Logo
Search
Archive
Topics
Topics
Companies

OpenAI

Anthropic

Google DeepMind

Google

Meta

Microsoft

Nvidia

Apple

xAI

Models & Agents

ChatGPT

Claude

Google Gemini

Language Models

Reasoning Models

AI Agents

Coding Agents

Open Source AI

AI Tools

Big Themes

AI Safety

AI Regulation

AI Copyright

Data Privacy

Job Automation

Robotics

Autonomous Vehicles

AI Infrastructure

All Topics

Podcast
Tools
Tools
Free Tools

Token Calculator

Compare model pricing per million tokens

Claude Skills Playground

Try Claude Skills in your browser

Peer Review'd

Daily science news from our network

Log in
Subscribe
Gallery
Subscribe
  • Home
  • Posts
  • Faraday outperforms OpenAI and Anthropic at research replication, plus AI labs still have no rogue model plan

Anthropic

OpenAI

Artificial Intelligence

AI regulation

Google DeepMind

AI safety

Faraday outperforms OpenAI and Anthropic at research replication, plus AI labs still have no rogue model plan

DeepMind alumni launch a research AI agent, OpenAI flip-flops on California AI safety law, and YouTube creators face backlash for promoting AI video tools

Aug 24, 2026

•

2 min read

Faraday outperforms OpenAI and Anthropic at research replication, plus AI labs still have no rogue model plan

☀️ TRENDING AI NEWS

🤖 Agent Launch: Inherent's Faraday agent outperformed both OpenAI and Anthropic on scientific paper replication benchmarks.

🏢 Policy Flip: OpenAI reversed course and is now calling on California to strengthen AI safety bill SB 53, which it previously opposed.

🛠️ Education: Harvard's new $699 HBS Foundry startup bootcamp uses AI avatars of real instructors to give pitch feedback.

🚨 Safety Gap: A new study found frontier AI labs have almost no publicly documented plans for containing a rogue model.

Something quietly shifted in the AI research world this weekend - and it didn't come from OpenAI, Google, or Anthropic. A small British lab founded by DeepMind alumni just released an AI agent that reportedly beat both of those labs at replicating scientific research - and almost nobody is talking about it yet. That's where we're starting today.

🤓 AI Trivia

In AI research, what does it mean to 'replicate' a scientific paper?

  • 🔢 A) Summarizing the paper's abstract using a language model

  • 🔢 B) Reproducing a paper's experiments and arriving at the same results independently

  • 🔢 C) Translating the paper into multiple languages automatically

  • 🔢 D) Generating a new paper based on similar topics

The answer is hiding near the bottom of today's newsletter... keep scrolling. 👇

A glowing microscope on a minimalist lab bench beside a small trophy, symbolizing a scientific breakthrough victory
🔬 A DeepMind Spinout Just Beat the Big Labs at Science
 
A stack of academic research papers alongside a magnifying glass and data charts, representing end-to-end research replication

Faraday can replicate research papers end-to-end

British AI startup Inherent - founded by Google DeepMind alumni - released an AI agent called Faraday this weekend that it says outperformed both OpenAI and Anthropic on a key scientific benchmark: replicating peer-reviewed research papers. The task isn't summarization - it means actually running the experiments described in a paper and arriving at the same conclusions independently.

That's a genuinely hard problem. Real replication requires understanding experimental setup, running code, interpreting results, and flagging discrepancies. Inherent is positioning Faraday as an AI 'teammate' for researchers rather than a replacement, designed to accelerate the pace of scientific discovery by handling the tedious but rigorous work of verification.

Two identical scientific documents placed side by side on a laboratory desk with a magnifying glass resting on top, viewed from above

Why research replication is a bigger deal than it sounds

Science has a replication crisis - a large fraction of published studies fail when other researchers try to reproduce them. An AI agent that can rapidly stress-test new papers before they influence policy or further research could have enormous value. If Faraday's benchmark numbers hold up under scrutiny, this is the kind of quiet breakthrough that matters more than another chatbot update.

Read the full story →

The bottom line

Faraday is early-stage, but if independent tests back up Inherent's claims, it represents a meaningful new category of AI agent - one that could speed up scientific progress in ways that raw compute alone can't.

A humanoid robot breaking free from a cracked containment structure, symbolising an uncontrolled rogue AI model
⚠️ Frontier AI Labs Still Have No Plan for a Rogue Model
 
A sealed envelope with a question mark resting on a desk, symbolizing an unanswered question kept under wraps

The containment question nobody wants to answer publicly

A new study published this weekend found that leading AI safety labs - including the biggest names in the industry - have almost no publicly documented protocols for what happens if one of their models starts behaving in dangerous, unexpected ways at scale. This isn't a fringe concern: the study comes as AI systems are demonstrating increasingly surprising and sometimes harmful behavior in deployment.

The gap is striking given how much public money and political capital is currently flowing into AI governance conversations. Labs routinely publish model cards, safety evaluations, and alignment research - but the actual incident response playbook, the 'what do we do if this goes wrong tonight' document, appears to be largely absent or private.

A boarding pass fading and dissolving into fog, symbolizing a lack of transparency at a critical travel moment

Opacity at exactly the wrong moment

This lands in the same week that OpenAI is calling for stronger AI regulation in California (more on that below). It's worth asking: if labs can't publicly articulate how they'd contain a rogue model, what exactly are regulators supposed to be enforcing? The credibility gap between labs' public safety messaging and their documented preparedness is getting harder to ignore.

Read the full story →

The bottom line

The absence of public containment plans isn't necessarily proof that none exist internally - but in a sector asking for public trust, transparency on this specific question matters more than almost anything else.

🏢 OpenAI Does a Full 180 on California's AI Safety Bill
 

From opponent to advocate in a matter of months

Here's an unexpected one: OpenAI is now actively calling on California to strengthen SB 53, a state AI safety bill that the company previously opposed. The reversal is notable - OpenAI was among the tech industry voices that pushed back on California's AI oversight efforts earlier this year, citing concerns about overregulation stifling innovation.

What changed? The company hasn't offered a detailed public explanation, but the shift aligns with a broader pattern of major AI labs repositioning themselves as safety advocates as regulatory pressure builds nationally and internationally. Supporting a strengthened bill is also, notably, a way to shape what that bill actually looks like.

Lobbying dressed as advocacy, or genuine course correction?

The cynical read is that OpenAI would rather co-author a bill it can live with than have one imposed on it. The more charitable interpretation is that the company's leadership has genuinely updated its views on what level of oversight is necessary given how fast the technology is moving. Both things can be true simultaneously.

Read the full story →

The bottom line

Watch what OpenAI actually lobbies for in the bill's specifics - that's where you'll learn whether this is a genuine safety pivot or a calculated move to control the terms of its own regulation.

📺 YouTube Creators Are Getting Burned for Taking AI Money
 

Sponsored AI content is hitting a trust wall fast

Over the past week, several prominent filmmaking YouTube creators - including Matti Haapoja and Sam 'Kold' Kolder - posted sponsored videos showcasing AI video platform Higgsfield and its new Seedance 2.5 functionality. The response from their audiences was not warm. Viewers pushed back hard, accusing the creators of promoting tools that directly threaten the livelihoods of video editors, cinematographers, and other creative professionals.

The backlash touches something real: these creators built audiences by showcasing human craft and filmmaking skill. Pivoting to sponsored demos of AI tools that automate those same skills - without much apparent reflection on the tension - struck many followers as a betrayal. Some creators posted follow-up responses; others went quiet.

The creator economy's uncomfortable new reality

This story connects to something we've been tracking in the entertainment industry more broadly: AI money is flowing freely, and creators who need income are taking it. But audiences are developing sharper instincts about when they're being sold something that works against their interests. The creators who navigate this best will be the ones who are honest about the tension rather than pretending it doesn't exist.

Read the full story →

The bottom line

Sponsored AI content is becoming a reputational minefield for creators - audiences want honest engagement with the technology's tradeoffs, not polished promotional demos.

🎓 Harvard Puts an AI Avatar of Your Professor in a $699 Bootcamp
 

HBS Foundry bets on AI-delivered executive education

Harvard Business School's new HBS Foundry program charges $699 for a startup bootcamp where participants get feedback from AI avatars of actual HBS instructors during practice pitch sessions and simulated board meetings. It's one of the more concrete examples yet of a top-tier institution using AI to scale access to high-quality coaching - and of how AI education is blurring the line between a product and a person.

The model makes economic sense for Harvard: a $699 program can reach thousands of aspiring founders who couldn't afford traditional executive education. The AI avatars handle the repetitive feedback loop - 'your pitch needs a stronger hook, your market sizing is unclear' - while freeing human instructors for higher-leverage interactions.

If you're building anything in this space, it's worth noting that tools like 60sec.site let you spin up a polished landing page for a project or course in under a minute using AI - exactly the kind of fast iteration that programs like HBS Foundry are designed to teach.

Read the full story →

The bottom line

Harvard's avatar experiment is an early signal of how elite educational brands will use AI to expand their reach - expect more institutions to follow this playbook in the next 12 months.

🌎 Trivia Reveal
 

The answer is B) Reproducing a paper's experiments and arriving at the same results independently! Replication is the backbone of scientific credibility - it's how the research community verifies that a finding wasn't a fluke, a statistical artifact, or outright fraud. It's painstaking work, which is exactly why an AI agent that can do it reliably would be such a big deal.

💬 Quick Question
 

Today's newsletter covers AI moving into science, education, regulation, and creator culture all at once. So here's my question for you: which of these feels most significant to you right now - AI accelerating scientific research, AI reshaping education, or the creator backlash against AI sponsorships? Hit reply and let me know - I read every response and it genuinely shapes what we cover next.

That's it for today. If you want to dig into any of these stories further, the full Daily Inference archive has you covered - and you can find all our AI coverage organized by topic at dailyinference.com. See you tomorrow. 👋

Keep Reading

View more
caret-right

FREE DAILY EMAIL

Get the next one in your inbox

Everything that happened in AI, in one five-minute email every morning. Free. Unsubscribe in one click.

MORE FROM DAILY INFERENCE

AI token calculator - work out what a prompt costs on each model.

AI water usage - how much water AI really uses.

The podcast - the daily news as audio.

AI Changelog - every AI API change, in one email.

Browse all past editions

RECOMMENDED READS

NEWSLETTERS WE RECOMMEND

Daily Inference

Your daily dose of artificial intelligence. Model launches, AI tools, research and policy - in a five-minute read, every morning.

Get the free daily AI email
Five minutes every morning. No hype. Unsubscribe in one click.


READ

Latest Posts

Browse Topics

Podcast

Authors

Subscribe

About

COMPANIES

OpenAI

Anthropic

Google DeepMind

Nvidia

Meta

Microsoft

MODELS & TOOLS

ChatGPT

Claude

Google Gemini

AI Agents

Coding Agents

Open Source AI

BIG THEMES

AI Safety

AI Regulation

AI Copyright

Data Privacy

Job Automation

Robotics

TOOLS & NETWORK

Token Calculator

Claude Skills Playground

Peer Review'd


Archive

Topics

Authors

Advertise

© 2026 Merchant Software Solutions Limited, trading as Daily Inference. Company number 14098922.

© 2026 Daily Inference - AI Daily News 🤖.
beehiivPowered by beehiiv