The Pulse — September 11, 2026

Models

DeepSeek V4.1 Flash

Sparse MoE model (8B active on input, 16B on output) using a new Causal Encoder-Decoder architecture.

First encoder-decoder in the V4 line; architectural shift worth tracking when eval results land.

InclusionAI Ling 3.0 Flash VL

124B total / 5.5B active MoE adding native visual perception to the Ling 3.0 Flash base.

Free tier multimodal option on OpenRouter; worth testing for budget-conscious vision tasks.

Research

Can LLMs Normalize Databases? A Benchmark and Multi-Agent Framework for Schema Normalization

Benchmark measuring LLM ability to reason about functional dependencies and lossless join decompositions for database normalization.

Tests structured-output reliability when outputs must satisfy database constraints; relevant for schema generation work.

REVA: Reusable Evidence View Aggregation for Context-Efficient RAG Serving

Post-retrieval compression that reuses compressed views across queries to reduce latency and KV cache cost.

Addresses the cost side of RAG serving; relevant for EverRaise and Mackinac where retrieval volume is high.

RAG-Safety-Bench: Reliable Evaluation of Retrieval-Augmented LLM Safety

Benchmark showing RAG can degrade safety when prompted for harmful content if retrieved documents introduce risk.

Safety evaluation gap for RAG systems; worth knowing before deploying user-query-driven retrieval.

Decoupling Readiness from Release for Tail-Aware Scheduling of Agentic LLM Workflows

Agent-turn scheduling policy that delays release of ready turns to reduce tail latency under contention.

Agentic workflow optimization; applies when we run multi-turn agent jobs under load.

Releases

LangChain Anthropic 1.7.2

Preserves invalid tool-use blocks instead of dropping them.

Relevant for EverRaise and SalesKick if we're using Anthropic models through LangChain.

Hacker News

Shopify is moving from React Native back to Swift and Kotlin

After years of React Native, Shopify is rewriting mobile apps in native code citing performance and platform integration.

Signals trade-offs in cross-platform mobile stacks; relevant if client projects consider React Native.

Hacker News discussion

Cognition launches SWE-2 model, rivaling Fable 5.1 and GPT-Astra

Cognition's new coding model claims performance on par with Fable 5.1 and GPT-Astra on SWE-bench and coding evals.

Another strong coding model entering production; worth benchmarking against Claude Code and Codex for PR review and feature work.

Hacker News discussion

OpenAI Agents API

Managed agent service with Codex harness, orchestration, long-running sessions, and context management.

Competes with Claude Managed Agents; relevant for SalesKick and Pipeline's agent workloads if we want fully managed execution.

Hacker News discussion

Neki – Sharded Postgres

PlanetScale's new sharded Postgres offering for horizontal scaling without application rewrites.

Scaling option for Postgres-backed projects when single-node limits are hit; worth watching for Pipeline clients.

Hacker News discussion

People on X

OpenAI: GPT-Live-1 in the API

Full-duplex voice API that listens while speaking, with custom voices and telephony support.

Production voice-agent primitive; relevant for EverRaise's Vapi integration and any future voice work.

OpenAI: Agents API public beta

Cloud agent orchestration with the Codex harness, handling sessions and context management.

Competes with Claude Managed Agents; worth evaluating for long-running agent workflows.

Nous Research: Hermes Agent subagent UI

Hermes now shows live subagent activity with manual steering and stop controls in CLI and desktop.

We run Hermes Agent for Botterfly; the subagent visibility helps debug multi-agent workflows.

ClaudeDevs: Desktop app pane management

Claude Code desktop app lets you pop out diffs and terminals into separate windows.

Quality-of-life for Claude Code workflows; multi-screen sessions become easier to manage.