HNHacker News
TopNewBestAskShowJobs

aleksiy123

899 karma · joined November 2, 2021

submissionscomments
aleksiy123··on Moving away from Tailwind, and learning to structure my CSS
that’s exactly the part that is anti locality of behaviour.

I don’t want/need cascade. I only care about components and building up from them. And I would rather have it be explicit over implicit and scoped/encapsulated.

Call it composition over cascade.

To be clear I think it’s possible to do this without tailwind. And tailwind has other out of the box features/opinions.

But it works well enough without too much friction.

aleksiy123··on Moving away from Tailwind, and learning to structure my CSS
Sorry maybe that wasn’t the best example. It’s not really about separation of files. But how they connect.

In sql your code may be in a seperate file but your app code is still clearly calling the sql. The inlining vs not inlining is just abstraction. You could use a function, or a separate file or not, a different language or not.

But there is a clear single call chain at the points where that behaviour is being applied and a single definition.

With css that’s not necessarily true. There’s a bunch of different rules that may or may not apply.

aleksiy123··on Moving away from Tailwind, and learning to structure my CSS
I think this assertion is where most of the conflict comes from.

There is a fair amount of people that disagree with the premise that it should be separated in that way (Including me).

I personally like this essay by the author of htmx on the topic

https://htmx.org/essays/locality-of-behaviour/

Also just better composition imo.

Practically I think this means components of scoped css, html, js.

People never seem to have the same complaint about mixing or separating app code and sql in the same way?

aleksiy123··on Mode collapse has a name, and he's selling cancer treatment advice on Amazon
Tbh I have no idea, I’m mostly thinking about it from what I can do when using the frontier models, so I don’t think such low level changes are available to me.

But another dumb idea I had was a set of random words inspired by Terry Davis godsay https://github.com/orhun/godsays

With a more appropriate wordlist appropriate. Call it muses.

aleksiy123··on Mode collapse has a name, and he's selling cancer treatment advice on Amazon
I’ve been thinking about how to get creativity out of llms, apart from temperature. The thing is even humans have a hard time with creativity.

Is it really surprising that llms don’t just one shot a unique story when they are all starting from roughly similar training data and state and a roughly 30 secs of processing time.

I had Gemini do some deep research for me around processes and frameworks to prompt ideation and creativity and they do exist. See SCAMPER and others.

Another interesting thing that comes up is using random decks of cards as prompts.

See Oblique Strategies, Deck of Lenses, the Story engine and similar.

I guess I still believe that even creativity is still fundamentally a type of search, as well as problem solving. Manipulating and or combining existing ideas in unexplored ways and breaking out of bias.

So I kinda want to experiment with these two approaches:

1. Longer running workflows that follow a framework and loop.

2. Some simple cli tools with these decks and a random draw to trigger interesting directions.

I think really just need to break llms out of their initial start state which is mostly the same for everyone.

And to run over longer horizons and so the higher level reasoning flows.

aleksiy123··on A History of IDEs at Google
code monkeys
aleksiy123··on A History of IDEs at Google
thats just a large project tho? not necessarily a mono repo.

A mono repo doesn't necessarily mean large compile times, because it depends on the projects and their dependencies within that repo.

aleksiy123··on I returned to AWS, and was reminded why I left
Went through a similar phase,

I think a mix of 2. and 3. is good for a small team or solo dev. Im throwing in a bit of homelab as well by adding some action runners and models on my desktop as well.

But cloudflare is great value for small teams. Not sure how it as at higher scale.

On the topic of env and config. It took me a while to get this write, and maybe overengineered.

But I invested a lot of time in trying to standardize env definitions, secrets manager, and per env config definition defined in my nx projects, and consumed by the commands or deployers. As well as pulumi for IaC.

I tried a couple of different approaches, but finally I just decided to use typescript as my config language. I use nx project.json but defined using typescript. And just define the env config as typescript functions to be injected to each command or deployment as a pure function of target env.

aleksiy123··on Bun's experimental Rust rewrite hits 99.8% test compatibility on Linux x64 glibc
It’s also just a useful exercise in general, especially for getting feedback for models and harnesses.

I’ve been thinking about setting up a non trivial project to use as a benchmark for any plugins and/or harness changes I make.

Having a prebuilt verification suite is great. You can use it to asses things like token usage, time, across different harnesses, models, plugins.

aleksiy123··on Agents need control flow, not more prompts
Yes but more declarative vs imperative.

I say what the llm says how.

aleksiy123··on Agents need control flow, not more prompts
It’s already here the question is just to what extent?

Are google search results modifying your software at runtime?

Take or agent chat for example, the output text is a ui, agents can generate charts and even constrained ui elements.

Isn’t that created and adapted at run time?

If you mean like agents live modifying your code. I think that’s pretty much here as well. Can read the logs and send prs.

The only thing is how fast that loop will execute from days or hours to mins or seconds, and what validation gates it needs to pass.

My git repo is pretty much self modifying personal software at this point, that I interface through the ide chat window.

But I don’t think we will ever lose the intermediary deterministic language (code) between the llm and the execution engine.

It would be prohibitively expensive to run everything through models all the time.

But I am starting to think we need a more precise language than English when talking with LLMs. That can do both precision and ambiguity when you need either.

aleksiy123··on Agents need control flow, not more prompts
There’s a whole range between completely random and completely rule based deterministic.

Somewhere in between that I guess is the varying levels of intelligence more likely able to make the “right” decision for anything you throw at it.

aleksiy123··on AlphaEvolve: Gemini-powered coding agent scaling impact across fields
Agreed, however imo there is def some problems unique to Google which is making the internal experience less than ideal.

Hoping they can figure it out sooner rather than later.

aleksiy123··on AlphaEvolve: Gemini-powered coding agent scaling impact across fields
Self improving, doesn’t necessarily imply singularity right?

There still could be hard constraints to make singularity intractable or just such a long time horizon it’s not practical right?

aleksiy123··on Agents can now create Cloudflare accounts, buy domains, and deploy
I was wondering if someone was going to allow payments through CLI at some point.

But jokes aside having a central place to manage billing and accounts for deploying infra across multiple providers is pretty awesome imo.

if they have a terraform provider even better. I wonder if also makes multi tenant architectures or environment isolation easier to provision as well.

aleksiy123··on Accelerating Gemma 4: faster inference with multi-token prediction drafters
Yes, but I think Google was playing that strategy from essentially day 1 or very early in this AI race, where as the others are there now because of their lack of access of compute.

The general narrative I would read on HN/others, was that Google would be able to outlast/outcompete OpenAI and Anthropic because Google had both more money and more compute. Playing the game of subsidizing their most capable models to capture market share longer than the VCs could.

But instead I feel like Google opted out of that much earlier. Shifting their focus on efficiency and scaling much much earlier. Flash and Gemma being where Google was actually ahead of the competition while everyone was focused on bigger more capable models.

In the last month the environment has changed, compute is constrained, costs for consumers are way higher than expected. Copilot pretty much imploded, and I'm guessing both Anthropic and OpenAI are starting to feel the squeeze.

My personal opinion was this was necessary because integrating AI into products like AI overview, search meant scaling to billions of users was a requirement right out of the gate. And theres not enough money/compute no matter who you are to use frontier models for that.

aleksiy123··on Accelerating Gemma 4: faster inference with multi-token prediction drafters
I’m starting to think that googles strategy is a bit different then the other frontier providers.

Focusing more on performance to compute efficiency over pure performance. And maybe that’s why Gemini is (seemingly) lagging behind?

Other providers hitting capacity and hitting the limits subsidising their inference.

Google strategy seems to be about scaling and distributing these models to their existing billions of users.

aleksiy123··on City Learns Flock Accessed Cameras in Children's Gymnastics Room as a Sales Demo
I feel like I’m missing something.

There is someone that is making the decision right?

Or are you just saying the person placing the cameras is decoupled from the person making the decision to aggregate them all.

But I still feel like the accountability is on who is giving the access to sensitive cameras.

aleksiy123··on The gay jailbreak technique (2025)
Yes, but more specifically putting them into a sort of contradiction of their beliefs or arguments.

Doesn’t even have to be correct, but it can be confusing and cause people to say something they don’t actually mean if they dont stop and actually think it through.

aleksiy123··on City Learns Flock Accessed Cameras in Children's Gymnastics Room as a Sales Demo
While I think this isn’t great.

Why is the camera there in the first place??

Presumably there are people that have access to it. And if you are demoing software that connects to cameras, then someone gave the sales guy access to those cameras.

I’m also assuming those probably weren’t the only cameras…

aleksiy123··on The gay jailbreak technique (2025)
Probably a mix of heuristics, keywords and simple ml model.

Then maybe a second gate with a lightweight llm?

Edit: actually Gcp, azure, and OpenAI all have paid apis that you can also use.

But I don’t think they go into details about the exact implementation https://redteams.ai/topics/defense-mitigation/guardrails-arc...

aleksiy123··on The gay jailbreak technique (2025)
Does this still work on newer models?

The reasoning on why it works is pretty interesting. A sort of moral/linguistic trap based on its beliefs or rules.

Works on humans as well I think.

aleksiy123··on Ghostty is leaving GitHub
Where is the arrogance? and I thought talk/slides where interesting.

And wdym even mean "lead the direction of tech". Its just people trying to build a product based on their views/vision.

Others are free to build their own competing visions? and everyones free to choose the platforms that they use.

aleksiy123··on Claude.ai unavailable and elevated errors on the API
My guesses are:

hardware capacity constraints is going to be the big one

Effective caching is another, I bet if you start hitting cold caches the whole things going to degrade rapidly.

The ground is probably shifting pretty rapidly.

Power users are trying to get the most out of their subscriptions and so are hammering you as fast as they possibly can. See Ralph loops.

Harnesses are evolving pretty rapidly, as well as new alternatives harnesses. Makes the load patterns less predictable, harder to cache.

The demand is increasing both from more customers, but also from each user as they figure out more effective workflows.

Users are pretty sensitive to model quality changes. You probably want smart routing, but users want the best model all the time.

Models keep getting bigger and bigger.

On top of that they are probably hiring more onboarding more, system complexity and codebase complexity is growing.

aleksiy123··on GitHub Copilot code review will start consuming GitHub Actions minutes
Yeah, I’ve been thinking in this direction as well.

My wife uses my old gaming desktop for her ux design work as well.

And I was thinking of using the gpu to run some tts models.

Now to just figure out a way to run it all on windows and have it auto start when she logs in.

aleksiy123··on GitHub Copilot code review will start consuming GitHub Actions minutes
Yeah, I did hertzner runner for a bit.

But I think the issue is that my situation (solo dev, mono repo) is just not right for a dedicated instance.

With only 1-2 runners, the pipeline is slow (low parallelism) and resource constrained. And at least 50% of the time its idle (I'm not working/sleeping).

I guess what I'm really looking for is for some kind of aggressive autoscaling, and aggressive caching.

I tried a couple of things (GHA, Dagger + Hertzner, Buildkite)

And Im just not too sure theres going to be any out of the box solution since my priority is essentially to minimize cost and maximize efficiency. Not really a great customer for any providers.

Im tempted to just get agent to build something out quickly with cloudflare workers + spot instances.

I also have some other nice to have requirements:

- ts/code over config

- locally runnable and testable

- preferably no lock in

- repeatable/reproducible

aleksiy123··on GitHub Copilot code review will start consuming GitHub Actions minutes
Anyone have good alternatives for ci/cd on the cheap for a solo dev?

I’m blowing through my 1000 mins in days.

Thinking to either pool some free tiers or figure something out with spot instances.

Also is it just me or is CI/CD tooling still sort of rough all around.

aleksiy123··on AI should elevate your thinking, not replace it
I think it’s because there is a communication and alignment inherently in the process now.

You went to a single entity (you) implementing an idea.

To two entities (agent) + (you) having to align through language.

My metaphor is it’s going from single process (1) to distributed system (2) with all the tradeoffs.

It’s just a whole different problem now that was originally only limited to other teammates.

aleksiy123··on Tendril – a self-extending agent that builds and registers its own tools
I guess the interesting part is the forces tool writing part.

Which kind of solves the when should we write a tool part by just saying always.

But I think the question is how will this scale. The real core issue I feel like I’ve been encountering is scaling complexity.

Reducing the number of tools without losing efficiency or capability.

Reducing duplication, abstracting, cleaning up, and maintaining knowledge and memory.

I think the issue for me has been threefold.

1. As the repo grows how does you make the agent keep understanding of it without excessive context pollution.

2. How do you maintain memory and knowledge over time.

3. How do you know the agent is performing better over time and not regressing as you evolve.

And what has somewhat been working for me is

A) trees or hierarchies.

Trees scale well. Folder structure but also in the form of just simple indices.

Logical structure and locality makes them even more effective.

B) caching.

Having the agents “cache” their thinking in the form of summaries, skills, tools.

Recursive summarization really helped with mono repo navigation for me.

But right now I still feel like I need to be constantly prompting them and I can’t quite close the feedback loop.

aleksiy123··on Using coding assistance tools to revive projects you never were going to finish
On the topic of procedural, one thing I experiment with is having the llm part of the procedural loop.

Sort of writing a narrative on top live.

Unfortunately, local models are still a bit slow and weak but was interesting to see what it came up with nonetheless.

← PreviousPage 4 of 16Next →