low_tech_punk

The documentation mentions error code "frontier_llm": The request could assist the development of competing AI models.

I'm very curious how do they know what requests could assist competing AI models.

You either "improve performance" or "ship more ___", never both.

Hint for mouse users, press and drag on them to hear sound and see visual effects. Mouse X/Y coordinates are inputs.

> Stream hundreds of games with XBOX Cloud Gaming

This might be the killer app if latency is low enough.

I like VSCode server in general, but it's too memory hungry. It would crash Raspberry Pi (1GB) by eating up all the RAM.

it's stupid fast!

Chinese and Russian space suits are almost identical in the photos!

iPhace Duo https://news.microsoft.com/surfaceduo/

This made history because Apple never copied Microsoft before. https://news.microsoft.com/surfaceduo/

I still enjoyed going through it, for what it is.

1. runjs promoting itself using music REPL as an example

2. author used LLM to fit a few music related ideas into the web audio API

3. they tried it with math two years ago: https://runjs.app/blog/equations-that-changed-the-world-rewr... They are clearly targeting js developers, which is one of the largest coding community

4. "The naming is not arbitrary, and the harmonic series explains why.", stop wasting tokens. Just say what it is please.

just tested it, it can be paired with a streaming setup called Voice Live. https://learn.microsoft.com/en-us/azure/ai-services/speech-s...

According to Artificial Analysis https://x.com/ArtificialAnlys/status/2095521214777442546 it is on the pareto frontier.

my hunch is that their hardware can't cache

which raises the question, is the model in the demo actually gpt-6? or it is gpt realtime 2.1? It's unclear how gpt-6 can interact at the realtime level and if so, how can developer get access to it?

What's the point of enlarging the screen into a room? In the 1979 Put That There demo, the user at least used his hand to point things. The model is impressive but the demo felt like a step back.

Original demo (fun ending) https://www.youtube.com/watch?v=RyBEUyEtxQo

No. You buy a minimum of $10 worth of credit, then use it at $1.49/M rate. There is no recurring charge.

There is a separate subscription based plan, which is sold out now.

Soniox website has a live comparison demo: https://soniox.com/compare-stt

For me, Gemini-3.5-Transcribe actually has slightly lower latency. Kudos to Soniox for both paying their competitor and letting them win. But yes, Soniox is much cheaper.

The tokens per second speed measurement is highly inflated nowadays because most of the tokens went into thinking. I wonder if there is a more realistic measurement for "effective speed", which accounts for thinking efficiency.

I started browsing the UK catalog and found a PAC-MAN edition: https://www.casio.com/uk/watches/casio/product.CA-53WPC-1B/

They already sell API access to the model on Google/Microsoft/Amazon cloud. OpenRouter is the same deal. Note that the labs are not giving out model weights to OpenRouter. They are just selling tokens.

It is decoupled from drivetrain. So, without torque sensor, how does it know when the rider wants to speed up vs. slow down?

Maybe Google is adding compute infra. I noticed today that the API has 500 tok/s throughput. It's quite compelling now.

> Three things matter today: Model. Data. Harness.

*knee jerk*

It's a classic scam. It is a 3.0 USB 7-Port Hub, not a 7-port USB 3.0 Hub. The number 3.0 is the model number of the hub, not the protocol.

According to the stats in this article https://www.ycombinator.com/investors, we can estimate 98.5% of the founders are not in OpenAI or Anthropic.

Is there a selection bias?

SELECT * FROM yc_founders WHERE employer IN ('OpenAI', 'Anthropic');

If there are 7k founders [1], the graph only shows 1.5% of the people from YC.

[1]: https://www.ycombinator.com/investors

Interestingly, I found the original nano banana also has the best latency/quality trade-off that new versions can't beat. This might be domain/prompt specific though. I wonder if there is some truth in the saying that something is either new or improved by never "new and improved".

Agree that good tool should be invisible. We want essential not accidental complexity in how the tool works.

But good tool should also be fun and makes us feel productive. We can't neglect the emotional aspects of designs. And at the end of the day, if a less productive tool makes us much happier, we will less likely be burned out. That is productivity in the long term.

Maybe only AI Agent doesn't care about the emotional aspects fro tool use, but that's a separate topic.

Also, it's not about steep learning curves. We want low floor, high ceiling tools. Some of the examples the author used are either low floor low ceiling, or high floor high ceiling. Neither is ideal.

(1) The original zig code was probably heavily AI coded with lower quality, hence the bag. There is a chance that a full rewrite in zig might do as well but we will never find out.

(2) The rewrite itself is a massively successful marketing move. It shows what Claude code can do and how little it costs compared to human engineering. But the question remains whether someone else, not knowing zig, rust, and TypeScript can pull this off.

If you click on the "create link" button multiple times, the menu would pile up vertically.