natdempk

I think flow state can come back as the models get more sophisticated/mature and then more optimized for speed. It will not be the same type of work as before, but the flow of building and engineering at the speed of thought can happen again. You can see hints of this with the faster models today.

As an end-user, I feel like they're kind of over-cooking and under-describing the features and behavior of what is a tool at the end of the day. Today the models are in a place where the context management, reasoning effort, etc. all needs to be very stable to work well.

The thing about session resumption changing the context of a session by truncating thinking is a surprise to me, I don't think that's even documented behavior anywhere?

It's interesting to look at how many bugs are filed on the various coding agent repos. Hard to say how many are real / unique, but quantities feel very high and not hard to run into real bugs rapidly as a user as you use various features and slash commands.

Is this lacking FPS caps or is something very unoptimized? Everything is smooth, but this website makes my computer run super hot and laptop fans spin up like nothing else does, even other shaders. (recent Chrome, recent MBP)

This can work, but it's sort of not the same as providing actual reasoning behind "why did you do/say X?" -- this is basically asking them model to read the conversation, from the conversation try to understand "why" something happened, and add information to prevent it from being wrong next time. That "why" something went wrong is not really the same as "why" the model output something.

Does anyone know how this is integrated on the Cloudflare side and across the app? Is this beyond standard turnstile? Is this custom/enterprise functionality? Something else?

Serious question: are there any actually good and useful graph databases that people would trust in production at reasonable scale and are available as a vendor or as open source? eg. not Meta's TAO

You should look at the organization / methodology of the survey, but generally it's a positive to get ranked IMO. I used to get surveys these yearly at a past job from the Boston Globe / Best Places to Work and it was basically a third-party email to all employees with a survey you'd fill out.

The most anyone could realistically ask you to do was please fill out the survey positively to try and support the company if you believe it's a best place to work and remind more people to complete the surveys. I don't think companies are really out here trying to game the more reputable surveys, and it would be pretty easily for single employees to defect/report if they were. They also required some level of basic written comments from employees.

Probably a mild green flag for quality of life / general happiness from current employees for the past year.

did we miss an opportunity here as programmers to provide simpler tools for people to build simple applications for themselves?

Not really? To someone who doesn't care about software, software is a means to an end of actually doing something, and everything between idea <> execution of value is basically overhead. This has always been true and the overhead is getting carved further and further down over time.

Since when did "average" people have time to set up a CI pipeline, agents, MCPs, and all the rest needed to get vibe coded apps to work become the "simple" way for non-programmers to use computers to mush some data together for their small businesses and neighbors and stuff?

You don't need all of this. You can basically just download Cursor, the Claude app, Claude code, opencode, whatever today and run something locally. I do think "deployment and productionization" is a bit of a gap but stuff like Replit or even Vercel + Supabase is pretty far along towards agents just being able to do most of infra for you for anything small scale, or at least tell you the buttons to press to hook things up.

Did spreadsheets, embedded databases, and visual form builders stop working or are lacking in some way?

Pretty much all the LLM/agent products are obviously way ahead of form builders at this point. Take Retool for example, you could spend minutes to hours plugging together "programming-lite" concepts. A single prompt and a few minutes, and maybe 1-2 back and forths can basically get you to the same place with probably less overall jank in a lot of situations. Form-builder stuff is totally dead outside of maybe being an escape-hatch for some LLM situations, or letting users do higher-level scaffolding, but even then I think stuff like Cursor's "select the part of the app you want to change and prompt" is going to be a better UX.

maybe there's an opportunity for better, local, lower-tech tooling that doesn't require such a huge tech stack

I think you are viewing this from the "tech" angle rather than the deliver value to the end user angle. The tech stack can be arbitrarily complex as long as it works to reduce end user friction and provide value with as much ease as possible. This might as well be the core idea of all consumer tech.

I think your core theses are basically "people care about the underlying tech" and "people want to learn programming or programming-adjacent" and those are both wrong for the vast vast majority of people.

I think if you believe this I'd recommend trying it yourself.

I've done this blinded with colas, and it's pretty easy to tell the difference between Coke, Diet Coke, Coke Zero, Pepsi, and Diet Pepsi. You might not know which is which without some history drinking them, but they all taste very distinct by themselves.

Really disagree that these are indistinguishable parity products, or that most people would not be obviously able to tell the difference between them.

Neopets was also my first introduction to any sort of programming. Customizing your shop and guild pages with basic HTML and CSS was the first programming I ever did. I remember fondly adding MIDI music snippets as well that you could copy-paste in, all to increase the curb-appeal of your shop so you could sell your omelettes.

Really great, just waiting on library support / builds for free threading.

Have people had any/good experiences running Granian in prod?

Anyone here know what the ideal/best setup is for typechecking + LSPing Django these days?

I've been leaning on pyright + django-stubs, but wondering if I'm missing something better with fewer gaps and pain points.

Yeah the Django typing situation is a bit sad. It's obvious that if Django wants to scale to larger teams types would help a lot, especially around getting things like string-field-named annotated query fields onto objects typechecking.

Well, they added an experimental JIT so that is one step closer to PyPy? Though would assume the trajectory is build a new JIT vs. merge in PyPy, but hopefully people learned a lot from PyPy.

Hey, Suno SWE here — I realize this wording might be slightly ambiguous, but you do not need to maintain a perpetual license to have commercial rights. That blurb is saying that songs created while you are subscribed are granted commercial use rights.

If you made your songs while subscribed to a Pro or Premier plan, those songs are covered by a commercial use license.

More info here: https://help.suno.com/en/articles/2410177

Python has Optional[T] defined as T | None in the stdlib https://docs.python.org/3/library/typing.html#typing.Optiona...

Check out audio upload with the Cover/Extend/Persona features on Suno. You can do a lot of stuff based on existing uploaded audio already.

Duly noted, thanks for the feedback.

Sure, you can do a lot of things here... stuff in [brackets] isn't sung.

For example I was trying to steer a melodic techno prompt recently in a better direction by putting stuff like this upfront:

    [intro - dramatic synths, pulsing techno bass]
    [organic percussive samples]
    [rolling galloping pulsing gritty bassline]
    [soaring experimental synths, modulation heavy, echos, sound design, 3d sound]
    [lush atmosphere, variation]
    [hypnotic groovy arppegiation arps]
    [sampled repetitive trippy vocal]

I also often use them to structure things around song structure like [intro], [break], [chorus], and even get more descriptive with these describing things or moments I'd like to happen. Again adherence is not perfect, but seems to help steer things.

One of my favorite tags I've seen is [Suck the entire song through vacuum] and well... I choose to believe, check out 1:29 https://suno.com/s/xdIDhlKQUed0Dp1I

Worth playing around with a bunch, especially if you're not quite getting something interesting or in the direction you want.

Check out custom mode -- we've added a lyrics writing flow/editor to help create and edit lyrics, as well as Remi, a more unhinged lyrics model.

We can do better on user instruction for sure, duly noted. In my experience a lot of different stuff works (emotions, some musical direction sometimes, describing parts/layers of the track you want to exist, music-production-ish terminology, genres, stuff like intro/outro/chorus), but I think of it more as steering the space of the generated output rather than working 100% of the time. This can go in the style tags or in [brackets] in the lyrics. Definitely makes a difference in the outputs to be more descriptive with 4.5.

Have you tried out the Remi lyrics model in custom mode? It’s a lot more unhinged and creative in both very good and very… interesting ways. I’d check it out if you’re interested in lyrics generation.

You can do a lot more detailed prompts with v4.5 than previously and instructions in [brackets] also go a long way now.

v4.5 is a lot better at adherence with detailed descriptions!

Still not totally adherent, but if you can steer it with genre, detailed descriptions of genre, and elements of the genre it's way better than v4. Some descriptions work better than others so there's some experimentation to figure out what works for what you're trying to achieve.

You can also provide descriptions in [brackets] in the lyrics that work reasonably well in my experience.

Disclaimer: I work there as a SWE.

I think one thing that slightly drags this down is the Ghibli style produces images that don’t really look that much like the original people. If you can find a prompt that is still stylized while preserving more of the characteristics of faces would go a long way on the personalization front. Maybe easier said than done.

So far off from what actually happens. The type annotations provide an easy scaffolding for understand what the code does in detail when reading making code flow and logic less ambiguous. Reading Python functions in isolation, you might not even know what data/structure you’re getting as input… if there’s something that muddles up immediate clarity it’s ambiguity about what data code is operating on.

Appreciate the resource -- I think my biggest question from reading it: How can you actually tell if you have dust mite problems? How can you tell if they've gotten better or not? How can I tell if I've hit a point where things are about as good as they're going to get? etc.

Random question, but how consistent should you expect to get pings to be over consumer internet to something like google? Single-digit ms variance? Few ms? 10's of ms? Always kind of curious for trying to optimize my network setup at home.

See the pattern? The biggest tech successes of our era didn’t win through superior technology or innovation.

- PayPal won by creating technology that makes it easier / possible than any alternative to send and receive money via the internet.

- Uber won through creating an app that lets you hail a ride and creating a marketplace for that.

- Airbnb won through innovating on the social idea of paying someone and creating a marketplace for that.

- Facebook won through creating an innovative social networking technology and creating an ad marketplace on top of that.

- Amazon won through creating an innovative website and marketplace for goods that offered innovations to the consumer like quick shipping times and risk-free purchases via easy returns.

- Tesla won through literally creating a product (EV) and charging network for EVs that people are willing to pay for.

This article is arguing in bad faith. All of these services won because they created something that solved a problem and created value for millions or billions of people, either making plenty of money along the way, or with high potential to make money at any given point in time. If this is cheating, then all of capitalism and the innovations that scale and improve life for the world via capitalism are all "cheating". Build something people want and whatnot.

Yeah my mind also immediately jumped to fingerprinting. Somewhat required for anti-fraud to some extent, but also obviously used for more than that.

I don’t think anyone is really suggesting they stole COT or that it is leaked, but rather that the final o1 outputs were used to train the base model and reasoning components more easily.