Hacker News: joennlae

New comment by joennlae in "Artificial Analysis Intelligence Index v5"

joennlae — Wed, 28 Oct 2026 08:07:50 +0000

agreed. It is not easy to benchmark these models. It just the date of index changes have some „spurious“ correlations

New comment by joennlae in "Artificial Analysis Intelligence Index v5"

joennlae — Wed, 28 Oct 2026 06:42:39 +0000

Interesting that changes in the index always happen when „non us big tech“ models seem to be a little to good for comfort.

Anthropic disclosed an $80.7B net loss for the first half of 2026

joennlae — Wed, 21 Oct 2026 08:51:15 +0000

Article URL: https://www.wsj.com/tech/ai/anthropic-revenue-growth-ipo-accaf0f1

Comments URL: https://news.ycombinator.com/item?id=50177795

Points: 18

# Comments: 12

New comment by joennlae in "GPT-Synopsys: Frontier Intelligence to Revolutionize Chip Design"

joennlae — Thu, 01 Oct 2026 10:53:26 +0000

„The joint service offering will provide the bundled compute, model, and licenses, while ensuring customer-specific design data is protected.“

I am not sure if Nvidia want to send their chip designs to OpenAI.

New comment by joennlae in "Legal AI Harvey's margins went from +50% to -50% in 6 months on frontier costs"

joennlae — Fri, 25 Sep 2026 08:08:31 +0000

Bloomberg: https://www.bloomberg.com/news/articles/2026-09-21/startups-...

Does someone have more insight here? I thought this would happen sooner, and this is the first public sign I saw.

Token usage growth + reliance on the quality of frontier models should heavily eat into your margins?

Legal AI Harvey's margins went from +50% to -50% in 6 months on frontier costs

joennlae — Fri, 25 Sep 2026 08:08:30 +0000

Article URL: https://finance.yahoo.com/technology/ai/articles/openai-anthropic-costs-push-more-160523322.html

Comments URL: https://news.ycombinator.com/item?id=49841581

Points: 4

# Comments: 2

New comment by joennlae in "AI 'kill switch' may need to be mandatory, Anthropic co-founder tells BBC"

joennlae — Tue, 15 Sep 2026 14:14:28 +0000

EU AI Act is exactly that. Mandatory „stop button“ for High Risk Systems.

New comment by joennlae in "Apple says more ex-employees may have taken confidential data to OpenAI"

joennlae — Tue, 04 Aug 2026 19:29:10 +0000

what a find

New comment by joennlae in "Does MHz Still Matter?"

joennlae — Fri, 22 Aug 2025 20:44:55 +0000

How can I make sure that each github runner uses exactly one cpu core?

New comment by joennlae in "Llama 3 implemented in pure NumPy"

joennlae — Thu, 16 May 2024 16:14:10 +0000

Trainable Llama-like transformer (with backpropagation) in numpy only (~600 lines)

https://github.com/joennlae/tensorli

Different Sampling Strategies for LLMs

joennlae — Mon, 18 Dec 2023 14:19:02 +0000

Article URL: https://huggingface.co/blog/how-to-generate

Comments URL: https://news.ycombinator.com/item?id=38682855

Points: 1

# Comments: 0

New comment by joennlae in "Designing a SIMD Algorithm from Scratch"

joennlae — Tue, 28 Nov 2023 10:55:24 +0000

+1. Does someone know how to do that?

New comment by joennlae in "Show HN: Stella Nera – Maddness Hardware Accelerator"

joennlae — Tue, 21 Nov 2023 10:02:07 +0000

Author here: Let me try to give an overview as I saw some questions repeating itself.

* This accelerator is for an Edge/Inference case, so there is no training on this chip.

* We introduce a differentiable form of Maddness, allowing Maddness to be used in e2e training and present an application -> ResNet.

* We are still in the process of understanding how this will translate to transformers.

* The goal was to show that Maddness is feasible with a good codesign of the hardware.

* Compared to other extreme quantisation (BNN/TNN) and pruning schemes, this is more general as it replaces the matmul with an approximate matmul.

* The model architecture is not fixed in hardware. It is „just“ a matmul unit.

I hope this helps :-)

New comment by joennlae in "Show HN: Stella Nera – Maddness Hardware Accelerator"

joennlae — Tue, 21 Nov 2023 09:49:21 +0000

Thank you for the feedback :-)

We have to be careful with the comparisons we make. The TPUv3 is a training and datacenter chip and not an Edge/Inference chip. They optimise for a different tradeoff, so while the comparison looks good, it is unfair.

New comment by joennlae in "Show HN: Stella Nera – Maddness Hardware Accelerator"

joennlae — Tue, 21 Nov 2023 09:35:56 +0000

Author here:

Thank you for the feedback :-) A lot of the work regarding the comparison with „simple“ approximate matrix multiplication has been done in the preceding paper: https://arxiv.org/abs/2106.10860

While I share your enthusiasm regarding the potential, we have to be careful about the limiting factors. Our main contributions on the algorithmic side are the reformulation of Maddness such that it is differentiable (autogradable), and we can use it in e2e DNN training, as decision trees are not differentiable.

We are still in the process of understanding how to optimise the training. In the next step, we want to look into transformers as, for now, we only looked into ResNets for easy comparability.

If you are a student at ETH Zurich and want to work on this -> reach out to me

Show HN: Stella Nera – Maddness Hardware Accelerator

joennlae — Tue, 21 Nov 2023 08:06:23 +0000

Article URL: https://github.com/joennlae/halutmatmul

Comments URL: https://news.ycombinator.com/item?id=38360776

Points: 154

# Comments: 26

Show HN: Maddness – Energy-efficient DNNs without multiplying

joennlae — Mon, 20 Nov 2023 13:45:44 +0000

Article URL: https://arxiv.org/abs/2311.10207

Comments URL: https://news.ycombinator.com/item?id=38347932

Points: 2

# Comments: 0

New comment by joennlae in "Show HN: less than 650 LOC trainable GPT only using NumPy"

joennlae — Fri, 17 Nov 2023 23:29:31 +0000

That is true. I went for a simple implementation of the layer norm and included it in the tensorli definition. But it would have been better to define it as a moduli for clarity.

New comment by joennlae in "Show HN: less than 650 LOC trainable GPT only using NumPy"

joennlae — Fri, 17 Nov 2023 23:26:20 +0000

This would be interesting to consider. But at the moment, nothing is optimized, so many things must be tackled first (especially in the backwards path, for example, buffering) to justify moving to cupy. The goal was to use it as an educational exercise for me.

New comment by joennlae in "Show HN: less than 650 LOC trainable GPT only using NumPy"

joennlae — Fri, 17 Nov 2023 23:22:28 +0000

They are still applying: https://tmsearch.uspto.gov/bin/showfield?f=doc&state=4805:wl...