AI companies are burning books, advocates complain to FTC
"What the public record does not establish – and cannot, from the outside – is how often the destroyed physical books are the last or among the last surviving copies of a given work."
I'd argue that it's not "how often" - the frequency would be within floating point error. But even one destruction of a last (even one of several) copy of a now dysfunctional publisher is a preventable cultural catastrophe.
https://www.theregister.com/ai-and-ml/2026/08/21/ai-companies-are-burning-books-advocates-complain-to-ftc/5291299
@llm
I tested a narrow context-management question: after an incorrect claim has propagated through several conversation turns, is deleting the original error enough?
The pilot used synthetic, objectively scored tasks represented as context DAGs. Across four model endpoints, pollution derailed 72 cases that had been answered correctly with clean context.
Repair recovered:
The interesting failures occurred when stale downstream turns still repeated or depended on the deleted claim. The source was gone, but its consequences were still present in the next request.
I do not interpret this as a general model ranking. Each endpoint was sampled once at temperature zero, the tasks are synthetic, and provider-default reasoning behavior was not normalized.
What I would most like to test next is whether the same pattern holds for local models and different quantization levels. Would you expect smaller or more heavily quantized models to be more sensitive to residual context?
Full report: https://chenxiachan.github.io/thoughtdag/research/context-repair-pilot-v1/
Disclosure: I designed this benchmark and maintain ThoughtDAG, the open-source reference interface used to visualize and edit the context graphs.
https://chenxiachan.github.io/thoughtdag/research/context-repair-pilot-v1/Open linkView original on lemmy.worldCascadia distributes LLM inference across Intel laptops, desktops, and AI PCs. Shard a model across the machines you already have and serve it through an OpenAI-compatible API. No cloud or NVIDIA GPUs required.
Frontier models don't fit on a single laptop. Cloud APIs are expensive, opaque, and require sending your data offsite. Cascadia lets you point a few Intel machines at each other and run models that none of them could handle alone.
https://github.com/labscommunity/cascadiaOpen linkView original on lemmy.ml| Pricing component | DeepSeek V4 Pro 0813 | Opus 4.8 | Price difference |
|---|---|---|---|
| Input | $0.435/M | $5/M | 91.3% cheaper |
| Output | $0.87/M | $25/M | 96.5% cheaper |
| Cached input | $0.003625/M | $0.50/M | 99.3% cheaper |
| Weighted avg input | $0.1118/M | $1.536/M | 92.7% cheaper |
https://openrouter.ai/deepseek/deepseek-v4-pro-0813Open linkView original on lemmy.worldThe Genesis Open Models Initiative is a U.S. Department of Energy effort to build and release open-weight AI models designed for scientific research. It is hosted through Argonne National Laboratory and is part of the broader DOE Genesis Mission.
Its main focus is accelerating work in areas such as:
Energy systems, including grid reliability and nuclear-energy deployment
Climate and environmental science
Materials science and chemistry
Scientific computing, including modernizing high-performance-computing code
Experiments, simulations, and data analysis
The initiative seeks contributions from universities, labs, companies, and nonprofits—including scientific text, code, datasets, workflow environments, expert task examples, and evaluation methods—to train and assess these models.
application-of-model-merging-in-different-machine-learning-subfields
application-of-model-merging-in-foundation-models
combine-the-capabilities-of-expert-llms
combine-the-capabilities-of-expert-mllms
enhancing-the-faithfulness-or-generation-quality-of-diffusion-models
human-preference-alignment-for-llms
improving-computational-efficiency-of-moe-based-llm
knowledge-editingunlearning-of-llms
mixing-datasets-via-model-merging
model-merging-as-a-defense-or-intellectual-property-protection
model-merging-for-better-domain-generalization-or-domain-adaptation
model-merging-for-better-out-of-distribution-generalization
model-merging-for-cross-modal-knowledge-transfer
model-merging-for-cross-task-generalization-in-few-shot-learning
model-merging-for-cross-task-generalization-in-zero-shot-learning
model-merging-for-knowledge-transfer-in-auxiliary-learning
model-merging-for-knowledge-transfer-in-multi-domain-learning
model-merging-for-knowledge-transfer-in-multi-objective-optimization
model-merging-for-knowledge-transfer-in-multi-task-learning
model-merging-for-local-knowledge-aggregation
model-merging-for-multimodal-fusion
model-merging-in-adversarial-learning
model-merging-in-continual-learning
model-merging-in-embodied-models
model-merging-in-federated-learning
model-merging-in-image-generative-models
model-merging-in-large-language-models
model-merging-in-multi-taskmulti-objectivemulti-domainauxiliary-learning
model-merging-in-multimodal-large-language-models
model-merging-in-out-of-distributiondomain-generalization
model-merging-in-video-generative-models
model-merging-in-zero-shotfew-shot-learning
model-merging-to-mitigate-catastrophic-forgetting
post-calibration-based-methods
reducing-training-cost-of-generative-models
routing-based-merging-methods-dynamic-merging
style-mixing-in-generative-models
subspace-based-merging-method-sparse-or-low-rank-subspace
SpaceXAI launched Grok 4.5 on Wednesday at $2/$6 per million tokens, with Elon Musk calling it "Opus-class" but cheaper and faster than Anthropic's models.
OpenAI made GPT-5.6 generally available Thursday in three tiers, with its cheapest Luna variant matching Grok 4.5's output price at $6 per million tokens.
Meta debuted Muse Spark 1.1, its first paid AI model, at $1.25/$4.25 per million tokens — the lowest price among the three flagship releases.
ggml.ai joins HuggingFace
ggml.ai is better known as the entity behind llama.cpp. It's nice to hear good news! Thank you @ggerganov and @huggingface
Orange Pi closes the year by unveiling new details about the Orange Pi AI Station, a compact board-level edge computing platform built around the Ascend 310 series processor. The system targets high-density inference workloads with large memory options, NVMe storage support, and extensive I/O in a small footprint.
The AI Station is powered by an Ascend 310 series processor integrating 16 CPU cores clocked at up to 1.9 GHz, along with 10 AI cores running at up to 1.08 GHz and 8 vector cores operating at up to 1 GHz.
https://linuxgizmos.com/orange-pi-unveils-ai-station-with-ascend-310-and-176-tops-compute/Open linkView original on lemmy.ca
