arnaudsm

Genuine question : why are hyperscalers like OpenAI and Oracle raising hundreds of billions ? Isn't their current infra enough ?

Naive napkin math : a GB200 NVL72 is 3M$, can serve ~7000 concurrent users of gpt4o (rumored to be 1400B A200B), and ChatGPT has ~10M concurrent peak users. That's only ~4B$ of infra.

Are they trying to brute-force AGI with larger models, knowing that gpt4.5 failed at this, and deepseek & qwen3 proved small MoE can reach frontier performance ? Or is my math 2 orders of magnitude off ?

This is exciting, exceptionally the firmware & binary blob foundations that are the biggest roadblock.

Concerning the UI, I wish we had another attempt at a web-based mobile OS. FirefoxOS was too early, but APIs are much more mature now, and WASM offers great performance for low-level stuff. I might work on this full time when I retire.

Why would you need bipedism for a flying object ? Wouldn't a quadripod + quadcopter setup would be much more stable and cheaper ?

Are you referring to the llama 4 cheating ? I cannot find any reference to llama 1 cheating

Yep, it's unfortunately the industry standard, but in my career I've seen it lead to mixing up ids and painful debugging.

My other favorite best practice with SQL is to have unique ID column names, so you never mix up IDs, and can use JOIN with the USING syntax:

  SELECT *
  FROM customer
  JOIN order USING (customer_id)

It's a great tool, but the mobile experience isn't great. Everytime the socket connection fails in the background, and i have to restart and refresh the app twice to get my results.

As a kid I didn't have a phone, but my parents gave me a cheap iPod Nano. I flashed a custom OS on it (RockBox), which used Morse code as the primary input, since iPods only have a single button.

Programming in lua and writing my diary in Morse code on this microscopic device was a lot of fun.

Morse was a clever way to bypass limitations, which is the definition of hacking itself.

We need numbers. Can someone benchmark the wattage difference of using Liquid glass vs disabled ?

I absolutely love Magic Lantern, and I wish similar initiatives existed on Sony and Nikon! I was forced to upgrade my Sony camera purely because of software limitations.

Firmwares should be open-source by law. Especially when products are discontinued.

Intel's single core performance has 3.4xed in 15 years (980X vs 285K)

Single core perf doubled every 8 years, multicore every 6 years, and GPUs every 3 years !

Web and GNU/Linux

Interestingly, this is 10x cheaper than the cheapest provider on OpenRouter : https://openrouter.ai/deepseek/deepseek-r1?sort=price

Inference is more profitable than I thought.

It's too late. As a developer, I'm pulling all my Android apps away from the Play Store.

If Google is hostile to me an my users, I prefer to dedicate my volunteer time to respectful plateforms instead.

The 25% error rate is for 100k+ vocabulary, it drops under 10% for smaller vocab. Meaningful bits per minute would be a better metric indeed.

I appreciate comments like yours that actually contribute to the debate. We need critical thinking and data. Not one-sided puff pieces out of context.

I was comparing Robotaxis with Waymo in Texas.

If you want to compare Teslas with consumer cars, the best metric we have is the fatality rate per mile. Tesla is #1.

I agree significant bits per minute are a better metric, I vulgarized a bit too much.

[flagged]

Most Android apps are crapware anyways. The only respectful apps that I know are open-source, and are being kicked out the of play store progressively.

I'm cancelling my Pixel 10 preorder.

Did you use o3 pro high, o3 high or o3 medium?

(OpenAI's naming is so confusing)

Why are so many AI demos about automating parenting?

Just spend more time with your kids, they want connection!

Who's the target of 120B open-weights models? You can only run this in the cloud, is it just PR?

I wish they released a nano model for local hackers instead

I feel the same, but cannot measure the effect in any context benchmark like fiction.livebench.

Are they aggressively quantizing, or are our expectations silently increasing ?

Meme benchmarks like this and Strawberry are funny but very easy to game, I bet they're all over training sets nowadays.

Even flagships like o3 & gemini_2.5_pro ?

I've lived the exact same scenario in a large public company. Large orgs and misaligned incentives are not exclusive to the public sector.

Cursor is a paid product and a company, VSCodium is an open-source project running on volunteers.

Microsoft claims their team banned the extension in "2 seconds" https://x.com/code/status/1943720372307665033?s=46

Concerning Apple, their review process is so hard and unjust, I've seen startups give up apps after months of work just because of that.

Maybe sandboxing and runtime-level permissions are a better compromise?

Context : Cursor, despite raising $900M, is a vscode fork that uses the open-vsx extension registry. It is maintained by european volunteers at a non-profit, and does not have the resources to check for supply-chain attacks like this.

Freeloading on (and blaming) volunteer infrastructure is irresponsible, especially when you have so much funding.