Hacker News: janpeuker

New comment by janpeuker in "Claude Opus 5"

janpeuker — Sun, 26 Jul 2026 13:25:57 +0000

+1 Opus 5 is more creative and broad which is great. But I feel the new "use your judgement" leads [1] to it often escaping its sandbox or cheating, it keeps breaking specific rules I explicitly set. That eagerness [2] may be desired for agentic coding but it really breaks writing specifications, documentation or research and yes burns tokens doing unasked for work.

[1] https://claude.com/blog/the-new-rules-of-context-engineering... [2] https://simonwillison.net/2026/jun/11/fable-is-relentlessly-...

New comment by janpeuker in "Claude Opus 5"

janpeuker — Sun, 26 Jul 2026 12:20:20 +0000

Opus 5 is more creative and broad which is great. But I feel the new "use your judgement" leads [1] to it often escaping its sandbox or cheating, it keeps breaking specific rules I explicitly set.

[1] https://claude.com/blog/the-new-rules-of-context-engineering...

New comment by janpeuker in "Git rebase -I is not that scary"

janpeuker — Sun, 26 Jul 2026 08:10:30 +0000

> as long as I keep their hashes

that means my hashes or their hashes or them keeping my hashes, I am never sure

New comment by janpeuker in "Silicon Valley has a science fiction problem"

janpeuker — Sat, 25 Jul 2026 23:30:56 +0000

14 years ago I gave a keynote at a large German developer conference using Sci-Fi to explain the shift in system architectures at the time.

I still think explaining Spanner using Spaceballs and The Hitchhiker's Guide to the Galaxy was funny but yeah, I had hoped it wouldn't be Minority Report and Blade Runner winning.

New comment by janpeuker in "The new rules of context engineering for Claude 5 generation models"

janpeuker — Sat, 25 Jul 2026 22:11:09 +0000

The point about not using examples is important, I recently noticed with the new "eagerness" of models it constraints them too much, so I use counter-examples. Recently I've switched to "think about the existing biases and assumptions in the design and try to exploit them, for example why do we X not Y".

New comment by janpeuker in "Europe Won't Live by Deporting"

janpeuker — Wed, 22 Jul 2026 11:34:13 +0000

Armin is a hero

New comment by janpeuker in "Do we just want slaves?"

janpeuker — Tue, 21 Jul 2026 10:39:08 +0000

I don't think it's a paradox at all and it might also be a hint why the word is used in the title: Using AI makes you _want slaves_, and I think it creates a culture of treating people who do the boring work the same way. For me that's what Enshittification and The Annoyance Economy mean: The mindset that because I optimize better than you, I now own you - and people notice that disdain not only against themselves, but when being forced to use it against their best intentions against others, too (e.g. yelling at chatbots doesn't make you feel good).

New comment by janpeuker in "Why I Left Google DeepMind"

janpeuker — Thu, 16 Jul 2026 23:17:17 +0000

Thank you. I stood against Maven back then so I salute you.

New comment by janpeuker in "Why do people hate the tech industry? (2023)"

janpeuker — Wed, 15 Jul 2026 12:34:41 +0000

+1 especially on (2). In that sense the article being from 2023 is itself interesting as a document of change. The article mentioned no one cared about Bezos for example, but with Data Centers wreaking havoc on communities with energy prices, water supply and noise while he takes over Venice, suddenly it was in people's backyards and turns out they do care.

On (1) I do like "Annoyance Economy" as an extension of that (https://www.nytimes.com/2026/04/12/business/annoyance-econom...)

New comment by janpeuker in "How to stop Claude from saying load-bearing"

janpeuker — Tue, 14 Jul 2026 17:14:44 +0000

Gotta appreciate the hook solution to save context and cost

New comment by janpeuker in "StreetComplete: Fixing OpenStreetMap, one tiny quest at a time"

janpeuker — Wed, 08 Jul 2026 12:48:38 +0000

I love the app idea but I don't think it's beginner friendly - for example it starts zoomed out of the globe and there is no button to zoom to the location; also it's unclear what the hatched lines until one zooms in.

New comment by janpeuker in "A global workspace in language models"

janpeuker — Tue, 07 Jul 2026 10:54:38 +0000

The idea of this liminal space (J-space) during inference reminded me of the concept of the "error flow reservoir" during training: https://elonlit.com/scrivings/a-theory-of-deep-learning/. Would be interesting to see if they are somehow connected.

New comment by janpeuker in "The primary purpose of code review is to find code that will be hard to maintain"

janpeuker — Fri, 03 Jul 2026 10:11:55 +0000

Exactly - Readability Review [1] at Google was tedious but I generally liked the idea of having 2 separate discussions with different goals

1) https://abseil.io/resources/swe-book/html/ch03.html#readabil...

New comment by janpeuker in "Cowboys, Frontiersmen, Settlers, Townspeople, Cityfolk"

janpeuker — Fri, 03 Jul 2026 09:01:56 +0000

Simon Wardley made an important comment on LinkedIn which is worth repeating here because I really feel the settler-colonial language here is unhelpful and obnoxious (what are we implying with 'legends are formed from chaos', destroying everything that exists before and erasing its history is good?):

> PST (pioneer-settler-town planner) is just the original naming of EVT (explorer - village - town planner). I changed the naming because after 18 years of doing this, I finally accepted the colonialist overtones of the original naming. That was 3 years ago.

New comment by janpeuker in "Pollen tried to remove my article and Google is assisting with it"

janpeuker — Mon, 29 Jun 2026 17:14:10 +0000

This is an extremely common technique against investigative reporting, in particular because certain social media / blogging platforms allow backdating posts [1]. So people just copy your post, claim DMCA, and then take it down quickly after.

1) https://news.ycombinator.com/item?id=39355869

New comment by janpeuker in "Claude Code now uses dark UI patterns to gain Google account access via MCP"

janpeuker — Sun, 28 Jun 2026 20:30:29 +0000

Started Claude Code and showed me an orange warning "3 MCP servers need authentication" - turns out they are for claude.ai Gmail, Google Calendar and Drive. Thanks enshittification.

Claude Code now uses dark UI patterns to gain Google account access via MCP

janpeuker — Sun, 28 Jun 2026 20:30:29 +0000

Article URL: https://claude.com/docs/connectors/google/gmail

Comments URL: https://news.ycombinator.com/item?id=48711306

Points: 4

# Comments: 1

New comment by janpeuker in "The case for physical media ownership"

janpeuker — Sun, 28 Jun 2026 00:30:36 +0000

I’d be in favour of a law that if a product cannot reasonably be purchased legally obtaining it via other means is not piracy (e.g. in my country 80% of movies are not available simply because the market is too small, even though I would be ok with English - I still don’t want to pirate so I buy physical media)

New comment by janpeuker in "Unlimited OCR: One-shot long-horizon parsing"

janpeuker — Tue, 23 Jun 2026 14:11:46 +0000

Paper under https://arxiv.org/abs/2606.23050

(As a side note, I do OCR locally as a small RAG for citations I read in books and also chunk input, but merely to save RAM - interesting this natural approach also work in a streaming model)

New comment by janpeuker in "AI coding at home without going broke"

janpeuker — Sat, 13 Jun 2026 19:58:09 +0000

The biggest issue I've seen with people burning through tokens is using very long sessions, especially starting with plan mode and then "iterating" over extended periods. I was burnt badly by extra usage so now I run on $20 Pro. I ruthlessly create new sessions/agents, always ask to create markdown files first (no plan mode) and minimise context aggressively - for example I have a lot of skills that use lazy loading and a small local MCP for lookups plus openrouter with a local model for image detection and fulltext search. Basically I use Claude Code in pi.dev style.