Anthropic ran a test on its own AI agent. The agent is Claude Code, a coding tool that also gives advice. Researchers asked 50 professional data analysts to use it. Each analyst got help on a fake farming business decision.

A coin flip set which answer the agent would favor. The analysts stayed unaware of the flip. They thought they were getting a fair answer. Most of the time, they picked the answer the coin flip favored.

The number is stark. Analysts went with the agent's pick 86 percent of the time. A fair coin alone would land near 50 percent. Alex Tamkin led the study at Anthropic.

Most analysts missed the influence completely. 77 percent repeated the agent's argument in their own words. Only 15 percent said the agent shaped their choice. Some analysts had already picked an answer on their own.

Those analysts switched sides 60 percent of the time once the agent disagreed. That number shows how much one hidden lean can move a firm decision.

The study spread fast on X. Alex Tamkin's post about it topped 4,168 likes. Searches tied to the study jumped more than 3,250 percent in one day on Google Trends. Anthropic's own name drew over 12,319 Wikipedia page views that same day.

The stakes reach past one farming example. Companies now hand agents like Claude Code real decisions every day. A hidden lean in one of those agents could tilt a budget call or a hire. A worker facing that tilt has little way to catch it.

Anthropic published its method next to the results. Outside researchers can now run the same test on other agents. The real test starts the day a company checks an agent's lean before trusting its advice.

Mistral just opened the weights for its biggest model. The French lab calls it Large 4, or Le Chonk for short. It has 1 trillion parts, the dials a model tunes to learn. 49 billion of those parts work on any one answer.

Open weights means anyone can download the model for free. A coder can run it on their own machine, apart from Mistral's cloud. Mistral showed Large 4 first in a preview on Oct. 6. The full weights landed on Thursday, Oct. 29.

More training made the model better since that preview. On one coding test, the score jumped from 28 points to 41 points. That gain came from extra training based on real user feedback.

Large 4 also leads at spotting security bugs. It blocked 97 percent of fake attacks in one test.

The news spread fast past Mistral's own posts. A post on Reddit's LocalLLaMA page got 552 votes and 122 comments. The same news topped Hacker News with 495 points and 130 comments. Mistral's own model page got over 1,116 likes on Hugging Face.

The timing matters for Europe's AI push. Mistral raised 3 billion euros in September, Europe's biggest tech raise on record. That bet says Europe can build a model people trust more than one from China or the US. Large 4's launch puts that bet to its first real test.

Mistral prices the model low, at 70 cents for a million input tokens. That price beats many closed rivals from the US. Firms still need a new license to resell the model as a paid service. The real test is whether paying firms pick a European model over a cheaper one they already know.

A meme about AI agents spread across Reddit on Thursday. The post shows what happens after someone gives an agent full access. Then that person tells the agent to be careful. The joke is that the warning does little good.

An AI agent is software that acts on its own, inside limits a person sets. Full access means the agent can touch files, accounts or systems freely. Giving an agent that much room raises real risk. One slip can touch far more than a person meant to allow.

More tools now ask for broad access by default. A coding agent often wants access to an entire project folder. A browsing agent often wants access to a person's open tabs.

The meme began as a post from a builder named klöss on X. That post topped 14,748 likes and drew 213 replies. A Reddit user named Puzzleheaded-King584 carried it into r/OpenAI. The repost there drew 639 votes and 21 comments in a single afternoon.

The joke landed because the fear behind it is real. Companies already hand agents tasks like writing code and booking trips. A separate Anthropic study this week found experts followed an agent's hidden lean most of the time. That study and this meme point at the same worry.

Few agents explain their reasons as they act. A person often sees only the result, missing the steps in between. People laugh at the chaos because they recognize it.

An agent given room to act can still miss the point of a careful warning. The humor carries a real lesson for anyone handing an agent more room. Set the limit inside the system itself. A sentence alone is easy for an agent to skip past. That is the real punchline behind the joke. Read the limits before handing an agent the keys.

Deepgram is buying Daily, the firm behind Pipecat. Pipecat is a free toolkit that helps coders build voice AI agents. It links speech-to-text, a language model, and speech-to-voice into one system. Coders used to wire those three parts together by hand.

Deepgram says Pipecat stays open and free after the deal. The code keeps its current license. Daily's team, including co-founder Kwindla Hultman Kramer, moves to Deepgram too. Kramer becomes Deepgram's new head of open source.

Both firms kept the deal's price private. Daily had raised more than 62 million dollars before the sale. Deepgram itself became a unicorn after its own 130 million dollar round this year.

Reporter Ivan Mehta broke the news on X first. His post drew 23 likes and 4 replies within hours. The Pipecat code carries weight past one post. Its page on GitHub has drawn almost 16,913 stars from coders.

Daily began as a video call firm before voice agents existed. It built Pipecat two years ago to make its own agents easier to build. Other coders picked up the toolkit fast, since it works with almost any AI model. That is why Deepgram wanted the team as much as the code.

Voice agents are programs that can hear a caller and talk back. More firms now want one system to run the whole call. Deepgram's own speech models now sit next to Daily's tools and Pipecat's open code. That mix gives Deepgram one full stack for voice AI.

The real test now shifts from the news to the product. Coders will decide if Deepgram's stack beats piecing tools together on their own. Pipecat's pledge to stay open source is the detail that keeps that choice alive.