moezd

I rotated over many LLM providers and tried out different harnesses, but I always come back to an honest Claude Pro subscription, plus the work account that gives me a GitHub Copilot Enterprise seat. I make sure to always have access to the agent, so it never runs on my laptop. I literally can't stand how GPT sounds confident and gets a crucial piece wrong every time. If I could afford the hardware, I would've probably set up a local LLM rig, but good ones that promise Fable-level inference are out of reach for me.

I haven't seen a lot of improvement by increasing the orchestration layers, honestly. Am I missing something?

Eh. We could've been better off by not inventing cars and creating wonderful concepts like jaywalking, highways literally tearing neighbourhoods apart, stupefying rush hour commutes, lead in gasoline... but who wants to get rid of their cars now, without any presuppositions about their current city's public infrastructure? The genie is out of the bottle.

I think we can make it sharper by not only 1:1 copying it, but focusing on a specific use case that bugs people. How about a backend-only replica that solves Terraform-specific bugs and makes a 1:1 compatible API that feels like a breeze? Then users can slap their own UI on top and basically own the product.

I can't believe we are rediscovering division of labour. Some tech bros are really that far gone, huh.

And how about CDNs that we use and love? Who's to say we don't serve our customers from France?

Doesn't France also have cloud infrastructure that offers similar products like AWS S3?

Exactly. Instead of getting a confident answer that I can store insulin at 95 F, I would like to know that it goes bad at 85. That 10 degree difference determines whether you store your medicine properly or not.

It's how they finally become profitable: You burn through your token quotas so you have to buy them per request now. Oh, and also Enterprise seats are selling like bubble gum.

It's the same addiction script again and again.

That's exactly how I imagined Ancient Romans got some of their deities: They heard of some people or some group in the marketplace as the praeco heralded the daily news, and if they were not fluent enough to understand the nuance, they came to believe that there's a group of people who solves their problems "in the name of X".

Say you're Matthew Boulton in 1775 who partnered with James Watt when he started popularizing his imrpoved steam engine. Could you have predicted that in 2026 we will be discussing about him here, on the Internet, (probably) on a smartphone? Putting down a number is the wrong discussion imho.

I'd like to imagine that a factory from 1880s looked a lot different than today, with IoT and robot arms and whatnot. Intellectual work will still be needed, always will. Just think about "computers" of the Space Race era. Their whole job was to sit down with a calculator and calculate the engineering approximations for an orbit that takes a rocket to the moon. Then they installed IBM monoliths and suddenly you needed a few computer scientists only, plus mechanics for overall maintenance of the machine and utilities. And noone ever stopped teaching electronics at all. In fact, that kick-started the Moore's law era.

Exactly. Factories will be built, but you will still need workers, project engineers, and actual architects. Automation is still not there yet, despite the elaborate harnesses of frontier models.

Noone is mad here, you're fighting with shadows. Go touch some grass. If you want to do science, get accepted in a research institute instead of schooling people over your made up standards, and get me an alternative story to argue against. You are still attacking the character and assuming that you have the authority and honestly, it's rather embarrassing for you at this point.

Aah, like a young grasshopper. Just what we needed, thanks Elon.

lend-and-lease, you say?

After a day of reading AI slop, reading what's essentially an MUN-style resolution, 100% human made, is like eye candy. Thank you for sharing this.

insert thank you meme

Who told you that this is a fact? It's a story, and as far as I can see, a convincing one. If you see otherwise, you should explain yourself in terms that don't resort to ad hominem, and appeals to authority that I assume you don't have either.

At scale you would drop pandas in favour of polar and eventually move to DuckDB. See, this is the blue collar mentality, arguing about whether a wrench or screwdriver helps here and how much WD-40 is enough. An engineer would be tasked to build the chassis entirely, making sure that complaints from previous versions are addressed properly.

Open models may complete these steps, so our models have been doing this for quite some time now. Better ban open models, I guess.

Oh, and, does anyone else realise that what that agentic work is basically how you set up a productive VPS from scratch? How is that supposed to be frightening?

If you still judge people by what kubectl command gives you a precise output or which pandas method is better, and expect them to fill a LoC or PR count quota, then yes. If you ask them to use their judgement and use dependencies wisely to squeeze out performance, absolutely not.

Does it spark joy, though?

Just when you think the golden age of hackers is over... Wasted food is bad, of course, but this is the kind of genius that military academies can never teach.

It's the obsidian story once again.

Assume we've settled near a volcano recently and you start seeing ornate jewelry being sold on the marketplace for the first time, shiny colors, straight edges. You like them, gift them to others. That's your average ChatGPT user.

Over time, you see these rocks being attached to the tip of a spear and other tooling. People report performance improvements while farming, cutting, but many also caution that these new sharp rocks can kill people. These are your IT people using AI agents at work.

And over there, on top of the mountain, in secret caves, there's a new religious order. They claim to use these stones to predict weather, heal diseases and prophesize about the future. That's your frontier lab people.

When will all this be over? Never, though in time the knowledge will eventually dissipate. When Iron Age leaders were able to hand out iron swords in bulk to their armies, people definitely learned how dangerous these are, but later it didn't stop people from using iron in building construction and railways eventually.

Let's just hope we don't need many civilization collapses to jump through these ages this time.

And one bonus question: Does the blade really incite to violence, or do you just want to destroy stuff?

Software engineers somehow got used to slop grenades, 5 figure PRs at full throttle and shift-left on every issue, but not all professions are taking it lying down, it seems.

700 potential papers dumped is not equal to 700 PRs with a LGTM.

Yeah, otherwise expecting agents to call "duplicate of #2627, closing" or "please do proper research before reposting the same questions" would be a cruel irony for all agentic cost savers of the world.

This is probably one step towards an agentic Stack Overflow. Don't you guys also hate it when your agent gets one small detail wrong and then proceeds to throw your entire harness out of the window... No? Oh well.

https://en.wikipedia.org/wiki/Franchise_(short_story) So it begins...

Yet another "AI must be stopped" post... If you're burned out, just say so. Don't philosophise. No, we're not buying that your latest model can almost replace all software engineers. It's us working with the harnesses, adding guardrails etc that make this agentic world possible for everyone.

26 bps at 1.5 m: And what exactly do we expect to carry with this communication channel? If you add horned antennas it becomes a ridiculous design just to gain a few more bytes. If you add antenna arrays then it's no longer cheap. Your transmitter's output power is basically hard limited by Boltzmann constant, so just increasing range means you will have a bad time decoding it. Digital tricks, cryogenics... maybe a few more bytes, but is it enough?

I'm really struggling to see if this can be useful for anything. Dicke radiometer reference is cool, though.