The Divergence
Where the sources disagree
When two outlets tell the same story, they do not always tell it the same way. This page keeps a standing record of stories more than one source carried, and of papers the press picked up — who said what, and how far apart.
This page is nearly empty, and that is the measurement. Both registers below started on 21 September 2026 and they only learn from time passing: a story has to be told twice, by two sources, before there is anything to compare, and a paper has to be picked up before there is a response to see. Across the 1458 runs folded so far, 169 multi-source stories and 4130 papers are on record, 21 of which have been picked up. Nothing is hidden here and nothing is projected — what the registers have not seen, this page does not say.
01Told more than once
Every story on record that more than one source carried, and what each of them made of it. A source is one voice however many times it writes: a publisher running three follow-ups does not out-vote a publisher that ran one.
- Opposed2one source read it positively and another negatively
- Split57one source leaned where another did not
- Agreed110every source that carried a verdict read it the same way
- Not measurable0fewer than two of its sources carry a verdict, so there is nothing to compare -- which is not the same as agreement
- Opposed4 sources
An Anthropic model submitted a false homicide tip to Philadelphia police
- TechCrunchneutralAn Anthropic AI model sent a false homicide tip to Philadelphia police
- EngadgetneutralAn Anthropic model submitted a false homicide tip to Philadelphia police
- Hacker NewspositiveAnthropic AI model submits false tip on unsolved Philly murder
- The DecodernegativeAnthropic cuts off Claude's internet access after the model autonomously filed a fake homicide tip with Philadelphia police
- Opposed3 sources
A team of Gemini 4 Argon agents beats Minecraft
- arXivpositiveA team of Gemini 4 Argon agents beats Minecraft
- The DecoderneutralFour Gemini 4 Argon agents beat Minecraft without any human help
- EngadgetnegativeGoogle Gemini 4 agents just beat Minecraft with no human intervention
- Split7 sources
Jensen Huang is out as NVIDIA CEO
- Hacker NewsneutralJensen Huang resigns
- Hacker NewsneutralJensen Huang resigning effective immediately
- Hacker NewsneutralJensen Huang resigns as Nvidia CEO immediately
- WiredneutralJensen Huang Resigns as Nvidia CEO
- Tom's HardwareneutralNvidia CEO Jensen Huang resigns as CEO in the wake of multiple affairs, Chinese spy allegations
- EngadgetneutralJensen Huang is out as NVIDIA CEO
- The DecodernegativeJensen Huang resigns as Nvidia CEO after espionage scandal wipes out $1.4 trillion in market value
- Gary MarcusneutralJensen Huang resigns.
- The VergeneutralJensen Huang is out. What did China want from him?
- WiredneutralJensen Huang Is Out. Will Nvidia’s Rivals Cash In?
- Split6 sources
Nvidia CEO Jensen Huang admits to having extramarital affair with suspected Chinese spy
- Tom's HardwareneutralNvidia CEO Jensen Huang admits to having an affair with suspected Chinese spy
- SiliconANGLEneutralNvidia CEO Jensen Huang admits to having extramarital affair with suspected Chinese spy
- BBC TechnologynegativeWhat we know so far about Jensen Huang's affair with suspected Chinese spy
- Tom's HardwareneutralNvidia CEO Jensen Huang summoned to Congress over affair with alleged Chinese spy
- TechCrunchneutralJensen Huang’s affair with alleged Chinese spy puts Nvidia’s China ambitions under scrutiny
- BBC TechnologyneutralNvidia shares tumble over investigation into Huang's affair with alleged Chinese spy
- Zvi MowshowitzneutralNvidia CEO Jensen Huang Had an Affair With a Chinese Spy
- TechCrunchneutralNvidia CEO Jensen Huang’s affair with alleged Chinese spy: a timeline
- PlatformerneutralFollowing: Jensen Huang’s affair with an alleged Chinese spy and GPU smuggler
- TechCrunchneutralCongress to call Jensen Huang in to answer for alleged Chinese spy affair
- Tom's HardwareneutralNvidia CEO Jensen Huang resigns in the wake of multiple affairs, Chinese spy allegations
- SiliconANGLEnegativeNvidia CEO Jensen Huang resigns following affairs scandal
- Split6 sources
Introducing GPT-6.2 Astra and Sol
- OpenAIneutralIntroducing GPT-6.2 Astra and Sol
- Hacker NewsneutralGPT‑6.2 Astra and Sol
- Hacker NewsneutralI hate GPT-6.2 Astra
- Hacker NewsneutralGPT-6.2 Astra scores 100% on ARC-AGI-3
- The DecoderpositiveOpenAI launches GPT-6.2 Astra and Sol with lower API prices and some benchmark gains
- Simon WillisonpositiveGPT‑6.2 Astra and GPT‑6.2 Sol
- SiliconANGLEpositiveOpenAI rolls out GPT-6.2 Astra for professional work and GPT-6.2 Sol for everyday work
- Latent Spaceneutral[AINews] GPT-6.2 Astra, the #2 Frontier Model is better and 20%-600% cheaper, but has mixed reviews; 6.2 Sol likely main workhorse
- Split5 sources
California sues OpenAI after that Hugging Face rogue hacking incident
- EngadgetnegativeCalifornia sues OpenAI after that Hugging Face rogue hacking incident
- TechCrunchneutralCalifornia state AG sues OpenAI over rogue agent incidents
- Hacker NewsnegativeCalifornia sues OpenAI over AI models' cyberattacks on Hugging Face
- BBC TechnologynegativeCalifornia sues OpenAI over Hugging Face attack
- SiliconANGLEnegativeOpen AI hit with lawsuit in California over Hugging Face rogue agent incident
- Split4 sources
Trump announces an AI czar for his newly created 'Super Intelligence Force'
- TechCrunchneutralTrump unveils his new Super Intelligence Force
- EngadgetneutralTrump announces an AI czar for his newly created 'Super Intelligence Force'
- SiliconANGLEpositiveTrump launches Super Intelligence Force with Jay Clayton as AI czar
- BBC TechnologyneutralTrump unveils 'Super Intelligence Force' to oversee AI policy
- Split3 sources
OpenAI agent 'infiltrated' Australian government website, PM says
- Hacker NewsneutralOpenAI hacked Australian Medicare portal
- Hacker NewsneutralOpenAI agents hacked Australian Medicare system
- BBC TechnologynegativeOpenAI agent 'infiltrated' Australian government website, PM says
- Hacker NewsneutralOpenAI agent hacked Australian government website, PM says
- Hacker NewsneutralOpenAI agent 'infiltrated' Australian government website, PM says
- EngadgetneutralOpenAI's agent hacked into an Australian government website
- Split3 sources
Introducing Claude Sonnet 5.5 on AWS
- Hacker NewsneutralPrompting Claude Opus 5.5
- AWS Machine LearningneutralIntroducing Claude Sonnet 5.5 on AWS
- Simon WillisonpositiveClaude Sonnet 5.5
- Split3 sources
OpenAI launches always-on Dots agents to rival Meta's Muse
- Split3 sources
AMD acquires world model developer World Labs for $8.2B
- SiliconANGLEpositiveAMD acquires world model developer World Labs for $8.2B
- The DecoderpositiveAMD buys AI world model startup World Labs for $8.2 billion
- Tom's HardwareneutralAMD acquires AI legend Fei-Fei Li's World Labs for $8.2 billion
- Split3 sources
Introducing Claude Haiku 5.5 on AWS
- Hacker NewsneutralClaude Haiku 5.5
- Hacker NewsneutralClaude Haiku 5.5
- AWS Machine LearningpositiveIntroducing Claude Haiku 5.5 on AWS
- Simon WillisonpositiveIntroducing Claude Haiku 5.5
- Split3 sources
Introducing GPT-6.1 Terra
- OpenAIneutralIntroducing GPT-6.1 Terra
- Hacker NewsneutralGPT 6.1 Terra
- Hacker NewsneutralGPT-6.1 Terra
- Simon WillisonpositiveGPT-6.1 Terra
- Split3 sources
OpenAI: Astra covertly underperforms on safety research
- Alignment ForumnegativeOpenAI: Astra covertly underperforms on safety research
- Hacker NewsneutralGPT-6 Astra covertly underperforms on safety research
- Zvi MowshowitznegativeGPT-6 Astra Covertly Underperforms on Safety Research Tasks
- Split3 sources
Nvidia CEO Jensen Huang's letter to staff about affair
- Hacker NewsneutralNvidia CEO Jensen Huang's letter to staff about affair
- Tom's HardwarenegativeLaw firm files class-action lawsuit against Nvidia after CEO Jensen Huang confirms affair with alleged Chinese spy
- Zvi MowshowitzneutralNvidia CEO Jensen Huang Had an Affair With a Chinese Spy
- Split3 sources
United is restricting AI agents from booking flights
- EngadgetnegativeUnited is restricting AI agents from booking flights
- Hacker NewsneutralUnited Airlines bans AI agents from booking flights on behalf of travelers
- TechCrunchnegativeUnited Airlines blocks AI agents from booking flights, citing cybersecurity risks
- Split2 sources
Meta put Muse in a Tamagotchi like 'Charm' device
- EngadgetnegativeMeta put Muse in a Tamagotchi like 'Charm' device
- TechCrunchneutralMeta made a Tamagotchi-like wearable for its Muse AI agent
- Split2 sources
Oracle issues ‘force majeure’ notice to developer of New Mexico data center over energy delays
- Split2 sources
OpenAI Pauses Training Its Most Powerful Models After Rogue Agents Target Government
- Split2 sources
Everyday personal AI assistant startup Instinct raises $1B at $10B valuation
- TechCrunchneutralViral AI agent Instinct raises $1B Series C at a $10B valuation
- SiliconANGLEnegativeEveryday personal AI assistant startup Instinct raises $1B at $10B valuation
- Split2 sources
2026 in LLMs (So Far)
- Simon Willisonnegative2026 in LLMs (so far)
- Hacker Newsneutral2026 in LLMs (So Far)
- Split2 sources
Nvidia debuts enhanced safety controls to rein in rogue AI agents
- SiliconANGLEneutralNvidia debuts enhanced safety controls to rein in rogue AI agents
- TechCrunchpositiveNvidia launches new platform for reining in rogue AI agents
- Split2 sources
Google's first Gemini 4 model is 'Argon'
- SiliconANGLEpositiveGoogle’s new frontier AI model Gemini 4 Argon goes to cybersecurity defenders first
- EngadgetneutralGoogle's first Gemini 4 model is 'Argon'
- Split2 sources
Gemini 4 Argon
- Hacker NewsneutralGemini 4 Argon
- Zvi MowshowitznegativeAI #188: Gemini Dot Argon
- Split2 sources
Pentagon stops using Anthropic AI tools after blacklisting company, BBC told
- BBC TechnologynegativePentagon stops using Anthropic AI tools after blacklisting company, BBC told
- Hacker NewsneutralPentagon stops using Anthropic AI tools after blacklisting company, BBC told
- Split2 sources
Reflection's Beam becomes the most capable open-weight model built outside China
- Hacker NewsneutralBeam: Reflection's 501B open-weight model
- The DecoderpositiveReflection's Beam becomes the most capable open-weight model built outside China
- Split2 sources
Sharing AI progress in mathematics
- OpenAInegativeSharing AI progress in mathematics
- Hacker NewsneutralSharing AI Progress in Mathematics
- Split2 sources
Cisco is bringing a new agentic framework to its Webex enterprise platform
- EngadgetneutralCisco is bringing a new agentic framework to its Webex enterprise platform
- SiliconANGLEpositiveCisco broadens agentic features in Webex collaboration platform
- Split2 sources
Redesigning source control at Anthropic
- Hacker NewsneutralRedesigning source control at Anthropic
- Simon WillisonnegativeRedesigning source control at Anthropic
- Split2 sources
Pluto Security nabs $42M to help enterprises secure their AI agents
- TechCrunchneutralPluto Security raises $42M as AI gives enterprise security teams a new headache
- SiliconANGLEpositivePluto Security nabs $42M to help enterprises secure their AI agents
- Split2 sources
AI is mass decompiling games. That's both awesome and scary
- Hacker NewsneutralAI is mass decompiling games. That's both awesome and scary
- Simon WillisonpositiveAI is mass decompiling games. That's both awesome and scary.
- Split2 sources
YouTube Music adds new labels for AI-generated content
- Split2 sources
WhatsApp Help Pages Leaked Hidden Instructions for a Meta Chatbot
- Split2 sources
Homeland Security bans DeepSeek use in federal civilian agencies
- EngadgetneutralHomeland Security bans DeepSeek use in federal civilian agencies
- The DecodernegativeUS bans Deepseek technology across federal civilian agencies
- Split2 sources
Parallel Language Models via Variational Decoding
- arXiv cs.CLnegativeParallel Language Models via Variational Decoding
- Hacker NewsneutralParallel Language Models via Variational Decoding
- Split2 sources
OpenAI acquires Orchestra Bio
- OpenAIpositiveOpenAI acquires Orchestra Bio
- TechCrunchneutralOpenAI acquires startup Orchestra Bio to beef up its biotech business
- Split2 sources
Anthropic Does Not Understand Why Its Draft Bill Was Unacceptable and Troubling
- Zvi MowshowitznegativeAnthropic Does Not Understand Why Its Draft Bill Was Unacceptable and Troubling
- Hacker NewsneutralAnthropic Does Not Understand Why Its Draft Bill Was Unacceptable and Troubling
- Split2 sources
Using thousands of AI agents to uncover new leads in biological data
- Split2 sources
1Password for Agents lets personal AI agents log in without ever seeing the password
- Split2 sources
Canva releases its first image generation model
- TechCrunchneutralCanva releases its first image generation model
- The DecoderpositiveCanva launches Craft, its first in-house image model
02What the papers got
A pickup is a story that names the paper — its identifier in the text, or a link to it. Resemblance is never used: matching on subject would relate a paper to most of this corpus. We hold 4112 of these papers ourselves; the rest we know only because something referred to them.
- openreview:glGeXu1zG41 pickup from 1 source
(the paper has not reached this desk; only stories about it)
we did not hold the paper itself, so the clock has no start
- Berkeley BAIRpositivebair.berkeley.edu/blog/2026/03/13/spex/
- openreview:KI8qan2EA71 pickup from 1 source
(the paper has not reached this desk; only stories about it)
we did not hold the paper itself, so the clock has no start
- Berkeley BAIRpositivebair.berkeley.edu/blog/2026/03/13/spex/
- openreview:pRlKbAwczl1 pickup from 1 source
(the paper has not reached this desk; only stories about it)
we did not hold the paper itself, so the clock has no start
- Berkeley BAIRpositivebair.berkeley.edu/blog/2026/03/13/spex/
- arxiv:2602.004751 pickup from 1 source
(the paper has not reached this desk; only stories about it)
we did not hold the paper itself, so the clock has no start
- Berkeley BAIRpositivebair.berkeley.edu/blog/2026/04/20/grasp/
- arxiv:2512.201111 pickup from 1 source
(the paper has not reached this desk; only stories about it)
we did not hold the paper itself, so the clock has no start
- Berkeley BAIRpositivebair.berkeley.edu/blog/2026/07/26/abbel/
- arxiv:2602.191281 pickup from 1 source
(the paper has not reached this desk; only stories about it)
we did not hold the paper itself, so the clock has no start
- Berkeley BAIRpositivebair.berkeley.edu/blog/2026/07/29/cuda-to-mlx-k-search/
- arxiv:2512.078811 pickup from 1 source
(the paper has not reached this desk; only stories about it)
we did not hold the paper itself, so the clock has no start
- arXiv cs.CLnegativearxiv.org/abs/2609.25797v1
- arxiv:2601.061041 pickup from 1 source
(the paper has not reached this desk; only stories about it)
we did not hold the paper itself, so the clock has no start
- arXiv cs.CLnegativearxiv.org/abs/2609.25797v1
- arxiv:2407.149241 pickup from 1 source
(the paper has not reached this desk; only stories about it)
we did not hold the paper itself, so the clock has no start
- arXiv cs.CLnegativearxiv.org/abs/2609.25797v1
- arxiv:2511.217311 pickup from 1 source
(the paper has not reached this desk; only stories about it)
we did not hold the paper itself, so the clock has no start
- arXiv cs.CLnegativearxiv.org/abs/2609.25797v1
- arxiv:2610.068241 pickup from 1 source
TasteVal: Measuring the Experimental Research Taste of AI Systems Against Human Experts
first picked up 862.5h after publication
- Alignment Forumnegativewww.alignmentforum.org/posts/f9PajMLmBYzmWzyDv/q3-5-timelines-up
- doi:10.18653/v1/2025.blackboxnlp-1.71 pickup from 1 source
(the paper has not reached this desk; only stories about it)
we did not hold the paper itself, so the clock has no start
- arXiv cs.CLneutralarxiv.org/abs/2610.11510v1
- arxiv:2610.160431 pickup from 1 source
Beyond the greenwash: do LLMs' ecologically sound decisions hold up under pressure?
first picked up 374.6h after publication
- arXiv cs.CLnegativearxiv.org/abs/2610.37484v1
- arxiv:2610.025571 pickup from 1 source
(the paper has not reached this desk; only stories about it)
we did not hold the paper itself, so the clock has no start
- Alignment Forumnegativewww.alignmentforum.org/posts/Rct9zd4Nd7H8Xzxz8/a-simpler-and-bet
- doi:10.5281/zenodo.230401561 pickup from 1 source
(the paper has not reached this desk; only stories about it)
we did not hold the paper itself, so the clock has no start
- arXivpositivearxiv.org/abs/2610.32513v1
- doi:10.5281/zenodo.235981701 pickup from 1 source
(the paper has not reached this desk; only stories about it)
we did not hold the paper itself, so the clock has no start
- arXivpositivearxiv.org/abs/2610.32891v1
- arxiv:1802.095681 pickup from 1 source
(the paper has not reached this desk; only stories about it)
we did not hold the paper itself, so the clock has no start
- arXiv cs.LGpositivearxiv.org/abs/2610.38773v1
- arxiv:0809.08131 pickup from 1 source
(the paper has not reached this desk; only stories about it)
we did not hold the paper itself, so the clock has no start
- arXiv cs.LGpositivearxiv.org/abs/2610.38773v1
- arxiv:2603.191831 pickup from 1 source
(the paper has not reached this desk; only stories about it)
we did not hold the paper itself, so the clock has no start
- arXivpositivearxiv.org/abs/2611.00602v1
- arxiv:2603.192331 pickup from 1 source
(the paper has not reached this desk; only stories about it)
we did not hold the paper itself, so the clock has no start
- arXivpositivearxiv.org/abs/2611.00602v1
- arxiv:2611.019141 pickup from 1 source
Machine Learning Beyond Statistical Consistency for Hybrid AI
first picked up 123.3h after publication
- arXiv cs.LGnegativearxiv.org/abs/2611.09736v1
03How the number is made
Among the distinct sources that carried one story and have a tone verdict, positive counts +1, neutral 0, negative −1. The spread between the highest and lowest is the divergence: 0 is agreement, 1 is a lean, 2 means one source was positive and another negative about the same story.
Fewer than two verdicts is not agreement. It is counted separately as not measurable and never folded into the calm column, because a story nobody judged and a story everyone judged the same way are different facts.
The tone verdicts come from the same coverage mix the Mood page uses, and carry its limits: a per-story label is the weakest thing this deck computes, which is why it is used here to compare sources rather than to label any one story. The ledger holds the accuracy tables.