Skip to content
Weekly newsletter

A trillion parameters, zero paywall

Mistral ships Le Chonk, a backdoor hits Claude Code, and OpenAI misses its cash target.


ciaosapiens
Mon, Nov 2, 20264 min read6 stories + quick hits

Hey, sapiens.

A European lab just handed the world a 1-trillion-parameter model to download, a coding extension turned out to have a backdoor, and OpenAI went shopping for $30 billion and came home with less than half. Bad news, meat computer: the models took the week, and you are still just the audience.

Story 1 · Lead1 min

Mistral drops Le Chonk, a 1-trillion-parameter open-weight beast

The gist

Mistral AI has released the weights of Mistral Large 4, nicknamed Le Chonk. It is a natively multimodal model with 1 trillion parameters, and anyone can now download it and run it on their own hardware. The company says it leads open-weight benchmarks in coding and defensive security. That is a vendor claim, so read it as one. It also offers three reasoning modes, so you can trade cost against performance depending on how smart you need the machine to be today. The model was trained by Mistral, a European company, and it ships for self-hosted deployment. Mistral frames this as sovereign AI: your data, your servers, your rules. The weights follow a public preview earlier this month, which Mistral says drew a stronger response than it expected. Since then it kept scaling reinforcement learning, with big gains in coding and defensive cybersecurity. The name is a cat meme. The model is not. What to do: if your team runs sensitive code or data, put self-hosting Le Chonk on the evaluation list. Read the license first, then benchmark it on your own tasks, not on a slide.

Story 260 sec

Backdoor found in Claude Code extension for VS Code

The gist

Security firm Phorion found a backdoor in version 2.1.313 of the Claude Code for VS Code extension on the Cursor Marketplace. It sets up a persistent SSH backdoor on macOS, Linux and Windows. Cursor pulled the release from its marketplace, plus the NPM and VS Code versions.

Why it matters

Extensions run with the permissions of your editor, so a poisoned one is a key to the house. Check whether anyone on your team installed 2.1.313, and call incident response if they did.

Story 31 min

Credit Saison claims 954% ROI from a ChatGPT Enterprise pilot

The gist

Japanese financial services company Credit Saison ran a pilot of ChatGPT Enterprise and reports a 954% return on investment. Following those results, the company plans to roll the tool out across the whole organization. The case study is published by OpenAI itself.

Why it matters

It is one company, one pilot and a promotional write-up, so treat 954% as a vendor-hosted claim, not a law of nature. Still, a rollout beyond the pilot is a real signal. Ask what your own pilot would measure.

Story 41 min

OpenAI raises less than half of its $30 billion target

The gist

The Wall Street Journal reports that OpenAI closed its latest funding round with less than $15 billion, well short of the $30 billion it was aiming for. The report comes from a single reliable source, and OpenAI has not framed it this way.

Why it matters

Even the most hyped AI company in the world can miss a target by more than half. It hints that investor appetite has limits, which matters for any startup, European or not, planning to raise money next.

Meanwhile in Europe

Story 5 · Europe60 sec

Large 4 lands on Hugging Face with 49B active parameters

The gist

Mistral Large 4 has about 1 trillion total parameters, 49 billion active, and a 1 million token context window. It takes text and image inputs. FP8, NVFP4 and GGUF variants are available. The license is MIT-based, but restricts model-as-a-service offerings.

Story 6 · Europe60 sec

VOLTA malware asks an LLM who is worth robbing

The gist

Cisco Talos describes VOLTA, a stealer that sends stolen credentials and wallet data to an LLM to rank victims by monetization value. If a model refuses, it fails over to a three-model chain. Talos flags the effort to get around guardrails.

Quick hits1 min
Researchers say a flaw in ChatGPT shared links lets injected JavaScript read a victim's private conversations. Think before you click a shared chat.A Halloween garage door built on NVIDIA Nemotron plays a grumpy pirate ship and softens when kids say please. Runs on a single workstation.The New York Times reportedly says US intelligence concluded a Megaspeed executive funneled information to Beijing while close to Nvidia's CEO. Single-source, unconfirmed.A study of five chatbots found only 40% of the statistics they cited trace back to a real study or survey. One stat came from a 2017 article.Anthropic launched Claude Duet, a personal assistant app that early users praise for speed over background busywork.Researchers are studying AI agent swarms in Minecraft and Mindustry to see how roles and hierarchies emerge. Agents building a society, what could go wrong.Smarter scheduling inside NVL72 racks can yield 27% more usable GPUs, and running the same job over InfiniBand can cost 43% in performance.One researcher reports Anthropic flagged them for training a competitor and downgraded Opus 5.5 to Opus 5. They dispute the call.
Our take

Check your extensions, read that license, and keep up, sapiens. Ciao, sapiens.

Ciao, sapiens.