Hacker News: gfysfm

What's so hard about continuous learning?

gfysfm — Mon, 23 Feb 2026 23:38:17 +0000

Article URL: https://www.seangoedecke.com/continuous-learning/

Comments URL: https://news.ycombinator.com/item?id=47130666

Points: 3

# Comments: 0

We are in the "gentleman scientist" era of AI research

gfysfm — Mon, 16 Feb 2026 03:41:31 +0000

Article URL: https://www.seangoedecke.com/ai-and-informal-science/

Comments URL: https://news.ycombinator.com/item?id=47030628

Points: 3

# Comments: 0

How does AI impact skill formation?

gfysfm — Tue, 03 Feb 2026 06:32:18 +0000

Article URL: https://www.seangoedecke.com/how-does-ai-impact-skill-formation/

Comments URL: https://news.ycombinator.com/item?id=46867305

Points: 1

# Comments: 0

New comment by gfysfm in "Ripple, a puzzle game about 2nd and 3rd order effects"

gfysfm — Sun, 04 Jan 2026 23:24:06 +0000

This is really cool! It's interesting that you can "cheat" by knowing the historical situation in advance. But I guess there's no way around that - if the situations involved were hypothetical, it would be unsatisfying when you guessed "wrong". It'd be neat to see really obscure examples drawn from history.

Grok is enabling mass sexual harassment on Twitter

gfysfm — Fri, 02 Jan 2026 05:55:56 +0000

Article URL: https://www.seangoedecke.com/grok-deepfakes/

Comments URL: https://news.ycombinator.com/item?id=46461854

Points: 20

# Comments: 1

How good engineers write bad code at big companies

gfysfm — Fri, 28 Nov 2025 19:59:32 +0000

Article URL: https://www.seangoedecke.com/bad-code-at-big-companies/

Comments URL: https://news.ycombinator.com/item?id=46082223

Points: 407

# Comments: 320

It's not surprising that 95% of AI enterprise projects fail

gfysfm — Tue, 18 Nov 2025 12:48:00 +0000

Article URL: https://www.seangoedecke.com/why-do-ai-enterprise-projects-fail/

Comments URL: https://news.ycombinator.com/item?id=45964975

Points: 2

# Comments: 0

Solving the inference problem for open source AI projects with GitHub Models

gfysfm — Thu, 24 Jul 2025 06:58:52 +0000

Article URL: https://github.blog/ai-and-ml/llms/solving-the-inference-problem-for-open-source-ai-projects-with-github-models/

Comments URL: https://news.ycombinator.com/item?id=44667770

Points: 1

# Comments: 0

Scan and resolve fixed GitHub issues and PRs with AI

gfysfm — Tue, 22 Jul 2025 02:45:25 +0000

Article URL: https://github.com/ashleywolf/continuous-ai-resolver

Comments URL: https://news.ycombinator.com/item?id=44642762

Points: 1

# Comments: 0

Show HN: OpenAI Codex for Free with GitHub Models

gfysfm — Wed, 25 Jun 2025 10:20:43 +0000

Article URL: https://gist.github.com/sgoedecke/2b4e8d5e6b21f536ea399f1728916ad5

Comments URL: https://news.ycombinator.com/item?id=44375587

Points: 7

# Comments: 0

The Illusion of "The Illusion of Thinking"

gfysfm — Sun, 08 Jun 2025 02:24:37 +0000

Article URL: https://www.seangoedecke.com/illusion-of-thinking/

Comments URL: https://news.ycombinator.com/item?id=44214118

Points: 6

# Comments: 1

New comment by gfysfm in "Why DeepSeek is cheap at scale but expensive to run locally"

gfysfm — Sun, 01 Jun 2025 11:49:11 +0000

Hi, I wrote the post! Also not a ML researcher, just an interested engineer, so I'm sure I got some things wrong.

> MoE should be the better tradeoff for the local/single-user scenario since the downside of batching being harder / less efficient doesn't matter.

What I meant was that the single-user scenario is going to get dramatically worse throughput-per-GPU, because they're not able to reap the benefits of multi-user batching (unless they're somehow doing massively parallel inference requests, I suppose).

> Is it really that the matrixes being multiplied are larger? My mental model is that the purpose of batching isn't to get larger input matrices. It's to move the bottleneck from memory bandwidth to compute.

As I understand it, you want larger input matrices in order to move the bottleneck from memory to compute: if you do no batching at all, your multiplications will be smaller (the weights will be the same, of course, but the next-token data you're multiplying with the weights will be 1xdim instead of batch-size x dim), so your GPUs will be under-utilized and your inference will spend more time doing memory operations and less time multiplying.

> The post has no numbers on the time to first token for any of the three providers.

I probably should have hunted down specific numbers, but I think people who've played with DeepSeek and other models will notice that DeepSeek is noticeably more sluggish.

New comment by gfysfm in "Strengths and limitations of diffusion language models"

gfysfm — Thu, 22 May 2025 18:20:04 +0000

Hi, I wrote the post. Thank you!

That’s how it does work, but unfortunately denoising the last paragraph requires computing attention scores for every token in that paragraph, which requires checking those tokens against every token in the sequence. So it’s still much less cacheable than the equivalent autoregressive model.