JOUNES // FORECASTS
Home  ·  Projects  ·  Essays  ·  GitHub  ·  LinkedIn  ·  Email
// RESEARCH-RATE FORECAST

Research-rate forecast

TimesFM 2.0 · zero-shot · 14-day horizon

What this is

A continuously-running pipeline (ReefResearch) reads AI-infrastructure research — arXiv abstracts, GitHub repos, HN threads, lobsters, OpenAlex — and extracts concrete engineering suggestions from each source. Those suggestions are clustered nightly by topic using BERTopic over their embeddings, producing themed clusters. Each cluster's daily emit rate is then fed through TimesFM 2.0 — Google's zero-shot time-series foundation model — to produce a 14-day forward forecast per cluster.

The numbers below are bullets per day: one bullet = one extractable engineering suggestion from one source. Rising clusters signal topics the AI-infrastructure community is converging on; declining ones are losing attention. Forecast made 2026-05-25, refreshed nightly as new research lands.

Caveats: zero-shot means no fine-tuning on this corpus — accuracy improves as history grows (meaningful calibration after ~30 days). This forecasts the pipeline's output rate, not the field's publication rate. Back-grading against actuals runs weekly; calibration is the artifact, not the headline numbers.

Last 7 days: 25.43 bullets/day  ·  Next 14 days: 24.08/day  ·  -5.3%
  1. trust sanitization untrusted prompt0.22/day+220.0%
    What’s being proposed
    • Prompt Flow Integrity (PFI) Frameworks Implementations of PFI that provide agent isolation and secure untrusted data processing to create "hard" guardrails.
    • Request-Flow Integrity Layer Building a verification shim to ensure that the destination capability remains invariant across the parsing chain.
    • Relational Governance Middleware Build a Semantic-Spatial-Relational (SSR) stack that embeds trust calibration and epistemic boundaries as structural mechanisms rather than post-hoc filters.
  2. global modality skip token0.2/day+195.9%
    What’s being proposed
    • Global Reasoning Units Integrate GloRe units to enable efficient coordinate-interaction space mapping via weighted global pooling and broadcasting, reducing the need for deep convolutional stacking.
    • Hybrid Perception Pipeline Integrate Set-of-Mark (SoM) prompting with the SeeClick paradigm of click-based attention mapping. This removes reliance on structured metadata (XML/HTML) while providing the LMM with "speakable" visual coordinates.
    • Dynamic Foveation Engine Implement a multi-stage pipeline to solve the resolution-token paradox. This requires a low-resolution global pass to identify Regions of Interest (RoIs), followed by high-resolution crops of those specific…
  3. routing models slms cloud0.74/day+160.3%
    What’s being proposed
    • Compute-Optimal Allocators Develop orchestration layers that adaptively allocate test-time compute based on prompt difficulty to maximize resource efficiency.
    • Consistent Hashing Router for Agent States Implement a gateway using consistent hashing—similar to the architecture in distributed-kv-cache—to map specific agent reasoning states to worker nodes. This ensures locality and can potentially increase inference…
    • Hybrid Disaggregation Orchestrator Build a scheduler that implements "partially disaggregated prefill", partitioning the initial prefill stage on lower-power cores and overlapping the remaining prefill and decode stages on high-performance nodes to…
  4. tokens bearer token proxy0.14/day+141.5%
    What’s being proposed
    • Zero-Trust MCP Proxies Replace bearer tokens—susceptible to log leakage and replay attacks—with capability-based tokens or RSA-based workload identities.
    • Capability-Sealed Mediation Replace the practice of passing bearer tokens via environment variables with a trusted broker like CapSeal.
    • Request-Scoped Capability Binding Build a mechanism to bind tuple-level access to ephemeral, cryptographically verifiable tokens. This removes reliance on SET ROLEand SET SESSION AUTHORIZATION, which are susceptible to "Confused Deputy" attacks.
  5. preference moe experts reward0.14/day+140.3%
    What’s being proposed
    • Dynamic Granularity Controller Develop a student-in-the-loop framework based on Gen-SSD that prunes teacher reasoning paths in real-time. This system must dynamically adjust the reasoning granularity of Chain-of-Thought (CoT) traces to match the…
    • Specialized Reasoning MoE Build a Mixture-of-Experts (MoE) student architecture that separates routing from specialized reasoning experts (e.g., math, tool planning, synthesis). This allows for hybrid-precision training where symbolic logic…
    • Precise state Injector Develop a protocol for the real-time injection of Knowledge Packsinto active reasoning trajectories. By leveraging the fact that contrastive deltas on cached values can nudge behavior without destroying coherence…
  1. syntax codebase representation0.08/day-85.3%
    What’s being proposed
    • Pliron-based IR Pipeline Implement the aforementioned pipeline. Using Pliron as an mlir-like framework allows for the preservation of structural metadata before lowering to LLVM IR.
    • A Strongly-Typed Codebase IR Develop an intermediate representation similar to SkIR that decouples codebase semantics from language-specific syntax.
    • Hub-and-Spoke Dependency IR Replace bilateral file-to-file adapters with a hub-and-spoke Intermediate Representation (IR). This IR should capture a "structural sketch" of the codebase—similar to the row/column summaries used for massive…
  2. subword gcd decoding grammars0.09/day-85.0%
    What’s being proposed
    • Subword-Aligned Constraint Engines Implement decoding algorithms like DOMINOor XGrammarto eliminate the performance overhead and accuracy degradation caused by misaligned sub-word vocabularies. By partitioning vocabulary into context-independent and…
    • Structure-Aware Transformers Construct an architecture that replaces standard subword tokenization with specific AST node granularity, mapping latent trajectories directly to syntactic structures to improve prediction accuracy over baseline NTP.
    • Optimized GCD Engines Build a Grammar-Constrained Decoding (GCD) engine that utilizes fast offline preprocessing to reduce the time required to align subword tokenizers with context-free grammars (CFGs).
  3. measurement race attestation performance0.07/day-83.8%
    What’s being proposed
    • Attestation Supervisor Develop a privileged monitor leveraging ptrace() to inspect and verify the allocation tags of processes in separate address spaces, providing a hardware-rooted audit trail of memory access.
    • Command Footprint Analyzer Develop a tool to compare the byte-count of a standard API dispatch against a pre-recorded execution pattern (similar to CUDA Graphs). This will reveal exactly how many bytes of "boilerplate" the driver inserts for…
    • Hybrid hardware Monitors Integrate LO-FAT's non-stalling control-flow attestationwith HardScope's run-time scope enforcementor LiteHAX's data-flow tracking. This would create a unified RISC-V primitive capable of detecting both ROP/JOP and…
  4. latent arc bfs agi0.04/day-74.5%
    What’s being proposed
    • Latent Recursion Scaffolds Replace linear Chain-of-Thought with hidden-space iterative refinement. Using Mamba-2 hybrid operators within a recursive scaffold allows models to achieve high performance on abstract reasoning tasks (e.g.…
    • BFS-Capable Latent Architectures Implement reasoning loops based on Coconut where the last hidden state is fed back as the next input embedding. Unlike the deterministic paths of standard Chain-of-Thought, this allows the model to encode multiple…
    • Contrastive Latent RL Frameworks Build systems like DeepLatent Reasoning (DLR) that shift reinforcement learning rollouts from discrete token spaces to the continuous latent manifold. Using a lightweight assistant to sample reasoning encodings and a…
  5. dependency graph ifg dissociation0.09/day-68.3%
    What’s being proposed
    • Information Flow Graph (IFG) Monitors Monitors that track agent-tool interactions via an IFG to enable real-time blocking of actions that exceed the least privilege required for a specific task.
    • State-Dependency Mapping Implement a graph-based tracking system to identify which subsequent steps depend on the output of a modified step. This prevents context amnesia or execution crashes when a tool call that created a necessary file is…
    • Information Dependency Graph (IDG) Debuggers Implement frameworks like GraphTracer that move beyond temporal sequencing to map how agents explicitly reference and build upon prior outputs. This allows for the localization of the exact "turn" in a trajectory…
Provenance: Published 2026-05-25
// GENERATED FROM A LIVE OBSIDIAN VAULT · CLOUDFLARE PAGES · DRAFTED WITH AGENTS
← back to Forecasts