the_king

I cannot wait for Starship to become a real thing, but you have to admit it's way behind schedule. The engines are awesome now.

This it is very impressive. But scrolling through the preprint, I wouldn't call any of it elegant.

I'm not blaming the model here, but Python is much easier to read and more universal than math notation in most cases (especially for whatever's going on at the bottom of page four). I guess I'll have one translate the PDF.

I just tested it on a very difficult Raven matrix, that the old version of DeepThink, as well as GPT 5.2 Pro, Claude Opus 4.6, and pretty much every other model failed at.

This version of DeepSeek got it first try. Thinking time was 2 or 3 minutes.

The visual reasoning of this class of Gemini models is incredibly impressive.

I think good names and a good file structure are the most important thing to get right here.

Totally understand, thanks for being a customer. I'm sorry we weren't able to make the web version as smooth as we wanted to.

We do plan to support Linux. This was probably a little bit of a blind spot for us - not realizing that anyone running a Linux desktop doesn't even have system voice tool to fall back on.

I think Aqua v1 had two problems:

1. The models weren't ready.

2. The interactions were often strained. Not every edit/change is easy to articulate with your voice.

If 1 had been our only problem, we might have had a hit. In reality, I think optimizing model errors allowed us to ignore some fundamental awkwardness in the experience. We've tried to rectify this with v2 by putting less emphasis on streaming for every interaction and less emphasis on commands, replacing it with context.

Hopefully it can become a tool in the toolbox.

you've got nothing to worry about here.

if firebase studio can make a todo app, i'd be surprised. this is the worst "vibe coding" tool i've ever used.

We can do that, with some help from the community.

thanks!

I share the same sentiment. I remember thinking in college how annoying it was that I was reading low-resolution, marked-up, skewed, b&w scans of a book using Adobe Acrobat while CS concentrators were doing everything in VS Code (then brand new).

but we do think voice is actually great with Cursor. It’s also really useful in the terminal for certain things. Checking out or creating branches, for example.

Aqua is in another league when it comes to accuracy. I just ran them side by side on a simple q to ChatGPT and here were the results...

Aqua Voice

  What is the first recorded eclipse in human history? I'm not asking when the first one occurred, but the first written record we have of an eclipse.

  What is the first recorded eclipse in human history i'm not asking 1 like the first occurred but the first ridden record we have of an eclipse

-"1" should be "when"

-"ridden" should be "written"

-No punctuation

thanks! We're working on iOS, but it's tough to get the ergos right given all of Apple's restrictions and neglected APIs.

I was excited to try this out because I've had a lot of trouble getting the Supabase integrations to work on Lovable and Bolt.new.

Sorry to say that Firebase Studio did an awful job. It did not successfully build even the first view of the app I asked for. It feels like I'm stepping back to release day of GPT-4.

Am I missing a switch to use the good Gemini 2.5 somewhere? I could tell from their response speed that I was not using a thinking model.

We're building something different, but there is some overlap. Aqua is built for max speed, while keeping accuracy high. To achieve that, inference runs in a datacenter (for now).

You can customize Aqua using custom instructions, similar to ChatGPT custom instructions, and get some Talon functionality from it:

In my own, I have:

1. Breaking the paragraphs with three or four sentences.

2. Don't start a sentence with "and".

3. Use lowercase in Slack and iMessage.

4. Here are some common terminal commands...

Thanks!

We're faster, more accurate, and have a streaming option. Aqua can go from key-up to paste in as little as 450ms. Flow was closer to 1000 in our tests.

Overall, you'll notice we make a few more tweaks to the output than Wisprflow.

For example, Aqua + Cursor is very powerful - we syntax highlight your transcript. The easiest way to see this is to use streaming mode (double press Fn) + deep context + cursor and try asking it to change something.

This also works in other "context rich" environments.

We're more accurate and much faster than Mac Whisper, even their strongest model (Whisper Cpp Large V3).

For that benchmarking table, you can use Whisper Large V3 as a stand-in for Mac Whisper and Super Whisper accuracy.

Shoot, might be due to AirPods mic init latency. AirPods work well on Desktop (though their mic quality isn't the best).

This is well written. I appreciated the line, "startups are a race to learn stuff."

Files-to-prompt has been a surprisingly useful tool for this kind of workflow.

https://github.com/simonw/files-to-prompt

Reminds me of some underrated Peter Thiel advice: almost always an extra hour invested in a project you’re already deeply into will have greater returns than starting some totally tangential thing from scratch.

The tangential thing will always feel better because you make rapid progress at the start, but the real returns come from completion.

It may turn out that in the “next epoch,” our stock market will be understood as our civilization's greatest work of art - our greatest cathedral to capitalism.

Sounds more like they were (and still are) being sloppy. “be better” is one thing. “runs without fatal crash” is what semi is talking about.

It’s not just that it’s immature software, they need to change how they do development.

I remember geohot saying something similar about a year ago

Yeah, they are here: https://withaqua.com/blog/benchmark-nov-2024

Yes! It's fully real time. This was a challenge, but something we decided not to compromise on.

Imagine a pen that was only showing you words at the end of a paragraph — that'd be crazy! But this is how most people use voice today because real time is so hard. Latency is 800ms-2000ms from speech.

I would love to see the dashboard that the team that made the decision was looking at.

I'd be interested to hear speculation by people who know about this as to what they think went wrong. Was it off course? Did the engines not relight in time? Did it not have enough fuel?

Anthropic's latest 3.5 sonnet is a cut above GPT-4 and 4.0. And if someone had given it to me and said, here's GPT-4.5, I would have been very happy with it.

Exactly my point. Saying something is the fastest ever is marketing code (at least to me) for minor improvement over the previous generation.

If you're 30% faster than the previous generation, I'd rather see that because my assumption is it's 5%.

That's only single core. I think Logic is pretty optimized to use multiple cores (Apple demoed it on the 20 core Xeon Mac Pro back in 2019).

But if the M1 isn't the bottleneck, no reason to upgrade.

Call me cynical, but when I see headlines like "up to 2x faster", I assume it's a cherry-picked result on some workload where they added a dedicated accelerator.

There's a massive difference between "pretty much every app is 80% faster" and "if you render a 4K ProRes video in Final Cut Pro it's 3x faster."

To run LLMs locally (Ollama/LLM Notebook), you want as much memory as you can afford. For actually training toy models yourself for learning/experiments in my experience it doesn't matter much. PyTorch is flexible.