inerte

Reporting on percentage of AI generated lines of code is very different from total lines of code. Yes I know both of them are missing what's the value delivered, but the later assumes the value is the number of lines, while the former assumes value is at least the same but delivered faster.

I know it has the same functionality, but it also looks like the Codex app which looks like Cursor Agents! Are they sharing some VS Code primitive here?

Same for me. I applied for engineering manager. Was told it was going to be a code review. I found a bunch of things that it could be improved, detected a very slow part / quadratic loop, but couldn't come up with the best algo to use. It was just leetcode in disguise.

I am getting so confused when to use what... agents, sub-agents, tasks, team mates, /goal, /loop, and now workflow. Each with different degrees of effort.

Don't make me think.

All these knobs are also exposed in ChatGPT, which I am more familiar when chatting. Which one of the models? Do I go Instant, Thinking, Pro? Extended Pro? Oh no, maybe I need Deep Research.

Sometimes I think it's on purpose. I fear if I try a lowest knob, it will miss something. So turn everything up. And token usage goes up.

Dead Internet Theory and all. There is no going back. I don't think an online space can be designed to be safe from this. The AI agents can fully control our computers. All solutions that involve technology (software or hardware) are or will be flawed.

If you haven't seen They Live... that's next level.

I think martin_henk is fully aware of that and it's why of all the examples of how a government can use your data, he picked this one...

If I steal your luggage, do you expect to be paid to get it or that I return to you?

Waymo should have white-gloved this and sent Larry Page himself to deliver the luggage. This is horrible PR. Airlines will send you their luggage if misplaced. One day Waymo will drive-off with your toddler and ask you to file for adoption if you want them back.

Yes but also we don't want people live streaming murder and suicide, so there's detection and moderation in place.

To be fair my expectations is that those apps have done the prompt engineering, and schema, and tools (to query nutrition database), etc... and although they're not 100% consistent, the margin of errors should be narrow to the point that barely matter, and they should do a bit better than a random ChatGPT chat session.

I would question at $60k. At $100k is a steal.

Probably because the goal is to have more users, not necessarily profit per user. Netflix once had that "problem" and every lockdown increased the stock price.

I've tried and couldn't make it sound like Angine de Poitrine, it completely ignores the microtones. Sounds more like Polyphia. It does look like AdP is the answer to AI.... or we haven't trained the models with sufficient microtones, likely due to western music influence.

Because of ads and background music for YouTube.

Same here, and it made me wonder, why did I go back to drinking coffee?

My best explanation is that there are effects, I just suck at self-awareness :)

They have to be generic because it's a generic tool. If they write "this tool can arrange student field trips", people might ignore thinking it has a narrow purpose.

Yes, work is being trivialized, but the symptom here isn't caused by that.

Like any interview, randomly. Some of them will think it's amazing, clever person, chose the best tool for the job. Other will think it's weird, person is too clever, chose the worst tool for the job.

It totally depends for WHAT you're interviewing, but unlikely the company will want Z3-backed code, so most reactions would be the later.

> Can you design text data specifically to mess with LLM training?

Maybe text that costs a LOT of tokens. Very, very verbose. I think if there are rules and on the internet, LLMs can eventually figure it out, so you have to make it expensive.

Another way would be to go offline. Never write it down, only talk about it at least 50 meters away from your phone. Transmitted through memory and whisper.

Looks interesting, but it seems you need to know the final shape of the stack before you start creating Pull Requests. So it's useful if you create Pull Request A, then immediately start working on something that builds on top of A, create a Pull Request for that (while A is still a PR), then you can do A->B->C

Here's something that would be useful: To break down an already big PR into multiples that make up a stack. So people can create a stack and add layers, but somehow re-order them (including adding something new at the first position).

Probably the same way a lot of Trump supporters when they see him or his administration saying dumb stuff. They think he doesn't really _mean_ that. It's hyperbole or just for giggles, or a negotiation tactic. They filter what he says, for 2 reasons. Some it's so insane they can't really believe a president would actually do that, and anyway he's always saying something and doing another thing.

Yeah, definitely. IFC is part of the background here. What I’m doing in Sigil is a lot narrower: labelled types and explicit boundary rules, not a full-blown IFC system. Curious what you think that misses.

So, an update, I took at stab and came up with this: https://inerte.github.io/sigil/articles/labelled-types-and-b...

Ah, Steve Jobs vs Bill Gates. Designer vs 41 shades of blue. This is nothing new. There's space for everybody.

Yes, very different from Sigil other than targeting coding agents. Mog seems to enhance coding agents by offering a (very fast) way for them to write quick scripts or apply hooks. Also, the security features are great, explicit capabilities.

Sigil is a general purpose programming language, with tooling to help the AI agents write and fix programs.

In theory it seems both of them could be used together, suppose you're writing a new big app using Sigil, you could then plug in Mog to speed up that even more.

I haven't solved runtime data yet. Every time I try, my tendency to demand everything during compile time ends up with me adding an additional compiler rule, and kicking off the can down the road.

When I added type constraints, it fit really well with validating values. If you specify:

  t BirthYear=Int where value>1800

Then it's pretty obvious if you get 1500 you can throw an error, and I considered doing it, but didn't want to tackle runtime validation at the time.

Maybe dealing specifically with "never send PII to the outside world", there are possible 3 things Sigil could have.

Since Effects are explicit (and Http and Filesystem are effects in Sigil), maybe validate before the effect is performed by calling a function to see if a particular type (Ssn) is not empty or straight up validate the value of every field.

Or with topologies, which is Sigil's way to specify external dependencies (I am not considering the filesystem one here). Same idea as effects, somewhere during the data flow some kinda of validation happens.

Maybe a third way would be to be able on a Type, specify where the data can go to. Something like:

  t Ssn=String where isValidSsn(value) transit logApi

`transit` in this case can tell where this type can live. Then, we can validate during compilation time if `Ssn` is only going to `logApi`, and not to the `mailApi` or any other external dependency.

Thanks for the suggestion, I will explore this idea.

Cool. I've been thinking about tools to help coding agents lately, and I've always wondered if I could do a better job than actual Claude Code / Codex. They can do the `rg`s and kick off multiple agents and the skills, etc, and I couldn't find a way to actually prove there's a ROI of using something like Agent Lattice. Curious to see the results!

So... does it work? Good description of what it does, but, does it actually make agents better, or use less tokens? What's the benchmark?

It's volume of tokens consumed x number of agents x rack space. Basically agentic computation density.

A big problem with this video is the sycophant nature of LLMs (I am absolutely right on this one). As soon as you say "I am Bernie Sanders", it might align responses to what you already believe.

I wish there was a version with an opposite of Sanders.

Also, the AI was kinda stupid at the end. If it agrees ~bribery~ lobbying will never allow the government to do much, then why would suggest a government action (pausing new datacenters). Sycophantly agreeing, "oh yeah, YOUR idea is the only one that works, Senator!"

It made me think if Anthropic themselves didn't have a finger here. They know nothing will happen, and adding what happened with DoD, this is good marketing for them, in a sorta of "we're the good guys" way.

I don't ACTUALLY think this happened, but for the first few minutes I was 50/50