
#opengeodata, #geolocation, Mac, Tourism and Moto Guzzi enthusiast.
Earth
Joined March 2010
- Tweets9.2K
- Following2.6K
- Followers2K
- Likes620
Bertalan Iván retweeted
The more I use Apple Containers, the more I like them!
My latest use: giving a coding agent read-only access to a codebase on the fly for investigation.
1. Start a disposable container. It takes about half a second 🔥
2. Mount one folder only, read-only
3. Run the agent inside (like pi). It can search, read, and reach the web with curl
4. Get the answer, destroy the container
An agent with a shell is powerful and risky: it could change files or read anything on your Mac 🤷🏻♂️.
The container here is the fence: total freedom, nothing to break outside. (we hope 🤣)
Bertalan Iván retweeted
This is a dream I had since I bought M5 Mac 128GB!
Prefill Accelerator (64K prompt, M3 Ultra 512GB + M5 Max 128GB)
- DeepSeek V4.1 Flash: 1,060 -> 3,123 tok/s (+195%)
- Qwen3.8 Flash-Next: 1,309 -> 3,666 tok/s (+180%)
- GLM-5.3 Flash: 540 -> 1,388 tok/s (+157%)
- MiniMax M3.1: 297 -> 569 tok/s (+92%)
oMLX 0.7.4.dev1 is out!
github.com/jundot/omlx/relea…
With the new Prefill Accelerator, connect an M5 Max to your big Mac and prefill up to 2.9x faster!
Prefill Accelerator (64K prompt, M3 Ultra 512GB + M5 Max 128GB)
- DeepSeek V4.1 Flash: 1,060 -> 3,123 tok/s (+195%)
- Qwen3.8 Flash-Next: 1,309 -> 3,666 tok/s (+180%)
- GLM-5.3 Flash: 540 -> 1,388 tok/s (+157%)
- MiniMax M3.1: 297 -> 569 tok/s (+92%)
- An M5 Max 64GB still adds 30-118%, and decode speed stays the same.
Faster DeepSeek V4.1 with A8 and buffer pool improvements
- Prefill on M5 Ultra is 43-48% faster, 32K goes 2,631 -> 3,779 tok/s (+44%).
- oQ4e builds now run INT8 Activation Prefill on M5.
Faster Qwen3.8 Flash-Next
- FP16 builds get the fused kernels. Decode on M3 Ultra goes 38.9 -> 63.0 tok/s (+62%), and +23% on M2 Max.
- Decode after long prompts is 3-5% faster on M5 Max.
Better for agents
- Parallel tool results from Claude Code reach the model in call order, so Qwen no longer pairs a result with the wrong call.
- Agent sessions on a cluster reuse the prompt cache across tool turns, 743 -> 9,266 cached tokens by the fifth turn.
- Switch oMLX models inside Codex and Copilot with /model.
Also in 0.7.4.dev1
- Ming-3.0 TTS with a 2x faster decode loop.
- Ask AI can use an oMLX server on another Mac.
- Check for model updates in the macOS app.
- German across the web dashboard and the macOS app.
- Batched decode with 6 requests on 8-bit models is 16% faster.
A big thank you to everyone who sent PRs, reported bugs and tested builds for this release, and a warm welcome to the five new contributors!
Bertalan Iván retweeted
WhatsApp Web takes 1 GB of RAM.
ZapFast is WhatsApp, native and fast: about 200 MB, opens instantly. Linux, macOS, Windows. Free and open source.
Fast to ship too: I released 3 versions in 5 days.
zapfast.rocks
Bertalan Iván retweeted
Hardens Linux servers with a comprehensive guide covering SSH, firewalls, and intrusion detection tools.
github.com/imthenachoman/How…
Bertalan Iván retweeted
Top 10 OpenCode Skill Repos:
1. Superpowers - github.com/obra/superpowers
2. Ponytail - github.com/DietrichGebert/po…
3. UI/UX Pro Max - github.com/nextlevelbuilder/…
4. Graphify - github.com/Graphify-Labs/gra…
5. Caveman - github.com/JuliusBrussee/cav…
6. Addy Osmani Skills - github.com/addyosmani/agent-…
7. Understand Anything - github.com/Egonex-AI/Underst…
8. Awesome Claude Skills - github.com/ComposioHQ/awesom…
9. Archify - github.com/tt-a1i/archify
10. Impeccable -github.com/pbakaus/impeccabl…
Bertalan Iván retweeted
Compositor (v1.4) is out! This update brings big GPU performance improvements throughout the app. Frames render 10x faster, it handles large files with ease, and everything feels buttery smooth. And the app is now just 11MB.
robbietilton.com/compositor#
don't throw away your old apple devices!
this is an iPad mini 1st gen from 2012!
it can't run anything, modern apps dropped iOS 9 long ago.
so what do you do with this hardware + Claude... heh
once you get ssh access, you can build any app you want.
mine is now a native Home Assistant wall panel.
(HA app doesn't support iOS 9, and the web dashboard breaks in old Safari)
and the best part:
you can grab one on ebay for ~$30-50,
same price as a popular esp32 IoT board :)
Bertalan Iván retweeted
We started Codex for OSS six months ago.
Today, we’re renewing the program with $100 Pro plans and doubling the number of grants (from 5,000 to 10,000) to reach more maintainers.
(If you received a grant previously, we encourage you to re-apply!)
openai.com/form/codex-for-os…
Bertalan Iván retweeted
iOS users in the EU, you can finally control your iPhone on the Mac!
Readers added context they thought people might want to know
Poster is talking about an app called Device Hub that’s part of the Xcode suite. iPhone Mirroring still isn’t available in the EU.
developer.apple.com/documentation/…
support.apple.com/en-my/120421
Bertalan Iván retweeted
71% da população mundial nunca usou AI generativa
28% só usou free chats
apenas 1% paga alguma assinatura de AI
e 0,14% usa AI pra escrever código
É por isso que microsaas vende. Aquele app basico de gerar foto com AI, ou de contar calorias, que o chat gpt free faz igual, vende.
O mercado de AI wrappers (aquele app que é só um promtp chamando a AI) tem mercado estimado em 300 BILHÕES de dolares em 2026.
Você e eu vivemos numa bolha. Você vê todo mundo aqui usando AI de forma avançada, e acha que o mundo inteiro é assim. E pior, você ainda cai na sindrome do impostor por isso, mas esquece que isso é 0,14% da população mundial.
Entao, bora construir microssas guys. Em breve vou trazer uma novidade sobre isso pra te ajudar, inclusive
Bertalan Iván retweeted
Connaissez-vous l'histoire de ce jeune ingénieur français qui a construit son propre modèle de langage ?
Il s'appelle Hugo Vergnes. Polytechnique, puis Stanford.
Pendant quatre ans il a construit une IA qui lit les mammographies pour dépister le cancer du sein. Pas une démo : un vrai dispositif médical, chez Whiterabbit.
Depuis juin dernier il est ingénieur machine learning chez Apple, sur la vidéo. Il vit en Californie (trop doué pour la France ? petite pique gratuite). Ceinture marron de judo, bleue de jiu-jitsu brésilien (ça n'a rien à voir, mais je le dis car j'ai l'info).
Et le week-end, voilà ce qu'il fait.
Il a entraîné son propre modèle de langage. Depuis zéro, tout seul, sans laboratoire derrière lui. 3,8 milliards de paramètres, 65 milliards de mots d'entraînement, 8 cartes B200 louées pendant 43 heures.
Facture totale : 998 dollars. Le prix d'un iPhone.
En 2019, OpenAI refusait de publier GPT-2 en expliquant qu'il était trop dangereux pour être mis entre toutes les mains. Sept ans plus tard :
-GPT-2 marque 0,2565 sur le benchmark CORE.
- Le modèle d'Hugo, fait en un week-end, marque 0,384.
Il a aussi battu nanochat, le modèle de référence d'Andrej Karpathy, l'ancien directeur IA de Tesla. Karpathy est à 0,310 pour à peu près le même budget et le même temps de calcul.
Les chiffres du run :
→ 3,8 milliards de paramètres, 65 milliards de mots
→ 8 GPU B200 loués, 43 heures, 998 $ au total
Et il a tout publié. Le code, la config, la méthode, son framework little-lm. Vous pouvez refaire exactement la même chose ce week-end.
Sa phrase résume tout : "Entre le « jouet nanoGPT » et le « besoin d'un laboratoire de recherche », il existe une vaste zone, peu décrite, où une personne disposant de quelques milliers de dollars peut entraîner un modèle significatif.".
Bravo à lui franchement. Un Polytechnicien qui montre à tout le monde que la barrière à l'entrée a fondu 🇫🇷
Et vous, vous vous y mettriez pour 1 000 balles ?
Bertalan Iván retweeted
Researchers spent 500 hours and $70,000 and found LG TVs logging plain text transcripts in standby, mapping every device in the house, and feeding it to LG's ad arm
Unplug the internet and it saves the files until you plug back in.
LG says its TVs don't record ambient conversations.
The evidence begs to differ.
216 million of these are sitting in living rooms.
Where are the regulators?
Writer: Daniel
Readers added context they thought people might want to know
The findings and video clip here are from Gamers Nexus' investigation, which receives no credit or link.
youtube.com/watch?v=6IFVTc…
x.com/gamersnexus/st…
Bertalan Iván retweeted
#ValentinaCera, consigliere segretario Ufficio di presidenza: tra le tematiche del concorso Ragazzi in Aula anche #bullismo, #cyberbullismo e #ambiente e #AI. Altre iniziative per gli studenti: il progetto dedicato al film “La sposa bambina” e lo spettacolo sulle Madri del ‘46
Bertalan Iván retweeted
La sfilata di nuovi modelli di AI continua.
📰"Lavoro e data center: perchè i signori dell'IA non sanno farsi amare" su @fattoquotidiano
#innovazione #AI @SGInnovazione @InnovaFiducia @MakerFaireRome @UnaTecnologia @INEST_digitaleu @lavoroformazion @womentech
Bertalan Iván retweeted
Molti mostrano solo mockup. Noi mostriamo siti live.
4 progetti negli ultimi mesi:
→ impresaedilegroupitalia.it
→ xenalogistic.it
→ mtgprintrun.com
→ comastridivision.com
Se hai un progetto da lanciare dopo l'estate, scrivici! marketing.evonbiz.com/r/jB/D…
Bertalan Iván retweeted
SEO helped Google find your pages.
AEO helped AI cite them.
WebMCP lets agents finish the job: research, buy, book, or request a quote.
Make your website agent-ready.
Bertalan Iván retweeted
Replying to @CoolestAK87
Webmcp-tool.com is the way. Put your url in get everything to fix beyond isagentic or isagentready
Bertalan Iván retweeted
The winner of an Anthropic hackathon open sourced his entire Claude Code setup. 68 subagents, 286 skills, 94 commands, MIT license
ECC turns Claude Code from one assistant into a full engineering team. It plans before it builds, writes the failing test first, then reviews its own work from a fresh context.
The agents, by job:
Planning. Turns a one-line request into a blueprint you approve before any code exists.
Review. Reads your diff in a clean context, with separate reviewers for Go, Python, TypeScript, Rust, Java and more.
Build repair. One resolver per toolchain, down to PyTorch and CUDA errors.
Security. An OWASP pass on your code, plus a scanner that audits your own agent config for injection risks.
Architecture. System design calls before they turn into migrations.
Domain work. Database queries, ML pipelines, end-to-end tests, docs.
The skills, by category:
Testing. tdd-workflow gates you from red to green, with eval-harness and verification-loop sitting on top of it.
Language packs. Idioms, testing and security for Python, Go, Rust, C++, Django, Laravel, Spring Boot, Next.js.
Context. search-first makes it read the docs before writing, iterative-retrieval keeps subagents from dragging your whole repo into the window.
Shipping. Docker, CI/CD, health checks, rollbacks, and migration patterns for Prisma, Drizzle and Django.
Work that is not code. Writing in your voice, market research with sources, pitch decks, slide decks.
Start with one plan and one rules pack. Installing all 286 skills at once is the fastest way to make it worse.































