Hacker News: jumpCastle

New comment by jumpCastle in "OpenAI o3 and o4-mini"

jumpCastle — Wed, 16 Apr 2025 20:06:06 +0000

It was a good benchmark until it entered the training set.

New comment by jumpCastle in "OpenAI o3 and o4-mini"

jumpCastle — Wed, 16 Apr 2025 20:02:57 +0000

But open source like aider

New comment by jumpCastle in "The Llama 4 herd"

jumpCastle — Sun, 06 Apr 2025 03:22:34 +0000

Deepseek v3 was FP8

New comment by jumpCastle in "The Llama 4 herd"

jumpCastle — Sun, 06 Apr 2025 03:19:50 +0000

It's quite similar to muP

https://github.com/microsoft/mup

New comment by jumpCastle in "Humans have caused 1.5 °C of long-term global warming according to new estimates"

jumpCastle — Sun, 17 Nov 2024 22:33:26 +0000

Stock prices cannot go up without if the planet is destroyed

New comment by jumpCastle in "Nvidia reportedly delays its next AI chip due to a design flaw"

jumpCastle — Sun, 04 Aug 2024 16:14:28 +0000

Attention was invented because Bengio lab had to be disciplined about a black box (google had more compute)

New comment by jumpCastle in "Exo: Run your own AI cluster at home with everyday devices"

jumpCastle — Tue, 16 Jul 2024 14:40:19 +0000

Aren't services like runpod solve half of these concerns?

New comment by jumpCastle in "Firefox 128 enables "privacy-preserving" ad measurements by default"

jumpCastle — Mon, 15 Jul 2024 13:21:15 +0000

Are the updates really slow?

New comment by jumpCastle in "Ilya Sutskever to leave OpenAI"

jumpCastle — Wed, 15 May 2024 01:30:47 +0000

Spend more time with my side projects.

New comment by jumpCastle in "2023 ACM Turing Prize awarded to Avi Wigderson"

jumpCastle — Wed, 10 Apr 2024 20:51:01 +0000

You can try his book. https://www.math.ias.edu/avi/book

New comment by jumpCastle in "GPT-4 Turbo with Vision is a step backwards for coding"

jumpCastle — Wed, 10 Apr 2024 14:59:07 +0000

You can use older models with the api.

New comment by jumpCastle in "How Chain-of-Thought Reasoning Helps Neural Networks Compute"

jumpCastle — Fri, 22 Mar 2024 13:57:56 +0000

Also the parameters are optimized also with loss of future tokens in the sequence.

New comment by jumpCastle in "Learning Theory from First Principles [pdf]"

jumpCastle — Sun, 03 Mar 2024 15:12:07 +0000

Work pretty well for all classes of problems?

New comment by jumpCastle in "Hi everyone yes, I left OpenAI yesterday"

jumpCastle — Wed, 14 Feb 2024 03:06:04 +0000

Rsu exist before ipo?

New comment by jumpCastle in "Memory and new controls for ChatGPT"

jumpCastle — Tue, 13 Feb 2024 23:50:17 +0000

Use api, embed history and retrieve.

New comment by jumpCastle in "OpenAI suspends ByteDance's account after it used GPT to train its own AI model"

jumpCastle — Sat, 16 Dec 2023 14:00:51 +0000

Use model output as training data. For better performance you can get some top log probs and minimize kl divergence.

New comment by jumpCastle in "Mistral: Our first AI endpoints are available in early access"

jumpCastle — Mon, 11 Dec 2023 15:09:34 +0000

Without open weights why would anyone care about them? By the time they could compete with gpt4 there's probably be gpt5 already.

New comment by jumpCastle in "GPT-4 generates simple app from Whiteboard photo"

jumpCastle — Wed, 27 Sep 2023 21:52:23 +0000

The future is the AI also writes the high level descriptions of stuff.

New comment by jumpCastle in "Exllamav2: Inference library for running LLMs locally on consumer-class GPUs"

jumpCastle — Wed, 13 Sep 2023 15:17:06 +0000

But you can fine tune gpt 3.5 turbo, so your comparison is not clear.

New comment by jumpCastle in "Discord.io breached, 760k user accounts for sale on darknet"

jumpCastle — Mon, 14 Aug 2023 22:56:40 +0000

I still use infinity for reddit though.