Tweets & Toots.

“#ai” 1 à 10 sur 30

• Lol at these tools, which are doomed to chase after the other models.
  I used an LLM for this post, and for almost everything I write, especially in English or Japanese, which aren't my first languages.
  But I don't use them to generate the text: I always write the first draft myself, and then feed it to an LLM to fix and enhance it.
  Having written on my blog for more than 20 years has given me a strong sense of what good writing means to me.
  I'm glad it's flagged as "100% Human Written" on Pangram :D #ai https://www.pangram.com https://files.mastodon.social/media_attachments/files/117/273/381/183/500/971/original/2b985be936a41264.png 15/09/2026 à 07:23 Permalien
• To be completely honest, given how much #ai is progressing, I have serious doubts about whether it still makes sense to pursue software development. It could very well turn out to be a dead-end career path… 25/08/2026 à 16:50 Permalien
• #ai Claude Desktop on Mac uses a VM as part of its security model:
  
  $ du -sh ~/Library/Application\ Support/Claude/vm_bundles
  
  This approach has its downsides too: https://github.com/anthropics/claude-code/issues/65577 23/08/2026 à 15:25 Permalien
• Rétro-ingénierie des .DAT du "Le Petit Robert 2011" (app PowerPC/i386 morte depuis Catalina) avec Claude #ai :
  
  - désassemblage du binaire
  - conversion en bundle .dictionary macOS
  - style calqué sur le dictionnaire français built-in
  - corrections d'après captures d'écran (espaces, doublons, audio, mots introuvables, etc.)
  
  Résultat : 54 037 articles, 9 176 sons 🤯 https://marcarea.com/static/tweets/png/ceccf53a5c1bc2ec.png 19/08/2026 à 18:39 Permalien
• So #ai text watermarking is hidden in token selection, not visibly added to text.
  
  - Kirchenbauer: make green tokens more likely than red using a secret key (bias the dice)
  - Aaronson: sample from the model's distribution while controlling the randomness with a secret key (control the dice)
  - SynthID: sample candidates, then use a secret key to determine how they compete (run a secret tournament)
  
  All require the secret key for authoritative detection…
  
  https://declaude.org/watermarking/ https://daringfireball.net/2026/08/anthropics_watermark_text_adulteration_in_claude_is_a_perversion_of_writing 17/08/2026 à 17:50 Permalien
• I've been playing with both Affinity and Pixelmator Pro for a couple of days.
  
  Both are mind-blowing and more than capable replacements for Adobe Photoshop and Illustrator.
  
  Affinity will be my go-to solution for vector graphics, while Pixelmator Pro will be my raster-all-the-things solution.
  
  The #ai powered features are impressive. 17/07/2026 à 20:24 Permalien
• "Introducing the Safari MCP server for web developers" https://webkit.org/blog/18136/introducing-the-safari-mcp-server-for-web-developers/ #ai 02/07/2026 à 09:24 Permalien
• "LLM Architecture Gallery" https://sebastianraschka.com/llm-architecture-gallery/ #ai 21/06/2026 à 10:01 Permalien
• "The Newest Instagram Exploit is the Goofiest I've Seen" https://www.0xsid.com/blog/meta-account-takeover-fiasco
  
  You'd think this kind of vulnerability wouldn't happen to the big players.
  
  According to the post, an attacker would:
  
  1. find the victim's username
  2. use a VPN near the victim's location
  3. claim to Meta's #ai support that the victim's account has been hacked and ask to send the verification codes to an arbitrary email
  4. use the codes to reset the password and take over the account
  
  https://xcancel.com/DarkWebInformer/status/2061253599758315527 01/06/2026 à 22:10 Permalien
• RT @daringfireball Google I/O Keynote in 54 Seconds #ai https://xcancel.com/ArtemR/status/2056961743142957143 https://daringfireball.net/linked/2026/05/21/google-io-54-secs 21/05/2026 à 20:28 Permalien