Hacker Newsnew | past | comments | ask | show | jobs | submit | giladd's commentslogin

Good test. But we can say the same for internet connection, or electricity, or fossil fuels. No? There are widespread infrastructure technologies without which we'd collapse regardless of LLMs.

But one thing we do have today, and presumably will still have in 20 years, are people with the knowledge of how Internet hardware, electrical systems, and fossil fuel systems work. For every layer in each system, we have at least one person who knows how it works.

In 20 years will we still have people with the knowledge of how each layer of a software system works?

There are already pieces of important software in use today that have no human holding the knowledge of how they function. We tend to regard such software as "running on a prayer", because if it breaks, we're screwed. And we all pray that our bank's software is not like this.

In 20 years, what percentage of software will be running with no human knowledge of its internals? And how many humans will even exist who have the capability to quickly gain that knowledge when the shit hits the fan?

I'm not saying we're doomed, but I do think we need to keep our collective eye on that since most businesses are opposed to "loser thinking" that eats into the profits.


Also good point, and an LLM is basically a trivial dependency from a "manufacturing" perspective.

Like compared to how much we depend on modern CPUs, that are in some way only producible by a single company, or the drug industry - a matrix of numbers plus anything that can do computations is something that can't realistically ever be stripped from humanity (without humanity itself getting wiped at the same time)

Like in a post-apocalyptic world it's far more likely that you will have slow (depending on when the apocalypse happens) LLM endpoint readily, but have no way to produce/get insulin.


Comments here say that these colorful techniques for info presentation are not needed when your audience is professional and already knows the context of the discussed topic. While true, many organizations don't have that, certainly not for all members at all levels!

Sometimes taking a marketing approach for passing information or ideas is the right call. Not all colleges are engaged, knowledgeable or even decent enough to listen during the meeting. You still have to work with the cards you're dealt. Being flashy is one way to do it.


Yes, the idea that presentation format is irrelevant is a silly one. No one is immune to "marketing", but if you are, great, the rest of your team or your stakeholders are not. Learning different ways to convey information and ideas to different audiences is a valuable leadership skill.

Most people don't understand the power of merchandising and how you can use it for fun and profit when it comes to ideas, data, process, change, etc.

The structural siloing pressure you're describing happens in every channel-based organizational chat (though admittedly, some got it worse than others). I'm working on a chat app that uses a different model to combat this. Instead of channels, every message has at-least-one tag that describes it. It's similar to specifying recipients in email, only that you don't need to know the actual people. The message is publicly readable by default, and routed to the inbox of anyone who follows any of its tags.

Hoping to make the repo public soon.


Sorry for commenting without actually learning the subject but from your description it feels like you are just send the message into the void without even knowing if someone reads it? Maybe you should call your project "message in a bottle".

I call it TagTree. Fun name though :)

Essentially yes, but the UI can warn you if no one follows any of the tags you've used. Also @mentions exist. And I'll probably allow users to see who follows what, just like channel memberships in other apps.

However, the ideal is to decouple senders from receivers. You don't always know who's relevant to the discussion, but you do know what you are discussing.


I wished that the trend would be opting for clearer, less verbose papers. Text length and complex language are no proxy for quality.

It's because doing high quality work requires deep understanding. It's that expertise that lets the user prompt accurately enough to hit the right spot in the model's vector space. That model _has_ the expertise, but it manifests only by giving it the right inputs, which differ across tasks.

With no feedback loop and reasonable token use?

That would be the case if competitors were in the same state.

Theoretically, we can construct systems that operate by alternative incentives. Small, self-sufficient economic cells that can grow and merge without the constant need for profit. I don't think it's feasible right now, but such a transition doesn't need to happen overnight.

Does Pi come with an auto mode that blocks dangerous actions? It didn't last time I checked, and these measures are critical to protect prod (and your personal files, etc) from a confused agent.

I gave the harness a try and it is refreshing, but the insistence on staying bare bones and not implementing basic functionality natively, together with no vetting process for popular 3rd party modules, made me return to Claude Code.


I would not trust auto mode in the first place.

Wrap your agent of choice in e.g. docker sandbox, nono.sh, Nvidia OpenShell, microsanbox or any other sandboxing tool and run it without any checks (i.e. "yolo mode", dangerously-skip-permissions etc.)

IMHO it makes the whole category of problems around permission fatigue, trusting the harness (or plugin) authors etc go away.


I did that. Customized Pi to run in Bubblewrap, made an action authorization flow, etc. But that's throwing the baby with the bathwater. This gives you a harmless, less capable agent. To get both safety and usefulness, Ant's solution of a fast classifier works really well.


Using Bubblewrap is pretty good. However, I moved to using Apple Containers with one local file directory mounted; When the container is not running, I do git operations outside the container. This is simple, and although not as convenient as running naked with no isolation, it is light weight and easy.


For me, its the other way around.

Strong sandbox + disabling permission checks gives me maximum capabilities, least annoyance and (relative) peace of mind that any prompt injection, malicious plugin,... won't wreck my shit or leak credentials of any kind


Well yeah, the whole point of pi is that you build it to your workflow. If you dislike the entire point of a project, it’s not for you. I run my agents in a sandbox. When Claude or Codex then double sandbox themselves, it gets really annoying.


I use https://pi.dev/packages/pi-warden with Jev (but can use other Decisions API), it would make sense to combine with pi-permission-system to get the CC two stage "default auto" mode.


I wrote pi-automode for exactly that reason. Works well enough for me, YMMV.

https://github.com/czottmann/pi-automode


You'd have to install an extension or write a function that intercepts tool calls.


gotgenes/pi-permission-system and the ai-guard authorizer can do a lot here, it feels very similar to cc auto mode.

I built a policy authorizer on top of pi-permission-system that looks kinda like nftables that models commands/subcommands/options as read/write and enfores policies depending on if the sub agent is read only or write, which really cuts down on the stupid decisions and nicely complements AI cmd review. Pi shines for making things like that possible.


You have time to watch over your agents?


like the auto mode classifier? if it doesn't then... that's a risky tradeoff


Sounds to me like it isnt basic functionality if it all works ;)


> bear bones

*bare


ty, edited


Does Pi support that natively?


Yes, in the sense that Ant credits some number of dollars worth of API tokens to your account and Pi has long been happy to use an Ant API key


Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: