Ville Säävuori

3K posts

Ville Säävuori banner
Ville Säävuori

Ville Säävuori

@uninen

Web dev (Python, Django, TypeScript, Vue), Founder, DJ. Documenting my learnings at https://t.co/jr0MQcSW6F. Find me at the blue thing: @uninen.net

Turku, Finland Se unió Ocak 2008
453 Siguiendo429 Seguidores
Tweet fijado
Ville Säävuori
Ville Säävuori@uninen·
I have two kinds of hobby code projects; 1) where I enjoy choosing my tools and writing everything as perfect and beautiful as I can, and 2) where I’m just enjoying hacking or learning something, and also the fact that I can write as much spaghetti as I want. I love both 😅
English
0
0
8
0
Ville Säävuori
Ville Säävuori@uninen·
@badlogicgames Haven't gotten this yet but its crazy that their Pro plan is now $194 per quarter! Got mine for 45 and then 90 (after I forgot to cancel, lol).
English
0
0
0
967
Quinn Slack
Quinn Slack@sqs·
Good reputation, good business, mass-market adoption AI companies: pick 2
English
8
7
89
8.8K
Ville Säävuori
Ville Säävuori@uninen·
@mitsuhiko Oneshotted a small extension that shows all the skills that have been loaded for the session realtime in the footer. Its been surprisingly useful for those occasions where the agent does something dumb and you can immediately see why. Never read the code; it Just Works 👌
English
0
0
1
377
Armin Ronacher ⇌
Armin Ronacher ⇌@mitsuhiko·
What's the smallest (but comparatively coolest) extension you built for pi?
English
74
2
104
34.3K
Ville Säävuori
Ville Säävuori@uninen·
OpenAI: We have the compute OpenAI compute: nope
Ville Säävuori tweet media
Tibo@thsottiaux

I don't know what they are doing over there, but Codex will continue to be available both in the FREE and PLUS ($20) plans. We have the compute and efficient models to support it. For important changes, we will engage with the community well ahead of making them. Transparency and trust are two principles we will not break, even if it means momentarily earning less. A reminder that you vote with your subscription for the values you want to see in this world.

English
0
0
0
61
Ville Säävuori
Ville Säävuori@uninen·
This is such a clusterfuck. Why, @AnthropicAI , why 🤦‍♂️ There probably is a way for them to fuck up their reputation and business even more somehow, but right now I genuinely don't know how. I don't recall a fall from grace this bad ever in my career. Total and utter shitshow.
Simon Willison@simonw

@TheAmolAvasare If I sign up for a new $20/month account today and roll the dice and get unlucky (1/50 chance) will I end up not being able to use Claude Code with it, or is the 2% just the chance that I'll see the new pricing grid during the signup flow?

English
0
0
0
80
Ville Säävuori
Ville Säävuori@uninen·
All I wish for Christmas is Vue support for Oxlint
English
0
0
1
39
Ville Säävuori
Ville Säävuori@uninen·
@tommyvedvik I have tons of skills for custom tools and processes that I don't want to copypaste repeatedly. Some skills indeed ARE just examples from a larger doc and then ref to that doc. Saves a lot of time bc I don't need to "read this [copypaste] and do x" but just "do x"
English
0
0
1
46
Tommy J. Vedvik
Tommy J. Vedvik@tommyvedvik·
I almost never use skills. Just provide it with context/docs etc for what I need. Is it just me? Imo a "skill" could in most cases just be a markdown endpoint on a website.
English
2
0
1
423
Ville Säävuori
Ville Säävuori@uninen·
Not sure if the thinking is not working or inherent to the model but it reliably disregards almost everything in AGENTS.md, and when forced to re-read instructions, remembers it for one turn and relapses again. Haven't seen quite anything like it since GPT-4 era models.
English
0
0
0
25
Ville Säävuori
Ville Säävuori@uninen·
My experiences after using Kimi 2.6 (high) for a handful of tasks: delightfully fast, absolutely TERRIBLE at following instructions, writes OK code eventually. Tried really hard to prompt it to do better, no avail. Useless for agentic work, outputs only slop w/o intervention 🤷‍♂️
English
1
0
2
81
Ville Säävuori retuiteado
Mario Zechner
Mario Zechner@badlogicgames·
pi tip: in capable terminals like ghostty, use CMD + up arrow key to jump to the top of final assistant messages in a turn, and user messages. very nice if the assistant spat out a long reply, like a plan.
English
12
5
207
12.5K
Ville Säävuori
Ville Säävuori@uninen·
This sums the current state of Claude Code; version update check --(what should be a) one http query to a cached static endpoint-- takes almost 8s 🤦‍♂️
Ville Säävuori tweet media
English
0
0
0
23
Ville Säävuori
Ville Säävuori@uninen·
@xyster Same here! Have wondered this for ages, it's not like this is somehow rocket science 🤦‍♂️
English
0
0
1
18
Steve💙🇨🇦
Steve💙🇨🇦@xyster·
Every MacBook I've owned has had super annoying grounding issues, where I feel like my palms are being lightly electrocuted. It seems like such a simple fix too; just add grounding to the power supply like every reasonable computer ever. My Dell 16 Pro, all metal, zero issues.
English
1
0
0
72
Ville Säävuori
Ville Säävuori@uninen·
@badlogicgames @mitsuhiko they probably use Jira. there space of worse solutions would be very small, yet you could ship extremely shitty UX and in your mind think it's good just because your POV. would explain so much!
English
0
0
0
16
Mario Zechner
Mario Zechner@badlogicgames·
@mitsuhiko also loses scroll position, which is famtastic during triage. it's like nobody at GH uses their own stuff.
English
1
0
0
941
Armin Ronacher ⇌
Armin Ronacher ⇌@mitsuhiko·
My new favorite GitHub issue list feature is that going back to the issue list, shows some things (that are read) briefly as unread.
English
1
0
24
4.4K
Ville Säävuori
Ville Säävuori@uninen·
Similar thoughts. I had it design two refactor plans (default settings, xhigh), one simple and one complex, in a repo that has *very* good LLM guides; it properly fucked up both. 4.6 would've done better, OTOH it also has had retarded days as well 🤷‍♂️
Machine Learning Street Talk@MLStreetTalk

Vibes on Opus 4.7: * Burned through my 5 hour token limits in about 90 minutes on my 20x Max plan with my normal usage. So I can't even work with this. And this is with them giving us "extra limits" to soften the token tax blow. You mean it's going to be even WORSE than this? * Objectively dumber and less coherent for me, spending more time doing less work. I don't want to have to wait for hours to finish important work and break my flow constantly. The model feels totally different in character Who is going back to Opus 4.6?

English
0
0
1
57
Ville Säävuori
Ville Säävuori@uninen·
@mitsuhiko Built a custom system prompt mechanism for my clanker and wanted to tweak it REALLY precisely but also did not want to lose the tools, skills, and docs sections. v1 I sliced up the built Pi version but it was messy. Now I have a carefully crafted custom prompt + full Pi blob 🤷‍♂️
English
0
0
0
81
Armin Ronacher ⇌
Armin Ronacher ⇌@mitsuhiko·
@uninen Would love to learn a bit more what kind of hackery you have in mind :)
English
1
0
0
99
Armin Ronacher ⇌
Armin Ronacher ⇌@mitsuhiko·
If something really bothers you about pi, give us feedback please :)
English
91
1
156
15.3K
Ville Säävuori
Ville Säävuori@uninen·
@mitsuhiko I wanted the building blocks that buildSystemPrompt() uses internally as functions so that I could've modified the current blob (personality/tools/guidelines/docs) more easier programatically. Current all or nothing is OK, but its fragile to hack.
English
1
0
1
94

Descubrir

@badlogicgames @sqs @mitsuhiko @AnthropicAI @tommyvedvik @slobbba @xyster @elonmusk