{
    "week": "2026-W47",
    "frozen_at": "2026-11-17T16:01:35Z",
    "data_through": "2026-11-15",
    "method": "v1",
    "resolution": {
        "window_weeks": 52,
        "wikipedia_daily_views": 500,
        "gdelt_articles_30d": 150
    },
    "predictions": [
        {
            "rank": 1,
            "concept": "weight-space programming",
            "slug": "weight-space-programming",
            "id": 19688,
            "aliases": [
                "WSP",
                "weight-space programming"
            ],
            "score": 14.9911,
            "signals": {
                "weekly": [
                    0,
                    0,
                    0,
                    0,
                    0,
                    0,
                    13,
                    7
                ],
                "mentions28": 20,
                "groups28": 19,
                "age_days": 13,
                "acceleration": 3.045,
                "propagation": 0.5
            },
            "summary": "Weight-space programming (WSP) has a frontier LLM read and write the raw parameters of another neural network directly, with no gradient descent, to diagnose bugs, repair behaviour, or build small networks from a task description. It traces to an MIT blog post and paper by Takehi et al., in which a GPT-6 Astra agent fixed reasoning loops in Qwen3.5-0.8B. Within two weeks, independent groups extended it to image-restoration diffusion models, clinical network auditing, quantum circuit angles, RL policies checked with model checking, tabular foundation models, and medical classifiers. Others produced a benchmark (WeightBench), a backdoor security study, and a discrete-optimization alternative. Two small tooling repos have also appeared."
        },
        {
            "rank": 2,
            "concept": "zero-knowledge proofs of machine learning inference",
            "slug": "zero-knowledge-proofs-of-machine-learning-inference",
            "id": 15995,
            "aliases": [
                "zkML",
                "verifiable inference",
                "zero-knowledge machine learning",
                "zero-knowledge proof of inference",
                "zero-knowledge proofs of machine learning inference"
            ],
            "score": 6.5495,
            "signals": {
                "weekly": [
                    0,
                    0,
                    0,
                    0,
                    0,
                    1,
                    1,
                    4
                ],
                "mentions28": 6,
                "groups28": 6,
                "age_days": 16,
                "acceleration": 0.405,
                "propagation": 0
            },
            "summary": "Zero-knowledge proofs of ML inference (zkML) let a prover show that a specific committed model produced a given output, without revealing the weights or inputs. The newer direction in these sources is scaling this to transformer LLMs with bit-exact semantics. One paper presents Sable, a small-field sumcheck argument with a single-GPU prover for integer-quantized transformers. Another builds a fully integer-arithmetic runtime so that inference is reproducible and therefore checkable. A third argues that generic inference proofs miss request and checkpoint binding for reward-model services. Others include a hardware-cost breakdown of proving, an unmaintained Lean research prototype under the anthropics account, and a Mac chat app advertising verifiable inference."
        },
        {
            "rank": 3,
            "concept": "NVFP4 quantization",
            "slug": "nvfp4-quantization",
            "id": 1286,
            "aliases": [
                "NVFP4",
                "NVIDIA FP4",
                "NVFP4 quantization",
                "4-bit floating point",
                "FP4 microscaling formats",
                "NVIDIA 4-bit floating point"
            ],
            "score": 6.4321,
            "signals": {
                "weekly": [
                    0,
                    0,
                    5,
                    2,
                    0,
                    1,
                    4,
                    2
                ],
                "mentions28": 7,
                "groups28": 7,
                "age_days": 41,
                "acceleration": 2.639,
                "propagation": 1
            },
            "summary": "NVFP4 is NVIDIA's 4-bit floating-point format (E2M1 values with a shared scale per 16-element block) that runs natively on Blackwell tensor cores. The sources show work moving past inference-only quantization toward training with it. Several groups apply it to RL post-training and report learner-sampler mismatch and the fixes they built for it. One comparative study, co-authored by Dan Alistarh, evaluates NVFP4 pretraining recipes at 8B and 30B MoE scale. Other papers cover smoothing and scale optimization, FP4 KV caches, LoRA fine-tuning on consumer GPUs, and per-tile format selection. Hobbyist repos ship vLLM and SGLang serving recipes for DGX Spark and RTX cards."
        },
        {
            "rank": 4,
            "concept": "recursive self-improvement",
            "slug": "recursive-self-improvement",
            "id": 11,
            "aliases": [
                "RSI",
                "LLM-driven code evolution",
                "recursive self-improvement",
                "self-evolving training pipeline",
                "self-improvement via synthetic data"
            ],
            "score": 5.6277,
            "signals": {
                "weekly": [
                    0,
                    2,
                    24,
                    25,
                    16,
                    19,
                    24,
                    8
                ],
                "mentions28": 67,
                "groups28": 66,
                "age_days": 42,
                "acceleration": 0.241,
                "propagation": 1
            },
            "summary": "Recursive self-improvement, long a theoretical idea about AI systems improving themselves, is showing up here as a concrete empirical method: agents that rewrite their own harnesses, workflows or training data across rounds using execution feedback. A formal-verification paper proposes diagnosing root causes instead of patching symptoms. One paper defines an 'RSI tax' to count the token cost of self-improvement. Another adds memory so evolved harnesses carry over to new benchmarks or models. Applications include visuomotor policies, image restoration, video generation and automated RL. Tooling is appearing too: a gym environment, an evaluation SDK, and a research atlas of evolved agent harnesses."
        },
        {
            "rank": 5,
            "concept": "diffusion language model",
            "slug": "diffusion-language-model",
            "id": 169,
            "aliases": [
                "DLM",
                "DLLM",
                "MDLM",
                "diffusion LLM",
                "masked diffusion",
                "discrete diffusion",
                "masked diffusion LM",
                "masked diffusion LLM"
            ],
            "score": 4.5239,
            "signals": {
                "weekly": [
                    0,
                    3,
                    24,
                    38,
                    34,
                    19,
                    18,
                    9
                ],
                "mentions28": 80,
                "groups28": 75,
                "age_days": 42,
                "acceleration": -0.503,
                "propagation": 1
            },
            "summary": "Diffusion language models generate text by starting from a masked or noisy token sequence and refining it over several denoising steps, rather than producing one token at a time from left to right. Most of the sources are arXiv papers that build on open models such as LLaDA and Dream. They address faster decoding through speculative branches and attention pruning, better-calibrated choice of which tokens to commit, remasking and self-correction, and post-training with ELBO surrogates. Practical activity is appearing too: a Mojo repo for local CPU inference, a 35B mixture-of-experts diffusion model aimed at agentic coding, and a Hacker News thread on Inception Labs' Mercury 3."
        },
        {
            "rank": 6,
            "concept": "Agent Client Protocol",
            "slug": "agent-client-protocol",
            "id": 987,
            "aliases": [
                "ACP",
                "Agent Client Protocol"
            ],
            "score": 4.4994,
            "signals": {
                "weekly": [
                    0,
                    0,
                    3,
                    1,
                    1,
                    1,
                    3,
                    6
                ],
                "mentions28": 11,
                "groups28": 11,
                "age_days": 40,
                "acceleration": 1.715,
                "propagation": 0.5
            },
            "summary": "Agent Client Protocol (ACP) is a JSON-RPC protocol, typically run over stdio, that lets any editor or client host any compliant AI coding agent, much as LSP decoupled editors from language servers. The sources show a broad, independent spread of clients. These include lightweight GPUI and Electron editors, a WhatsApp-style desktop chat app for agents, a browser desktop, a self-hosted remote-control server, a reverse-engineering tool, and an Xcode bridge for Claude Code. Several repos wrap Google's Antigravity CLI over ACP to expose OpenAI and Anthropic-compatible APIs. One experiment proposes ACP over HTTP, and a blog post runs a Docker agent in a sandbox under Zed. One tag-stuffed repo is noise."
        },
        {
            "rank": 7,
            "concept": "specialized decision model",
            "slug": "specialized-decision-model",
            "id": 3789,
            "aliases": [
                "decision model",
                "specialized decision model"
            ],
            "score": 4.2459,
            "signals": {
                "weekly": [
                    0,
                    0,
                    2,
                    1,
                    2,
                    0,
                    5,
                    1
                ],
                "mentions28": 8,
                "groups28": 8,
                "age_days": 38,
                "acceleration": 1.135,
                "propagation": 1
            },
            "summary": "Decision models are LLM-based systems that, given a prompt and a fixed set of options, return a probability for each option in one forward pass instead of generating text autoregressively. The draw is lower cost and latency, no output parsing, and probabilities that can be calibrated. Much of the work centers on one model, Jev, described as a \"System One\" model, but independent groups are now testing the approach on entity matching (University of Mannheim), Earth observation (MBZUAI), economic surrogates, clinical extraction and digital nudging. New models such as Zuse, for European languages, and Drex are also appearing. A Hugging Face-based fine-tuning library and a make-or-buy economics blog post suggest early practitioner interest."
        },
        {
            "rank": 8,
            "concept": "EEG visual decoding",
            "slug": "eeg-visual-decoding",
            "id": 6614,
            "aliases": [
                "EEG decoding",
                "EEG image decoding",
                "EEG visual decoding",
                "visual brain decoding",
                "brain-to-image generation",
                "EEG signal classification",
                "EEG-to-image reconstruction",
                "electroencephalography decoding"
            ],
            "score": 4.0897,
            "signals": {
                "weekly": [
                    0,
                    0,
                    2,
                    0,
                    1,
                    0,
                    4,
                    2
                ],
                "mentions28": 7,
                "groups28": 7,
                "age_days": 39,
                "acceleration": 1.658,
                "propagation": 0.5
            },
            "summary": "EEG visual decoding aims to classify or reconstruct what a person sees, or recalls, from non-invasive scalp EEG, typically by aligning EEG embeddings with pretrained vision-language features and conditioning a diffusion or flow-matching generator. The recent papers target specific weak points. A prototype-alignment method avoids dense image targets. Another work does zero-shot cross-subject reconstruction via subspace calibration. Others combine flow matching with wavelet encoders or directly supervise the generator rather than only the alignment. New benchmarks cover video recall and controlled face latents. Contributors include groups at USC, Shanghai Jiao Tong, Helsinki and Korea University. The one code repository is a generic EEG classification library."
        },
        {
            "rank": 9,
            "concept": "pixel-space diffusion",
            "slug": "pixel-space-diffusion",
            "id": 1571,
            "aliases": [
                "pixel-space diffusion",
                "pixel-space flow models",
                "pixel-space diffusion model",
                "pixel-space video diffusion"
            ],
            "score": 4.0804,
            "signals": {
                "weekly": [
                    0,
                    0,
                    4,
                    2,
                    1,
                    3,
                    5,
                    3
                ],
                "mentions28": 12,
                "groups28": 12,
                "age_days": 41,
                "acceleration": 0.924,
                "propagation": 0.5
            },
            "summary": "Pixel-space diffusion trains image and video generators directly on raw pixels instead of inside a pretrained VAE latent space. That avoids lossy compression and two-stage training. What is new in these sources is that the approach is being made to work at scale. Iris-3B is a 3B open text-to-image model with weights and an Apple Silicon port. A two-person lab released a 2B Pyramid-JiT model trained for about $21K and posted it on Show HN. Several papers study why v-prediction fails, add adaptive and multi-scale tokenization, or convert pretrained latent video models (PixVid, JVT, U-JiT) into pixel space."
        },
        {
            "rank": 10,
            "concept": "on-policy distillation",
            "slug": "on-policy-distillation",
            "id": 358,
            "aliases": [
                "OPD",
                "on-policy distillation",
                "on-policy knowledge distillation",
                "Open Publication Distribution System",
                "multi-objective on-policy distillation"
            ],
            "score": 3.8006,
            "signals": {
                "weekly": [
                    0,
                    3,
                    39,
                    45,
                    27,
                    20,
                    18,
                    14
                ],
                "mentions28": 79,
                "groups28": 78,
                "age_days": 42,
                "acceleration": 0.197,
                "propagation": 0.5
            },
            "summary": "On-policy distillation trains a student model on trajectories it generates itself, with a teacher scoring or supplying token-level distributions over those student-produced sequences, rather than imitating fixed teacher outputs. The sources show it being treated as a standard post-training step for reasoning LLMs and then extended in many directions: sparse teacher interfaces for long traces (ResOPD), multi-teacher merging (Janus-MOPD), block diffusion and few-step image diffusion students, speculative decoding drafts (a Qwen repo), agent memory and ReAct behavior, multimodal perception, and backdoor transfer from teacher to student. Activity is almost entirely arXiv papers from many independent academic and industry groups. A few hits, such as an OPDS e-book server, are acronym collisions."
        }
    ],
    "baseline": [
        {
            "rank": 1,
            "concept": "on-policy distillation",
            "slug": "on-policy-distillation",
            "id": 358
        },
        {
            "rank": 2,
            "concept": "graph neural network",
            "slug": "graph-neural-network",
            "id": 423
        },
        {
            "rank": 3,
            "concept": "tabular foundation model",
            "slug": "tabular-foundation-model",
            "id": 1083
        },
        {
            "rank": 4,
            "concept": "linear probing",
            "slug": "linear-probing",
            "id": 1741
        },
        {
            "rank": 5,
            "concept": "contrastive language-image pretraining",
            "slug": "contrastive-language-image-pretraining",
            "id": 503
        },
        {
            "rank": 6,
            "concept": "catastrophic forgetting",
            "slug": "catastrophic-forgetting",
            "id": 278
        },
        {
            "rank": 7,
            "concept": "world action model",
            "slug": "world-action-model",
            "id": 190
        },
        {
            "rank": 8,
            "concept": "on-policy self-distillation",
            "slug": "on-policy-self-distillation",
            "id": 344
        },
        {
            "rank": 9,
            "concept": "parameter-efficient fine-tuning",
            "slug": "parameter-efficient-fine-tuning",
            "id": 398
        },
        {
            "rank": 10,
            "concept": "diffusion language model",
            "slug": "diffusion-language-model",
            "id": 169
        }
    ],
    "prev_hash": "0000000000000000000000000000000000000000000000000000000000000000"
}
