brap

I get what you’re saying, and you’re right, however I can also see where they’re coming from:

Empirically, agents (especially the coding CLIs) seem to be doing so much better with files, even if the tooling around them is less than ideal.

With other custom tools they instantly lose 50 IQ points, if they even bother using the tools in the first place.

I couldn’t put this into words so thank you for that

I hate how agreeable these things are. When I need it to review something I wrote I have to explicitly pretend that I’m the reviewer and not the author. Results change dramatically.

Wegovy/Ozempic didn’t do anything for me for months. Then my doc put me on Tirzepatide+Phentermine combo and I forgot what being hungry even feels like.

Doesn’t something like Wix take care of all of this?

> Great, I'll get started.

*does nothing*

I actually don’t think golfing is such a bad thing, granted it will first handle the low hanging fruits like variable names etc, but if you push it hard enough it will be forced to think of a simpler approach. Then you can take a step back and tell it to fix the variable names, formatting etc. With the caveat that a smaller AST doesn’t necessarily mean simpler code, but it’s a decent heuristic.

I think these are 2 independent axis:

1. Destructive by accident 2. Destructive because it was prompt-injected

And

1. Fucks up filesystem 2. Fucks up external systems via credentials

This but unironically

Good riddance

I think it’s worse actually, I think the point is “I’m too chill and down to earth, I’m just like you fellow kids”

It feels incredibly performative. Same with sama.

Because they’re not running a charity

So you give it approval to the secret once, how can you be sure it wasn’t sent someplace else / persisted somehow for future sessions?

Say you gave it access to Gmail for the sole purpose of emailing your mom. Are you sure the email it sent didn’t contain a hidden pixel from totally-harmless-site.com/your-token-here.gif?

I had it coding autonomously for about an hour (including lots of tool wait time) on a difficult task, and it actually produced good results.

What’s most surprising is that I had it follow a strict loop/workflow and it did that perfectly. Normally these things go off the rails after a while with complex workflows. It’s something I have to usually enforce with some orchestration script and multiple agents, but this time it was just one session meticulously following orders.

Impressive, and saves a lot of time on building the orchestration glue.

Every time this is brought up I have to go read up on it again, assuming I just didn’t get it last time.

And every time I reach the same conclusion: it's a WhatsApp/Telegram/etc wrapper for LLMs.

Until next time!

Whenever I see instances like this I can’t help but think a human is just trolling (I think that’s the case for like 90% of “interesting” posts on Moltbook).

Are we simply supposed to accept this as fact because some random account said so?

Wow, I thought it was bad, then I got to last paragraph which was really bad.

I don’t see it as the author being lazy, actually the opposite, I see it as being performative and a tryhard. Either way it’s annoying and doesn’t make me want to read it.

After looking into it, as I suspected, the author seems to make his living by selling people the feeling that they’re in the cutting edge of the AI world. Whether or not the feeling is true I don’t know, but with this in mind this performance makes sense.

I see a lot of folks are excited about this but I genuinely don’t quite get it, pretty sure I’m missing something. ELI5, why is this useful?

>prompt isolation

You mean isolating data from instructions? As far as I know this is not a solved problem, or is it?

They really should start using better models for autocorrect and autocomplete. My autocomplete is total bs

Ok Dario

All of this performative bullshit coming out of Anthropic is slowly but surely making them my least favorite AI company.

We get it guys the very scary future is here any minute now and you’re the only ones taking it super seriously and responsibly and benevolently. That’s great. Now please just build the damn thing

Anthropic seems to be very busy producing a lot of this kind of performative nonsense.

Is it for PR purposes or do they genuinely not know what else to spend money on?

I learned that every step that can be solved reasonably without an LLM, should be solved without an LLM. Reliability, cost, performance, etc.

I try to transfer as much work as I can out of LLMs and into deterministic steps. This includes most of the “orchestration” layer which is usually deterministic by nature.

Sprinkle a little bit of AI in the right places and you’ll get something that appears genuinely intelligent. Rely too much on AI and it’s dumb as fuck.

Make their tasks very small and simple (ideally, one step), give them only the context and tools that they need and nothing else, and provide them with feedback when they inevitably mess up (ideally, deterministically), and hope for the best.

One thing that’s always missing from these compound interest calculators is multiple assets with different rates, and different rates over time (e.g between X date and Y date use Z rate, etc). I didn’t quite figure out the right UI for the second one.

BEADS

Do I have poor history?

You said “best results”, but I imagine that the theoretical “best” may not necessarily be the most aesthetically pleasing in practice.

For example, limiting output to a small set of characters gives it a more uniform look which may be nicer. Then also there’s the “retro” effect of using certain characters over others.