sthuck

Im sure there are more reasons but updates to prompts and tools coming from the server side is a major one.

It's typed so you can build some governance around it, by allowing only some tools or parameters for your org (this is a pretty weak point, but still)

A skill has one giant description from the frontmatter loaded into the context, where MCP loads a smaller one for every tool. Not necessarily better, the skill approach is often better actually, but sometimes the MCP approach fits more

Expectations are a funny thing, a year ago I thought all software industry will cut at least 30% in a year. It's very far from happening. A big change is obviously coming but now I think I'm good enough to last at least the next 5 years, which suddenly feels like a long time if you see it coming.

The truth is despite these very impressive improvements, most impressive work done by agents require many iterations running in a loop, with tens of thousands of dollars in API pricing. And it's still far from being always reliable. Somewhere along the way hardware will get better, energy will be cheaper, the market will be flooded by chips. There a physical world issues that limit all of these for now, thank god. I think 5 years is a good number.

The real damage is that enterprise work became unbearable. Slop code with slop code review, and overly verbose emails with repetitive presentations. And on the other hand, I now enjoy "coding" for myself like I'm 16 again. All I want to do is sit at home and build apps for myself and family. I barely go to work

In the world of web apps, I find the agent's ability to write good e2e tests to be a real game changer. Turns out with enough rigor you can write pretty stable mostly not flaky e2e tests. And even flaky ones are fixed quickly due to a fuck ton of assertions at every step.

Test code looks like a mess, even more than the usual LLM code. takes a while to let it go. Test report looks beautiful though.

Whenever it gets back I recommend everyone to use it. My go to prompt for demoing it to others is "I eat poop off the floor"

And I'm scared to hardcode some negligible config value that will never change

I played with it for 10 minutes, very impressive. Couldn't get vscode to work, will try again later with different apps. (vscode is not that important anyway as it has a remote mode)

I understand (well, "guess" more than understand when it comes to graphic stack in Linux) that it being it's own compositor probably means running a full desktop environment in it is infeasible, but a good way to run software remotely is painfully missing in Linux right now, even one app at a time. The built in servers in kde/gnome are problematic in general and require the monitor to be turned on.

I might be mistaken, but both need a physical monitor turned on. If your monitor goes to sleep you can't connect/see a black screen.

The kde one doesn't support remote user login, and while the gnome one does on paper, I never got it to work. The remote connection situation in Wayland is a major regression.

Fair enough, I probably wrote most people just as an automatic phrase when it should be "some people"

Check my math here, but 1 is larger than 0.

The maximum battery save mode on Samsung (or maybe on all android devices?) does just that - you can set the limit on which apps can be opend. Turn the display to greyscale. Homescreen only shows like 8 apps.

I think for most people, just putting an extra step between you and whatever toxic app(s) you have can be enough. It mostly works for me, and as an added bonus you get insane battery life. I would try that before buying a new phone.

I kinda miss that in the early 2000's kde and gnome shipped with a fuck ton of window decorations based on all those (then-not-so) old OS. Teenager me had fun switching them every day and playing with windowing behavior (focus follows mouse! hover to select and only one click needed!). I wonder what techy kids today do to explore and have fun.

Speaking of the early 2000's, man, Aqua was such a good design. I appreciate the nextstep paradigm and design, but Aqua was just so futuristic, in a good way.

I don't think article misses it, it's exactly the point it makes

Very very strictly speaking relying on models in it's essence is not the problem I think. There is enough "meat" there you can build a nice small profitable company.

Those tools are better than vanilla agents by dedicating expensive human time on evaluating and fine tuning models. You can also build various integration, management and reporting features to add value. If you freeze model progress today, or 12 months ago when most of those companies started, it's a viable business I think.

But any gains you make on the first part will be lost to newer models, and the 2nd part is not as valuable when llms allow people to build fairly complicated features quickly.

I don't if worthless but all those companies have very limited time to gather customers and at least make themselves valuable for an acquisition

groq is a series E hardware startup founded in 2016. It took them this long to be a potential threat, I'm not sure they are even an actual threat.

Even if this purchase causes 100 new hardware startups to be funded tomorrow, nVidia is perfectly fine with that. Let's see how many survive 5 years down the line

It should work if the library was compiled with deceleration map option. Which most libraries are not and it's a shame.

It was added like 3 years ago which was probably a bit too late, not even sure why it's not the default. (File size?)

The best ”algorithm” for discovering new music was digging through profiles on last.fm back when the social functions of the site were still active. Sure, it was a lot of manual work, but the results were amazing. It wasn't completely blind, I found that people I had high similarity with, it was more likely I'll like what they like, even across different genres. Sometimes people were nice and took the effort to recommend based on my profile. I got introduced to varied music, different genres and even a bit from different countries.

The worst was Pandora, which did recommendations based on breakdown of musical instruments and elements in the song. It did what it aimed to do pretty well, only it was a bad idea. It gave you a lot of uninspiring music that sounded like a bland copy of something you actually liked.

Spotify's recommendations are not super awful, but definitely feel closer to Pandora's style. I wonder why is the result like that even though I'm sure they train their model based on listening history.

I'm using it for a hobby project, and pretty pleased.

My personal maybe somewhat "stubborn old man" opinion is that no node.js orm is truly production quality, but if I were to consider one I think I would start with it. Be aware it has only one (very talented) maintainer as far as I recall.

That definitely helps and worth doing. On Mac though I guess you need to move the entire development to containers due to native dependencies.

2003 me thought Wine is a dead end project and a waste of developer time. Granted valve put a lot of effort into Proton but they wouldn't even have considered it without the massive amount of work done before, kudus to all the non cynical wine devs

Css in js was like a fever dream that lasted 2-3 years and seems to mostly go away. It's a good example as to how the frontend world just seems to make bad decisions.

Like if you take React's server components, it has a ton of problems and gets excessive focus from react devs, but fundamentally I can agree on what's its trying to solve. I understand the need, even if i disagree in almost anything else regarding it. I still don't know what the css in js phase was about.

Where I live, the early Facebook apps were the killer feature of the early days of Facebook.

Genuinely asking, what other software company did something like this back then? I'm tempted to say (tounge in cheek) FB was the original PaaS, but maybe im not old enough

Look I agree but one can also manage with an always on pc and an external hard drive instead of a homelab. It's part hobby part learning experience.

Also if you have kids 0-6 you can't schedule anything relaibly

The part about buying a misleading CPU really hit home with me. Man all those small retailers in the 90s were living the dream with their unaware customers.

My parents bought an AMD 486 100mhz at the end of 1995, it was still a viable low end machine for that time but they paid as if it was a Pentuim. Had to somehow make it work up until 2001 with that machine. Don't know if I had a career in tech if I wasn't forced to tinker with it.

look the peace is definitely an achievement and much much better than the alternative. But it's incredibly fragile peace held only the Egyptian Army control and Sisi's dictatorship.

Most of the population still despises Israel, another revolution in Egypt and its a matter of time until there is another war. Now with western equipment for both sides!

Generally those smaller models do well when most of the data comes from the prompt. Summarising, rephrase and change tone, categorize. Ask them a question, or to generate code and you'll get hallucinations

I'm purely guessing here, but also considering I read him and Linus both say "we have enough kernel developers", I think it's likely they don't want to encourage low quality contributions from new developers.

I mostly think it's an interesting concept that can allow many interesting user experiences.

At the same time, it is a major risk for browser compatibility. Despite many articles claiming otherwise, I think we mostly avoided repeating the "works only on IE6" situation with chrome. Google did kinda try at times, but most things didn't catch on. This I think has the potential to do some damage on that front.

Linux desktop efforts need a BDFL.

Too many duplicated efforts, the wayland migration while needed was a complete mess, and actually the worst time I had using Linux since like 2005. This kind of complete breakage of user experience is unimaginable even in MacOS and it's hostile approach to power users.

I think otherwise it's going to continue being a frustrating two steps forward one step back.

A lot less sales and cuatomer support people

I'm still looking for a conclusive list on every possible way to create an object on c++