And for a long time (pre-AI) youtube was the biggest load on Google's entire infra. The number i recall was ~30% of all Google's cpu utilization was for youtube, and google spent a lot of effort optimizing it.

For example, it writes the whole front-end twice, once is claude design and then later it has to read it again and re-implement it in code. Also a lot of stuff (e.g. Claude.md, skill files, etc) are not supported, and they have their own set of ui-design and design systems, which claude code doesn't support.

I think Claude Design is a wonderful product, i'm just pointing that it is an independent product to Claude Code, heck even today it works better with Codex than Claude Code (that is how i use it with my own browser-use agent, i tell it to browse the design in Claude.ai/design and re-implement the design, works much better than downloading the zip file and asking the model to implement in that way.

He also saw LLM would replace search before anyone else, and that is something to look at the Lamda or GPT-1's output and think: yeah this will answer all of our questions one day.

- The price is actually competitive. They want to compete with Meta's Orion. However the product is ... lacking. None of the demos actually show-cased the dual wavelength display. The current usecases are available today at Meta Rayban Display for more cheaper and more polish.

- Dual snapdragon, and i assume one is dedicated to CV alg / scene-understanding/slam, without an external puck, i wonder how the thermal performance would look like.

- Looks very ugly, C'mon. This is the same team that desigend the original spectacle? where is Evan?

- I liked the Los Angeles text on the side. well-done.

Standard approach for training MM-LLMs is we train the encoder first, there are O(2-10B) good images on the internet, so encoder needs to see each image O(10-100) times, that is O(100T) tokens, which is more than the entire pre-training budget for most runs. That is the reason we train the encoder separately (smaller model, 2B active vs 30B or 200B active LLM); there is nothing magical about training the encoder and LLM together, it is just more token-efficient to train the image modality first.

* in our experience, in our evals and codebase, 4.6 was a bad model. This is over 60k developers, so statistically significant.

In my experience, Opus 4.0 was fantastic, major jump from 3.7. it was creative, super slow and expensive, and would sometime forget what it was doing, but it was getting the job done.

4.1 they made it much faster, so a lot of infra improvements.

4.5 was the time it could work on longer task, didn't make a lot of obvious mistakes of 4.0, and i think this was about the time the opus went mainstream, and all of the anthropic's compute crisis began, so instead of making the model better they tried to optimize it to reduce cost instead.

4.6 was such a bad model, they switched to adaptive thinking and it had so many bugs. poor api design, benchmaxxed and poor real-world results. i switched back to 4.5.

4.7 they just fixed the bugs they added in 4.6. Better than 4.5.

haven't fully tested 4.8 yet.

https://www.google.com/search?&udm=web&q=hackernews

they added some new stuff, like remote control to wherever the desktop codex app is running, but these companies need to work much more on their press releases.

Claude code, on the other hand, is the most subsidized one, both for consumers (through max subscription) and for enterprises (token discounts). It is also heavily optimized for cost, specially token caching and reduced thinking, at the expense of quality.

The Openclaw and personalized/proactive agents caused a sudden surge on Mac-Minis.

Comments URL: https://news.ycombinator.com/item?id=47980746

Points: 5

# Comments: 2