I build with AI and write about what actually works: agents, dev tools, and Portugal's AI scene. VP of Engineering at @LayerXlab

Portugal
Joined December 2008
Spotted an international bus (from Ukraine, in Portugal) with a @Starlink on the back 👀
1
14
1,257
yup, this seems to be out of this world
This quoted post is unavailable.
23
thanks @ClaudeDevs 🙏 🇺🇸🛫🇵🇹
5
162
i'm dreaming?! 🫠
Real life, baby
1
286
Rockstar Games
@RockstarGamesOct 15
Watch the Grand Theft Auto VI Launch Trailer, featuring the debut of “Flower in a Hurricane" from original score composers Justice feat. Tame Impala, SZA, and King Gizzard & The Lizard Wizard. Grand Theft Auto VI arrives November 19: rockstargames.com/VI/ps5
55
Most scenes already fit under the 1568px limit, so the gain from tiling came from having less to look through, pretty much how humans play it. With zoom the confidence score starts to mean something: on the 3x3 setup, 20 of the 30 answers above 0.5 were right, and it never confidently "found" a character who wasn't in the scene. My favorite fail: on level 4 it boxed the same patch of yellow and black stripes as Odlaw in all 3 runs. The real one is peeking out behind a soldier on the other side of the page.
1
32
Jev 1.14 from @typesafeai can now draw boxes on images, so obviously I made it play Where's Waldo 🔍 10 scenes, 4 characters (Waldo, Wenda, Wizard, Odlaw), 5 search strategies, 3 runs each. A hit is when the center of the box lands on the character: Whole scene: 10% 2x2 grid, pick a tile, then box: 25% 3x3 grid with the tile enlarged: 24% Random guessing: ~0.3%
1
1
144
Apple HomeView Won't Be Available in the European Union macrumors.com/2026/10/13/app…
2
135
Opus 5.5 tops the new FrontierCode leaderboard from @ArtificialAnlys at 52%. No surprise here. Opus hasn't let me down in my own work, and the feedback I hear and read from other engineers says the same. FrontierCode, built by @cognition with open source maintainers, grades whether a maintainer would actually merge the change. Large diffs, files touched outside the task and pointless tests all count against the model, and failing any blocking check scores zero. Opus passed those checks in 58% of its attempts. The other number worth a look: GPT-6.1 Sol scores 49% at $0.81 per task, level with Fable 5.1, for less than a tenth of the cost per task.
2
2
188
A Meta lançou hoje dois modelos e só o mais pequeno chega à API aberta. O Muse Wonder está disponível desde hoje na Meta Model API e no Muse Code, o produto de programação da Meta. O Muse Wonder Full, o maior e o que aparece no topo dos gráficos de benchmark da própria Meta, fica reservado a quem subscreve o Muse Code. Quem quiser testar pela API fica com o modelo mais pequeno e sem forma de confirmar os resultados do Full. E esses resultados são todos reportados pela própria Meta, ainda sem verificação independente. Vamos ver e colocar isto em testes reais 👀
2
1
1
163
Pus o Jev da @typesafeai contra o Clef e o Clef-flash da @Cloudflare em 300 jogos de xadrez, com o Stockfish (o motor de xadrez de referência) a avaliar cada jogada. O Clef ganhou ao Jev por 57 a 43, e o Clef-flash empatou com o Clef a um terço do preço. Para quem constrói com estes modelos, o número mais útil é a confiança: acima de 0,8, o Clef jogou a jogada do Stockfish 98% das vezes, em 112 jogadas. O Jev chegou a esse nível 4 vezes em 100 jogos. Entre 76 e 83 de cada 100 jogos acabaram empatados, quase todos por repetição (a mesma posição três vezes). Estes modelos leem a posição uma vez e escolhem sem planear. Perante a mesma posição dão sempre a mesma resposta, e as peças andam para trás e para a frente até ao empate.
1
2
6
359
— These credits also apply in third-party harnesses that accept a Claude API key.
ClaudeDevs
@ClaudeDevsOct 8
Replying to @ClaudeDevs
Credits can be used with the Messages API, Claude Managed Agents, and the Agent SDK, but cannot be used with interactive Claude Code sessions. These credits also apply in third-party harnesses that accept a Claude API key.
75