• Anthropic's Claude Code developer tool leaked.  (Dev.To)
  
  Which...  Well, so what?  You can download it.  Countless thousands of people have.  I have.  Anyone who wanted to put in the effort to pick it apart could have done so.
  
  Anthropic left a debug option set it one release and that made all the source files visible, but that just made it easier.
  
  The real brains - Anthropic's AI models like Sonnet and Opus - run safely on their servers and haven't leaked anywhere.
  
  If you're interested though it's available on GitHub.
  
  
  
• If you want to run your own LLM and not just local tools that talk to a remote server somewhere Bonsai from PrismML might be of interest.  (PrismML)
  
  Because the 1.7 billion parameter model runs in 240MB of memory - yes, M, not G - and churns through 130 tokens per second on an iPhone 17.
  
  Which uses noticeably less power than a rack full of high-end graphics cards.
  
  Bonsai 8B uses 1.15GB of RAM.
  
  While it doesn't lead in test scores, it's being tested against 16GB models, which require an entirely different class of hardware.  It would be interesting to see how a 70 billion parameter model would perform on the same tests if it's possible to perform the same trick - quantising the model down from half-precision (16 bits) per parameter to 1 bit with error correction.

• The IRGC has issued threats not just against the US and Israel and the Arab world - that's old hat - but now against Nvidia, Microsoft, Apple, Google, Meta, IBM, Cisco, and Tesla.  (Tom's Hardware)
  
  Gotta warn you guys, Tesla has a friend with a global lock on orbital space lasers.  Might want to tread carefully.
  
  
  
• Fujitsu is developing a new AI processor - an NPU - aimed at the upcoming 1.4nm process node.  (Tom's Hardware)
  
  That's not the interesting part.
  
  The interesting part is that this will be manufactured at the Rapidus plant in Hokkaido.  With Rapidus ramping up production Japan has leapt directly from 40nm chip fabrication to 2nm, largely catching up with Taiwan, Korea, and the United States, and leaving China trailing behind.
  
  
  
• Looks like we won't be rid of OpenAI just yet.  (Tech Crunch)
  
  They just raised $122 billion.  Apparently from wealthy masochists.
  
  
  
• GMKtec's new Evo T2 mini-PC is now shipping with Intel's Panther Lake CPU and it's class-leading graphics performance coupled with 64GB of soldered RAM.  (Liliputing)
  
  Just two problems: First, it costs $1899, and the Minisforum model I bought around Christmas cost about $600.  Sure, the integrated graphics on that model are a lot slower than Panther Lake, but not that much slower.  (This site suggests the difference is around 25%.)
  
  Second, it's out of stock.
  
  
  

Musical Interlude

Posted by: cxt217 at Thursday, April 02 2026 02:36 AM (ZLF73)

Posted by: PatBuckman at Thursday, April 02 2026 05:29 AM (s6adZ)

Hide Comments | Add Comment