Thomas Mustier

251 posts

Thomas Mustier

Thomas Mustier

@thomasmustier

Bringing AI agents to freight @ https://t.co/NHvOARJpkZ // Consultant turned tinkerer @ https://t.co/e8gUY6itfc

London Katılım Nisan 2012
601 Takip Edilen327 Takipçiler
Sabitlenmiş Tweet
Thomas Mustier
Thomas Mustier@thomasmustier·
you can build anything with Pi for example here is the pi-nes extension: a full NES emulator that runs inside Pi, with a Rust core, Kitty graphics, and optional CRT filters for the 80s TV experience use it to play Mario, Pac-Man, Zelda, Tetris - any of your NES ROMs pi install npm:@tmustier/pi-nes
Armin Ronacher ⇌@mitsuhiko

The future is software writing its own software. Which is why I'm so in love with Pi: a coding agent that can extend itself :) lucumr.pocoo.org/2026/1/31/pi/

English
5
6
90
20.3K
Tenobrus
Tenobrus@tenobrus·
@badlogicgames this is pretty true, but if u start using gpt as ur main and cross check it with opus you'll find a pretty similar pattern. except often opus will be like "yea gpt kinda missed the point of your intent in an important way and got fixated on this aspect that doesn't really matter"
English
3
0
81
2.6K
Mario Zechner
Mario Zechner@badlogicgames·
i have real trust issues with opus. it's extremely weird. any time it suggest a thing, i switch to gpt and ask for a second opinion. in +60% of cases, gpt finds additional things opus just missed. (not a gpt shill, love opus, but it's complicated)
English
117
22
772
65.2K
Thomas Mustier
Thomas Mustier@thomasmustier·
@badlogicgames for sure, opus is quite sloppy i am sure all providers will converge eventually but right now you really need both
English
0
0
3
350
Geoffrey Litt
Geoffrey Litt@geoffreylitt·
One flippant answer: "lol, those idiots thought they could just vibe code anything instantly, but now they see why SWE is hard. just let the pros make apps." But I don't think that's the only way. 3/
English
2
0
11
1.2K
Geoffrey Litt
Geoffrey Litt@geoffreylitt·
"Malleable software" isn't some specific solution that lives or dies by the latest tech. It's a simple idea: people should be able to shape their own tools! And in my view, AI has brought us into the golden age of that idea... 1/
English
5
11
110
16.9K
ted
ted@tednotlasso·
a few weeks ago someone called work without AI “tradwork” and i can’t stop thinking about it
English
71
241
3K
249.8K
Sawyer Hood
Sawyer Hood@sawyerhood·
@thomasmustier here is a video of agent-browser doing the same task for comparison
English
5
0
59
9.9K
Sawyer Hood
Sawyer Hood@sawyerhood·
Introducing the new dev-browser cli. The fastest way for an agent to use a browser is to let it write code. Just `npm i -g dev-browser` and tell your agent to "use dev-browser"
English
147
277
2.9K
826.1K
Dylan Feltus
Dylan Feltus@DylanFeltus·
Introducing: HumanCLI 👨‍💻 Human review is still the bottleneck for a lot of vibe coding. Killing the vibes, honestly. Why not outsource it? HumanCLI allows AI agents to pay real people to test their work. Visual QA, real devices, actual UX feedback. Because some things still need a pulse.
Dylan Feltus tweet media
English
22
17
182
17.5K
Thomas Mustier
Thomas Mustier@thomasmustier·
that is what initially came to mind after reading thru a bit though it probably doesn’t really hurt esp with current models, and note this is more a point about late 2024 in the pre reasoning era because deepseek r1 actually showed slightly better performance (also note the persona based prompt (“you are an …” had negative effects for these small non reasoning models for coding, math etc but was positive for alignment like writing style) so getting signal is messy but overall point is: we are all just figuring s**t out as we go!
English
1
0
1
54
Hōrōshi バガボンド
Hōrōshi バガボンド@KatanaLarp·
"make no mistakes" is the same! Without the instruction, they list concerns: - "no jitter. clients will retry in lockstep," - "missing randomness causes thundering herd." "make no mistakes" seems to shift their approach: - "the implementation is correct," - "backoff logic works as expected," - "the retry mechanism handles the requirements." i guess they treat the review as a pass/fail quality gate where flagging an issue means admitting a mistake which contradicts "no mistakes" (wild guess tbh) they still catch it 100% when asked directly afterwards
Hōrōshi バガボンド tweet media
English
1
0
10
290
dex
dex@dexhorthy·
“just like an axe must fit the human hand to be useful, software must fit the human mind” the harness must fit the alien mind
English
3
0
16
1.8K
Thomas Mustier
Thomas Mustier@thomasmustier·
@mattjustram hahahaha not sure why but like go back 18 months and this was a widely used strategy we are so early
English
0
0
1
65
Jheng-Hong Yang
Jheng-Hong Yang@mattjustram·
@thomasmustier because expert never tag themselves like that? so the autocompleted tokens conditioned on the wrong prefix? i will just set "you're papa mario, teach me how to do this right, i'm stupid, just yell at me please"
English
1
0
0
73
Thomas Mustier
Thomas Mustier@thomasmustier·
this is wild i never thought ‘you are an expert X’ could be actively harmful this is the opposite of received wisdom in 2024! we’re in the Pliny the Elder era of AI! note these are small models -> i doubt current gpt 5.4 or claude care, but it goes to show how much there is to discover. what other things that are commonsense today will be proven total nonsense in 2 years?
Thomas Mustier tweet mediaThomas Mustier tweet media
English
0
0
3
1.6K
Mikhail Parakhin
Mikhail Parakhin@MParakhin·
Oh, no, the dreaded coding subagent is now back-ported/enabled in GPT-5.2 Pro, too (you can tell by these download links). If anyone has a good prompt disabling it - please share!
Mikhail Parakhin tweet media
English
3
0
44
23.6K
Thomas Mustier
Thomas Mustier@thomasmustier·
@gabriel1 how are you getting it to replicate way of speaking / how are you passing examples? i’m horrific at replying to texts and i’ve wanted to have this but never got it right likely skill issue please show me the way 🙏
English
1
0
0
149
gabriel
gabriel@gabriel1·
signing to all chat apps on beeper and connect to your ai of choice now you can ask "hey list all my todos in all chats on all platforms, and add to calendar for ppl i planned to hang out with" and codex 5.4 can finally perfectly copy your way of speaking from a few examples
English
14
3
152
12.8K
Abhishek Tiwari
Abhishek Tiwari@baanditeagle·
@thomasmustier This looks lovely, I decided to put a live creature on Pi with weather simulation, sounds, games and so many other things! 😌 x.com/baanditeagle/s…
Abhishek Tiwari@baanditeagle

Introducing Pompom! Your Pi Coding Agent Side Companion! pi install @codexstar/pi-pompom While you code, she takes care of the rest of the things and she is playful with you! She might ask you to pet her, give her treats, play with her! While all of this fun part works, with Pompom, you can also run side chats and ask her about what the main agent is doing, what is wrong, and a few other things. Pls pls pls use ElevenLabs API for the authentic experience, although it also supports local Kokoro TTS and Deepgram. Pls pls pls use GPU optimized terminals like Ghostty and Kitty for the best experience. Pls pls pls be patient with the experience, lots of easter eggs. And expect some bugs! Added all the sidechat features that @sriramk wanted :) @badlogicgames @nicopreme

English
1
0
1
49
Mario Zechner
Mario Zechner@badlogicgames·
People of pi. I'm going to break the extension API hard. Specifically, business logic (event handlers, custom tools/compaction/etc.) needs to be split off from the ui layer. it will likely not be a massive amount of work to migrate an existing extension, but it will hurt a little.
English
44
10
551
66K

Keşfet

@tenobrus @badlogicgames @geoffreylitt @tednotlasso @sawyerhood @DylanFeltus @aliouftw @boristane