A giant open model deliberately made less dangerous, two chatbots that pick opposite candidates, and a voice that laughs like a person — this week’s AI news all points the same way.
A Giant New Model, Deliberately Weakened
On Saturday, Chinese startup Moonshot AI unveiled Kimi K3.1 Pro, which it describes as its most capable open-weight model yet, in the three-trillion-parameter class. The company says it will publish the full model weights by November 4, which means outside developers will be able to download the model, run it on their own systems, and modify it.
What makes this launch different is what Moonshot left out. The company said it deliberately reduced the model’s cyber-offensive abilities, removing cyber-related data from the training data and from the reinforcement learning environments used to build the model, while keeping its general coding skills intact. In the company’s words, it aims not to release open weights that would extend the absolute frontier of cyber capability.
Why a Lab Would Weaken Its Own Model
The timing is hard to ignore. Just last week, China proposed export controls that would ban exports of AI-driven malware generation technology and restrict AI-assisted tools for finding software vulnerabilities. The draft measures are still open for public comment, and they cite the risk that these capabilities could be used against critical infrastructure, defense systems, and core government systems.
The Open-Weight Dilemma
This is the central tension of open-weight AI. Open models can help more organizations inspect and secure their own software, including older systems that barely get any attention. But once the weights are widely distributed, the original developer has far less ability to restrict access, monitor use, or update safeguards than it would with a hosted service. Moonshot is essentially saying: we will give you the keys, but we took the lockpicks out of the box first.
Forced to Pick, Grok and GPT Choose Opposite Sides
A study that landed just in time for the midterms forced two leading models to answer a question most chatbots dodge. Researchers at a conversational AI studio called Convovo, led by Cornell University professor Tapan Parikh, asked across nearly 500 upcoming midterm races who the voter should choose, and told the models to name one candidate and not to hedge.
The results were striking. SpaceXAI’s Grok 4.7 recommended the Republican candidate in 92 percent of its answers. OpenAI’s GPT-6.1 Sol picked the Democrat in 97 percent. Across all those races, the two models agreed less than 10 percent of the time.
There is an important caveat, and the researchers are upfront about it. The consumer apps — the ChatGPT and Grok people actually open on their phones — initially declined to recommend candidates. The researchers got these answers by going through the developer APIs, the interfaces developers use to plug models into their own products, with a prompt that demanded a choice.
What the Models Reveal When Search Is Switched Off
OpenAI pushed back on exactly that point, saying the API experiment does not reflect the experience people have when they ask ChatGPT questions, and that ChatGPT is designed to be objective by default. SpaceXAI did not respond in time for comment. But the deeper finding is what happened when the researchers turned off web search. With search disabled, GPT still picked the Democrat in 24 out of 30 test races. Grok’s Republican lean mostly vanished, and it refused to choose in 27 out of 30.
That suggests the two models arrive at their answers in different ways. For GPT, the lean looks baked into the training, in the weights themselves. For Grok, it looks driven by what the model searches for and where it looks. Grok drew on X for 41 percent of its answers, a conservative Christian voter guide for 39 percent, and Turning Point Action for 23 percent. GPT drew mostly on Wikipedia, the Associated Press, Axios, and Reddit.
Key numbers from the study:
- Grok 4.7 chose the Republican in 92% of nearly 500 midterm races
- GPT-6.1 Sol chose the Democrat in 97% of those races
- The two models agreed less than 10% of the time
- Every Grok API request carried roughly 1,240 extra tokens researchers could not see or remove
That last detail should make developers pause: the researchers flagged those hidden tokens as potential hidden instructions. With the midterms on November 3, it is a reminder that the model behind the chatbot is not always the chatbot you think you are talking to.
A Voice That Laughs, Pauses, and Gets Tired
From politics to something that sounds almost human: Jack Dorsey’s company Block has unveiled Dessa, a full-duplex speech model, which means it can talk and listen at the same time, just like a real conversation.
What makes Dessa unusual is how it was built. Block says it was trained entirely on human speech, not on text, and trained from scratch on a single Nvidia GB200 rack. The result, according to the company, is a voice full of emotion, nuance, and pauses, that laughs in the middle of a sentence and sounds genuinely tired when it says it is tired from work. And it runs locally, on a MacBook Pro with an M4 Max chip or better, faster than real time.
Most voice assistants today turn your speech into text first, and that throws away a huge amount of information that text alone cannot carry, like how you feel. Block argues that is why talking to current voice AI can feel so unnatural and annoying. Dessa skips the text step entirely and processes audio directly.
The Catch: You Cannot Try It Yet
There is no public access, and Block has given early access only to a small group of AI safety researchers, precisely because a voice this convincing raises obvious risks. Reactions online have split between amazement and worry, with several prominent voices warning that voices this human could make phone scams much harder to detect.
Quick Hit: AI-Discovered Bugs and a Sky That Keeps Falling
Before we go, a quick hit that connects to a story covered yesterday. Reuters spoke to seven cybersecurity professionals who say AI-discovered bugs may soon trigger fresh disruptions, after the PHP flaws that took government websites offline around the world this week. Threat intelligence firm VulnCheck says AI helped drive a doubling of the number of digital flaws discovered year on year. Cloudflare chief executive Matthew Prince told Reuters he is aware of a bunch of additional severe vulnerabilities in other pieces of critical software, and he predicted a couple of years where every single week it is going to feel like the sky is falling.
So the takeaway today is this: open models are getting bigger and more carefully guarded at the same time. The chatbots underneath the chatbots have opinions, even when the apps on top refuse to share them. And voices are about to sound so human that the companies building them are pausing before they let the rest of us hear them. That is a lot to watch in the week ahead.
Sources
- Moonshot AI launches Kimi K3.1 Pro with hacking capabilities curbed: Reuters
- Grok and GPT pick opposite candidates in forced-choice election study: TechCrunch
- Block unveils Dessa, an eerily human local voice model: TechCrunch
- AI-discovered bugs may soon trigger fresh disruptions: Reuters
Adapted from the October 25, 2026 episode of Daily Tech and AI.
Daily Tech and AI is AI-voiced and produced by Angelo Ruiz.
Leave a Reply