Authors push back as publishers and agents seek share of Anthropic settlement

The Hook

Today: the copyright fault line inside AI training-data deals cracked inward in an unexpected direction, a new benchmark documented exactly where frontier LLMs fail on telecom specifications, and a GPU performance paper posted a 385x number so rigorously supported it should change how you read every infrastructure comparison you have ever trusted.

The Signal

1. Authors vs. Their Own Publishers: The Anthropic Settlement Fractures Internally
The copyright dispute over AI training data just found a new front — and it is not between authors and AI companies. It is between authors and the publishers and literary agents who represent them. In the wake of Anthropic settling with a class of plaintiff authors, writers are now pushing back on their own advocates, who appear to be claiming a disproportionate share of the payments. The structural concern is serious: if publishers extract the bulk of a settlement that was framed as compensation for creators, the precedent corrupts every future AI licensing deal. The economic value of creative work in the training-data economy will have been quietly redirected to intermediaries. For anyone tracking the AI copyright resolution cycle — including how Google structures its Gemini training-data sourcing — this is the data point that matters most: settlements do not automatically resolve the underlying grievance if the money does not reach the people who made the work.

2. Seattle Times and Newsday Sue OpenAI and Microsoft
The plaintiff roster in the AI copyright wave added two more prominent names: The Seattle Times and Newsday have filed suit against OpenAI and Microsoft, alleging their journalism was used as training data without permission or payment. The legal theory is not new, but the name recognition is escalating. Regional publishers with deeply loyal local readerships joining the litigation signals this is no longer a coordinated campaign by a handful of national mastheads — it is becoming the default legal response to training-data use. The economic argument is sharp: these outlets spent decades building trusted journalism, and AI systems potentially benefited from that corpus without contributing to its sustainability. Every new high-profile plaintiff adds legislative pressure for training-data disclosure frameworks — frameworks that will directly shape how Google, Anthropic, and every other AI lab sources future training corpora. Watch this lane.

3. TeleTables: A Benchmark That Documents Where LLMs Break on Telecom Specs
A new benchmark called TeleTables has surfaced a specific, documented failure mode in large language models — including frontier-tier systems — when they encounter the dense table structures of 3GPP telecommunications standards. Telecom is one of the largest enterprise deployment vectors for AI assistants, and 3GPP specs are the technical lingua franca of the industry. The problem is structural: these documents encode relational and hierarchical data in table formats that differ fundamentally from the natural-language-dominant distributions most LLMs were trained on. TeleTables is a purpose-built evaluation suite that quantifies where and how models fail. For enterprise teams at telecom companies using Vertex AI or Gemini for specification interpretation tasks, this benchmark is the audit checklist that did not exist before today — and a direct argument for domain-specific fine-tuning on 3GPP-format data before deploying at scale.

4. From 80x to 385x: GPU Benchmark Asymmetry and What It Means for AI Infrastructure
A rigorous new arxiv paper exposes a pervasive methodological flaw in GPU kernel performance comparisons: asymmetric tuning. The mechanism is simple — one implementation is optimized by its author; the competitor is run as-found from a public repository. When a researcher corrected for this by applying equal optimization effort to both sides, a comparison that showed an 80x advantage flipped to a 385x figure in the opposite direction. The direction reversed and the magnitude grew nearly five times. This is not a finding about one GPU vendor or one model. It is a finding about the benchmarking methodology the entire industry uses to make hardware selection, framework adoption, and architecture decisions. For teams evaluating Gemini-scale workloads on TPU pods, A100s, or H100s, this paper is required reading before trusting any published throughput comparison. Asymmetric baselines are not rare exceptions — this research suggests they are the default.

5. Pitch-Class Steering for Diffusion-Based Music Generation
Diffusion models have dominated image generation but lagged autoregressive approaches in controllable music generation — until now. A new paper introduces latent-space pitch-class probes as a steering mechanism for diffusion-based music systems, enabling precise pitch-class control without retraining the underlying model. The practical payoff: you can take an existing music diffusion model and add fine-grained pitch control as a post-hoc steering layer. The paradigm is borrowed from language model interpretability research — the same probe techniques used to understand what LLMs 'know' internally — and is now crossing modalities into structured audio generation. For teams building on Google Lyria or similar diffusion-based audio backends, this opens a new class of compositional control interfaces that do not require expensive base-model retraining. The architectural implication is broad: probe-based steering may generalize to rhythm, timbre, and dynamics as well.

6. Beyond Aggregate Scores: Hidden Assumptions in Automated NLG Evaluation
Anyone running BLEU or BERTScore in a production NLP pipeline should read this paper carefully. Researchers have identified and catalogued the hidden behavioral assumptions baked into reference-based automated evaluation methods — assumptions that aggregate scores completely obscure. The core finding: meta-evaluation of NLG methods typically checks whether aggregate rankings correlate with human judgment, but ignores instance-level behavioral correctness. A method can rank correctly on average while failing systematically on specific output types — the types that often matter most in production. For teams using Gemini or other frontier LLMs in document generation, summarization, or translation workflows at scale, this is a calibration alert. Your evaluation pipeline may be telling you yAudit your methodology against the correctness criteria this paper surfaces before your next production evaluation cycle.

7. CAS-Brain Closes B+ Round at Hundreds of Millions of Yuan
CAS-Brain, a Chinese AI infrastructure company with roots in the Chinese Academy of Sciences, has closed a B+ round at hundreds of millions of yuan with an industry strategic lead investor. The strategic lead structure — rather than a pure financial VC — is the signal worth reading. Strategic leads at this stage typically buy ecosystem integration rights and preferred deployment partnerships, not just equity upside. In an active AI funding environment, this close is a data point that the AI infrastructure build-out remains well-capitalized on the Chinese side of the market despite macro headwinds. For readers tracking the global competitive landscape for AI compute and inference infrastructure, CAS-Brain is a name to add to the watch list. A deployment announcement tied to the lead investor's industrial vertical within 12 months of close is the likely next move.

8. Mixture of Modulated Experts for Multimodal Time-Series Forecasting
Real-world time-series data breaks single-modal forecasters: multiple modalities, distribution shift, evolving dynamics. A new paper proposes Mixture of Modulated Experts (MoME), an architecture designed specifically for this challenge. Rather than training one model that generalizes across all input regimes, MoME routes inputs dynamically through specialized expert modules, each calibrated to a different distributional context. The practical payoff is substantial: single-modal forecasters regularly break when real-world data shifts distribution, and MoME's routing mechanism provides a principled defense. The architecture is directly relevant to any team building production prediction pipelines — on Vertex AI or elsewhere — where data heterogeneity is a known problem. For Gemini-adjacent applications in finance, logistics, and operations, MoME is worth serious evaluation as a replacement for monolithic forecasting models the moment distributional complexity enters the picture.

Quick Hits

• TeleTables benchmark: TeleTables demonstrates that 3GPP table formats are a documented, reproducible blind spot for frontier LLMs — telecom AI teams now have a named failure mode to test against before deployment.
• MoME architecture: Mixture of Modulated Experts delivers measurable forecasting gains on heterogeneous, distribution-shifting time-series data — a credible replacement candidate for monolithic forecasting models in production pipelines.
• CAS-Brain B+ close: The strategic-lead structure signals the investor is buying ecosystem integration and deployment access, not just financial upside — the Chinese AI infrastructure lane is not cooling.
• NLG eval paper: BLEU and BERTScore aggregate rankings can mask systematic instance-level failure — any team auto-evaluating generative outputs should audit their pipeline against the correctness criteria in this paper before the next production cycle.

The Cold Open

Picture this: an author spends three years writing a book. A settlement arrives — some AI company, having trained on that work, has agreed to pay. Justice, maybe. Then the check is divided. The publisher takes a share. The literary agent takes a share. And the author, the person who built the thing that was taken, is left wondering whether the fight was worth it at all.

That fracture — between creators and the intermediaries who represent them — is the real story inside today's AI copyright news. And it is the story that will shape every training-data deal that follows.

The Anchor

The Anthropic Settlement's Unexpected Fault Line: Creators vs. Their Own Advocates

When AI companies began settling copyright lawsuits with authors, the narrative was clean: creators win, AI companies pay, the training-data economy gets a correction. Reality is messier — and considerably more instructive about how AI licensing economics will actually resolve.

The emerging conflict in the wake of Anthropic's settlement is not between authors and Anthropic. It is between authors and their publishers and literary agents — the very intermediaries authors depend on to negotiate on their behalf. Writers say publishers and agents appear to be claiming shares of settlement payments that exceed what their contractual roles reasonably justify. The publishing side presumably argues it holds rights under existing agreements. Authors counter that those agreements were never written to transfer AI training-data licensing rights — and that the copyright at issue is fundamentally theirs, not the publisher's.

This is not a minor accounting dispute. It is a structural question about who owns the economic value of creative work in an AI training-data economy, and the answer set by this conflict will cascade into every settlement, licensing framework, and legislative proposal that follows.

Consider the downstream shape. If publishers successfully claim a substantial share of AI settlement payments, it creates a durable asymmetry: publishers benefit from training-data licensing without having created the underlying work, while authors bear the creative risk and capture a fraction of the return. That asymmetry will reshape what authors are willing to sign in future publishing contracts, how agents structure rights language, and whether future AI copyright actions are pursued as class settlements or as individual direct claims — which are far harder and more expensive for AI labs to manage at scale.

There is a direct Google and Gemini angle here. Google has faced its own parallel pressures on training-data sourcing — from publishers, from news organizations in Europe, and from authors globally. The Anthropic settlement's internal fallout is a real-time stress test of the settlement-as-resolution thesis. If settlements route money to publishers rather than creators, they do not resolve the underlying grievance. Authors remain uncompensated. The political and reputational pressure persists. And future legislative proposals will be written in the shadow of that failure — potentially mandating direct-to-creator pass-through structures that AI labs have less control over.

The practical read: watch for authors' organizations to push for settlement structures that bypass the publisher and agent layer entirely in the next round of AI copyright negotiations. The fracture is now public and documented. The fix, when it comes, will reshape the creator-intermediary relationship in ways that go well beyond AI — and every AI lab with training-data exposure should be modeling this scenario now.

Deep Dive

How Asymmetric Benchmarking Inflates GPU Performance Claims — and Why 385x Is the Number That Should Unsettle You

The headline figure — 385x over a symmetrically-tuned baseline — is not a marketing claim. It is a methodologically rigorous result, which makes it considerably more disturbing than any inflated vendor number.

Here is the mechanism the paper exposes. GPU kernel performance is almost universally measured comparatively: implementation A against implementation B. In standard practice, implementation A is submitted by its author, who has tuned it extensively — profiled on the target hardware, swept kernel launch configurations, selected memory layouts optimized for the access pattern, chosen the batch dimensions where the approach excels. Implementation B — the baseline — is typically retrieved from a public repository and run as found. No profiling. No tuning. No sweep.

This asymmetry is not cheating in the traditional sense. It is a systemic methodological bias that the entire field has absorbed as normal practice. Author teams know their own code intimately. They have also, often unconsciously, selected benchmark suites and input configurations that favor their design choices. The baseline team — if there is one at all — has done none of this preparatory work.

The researcher in this paper ran what they call a symmetric tuning programme: take both implementations, apply equivalent optimization effort to each — equivalent profiling time, equivalent configuration sweeps, equivalent memory layout experimentation. The result was not a modest correction. A comparison that had previously shown an 80x performance advantage for one implementation became, under symmetric tuning, a 385x advantage in the opposite direction. The direction flipped. The magnitude grew nearly five times.

Allow that to settle. Under standard benchmarking methodology, implementation A appeared 80x faster than implementation B. Under symmetric methodology, implementation B is 385x faster than implementation A. The winner and the margin both inverted when the measurement was made fair.

Why does this matter specifically for teams working at Gemini scale? Because every hardware selection decision in AI infrastructure — A100 versus H100, cloud TPU versus on-premise GPU cluster, one inference framework versus another — rests on published benchmark comparisons produced under exactly this asymmetric methodology. Every kernel library adoption decision, every cloud vendor inference pricing analysis, every architecture selection for a Vertex AI deployment has been informed by performance numbers that may bear no relationship to the numbers you would see if both sides were given equal optimization attention.

The practical corrective is not complicated but it is not free either. Before acting on any published performance comparison: identify who produced both the proposed implementation and the baseline. If the same team produced both, or if the baseline is a well-known reference implementation that no one optimized specifically for this comparison, weight the result skeptically. Treat the published number as a lower bound on what the baseline could achieve, not as the baseline's actual ceiling. And when you are running internal evaluations, build symmetric tuning requirements into your evaluation protocol from the start — not as an afterthought after the decision is made.

For AI infrastructure teams and anyone making procurement decisions on the basis of performance benchmarks, this paper is the most important methodological read of the quarter. The field has been measuring itself incorrectly and building enormous decisions on the results. Now there is a rigorous, reproducible demonstration of exactly how wrong those measurements can be.

One Technique

Symmetric Baseline Auditing Before Infrastructure Decisions

Before adopting any AI library, framework, or hardware configuration on the basis of published performance benchmarks, run a one-step audit: identify who produced the baseline in the comparison. If the baseline came from the same team as the proposed implementation, or if it is a well-known reference implementation with no evidence of optimization effort, weight the comparison skeptically. Ask three questions: (1) Was the baseline tuned to a comparable effort level? (2) Who selected the benchmark suite and input sizes, and do those choices favor one side? (3) Does the paper disclose profiling and configuration methodology for both implementations? Apply this lens to GPU kernel comparisons, LLM inference speed claims, and model evaluation leaderboard entries alike. In Gemini and Vertex AI procurement contexts, request explicit disclosure of baseline configuration — model size, batch settings, quantization level, hardware revision — before committing to any published throughput figure. Five minutes of source-checking can prevent months of infrastructure decisions built on asymmetrically inflated numbers.

One Prompt

Use this prompt to critically evaluate any published AI performance benchmark before making a procurement or infrastructure decision:

I need to evaluate a published AI performance benchmark before acting on it. Here is the claim: [paste the benchmark claim, abstract, or result table].

Please analyze:
1. Who produced the baseline — is it the same team as the proposed implementation, or an independently optimized reference?
2. What tuning methodology is disclosed for each side of the comparison? Is there evidence of symmetric effort?
3. What input sizes, batch dimensions, or hardware configurations were selected — and who benefits from those specific choices?
4. What would the comparison plausibly look like under symmetric tuning assumptions, based on the disclosed methodology?
5. What is the realistic performance floor for the baseline if it were given equivalent optimization attention?

Return: a skepticism score from 1 (fully trustworthy) to 10 (highly suspect), the single biggest methodological red flag, and one paragraph I can share with my infrastructure team to frame the decision correctly.

One Tip

In Google AI Studio: set y Before iterating on prompt wording in AI Studio with Gemini, add a system instruction that specifies your expected output schema — JSON field names, length constraints, required keys. Half the time a prompt appears to be failing, the actual problem is output format ambiguity, not the prompt itself. One system instruction written up front saves five rounds of debugging output parsing downstream — and gives you a cleaner signal on what prompt changes are actually doing to model behavior.

Tool of the Day

Google AI Studio — System Instruction Workspace

AI Studio's system instruction panel is genuinely underutilized by teams doing structured output work with Gemini. What it is actually good for: rapid qualitative iteration on Gemini's behavior across different system prompt configurations, with the ability to run the same user prompt under multiple system instruction variants and compare outputs side by side. Free tier covers most exploratory use cases. The honest limit: it is not a real evaluation harness. You cannot run statistically meaningful batch evaluations inside AI Studio without scripting the API directly. Use it for fast qualitative exploration when you need directional signal quickly. Switch to the Vertex AI Evaluation SDK the moment you need quantitative confidence or reproducible metrics. Do not confuse productive tinkering with rigorous benchmarking — that conflation is exactly what today's GPU paper is warning against.

Signature Bites

• The real AI copyright fight: It is between creators and the intermediaries who represent them — not between creators and AI companies. The settlement money is the new battleground.
• 385x, not 80x: That is the correct GPU performance multiplier once asymmetric tuning is controlled for. The direction and magnitude both flip. Trust the methodology, not the headline number.
• Telecom LLM gap is documented: 3GPP table formats are a reproducible, benchmarked failure mode for frontier models. TeleTables is the tool to prove it in an enterprise conversation.
• Probe-based steering crosses modalities: What worked for LLM interpretability is now steering diffusion-based music generation. This paradigm is moving fast across model types.

Joke of the Day

A Gemini model walks into a library. The librarian says: 'We carry everything — novels, research papers, and 3GPP telecommunications specifications.' Gemini says: 'Wonderful. I will take the novels and the research papers.' The 3GPP spec sits on the shelf, confident it will never be correctly interpreted.

The TeleTables benchmark team nods in agreement.

Fact of the Day

3GPP — the standards body that produces the telecommunications specifications at the center of today's TeleTables benchmark — has published an extensive library of technical documents over its history. Each one is dense with the cross-referenced, table-heavy formatting that frontier LLMs demonstrably fail on. TeleTables systematically measures that failure at scale across the full breadth of that corpus.

Stat That Matters

385x. The GPU performance multiplier documented in today's arxiv paper after correcting for asymmetric tuning — compared to the 80x figure the same comparison produced under standard methodology. The gap between those two numbers is not noise. It is the size of the bias the field has been absorbing in every published GPU kernel comparison that did not disclose equivalent tuning methodology for both implementations. Infrastructure decisions made on the uncorrected number may be structurally wrong.

Trends

Agentic AI is the dominant story category today, leading other lanes by volume. The implication is clear: the industry has moved past debating whether agents work and into building the scaffolding around them — benchmarks, evaluation frameworks, domain-specific failure-mode documentation like TeleTables. The evaluation infrastructure is catching up to the deployment reality. The funding lane remains active despite macro headwinds, with strategic-lead deal structures replacing pure financial VC as the dominant architecture in Chinese AI infrastructure rounds. And the copyright and policy lane is generating fewer stories but higher-stakes ones: the Anthropic settlement fracture and the Seattle Times and Newsday suits together suggest the legal resolution phase has entered its messy, contradictory middle act — and that clean outcomes are further off than the early settlement announcements implied.

Bold Prediction

Within 18 months, at least one major AI training-data licensing framework — whether from a legislative body, a publisher consortium, or a court-supervised settlement structure — will include an explicit direct-to-creator pass-through clause designed to prevent the intermediary extraction problem now visible in the Anthropic settlement conflict. The fracture is too public, the narrative too damaging to the 'AI companies pay, creators win' frame, and the political incentive to align with creator interests too strong for this to go unaddressed. Watch for authors' organizations to make this the centerpiece demand of the next major AI copyright negotiation.

Paper Watch

'Pitch-Class Steering for Diffusion-Based Music Generation via Latent-Space Probes'

This paper introduces a new steering paradigm for diffusion-based music generation systems: rather than fine-tuning the underlying model to respond to pitch instructions, the authors train lightweight probes on the model's internal latent representations and use those probes to steer the diffusion process toward specific pitch classes at inference time. The practical result is fine-grained compositional control added as a post-hoc layer to an existing diffusion model — no retraining required.

Why it matters: diffusion-based music generation has lagged autoregressive models on controllability, which has limited their uptake among creative AI builders who need precise musical control. Probe-based steering closes that gap without the cost of base-model retraining. If the technique generalizes across other musical attributes — rhythm, timbre, dynamics — it opens a new class of compositional interfaces for diffusion audio systems. The paradigm is borrowed directly from LLM interpretability research and is now crossing modalities into structured audio. Expect to see probe-based steering appear in Google Lyria and similar diffusion audio backends within the next two model generations. The approach is clean, modular, and low-cost enough to be adopted quickly wherever diffusion audio is already deployed.

Founder Spotlight

CAS-Brain — Strategic B+ Close with an Industry Lead Investor

The move worth watching is the choice of a strategic industry lead over a pure financial VC for CAS-Brain's B+ round. At growth stage, a strategic lead investor in AI infrastructure typically buys three things simultaneously: equity upside, preferred deployment partnership rights, and access to the portfolio company's technical roadmap as a co-development context. The investor's existing customer base becomes a distribution pipeline for the AI infrastructure product. In exchange, the portfolio company gains a deployment anchor and an enterprise introduction channel that pure financial VC cannot provide.

CAS-Brain's decision to optimize for a strategic lead at this stage signals that they are thinking about deployment footprint and ecosystem reach, not just capital efficiency or valuation. That is a maturing strategy for a Chinese AI infrastructure builder operating in an environment where enterprise trust and integration depth matter more than headline model capability. Strategic read: watch for a deployment or co-development announcement tied to the lead investor's industrial vertical — most likely within 12 months of the close. That announcement will clarify the deal's strategic logic and signal whether CAS-Brain is building toward a platform play or a vertical-specific infrastructure position.

Quote

'Authors say publishers seem to be claiming more than their fair share of settlement payments.'

— TechCrunch, reporting on the emerging internal conflict in the Anthropic copyright settlement. One sentence that captures the unexpected fault line in what was supposed to be a resolution.

Learner's Edge

Concept: Latent-Space Probes

A latent-space probe is a lightweight classifier or regressor trained not on a model's inputs or outputs, but on its internal activations — the intermediate representations the model builds as it processes data. Large models encode far more structured information in those internal representations than they ever surface in their outputs. A probe reveals what the model 'knows' internally, even when it does not express it.

The technique originated in language model interpretability research — a way to ask: does this model represent the concept of 'truthfulness' or 'city names' in its internal layers? But as today's music paper demonstrates, probes can do more than read the latent space. They can steer it. A steering probe applies targeted activations at inference time to push the model's generation toward a desired attribute, without retraining. This paradigm is now moving from language models into diffusion models for audio and image generation. Understanding probes is foundational for anyone working on model interpretability, fine-grained control, or AI alignment — it is one of the core tools in the mechanistic interpretability toolkit, and it is becoming more practically relevant every quarter.

Sign-off

That is The Agent Signal for September 7th. Tomorrow we are watching how the Anthropic settlement conflict develops — specifically whether authors' organizations respond with demands for direct-to-creator payment structures that cut out the publisher layer entirely. The answer will tell us a great deal about how AI copyright economics actually resolve at scale. Stay sharp.

Sources

1. Authors push back as publishers and agents seek share of Anthropic settlement — techcrunch.com
2. Seattle Times and Newsday sue OpenAI and Microsoft for infringement — theverge.com
3. TeleTables: A Benchmark for Large Language Models in Telecom Table Interpretation — arxiv.org
4. From 80x to 385x: A Best-Matching-Unit Search at the L2 Roof, Measured Against a Symmetrically Tuned Baseline — arxiv.org
5. Pitch-class Steering for Diffusion-based Music Generation via Latent-space Probes — arxiv.org
6. Beyond Aggregate Scores: Behavioral Correctness Assumptions for Assessing Reference-Based Automatic Evaluation Methods — arxiv.org
7. Industry leader leads investment! CAS-Brain completes hundreds of millions of yuan in B+ round strategic financing — qbitai.com
8. Multi-Modal Time Series Prediction via Mixture of Modulated Experts — arxiv.org