Hacker News: gpjt

New comment by gpjt in "Ask HN: What are you working on? (August 2026)"

gpjt — Sun, 09 Aug 2026 18:53:08 +0000

I'm adding MoE support to the GPT-2 code from Sebastian Raschka's "Build a Large Language Model (From Scratch)". Just the minimal changes. Planning to do (or at least start!) a from-scratch base model training run on my own hardware before the end of the month.

A Quick(ish) Chinchilla Check

gpjt — Fri, 07 Aug 2026 19:38:55 +0000

Article URL: https://www.gilesthomas.com/2026/08/chinchilla-check

Comments URL: https://news.ycombinator.com/item?id=49215301

Points: 1

# Comments: 0

New comment by gpjt in "I use AI on this blog"

gpjt — Fri, 31 Jul 2026 19:37:04 +0000

Lovely bit of inadvertent editorialising by the submission form there :-) Title should be (as it is on the page) "How I use AI on this blog".

I use AI on this blog

gpjt — Fri, 31 Jul 2026 19:31:43 +0000

Article URL: https://www.gilesthomas.com/2026/07/ai-use

Comments URL: https://news.ycombinator.com/item?id=49127589

Points: 1

# Comments: 1

Why do OpenAI's GPT-2 weights beat mine? Part three: testing overtraining

gpjt — Fri, 31 Jul 2026 00:16:57 +0000

Article URL: https://www.gilesthomas.com/2026/07/why-do-openai-gpt2-weights-beat-mine-3-overtraining

Comments URL: https://news.ycombinator.com/item?id=49117559

Points: 2

# Comments: 0

Why do OpenAI's GPT-2 weights beat mine? Part two: the bugfix

gpjt — Thu, 30 Jul 2026 18:13:31 +0000

Article URL: https://www.gilesthomas.com/2026/07/why-do-openai-gpt2-weights-beat-mine-2-the-bugfix

Comments URL: https://news.ycombinator.com/item?id=49113601

Points: 8

# Comments: 0

Why do OpenAI's GPT-2 weights beat mine?

gpjt — Wed, 29 Jul 2026 14:55:11 +0000

Article URL: https://www.gilesthomas.com/2026/07/why-do-openai-gpt2-weights-beat-mine-1-intro

Comments URL: https://news.ycombinator.com/item?id=49098367

Points: 4

# Comments: 0

Benchmarking Qwen 3.6 35B MoE (3B active) on an RTX 3090

gpjt — Fri, 24 Jul 2026 22:33:55 +0000

Article URL: https://www.gilesthomas.com/2026/07/benchmarking-qwen-3-6-35b-moe-rtx-3090

Comments URL: https://news.ycombinator.com/item?id=49042436

Points: 16

# Comments: 0

Building intuition about LLM parameter counts

gpjt — Fri, 10 Jul 2026 20:37:13 +0000

Article URL: https://www.gilesthomas.com/2026/07/llm-parameter-counts

Comments URL: https://news.ycombinator.com/item?id=48864905

Points: 2

# Comments: 0

Poppy the training box, part 1: the beginnings

gpjt — Thu, 09 Jul 2026 00:13:57 +0000

Article URL: https://www.gilesthomas.com/2026/07/poppy-the-training-box-1-the-beginnings

Comments URL: https://news.ycombinator.com/item?id=48839233

Points: 3

# Comments: 0

From bigrams to GPT-2, one component at a time (in Jax)

gpjt — Wed, 08 Jul 2026 17:52:25 +0000

Article URL: https://www.gilesthomas.com/2026/07/llm-from-scratch-34b-building-and-training-gpt-2-small-in-jax

Comments URL: https://news.ycombinator.com/item?id=48835002

Points: 1

# Comments: 0

Building a Jax training loop for an LLM training run

gpjt — Tue, 30 Jun 2026 17:59:32 +0000

Article URL: https://www.gilesthomas.com/2026/06/llm-from-scratch-34a-building-a-jax-training-loop-for-an-llm-training-run

Comments URL: https://news.ycombinator.com/item?id=48736594

Points: 2

# Comments: 0

Thoughts on Role Confusion

gpjt — Wed, 24 Jun 2026 19:14:21 +0000

Article URL: https://www.gilesthomas.com/2026/06/role-confusion

Comments URL: https://news.ycombinator.com/item?id=48664457

Points: 3

# Comments: 0

New comment by gpjt in "Ask HN: Anthropic banned me from using Claude Code and I don't know what to do"

gpjt — Tue, 23 Jun 2026 13:31:59 +0000

In the EU, at least one of the problems is regulatory/tax for digital services. They need to charge you the VAT rate for the country where you are based. For that, they need two pieces of evidence about your location. There are various things they can use for that -- telephone number, IP address, card billing address, and so on. If they can collect two that indicate the same country, they're safe -- but if all of them point in different directions then they could get in trouble during a tax audit.

Of course, for larger transactions you'd expect that a human in the loop could work with you to get the right info so that they would be covered. But I guess for Microsoft, their definition of "larger" might be more than a few grand...

New comment by gpjt in "10Gb/s Ethernet: switching to a Broadcom SFP+ module"

gpjt — Wed, 17 Jun 2026 02:36:53 +0000

OP here -- yes indeed. I ever do a new series of posts on upgrading my network to anything faster, the first one will probably be titled something like "25Gb/s Ethernet: how much it costs to rip out your CAT-6A and replace it with fibre".

Flax debugging: making a hash of things

gpjt — Wed, 17 Jun 2026 02:11:12 +0000

Article URL: https://www.gilesthomas.com/2026/06/hashing-jax-parameters

Comments URL: https://news.ycombinator.com/item?id=48564929

Points: 2

# Comments: 0

New comment by gpjt in "10Gb/s Ethernet: switching to a Broadcom SFP+ module"

gpjt — Wed, 17 Jun 2026 00:29:25 +0000

OP here -- yes, 100%! Within my study it's DACs all the way. I only use 10GBASE-T where it's the only option: through the wall cabling, and from the connector on my ISP's crappy router to the mini-PC that acts as the real router.

New comment by gpjt in "10Gb/s Ethernet: switching to a Broadcom SFP+ module"

gpjt — Wed, 17 Jun 2026 00:27:49 +0000

Thanks for making the comment! It saved me having to bolt a USB fan over the old, misbehaving SFP+ module or something similarly silly...

10Gb/s Ethernet: switching to a Broadcom SFP+ module

gpjt — Tue, 16 Jun 2026 17:48:42 +0000

Article URL: https://www.gilesthomas.com/2026/06/10g-ethernet-switching-to-broadcom-sfp-plus

Hacker News: gpjt

New comment by gpjt in "Ask HN: What are you working on? (August 2026)"

A Quick(ish) Chinchilla Check

New comment by gpjt in "I use AI on this blog"

I use AI on this blog

Why do OpenAI's GPT-2 weights beat mine? Part three: testing overtraining

Why do OpenAI's GPT-2 weights beat mine? Part two: the bugfix

Why do OpenAI's GPT-2 weights beat mine?

Benchmarking Qwen 3.6 35B MoE (3B active) on an RTX 3090

Building intuition about LLM parameter counts

Poppy the training box, part 1: the beginnings

From bigrams to GPT-2, one component at a time (in Jax)

Building a Jax training loop for an LLM training run

Thoughts on Role Confusion

New comment by gpjt in "Ask HN: Anthropic banned me from using Claude Code and I don't know what to do"

New comment by gpjt in "10Gb/s Ethernet: switching to a Broadcom SFP+ module"

Flax debugging: making a hash of things

New comment by gpjt in "10Gb/s Ethernet: switching to a Broadcom SFP+ module"

New comment by gpjt in "10Gb/s Ethernet: switching to a Broadcom SFP+ module"

10Gb/s Ethernet: switching to a Broadcom SFP+ module

Jax: Commitment Issues