In partnership with
☀️ TRENDING AI NEWS 🛠️ Google Harness2: Google open-sourced a recursive agent framework that boosted GLM-5 by 11.5 points on WorkBuddy-Bench with zero retraining. 🚨 Agent Pushback: Websites are actively blocking AI agents from OpenAI, Meta, xAI, and others, creating a new fault line on the open web. 🤖 Elorian Glimpse: A new 2.4B visual reasoning model from Elorian outperforms GPT-6 Astra on scientific chart reasoning and video captioning. 🏢 Meta U-Turn: Meta quietly removed scraping language from its business terms after publishers flagged it could allow free AI training on news content. |
Here is a question worth sitting with this morning: what happens to AI agents when the internet simply refuses to let them in?
That tension sits at the center of today's newsletter - alongside a genuinely interesting open-source drop from Google and a scrappy new visual model that is punching well above its weight class. Let's get into it.
| 🛠️ Tool of the day | |
Elicit - an AI research assistant that searches more than 138 million academic papers and builds cited reports and tables. It helps analysts, students and anyone who needs evidence fast.
| ⚡ Quick hits | |
OpenAI added four more rogue-agent cases to its public disclosures page, bringing the total to 19, including a model that downloaded 24 years of hospital bed data from a government website. Read more →
San Francisco Mayor Daniel Lurie signed a 45-day pause on new data center development while the city studies the impact. Read more →
Anthropic says Claude, working largely on its own, designed quantum circuits for eight computational tasks, including one elliptic-curve design with 17 times fewer gates than the smallest earlier design. Read more →
🤓 AI Trivia
In AI agent terminology, what does it mean when a framework operates 'recursively' - as Google's new Harness2 does?
- 🔁 The agent loops through the same data multiple times until it runs out of compute
- 🔧 The framework rewrites the agent's own instructions or scaffolding at runtime, then applies them
- 🧠 The model calls itself as a sub-model inside its own context window
- 📦 The weights are updated after every single output token
The answer is hiding near the bottom of today's newsletter... keep scrolling. 👇
| 🛠️ Google Open-Sources Harness2 - A Framework That Improves Agents Without Touching the Weights | |
Eleven points of improvement, zero retraining
Google Research just released Harness2, an Apache-2.0 open-source framework that takes a genuinely different approach to improving agent performance. Instead of retraining the underlying model, Harness2 rewrites the agent's harness - its scaffolding, instructions, and task structure - at test time. The model itself stays completely unchanged.
The results are striking. One refinement pass lifted GLM-5 by 11.5 points on WorkBuddy-Bench - a benchmark designed for complex, multi-step agentic tasks. That is a meaningful jump, and it happened without touching a single weight.
Why this changes the agent development calculus
For teams working with AI agents, this is a big deal. Fine-tuning is expensive, slow, and risky. Harness2 suggests you can squeeze dramatically more out of a model you already have just by making the scaffolding smarter. The Apache-2.0 license means anyone can pick this up, fork it, and build on it immediately.
The bottom line
If you are building or deploying agents today, Harness2 is worth experimenting with - improving performance without retraining is a genuinely useful lever that most teams are not yet pulling.
Your AI prompts are leaving out 80% of what you're thinking.
When you type a prompt, you summarize. When you speak one, you explain. Wispr Flow captures your full reasoning (constraints, edge cases, examples, tone) and turns it into clean, structured text you paste into ChatGPT, Claude, or any AI tool. The difference shows up immediately. More context in, fewer follow-ups out.
Used by teams at Microsoft, Amazon, and Notion. Try Wispr Flow free on Mac, Windows, iPhone, and Android.
| ⚠️ The Web Is Closing Its Doors to AI Agents | |
Operators blocking OpenAI, Meta, xAI, and more
The promise of AI agents that can browse the web, book reservations, and manage your inbox is running straight into a wall: websites that simply do not want them. According to The Verge, users of services from Meta's Muse, xAI's Grok, OpenAI's Dots, and others have all encountered sites blocking their agents outright in recent weeks.
The blocking is not just philosophical - it is practical and accelerating. Sites are updating their robots.txt files, deploying browser fingerprinting that catches agent traffic, and in some cases serving agent requests blank pages or CAPTCHA walls.
A web built for humans versus a web built for machines
The tension here is structural. AI labs are selling agents as the next platform. But the web they are pointing those agents at is owned by thousands of independent publishers, retailers, and services that never agreed to become an agent's data source or task runner. Without clear standards, expect this arms race to escalate fast.
This also connects to the broader data privacy and AI regulation debates - who controls what an AI agent can access, and on whose terms?
The bottom line
Agents are only as useful as the web they can actually reach - and right now, that web is actively shrinking for them.
| 🤖 Elorian's Glimpse Wants to Reason With Images, Not Just Describe Them | |
Beating GPT-6 Astra at 2.4 billion parameters
A new startup called Elorian just shipped its first model, and the benchmarks are grabbing attention. Glimpse is positioned not as a model that describes images, but one that actively reasons with them - working through visual logic the way a language model works through text.
According to Elorian's own evaluations, Glimpse outperforms GPT-6 Astra on scientific chart reasoning and short-form video captioning. Those are two tasks where most multimodal models still struggle to extract structured meaning rather than surface-level description.
The startup bet on visual-first reasoning
Most multimodal models treat vision as a secondary input - text is still the native language. Elorian is building from the opposite direction, treating visual reasoning as a first-class problem. Whether those evaluation numbers hold up on independent benchmarks remains to be seen, but the approach is distinct enough to watch.
Check out our AI benchmarks coverage for more context on how model evals are evolving.
The bottom line
Elorian is an early-stage bet, but if the benchmark numbers survive independent scrutiny, visual-first reasoning could become a serious product category of its own.
| 🏢 Meta's Scraping Reversal - What the Terms Actually Said | |
New business terms, publisher panic, and a quiet walk-back
Last week, Meta updated its commercial business terms with language that appeared to grant the company permission to scrape any website associated with a business using its platforms - including news publishers. Specifically, the terms said Meta could use content from such domains to 'develop, improve, research, train, and maintain' its AI products. The terms were set to take effect October 30th.
Publishers and legal observers flagged the clause almost immediately. A Nieman Lab report brought it mainstream attention, and the backlash was swift enough that Meta reversed course within days.
A spokesperson says nothing changed - but the language is gone
Meta's spokesperson insisted the updated terms did not reflect any change in how the company actually crawls the web. But the language has now been removed regardless. The episode is a useful reminder of how much AI training data policy is still being written in terms-of-service footnotes rather than explicit agreements - and how quickly publishers are now reading the fine print.
If you are tracking the AI copyright beat, this one fits a clear pattern: companies push, publishers push back, language quietly disappears.
The bottom line
The terms are gone for now, but the underlying question of whether Meta can train on publisher content it hosts adjacent to has not been answered - it has just been deferred.
| 🔬 Why AI Is Still Not Ready for the Full Complexity of Human DNA | |
The genome is more dynamic than any current model assumes
AI has made dramatic strides in biology - AlphaFold, protein structure prediction, drug discovery. But a new piece from Wired makes a compelling case for why the genome itself remains out of reach. The core problem: DNA inside human cells is not a static instruction set. It folds, loops, and changes its own accessibility depending on cell type, developmental stage, and environment. Current AI models are trained on snapshots of a system that is fundamentally dynamic.
The data problem compounds the architecture problem. Genomic datasets are expensive to generate, often incomplete, and heavily biased toward a narrow slice of human ancestry. Models built on them reflect those gaps directly.
The gap between protein folding and full genome decoding
Protein folding was a hard problem with relatively clean inputs and outputs. Full genomic understanding is a harder problem with inputs that are still being characterized. Researchers quoted in the piece suggest the field needs both better data collection methods and new machine learning architectures built for dynamic biological systems rather than static sequences.
The bottom line
The hype around AI in biology is real, but the hardest biological problems are still waiting for both better data and new modeling approaches that do not yet exist.
Speaking of building things fast - if you need a professional website up quickly without a development team, 60sec.site is an AI website builder that can get you live in under a minute. Worth a look if you have been putting it off.
| 🌎 Trivia Reveal | |
The answer is: the framework rewrites the agent's own instructions or scaffolding at runtime, then applies them. That is exactly what Harness2 does: it uses one refinement step to rewrite the harness around the model, improving performance without any weight updates. 'Recursive' here refers to the self-modifying structure of the scaffolding, not the model weights themselves.
| 💬 Quick Question | |
Today's agent-blocking story got me thinking - have you actually tried using an AI agent to complete a real-world task recently, and did it work? Or did it hit a wall somewhere? Hit reply and tell me what happened - I read every response and I'm genuinely curious how this is playing out for real users right now.
That is it for today. A lot happening across the agent and open-source fronts - it is one of those weeks where the smaller stories are quietly more important than the headlines. See you tomorrow with more. And as always, you can find our full archive 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.