The LoubLab blog
AI news, new tools, language models, and practical skills — in Arabic. Something new every day.
Moonshot's Kimi K3: 2.8T open weights from China
China's largest AI model yet claims it beats Claude Opus 4.8 and GPT-5.5 on coding benchmarks. Here's what that means for builders and everyday users.
Read article →Linus Torvalds: judge AI code by the patch
Linus Torvalds pushed back on calls to ban LLM-generated patches from open-source projects. His argument is simple: review the technical result like any other contribution.
Read article →PixVerse raises $439M — video AI expands
Singapore-based PixVerse reportedly raised $439M and passed a $2B valuation. The signal is not just the money: serious video-model competition is spreading beyond the usual US-China shortlist.
Read article →Nous Research is reportedly raising at a $1.5B valuation
TechCrunch reports that the maker of the open-source Hermes agent is in talks for new funding at a $1.5B valuation. The interesting part is the commercial path for an open-source agent lab.
Read article →Clawk gives your coding agent its own disposable Linux machine
Clawk is an open-source CLI that runs Claude Code, Codex, or a shell inside a disposable Linux VM, keeping the host machine outside the agent's reach.
Read article →Soofi S 30B: Europe's first serious open-source MoE model
A 30B German-English Mixture-of-Experts hybrid model trained on 27T tokens is releasing weights, code, and full data accounting. It's the most transparent sovereign model in Europe so far.
Read article →Apple's M7 Ultra could turn memory into an AI advantage
The Verge reports that technology from Apple's cancelled car project may feed into a future M7 Ultra with a much larger memory ceiling. That could matter for running bigger AI workloads locally.
Read article →Newer Claude models call tools worse — Armin Ronacher has the receipts
The creator of Flask says Opus 4.8 and Sonnet 5 emit more malformed tool calls than older Claude models. If you build agents, this hurts directly.
Read article →Claude Code leaked a stranger's session into a corporate workspace — what went wrong
An Enterprise ZDR user opened Claude Code and saw responses from someone else's session — a Minecraft temple build. The bug is now public, and it raises real questions about isolation promises.
Read article →Claudoro puts a Pomodoro timer inside Claude Code
Claudoro is a small open-source utility that puts Pomodoro work and break cycles in the Claude Code statusline, so your focus timer stays in the terminal.
Read article →GPT-5.5 Codex cuts its own thinking short — proof inside
A new telemetry analysis of 390k Codex replies shows GPT-5.5 stops reasoning at the same token counts over and over. If your tasks feel dumber, this is probably why.
Read article →Meta AI watermarks are not a full authenticity guarantee
A technical teardown of Meta's Stable Signature argues that invisible watermarks can be inconsistent and removable. Provenance is useful, but a single watermark should not decide what you trust.
Read article →Alibaba bans Claude Code — what it means for builders
Alibaba reportedly classified Claude Code as high-risk software and is pushing employees toward Qoder. The bigger story is the growing split between AI tools in China and the US.
Read article →GPT-5.5 Codex is cutting its own thinking short at 516 tokens
A GitHub analysis of 390k Codex responses finds GPT-5.5 disproportionately ends reasoning at exactly 516/1034/1552 tokens. Here's why that happens and what to do about it.
Read article →Gemini Spark is on Mac now — Google's 24/7 agent that works while you sleep
Google brought its 24/7 agentic assistant Gemini Spark to macOS. It sorts your files, builds Google Workspace docs from local files, and runs multi-step tasks. US-only for now, needs a Google AI Ultra plan.
Read article →Anthropic opens a HackerOne program for Claude Fable 5 jailbreaks
A jailbreak-severity scoring framework plus a public HackerOne disclosure program for Fable 5. The part that matters for everyday users is the new public reporting channel.
Read article →Anthropic launches Claude Science: a research app that reads the literature for you
Claude Science searches peer-reviewed papers, runs analyses on datasets you upload, and hands back a citable summary. First Anthropic product built outside the chat window.
Read article →Claude Fable 5 is back — what changed and how to try it
Anthropic's flagship model returned worldwide on July 1 after a two-week US export-control pause. Here's what Fable 5 actually unlocks and how to access it today.
Read article →