out of it (or at least strongly feel they do). Not sure how one would migrate that.

Yes, that's exactly how I meant it. I might _rarely_ peruse some code if I'm really curious about it, but by and large I just trust the developers of the software I use and don't really care how it works. I care about what it does.

Users typically don't read code, developers (of the software) do.

If it's not worth reading something where the writer didn't take the time to write it, by extension that means nobody read the code.

Which means nobody understands it, beyond the external behaviour they've tested.

I'd have some issues with using such software, at least where reliability matters. Blackbox testing only gets you so far.

But I guess as opposed to other types of writing, developers _do_ read generated code. At least as soon as something goes wrong.

Because even if an organisation hasn't rolled out generative AI tools and policies centrally yet, individuals might just use their personal plans anyway (potentially in violation with their contract)? I believe that's called "shadow AI".

It wouldn't keep them from equipping _new_ models with additional sensors, spinning a story around how this helps them train the camera-only AI, or whatever.

Fairly cynical indeed. Though I must admit that Anthropic's software - not the models, the software they build - seems to be generally plagued by quality issues. Even the dashboard is _somehow_ broken most of the time, at least whenever I try to do something.

Ah, got it. Very different from where I suspected the issue then.

> Anyway, it's not related to CoPilot, but because Notepad makes links clickable now...

True, not related to CoPilot, but if I understand your conclusion right (which I'm not sure about), it's not _just_ that links are clickable now, it's because Notepad actually does something with the links. Otherwise it'd be a browser vulnerability, and Notepad couldn't seriously be blamed.

I think most of us - if not _all_ of us - don't know how to use these things well yet. And that's OK. It's an entirely new paradigm. We've honed our skills and intuition based on humans building software. Humans make mistakes, sure, but humans have a degree and style of learning and failure patterns we are very familiar with. Humans understand the systems they build to a high degree, this knowledge helps them predict outcomes, and even helps them achieve the goals of their organisation _outside_ writing software.

I kinda keep saying this, but in my experience:

1. You trade the time you'd take to understand the system for time spent testing it.

2. You trade the time you'd take to think about simplifying the system (so you have less code to type) into execution (so you build more in less time).

I really don't know if these are _good_ tradeoffs yet, but it's what I observe. I think it'll take a few years until we truly understand the net effects. The feedback cycles for decisions in software development and business can be really long, several years.

I think the net effects will be positive, not negative. I also think they won't be 10x. But that's just me believing stuff, and it is relatively pointless to argue about beliefs.

It's not a complaint about LLMs, it's a complaint about my own behaviour. Maybe I didn't make that clear enough.

Maybe like one of the siblings said, I just shouldn't jump right into execution even if that barrier has been lowered. But it sure is tempting for me.

Yeah the latter. Doesn't happen much with super obvious stuff, but especially when I'm a bit fuzzy on requirements, it can be just endless iterations.

What kills me personally is that I'm constantly 80% there, but the remaining 20% can be just insurmountable. It's really like gambling: Just one more round and it'll be useful, OK, not quite, just one more, for hours.

I feel agentic development is a time sink.

Previously, I'd have an idea, sit on it for a while. In most cases, conclude it's not a good idea worth investing in. If I decided to invest, I'd think of a proper strategy to approach it.

With agentic development, I have an idea, waste a few hours chasing it, then switch to other work, often abandoning the thing entirely.

I still need to figure out how to deal with that, for now I just time box these sessions.

But I feel I'm trading thinking time for execution time, and understanding time for testing time. I'm not yet convinced I like those tradeoffs.

Edit: Just a clarification: I currently work in two modes, depending on the project. In some, I use agentic development. In most, I still do it "old school". That's what makes the side effects I'm noticing so surprising. Agentic development pulls me down rabbit holes and makes me loose the plot and focus. Traditional development doesn't, its side effects apparently keep me focused and in control.

> Ads do not influence the answers ChatGPT gives you.

I wonder if this is a don't-break-product-value thing, or just compliance (ads need to be clearly labeled, but OpenAI seems like it has the risk appetite to ignore that kind of thing).

(1) is not something the typical employee can do, in my experience. They're expected to work eight hours a day. Though I suppose the breaks could be replaced with low effort / brain power work to implement a version of that.

I suppose ultimately, the external behaviour of the system is what matters. You can see the LLM as the system, on a low level, or even the entire organisation of e.g. OpenAI at a high level.

If it's the former: Yeah, I'd argue they don't "learn" much (!) past inference. I'd find it hard to argue context isn't learning at all. It's just pretty limited in how much can be learned post inference.

If you look at the entire organisation, there's clearly learning, even if relatively slow with humans in the loop. They test, they analyse usage data, and they retrain based on that. That's not a system that works without humans, but it's a system that I would argue genuinely learns. Can we build a version of that that "learns" faster and without any human input? Not sure, but doesn't seem entirely impossible.

Do either of these systems "learn like a human"? Dunno, probably not really. Artificial neural networks aren't all that much like our brains, they're just inspired by them. Does it really matter beyond philosophical discussions?

I don't find it too valuable to get obsessed with the terms. Borrowed terminology is always a bit off. Doesn't mean it's not meaningful in the right context.

I can't speak for parent, but I use gptel, and it sounds like they do as well. It has a number of features, but primarily it just gives you a chat buffer you can freely edit at any time. That gives you 100% control over the context, you just quickly remove the parts of the conversation where the LLM went off the rails and keep it clean. You can replace or compress the context so far any way you like.

While I also use LLMs in other ways, this is my core workflow. I quickly get frustrated when I can't _quickly_ modify the context.

If you have some mastery over your editor, you can just run commands and post relevant output and make suggested changes to get an agent like experience, at a speed not too different from having the agent call tools. But you retain 100% control over the context, and use a tiny fraction of the tokens OpenCode and other agents systems would use.

It's not the only or best way to use LLMs, but I find it incredibly powerful, and it certainly has it's place.

A very nice positive effect I noticed personally is that as opposed to using agents, I actually retain an understanding of the code automatically, I don't have to go in and review the work, I review and adjust on the fly.

That's a job for a multi agent system.

I can't count the times I've told clients and prospects to _not_ hire us to build something they wanted. Because they could just use off the shelf solutions that were cheaper financially, at least in the short to mid term, and much, much cheaper in terms of opportunity costs. I struggle to put even billed hours into something that doesn't make sense to me.

Of course some overdo it. I've seen companies with more random SaaS tools than staff, connected with shaky Zapier workflows, manual processes, or not at all. No backups, no sense of risks, just YOLOing. That's OK in some cases, in others really not.

I suppose it does need some engineering thinking to find the right things and employ them in a good way. Unfortunately even developers commonly lack that.

It's not tool use with natural language search queries? That's what I'd expect.

I _think_ the idea is that the first one to hit self improving AGI will, in a short period of time, pull _so_ far ahead that competition will quickly die out, no longer having any chance to compete economically.

At the same time, it'd give the country controlling it so much economic, political and military power that it becomes impossible to challenge.

I find that all to be a bit of a stretch, but I think that's roughly what people talking about "the AI race" have in mind.

Well, wouldn't you be on board?

1. Does xAI seem more valuable than X?

2. Does SpaceX seem more valuable than xAI+X?

Not sure a lot of people would say "no" to either of these questions.

The only other question that I think is worth asking for investors, is how much stock in the acquiring company they get for their stocks in the acquired company. If the valuation of the acquired company in the deal is... optimistic enough, that seems like a no brainer.

First thought: In my experience, this is a muscle we build over time. Humans are pretty great at pattern detection, but we need some time to get there with new input. Remember 3D graphics in movies ~15 years ago? Looked mind blowingly realistic. Watching old movies now, I find they look painfully fake. YMMV of course.

Second thought: Does it _really_ matter? You find it interesting, you continue reading. You don't like it, you stop reading. That's how I do it. If I read something from a human, I expect it to be their thoughts. I don't know if I should expect it to be their hand typing. Ghost writers were a thing long before LLMs. That said, it wouldn't even _occur_ to me to generate anything I want to say. I don't even spell check. But that's me. I can understand that others do it differently.

Just like how writing helps memorisation. Our brains are efficient, they only do what they have to do. Just like you won't build much muscles from using forklifts.

I've seen multiple cases of... inception. Someone going all in with ChatGPT and what not to create their strategy. When asked _anything_ about it, they defended it as if they came up with it, but could barely reason about it. Almost as if they were convinced it was their idea, but it really wasn't. Weird times.

Perfection.

You're absolutely right!

Sorry, couldn't resist :P But I do, in fact, agree based on my anecdotal evidence and feeling. And I'm bullish that even if we _haven't_ cracked how to use LLMs in programming well, we will, in the form of quite different tools maybe.

Point is, I don't believe anyone is at the local maximum yet, models changed too much the last years to really get to something stable.

And I'm also willing to leave some doubt that my impression/feeling might be off. Measuring short term productivity is one thing. Measuring long term effects on systems is much harder. We had a few software crises in the past. That's not because people back then were idiots, they just followed what seemed to work. Just like we do today. The feedback loop for this stuff is _long_. Short term velocity gains are just one variable to watch.

Anyway, all my rambling aside, I absolutely agree that LLMs are both revolutionary and useful. I'm just careful to prematurely form a strong opinion on where/how exactly.

Parent didn't mention Simon Willinson, and neither me nor parent appear to imply that _all_ people posting positively about LLMs are paid influencers, that'd be a ridiculous claim. It's just that there _are_ paid influencers, at every level, down to non-famous people getting a few bucks, and that's worth knowing.

Here's one thing I quickly found on one of Anthropic's campaigns on LinkedIn: https://www.favikon.com/blog/inside-anthropic-influencer-mar...

Their Readme.md is weirdly obsessed with "2 hours":

"before Claude Opus 4.5 started doing better than humans given only 2 hours"

"Claude Opus 4.5 in a casual Claude Code session, approximately matching the best human performance in 2 hours"

"Claude Opus 4.5 after 2 hours in our test-time compute harness"

"Claude Sonnet 4.5 after many more than 2 hours of test-time compute"

So that does make one wonder where this comes from. Could just be LLM generated with a talking point of "2 hours", models can fall in love with that kind of stuff. "after many more than 2 hours" is a bit of a tell.

Would be quite curious to know though. How I usually design take home assignments is:

1. Candidate has several _days_ to complete (usually around a week).

2. I design the task to only _take_ 2-4 hours, informing the candidate about that, but that doesn't mean they can't take longer. The subsequent interview usually reveals if they went overboard or struggled more than expected.

But I can easily picture some places sending a candidate the assignment and asking them to hand in their work within two hours. Similar to good old coding competitions.

Fair, didn't think that one through :D

Yeah, driving kids to stuff is my use case too. I did the cafe thing for a while, but I didn't like the overhead of going there, ordering etc. For some reason I'm pretty productive in weird environments (trains, airplanes etc), so it's always a good opportunity to just get some stuff done I'd otherwise be distracted from. I'd probably sit outside in a bench if I was in a warmer, dryer climate.