parentheses

Agree on this point. Recent anecdotal testing I did found Whisper is still better than Parakeet.

Ghostty has crashed nightly for me with ~10 terminals open across a few windows. So, I haven't been able to run it nor would I want to embed it inside anything I daily drive.

You're right in my neighborhood. Signed up! Fantastic!

Building a team to operate based on your own personal preferences is selfish leadership... or even dictatorship.

There's a very strong "focus culture" which relies on the idea that work is not done in meetings. This is wrong. Progress comes in many forms.

With AI it feels writing software that is open is less attractive. It's hard to trust OSS made recently b/c you can tell if someone knows what they're doing and even spent any time on quality. Also, often times people don't reach for software others make (unless it's boring and old stuff, in which case this advice doesn't apply.)

IMO types are the main lever you can use other than procedural abstraction. I feel that Haskell gives you both in a way that marries them for maximum constraint-building. Constraints that prevent illogical or illegal programs are the bread and butter of reliable software.

I've been poking at running LLMs in the browser. It feels like we're definitely close (<1 year) to seeing real use cases there.

Ubiquity and coverage of devices is what will take longest. Largely dependent on how well we can shrink models with similar performance and how much we can accelerate mobile devices. This feels like it's but further (<3 years?)

Reading only the abstract: LLMs prefer output of their own generation over humans or even other models.

This is a very good reason to avoid using model-generated data to train future models. We'd be deepening this bias by continuing to do that, essentially forcing society to reshape their output using LLMs to increase engagement. This feels like a form of enshittification that doesn't just touch one product but all of society.

I have been pondering this for a while. Cat's out of the bag.

Maybe the better way to author your work is to:

1. Write what you want

2. Loop through a random set of "tumbler" skills that preserve meaning

3. Finally pass the output through a "my style" skill that applies what you about

In order for this to work the "my style" would have to be a very common-place style.

Legit first post I've wanted to upvote in a long time.

This type of thing requires an economic driver to monetize the service.

I'd have a strong inclination to run such software if I knew that I was both helping host repos and getting paid.

I feel like it's not news that a company with (probably) millions of DAU is not able to handle a single case like this one.

At the same time, it's clear that after this happened, Anthropic took action. 3 DAYS AGO! (https://news.ycombinator.com/item?id=47954655)

That's before this comment was made on the issue:

https://github.com/anthropics/claude-code/issues/53262#issue...

I'm surprised Anthropic didn't also say this on the issue. Weird that they wouldn't. It seems to have made for unnecessary bad PR.

It feels to me that Anthropic is less focused on quality, and more focused on PR stunts/flash. My experience with Claude is always "it's pretty and feels cool", where-as codex feels like "solid and boring". I realize I'm probably biased. Am I alone in this thinking?

The timing makes me wonder if this is a direct response to Deepseek V4 having performance comparable to SOTA models.

I built a small tool that makes one cleanup commit at a time, keeps it only if tests pass, and moves on. The article is about how it grew from that basic loop into taste files, staged migrations, and a way to keep repos getting a little cleaner in the background.

https://github.com/bigH/continuous-refactoring

Been tinkering on my personal site and wanted to add some AI features without being the one paying for tokens on every visit. So I went looking at how close browser-side inference actually is to practical. Closer than I expected.

The post is less a tutorial and more me walking through what it felt like. Kicking tires on random models, getting one to actually run, then doing the small unglamorous work to turn "demo" into something I'd put in front of a reader.

Related: https://news.ycombinator.com/item?id=47851111

I added live translation to this site, wrote about the weird edges, then made this little playground to mess with them directly. Proper nouns, tags and markdown syntax were always coming through weird, so this helped me investigate.

TranslateGemma was one of the few models I tried. Something interesting I noticed is that each model has a different expectation of interface (fields passed in, naming of those fields and such.)

HN Post of Article:

https://news.ycombinator.com/item?id=47808030

Been tinkering on my personal site and wanted to add some AI features without being the one paying for tokens on every visit. So I went looking at how close browser-side inference actually is to practical. Closer than I expected.

The post is less a tutorial and more me walking through what it felt like. Kicking tires on random models, getting one to actually run, then doing the small unglamorous work to turn "demo" into something I'd put in front of a reader.

This tool is not just used for safety. ;)

You can spoof or disappear a mashed file. You can trigger vulnerabilities by breaking internal assumptions of a program.

I am only trying it out for now. But what I'd love is for things to be quite automatic. I join a zoom, get a prompt "what do you want? record, record+transcribe or record+transcribe+summarize (or use some custom prompt)"

I just downloaded and I'm presented with a seemingly required google login.

I really appreciate that this is free, but I do feel like the privacy-first approach is incompatible with requiring google login.

edit: FWIW, I bought MacWhisper and would buy this if it didn't require the Google login.

Fully agree to this. I find the cost of cloud providers is mostly driven by architecture. If you're cost conscious, cloud architectures need to be up-front designed with this in mind.

Microservices is a killer with cost. For each microservices pod - you're often running a bunch of side cars - datadog, auth, ingress - you pay massive workload separation overhead with orchestration, management, monitoring and ofc complexity

I am just flabbergasted that this is how we operate as a norm in our industry.

I think this is the kind of investigation that AI can really accelerate. I imagine it did. I would love to see someone walk through a challenging investigation assisted by AI.

What the commenter is saying pertains to the _decision_ to use Android. That is why this is happening. That is NVidia.

I look at this as the equivalent of writing a MUD as you ladder up to greater capabilities. MUDs are a good educational task.

Similarly AIs are just putzing around right now. As they become more capable they can be thrown at bigger and bigger problems.

This dynamic would create even more gate-keeping using credentials, which is already a problem with academia.

Totally agree!

I feel like this means that working in any group where individuals compete against each other results in an AI vs AI content generation competition, where the human is stuck verifying/reviewing.

It feels generally a bit dangerous to use an AI product to work on research when (1) it's free and (2) the company hosting it makes money by shipping productized research

I suppose in that position your head has lower elevation, allowing for better circulation.

I got some errors trying to run this on my MBP. Claude was able to one-shot a fix.

``` Loaded speech tokenizer from ~/.cache/huggingface/hub/models--Qwen--Qwen3-TTS-12Hz-1.7B-VoiceDesign/snapshots/0e711a1c0aa5aad30654426 e0d11f67716c1211e/speech_tokenizer Fetching 11 files: 0%| | 0/11 [00:00<?, ?it/s]Fetching 11 files: 100%|| 11/11 [00:00<00:00, 125033.45it/s] The tokenizer you are loading from '!/.cache/huggingface/hub/models--Qwen--Qwen3-TTS-12Hz-1.7B-VoiceDesign/snapshots/0e711a1c0aa5aad30654426e0d11f67716c1211e' with an incorrect regex pattern: https://huggingface.co/mistralai/Mistral-Small-3.1-24B-Instr.... This will lead to incorrect tokenization. You should set the `fix_mistral_regex=True` flag when loading this tokenizer to fix this issue. ```

With the recent success of AI, I feel the more insidious issue is preventing the use of AI in reading paper ballots. There's a lot of room to engineer bias.