In partnership with
☀️ TRENDING AI NEWS 🤖 Agent Scale: Microsoft open-sourced Agensh, a leaderless harness that coordinates up to 1,024 AI agents simultaneously. 🏢 Political Bias: A new study found Grok picks Republicans and GPT picks Democrats in over 90% of 500 election races tested. 🛠️ Regulation: Pennsylvania Governor Shapiro signed the first US law requiring labels on AI-generated advertising. 🚨 Cybersecurity: Cognition released an open-source AI plugin that auto-patches critical RCE flaws in PHP, Node.js and GitLab repos. |
Two AI models walk into a ballot booth - one pulls the lever for Republicans, the other for Democrats, and they disagree on more than 90% of races tested. That's not a joke. It's one of three stories today that quietly reframes how much the AI tools we use are already shaping the world around us.
| 🛠️ Tool of the day | |
Perplexity - an AI answer engine that searches the web and gives you a clear answer with its sources listed. It helps researchers, analysts and anyone who wants fast, checkable answers. It has a free version.
| ⚡ Quick hits | |
TechCrunch rounded up AI agents you can use by text message, including Instinct, which reached a $10 billion valuation after a $1 billion round in September. Read more →
A viral video of Nvidia CEO Jensen Huang playing with Pollen Robotics' open-source Microduck robots led to more than 20,000 preorders within a day. Read more →
TechCrunch published a running list of AI products and startups that shut down or fell short, including OpenAI's Sora, the Humane AI Pin and the Rabbit R1. Read more →
🤓 AI Trivia
Microsoft's new Agensh framework can coordinate up to 1,024 AI agents at once - but how do they stay in sync without a central orchestrator?
The answer is hiding near the bottom of today's newsletter... keep scrolling. 👇
| 🤖 Grok Picks Republicans, GPT Picks Democrats - And They Agree on Almost Nothing | |
90% disagreement across 500 races is not a rounding error
A new study tested xAI's Grok and OpenAI's GPT by forcing them to pick candidates in nearly 500 US election races. The result: Grok overwhelmingly recommended Republican candidates while GPT favored Democrats - and the two models disagreed in more than 90% of contests. That's not a minor tilt. That's a systematic divergence baked into the models themselves.
When the Training Data Has a Party Affiliation
Neither model is supposed to endorse political candidates - both companies have policies against it. But when researchers applied pressure and forced a choice, the underlying tendencies surfaced hard. It raises an uncomfortable question: if millions of voters casually ask AI assistants for election guidance, does the model's political gravity nudge them without anyone realizing it?
The bottom line
This study is a reminder that AI models are not neutral mirrors - they reflect the data and fine-tuning choices of the labs that built them, and those choices have real political weight during election season.
Master Claude AI (Free Guide)
The professionals pulling ahead aren't working more. They're using Claude.
Our free guide will show you how to:
Configure Claude to be the perfect assistant
Master AI-powered content creation
Transform complex data into actionable strategies
Harness Claude’s full potential
Transform your workflow with AI and stay ahead of the curve with this comprehensive guide to using Claude at work.
| 🛠️ Microsoft Open-Sources Agensh: 1,024 Agents, Zero Central Command | |
A leaderless swarm that scales from 1 to 1,024 workers
Microsoft Research just open-sourced Agensh, a multi-agent harness that coordinates up to 1,024 simultaneous AI agents - with no central orchestrator telling them what to do. Instead, self-organizing worker agents share three things: a Git repository, a group chat channel, and a findings board. That's it. No manager. No bottleneck.
ProgramBench Results on a 6-Hour Budget
The team tested Agensh on ProgramBench under a hard 6-hour compute budget, scaling from a single agent all the way to 1,024. The architecture is deliberately flat - any agent can read and write to the shared workspace, which means the system keeps working even if individual agents fail or stall. For developers building complex pipelines or long-horizon coding tasks, this is a meaningful structural shift away from hub-and-spoke agent designs.
If you've been following our coverage of AI agents evolving from single-step tools into full autonomous workflows, Agensh is the next logical step - a framework built for genuinely distributed, parallel intelligence.
The bottom line
Open-sourcing a 1,024-agent framework removes a major barrier for teams that want to experiment with large-scale agent coordination without building the infrastructure from scratch.
| ⚠️ Pennsylvania Passes America's First AI Advertising Disclosure Law | |
Every AI-generated ad must now say so - in the same format as the ad itself
Pennsylvania Governor Josh Shapiro signed the Artificial Intelligence Deceptive Advertising Act into law last Thursday, making Pennsylvania the first US state to legally require disclosures on AI-generated ads. The scope is broad: still images, video, audio, and text ads that have been 'significantly modified or generated by artificial intelligence' must carry a label - and that label must appear in the same format as the ad itself.
That last clause is the sharp edge of the law. A TV commercial can't bury the disclosure in a website footnote. A radio ad has to say it out loud. The requirement applies across TV, online platforms, print, and social media - which means every major ad channel is covered.
A Template Other States Are Watching
This comes directly in the wake of the Spain deepfake election ad controversy we covered yesterday - a reminder that deepfakes in political advertising are no longer hypothetical. Pennsylvania's law won't stop bad actors overnight, but it creates a legal framework other states can copy. With the US still lacking a federal AI advertising standard, expect more states to follow quickly.
The bottom line
If your brand runs AI-assisted creative in Pennsylvania, you now have a legal obligation to label it - and the burden of proof for what counts as 'significantly modified' will be tested in court soon enough.
| 🚨 The PHP Meltdown: How AI Turned a Bug Into a Vulnerability Tsunami | |
Security researchers used AI to find the flaw - then attackers used AI to exploit it
Cybersecurity experts are calling this 'Black October.' A severe remote code execution flaw was discovered in the PHP programming language - which powers a huge chunk of the web - and AI tools played a double role in the disaster. Researchers used AI to uncover the years-old vulnerability faster than traditional methods would allow. Then attackers used AI-assisted tools to begin exploiting it almost immediately after disclosure.
The Guardian quoted cybersecurity executives warning of a coming 'vulnerability tsunami' - the idea that AI dramatically compresses the time between discovery and exploitation, leaving defenders less room to patch. The same tools that help security teams find bugs faster also hand attackers a faster on-ramp.
Cognition's Open-Source Response
On the defensive side, Cognition released the Devin Vulnerability Fixes plugin - three MIT-licensed AI agent skills that automatically inventory, patch and test PHP, Node.js and self-managed GitLab deployments hit by this week's critical RCE flaws. The plugin opens one pull request per affected repo - a practical, targeted tool that any team running those stacks can deploy right now.
The bottom line
AI is simultaneously the best new tool for finding vulnerabilities and the force multiplier that makes exploiting them faster - and the PHP crisis is the clearest proof of concept yet that security teams need AI-native defenses, not just AI-augmented offense.
| 🏢 Jack Dorsey's Block Built a Voice Model Trained on Nothing But Human Conversation | |
Dessa skips the text pre-training entirely - and it shows in the output
Jack Dorsey's payments company Block announced an AI voice model called Dessa, and its origin story is unusual: it was trained 'from scratch on nothing but human conversation.' No text pre-training. No web scrape. Just speech. The result is a model that Block says sounds eerily human - and can run locally on a MacBook M4 Max.
Block shared sample conversations to demonstrate the model, and reviewers noted the exchanges have the slightly awkward, naturalistic rhythm of real human dialogue rather than the polished cadence most voice AI aims for. That might actually be the point - familiar friction can feel more trustworthy than synthetic perfection.
Speaking of building things that actually work on local hardware - if you're prototyping a product around voice AI or any AI feature, 60sec.site is worth a look. It's an AI-powered website builder that gets you from idea to live site in under a minute - no dev time required.
The bottom line
A voice model that runs locally on consumer hardware and skips text pre-training entirely is a genuinely different architectural bet - and if Block can productize it, Dessa could give privacy-focused developers a compelling alternative to cloud-dependent voice APIs.
| 🌎 Trivia Reveal | |
The answer is: they share a Git repo, a chat channel, and a findings board! Microsoft's Agensh deliberately avoids any central orchestrator. Instead, all 1,024 agents coordinate through these three shared surfaces, which keeps the system resilient if individual agents stall or fail. No manager required.
| 💬 Quick Question | |
The study showing Grok picks Republicans and GPT picks Democrats got me thinking - do you ever ask AI tools for opinions on anything political or civic? Or do you keep AI strictly off that territory? Hit reply and let me know your take - I read every response and I'm genuinely curious where people draw the line.
That's it for today - see you tomorrow with more from the fast-moving world of AI. And if a friend forwarded this to you, you can catch every edition at dailyinference.com.
| 🎁 Share Daily Inference | |
Know someone who wants to keep up with AI in 5 minutes a day? Share your link. 1 referral unlocks the AI Tools Starter Kit. 5 referrals unlock the AI Insider Briefing.