The Pulse, Tuesday 9 September 2026
OpenAI claims a solution to the Navier-Stokes Millennium Prize problem with an AI-generated proof in Lean, drawing immediate scrutiny from mathematicians. Mistral raised €3B, the largest equity round in European tech history. Meta launched Muse, a personal AI agent with long-term memory; Addy Osmani joined Anthropic to work on Claude Code.
Models
Inception: Mercury 2.5
Diffusion LLM that generates and refines multiple tokens in parallel rather than sequentially, 4¢/M prompt tokens.
Parallel token generation could reduce latency for longer agent responses if the approach generalizes beyond the vendor's benchmarks.
Nex AGI: Nex-N2.5-Pro
Free agentic coding model with visual feedback loop, multimodal input, 262k context.
Another free agentic coder joins the Claude Code / Codex competitive space; worth testing against our coding workflows.
Research
When Does Memory Help? A Cost-Aware Evaluation of Long-Term Memory in Tool-Using LLM Agents
MERIT benchmark measures whether remembered facts change what an agent does, not just conversational recall accuracy.
Directly relevant to our agent memory work in Hermes and Botterfly—evaluates if memory improves decisions, not just recall.
MemForest: Efficient Agent Memory Management via EventTree Partitioning and Progressive Merging
Memory compression framework to reduce storage and retrieval costs during long-term agent inference.
Compression that preserves retrieval quality matters for production agents with accumulated context; applicable to our memory layer.
PARSER: Read in Parallel, Reason in Depth for Long-Context LLM Agents
Decouples document reading from reasoning, processing chunks in parallel to reduce latency and evidence-placement sensitivity.
Parallel reading would help our retrieval-heavy flows in Pipeline's clients and SalesKick's call grading where documents are long.
Q2D-Web: A Large-Scale Benchmark for Retrieval in Agentic RAG Systems
Benchmark pairing large-scale corpus with agent-reformulated search queries based on real conversations.
Tests retrieval in the way we actually use it—agents reformulating queries across conversation turns in production RAG.
Releases
Node.js v24.21.0 'Krypton' (LTS)
LTS release with NSS 3.126 root certificates, OpenSSL 3.5.8, Undici 7.29.1, and performance improvements to net.BlockList.
Our Next.js services and edge functions depend on current Node LTS; certificate and dependency updates matter for security.
langchain-openai 1.6.1
Fixes Azure AD auth with OpenAI 3.8, adds async tools support and configuration_update, routes gpt-5.6-sol correctly.
We use LangChain's OpenAI integration for SalesKick; async tools and routing fixes reduce integration friction.
Langfuse v4.32.0
Alerts when a labeled preview is not serving, export data freshness metrics, fixes dataset metric reads when preview enabled.
We rely on Langfuse observability for agent calls; alerting on broken previews and export freshness helps catch stale monitoring.
Hacker News
On the Navier–Stokes Millennium Prize Problem
OpenAI claims an AI-generated solution to the Navier-Stokes problem with a formal proof in Lean; 1022 comments, significant mathematician skepticism.
If verified, this reshapes what reasoning models can tackle; if not, it's a cautionary tale for publishing AI-generated proofs without peer review.
Mistral raises €3B
Series D at €21B post-money valuation, three years after launch, largest equity round in European tech history.
Mistral's open-weight models are on our evaluation list for cost-sensitive workloads; continued funding keeps them in the race.
Muse – Meta's personal AI agent
Personal agent with long-term memory and proactive task suggestions, powered by Muse Spark 1.3.
Meta entering personal agent space with memory shows the category maturing; patterns here apply to our assistant and scheduling work.
I-have-ADHD: A skill to stop coding agents from burying the answer
Hermes skill forcing agents to lead with the answer and put context second; 360 points, resonated with agent users.
This is a Hermes skill gaining traction in the wild—validates our agent framework and shows what users actually want from output structure.
People on X
Addy Osmani joins Anthropic
Addy Osmani (ex-Google Chrome, developer tooling) announced he joined Anthropic to work on Claude Code.
High-profile dev tooling hire signals Anthropic is deepening Claude Code investment; we track that product closely for coding workflows.
GPT-Image-2.5 Flare and Sunburst
New image models in the OpenAI API with sharper detail, stronger style adherence, more control over edits.
We use OpenAI's image models for content generation in campaigns; better style adherence reduces iteration cycles.
Hermes in Omarchy
Hermes Agent now has first-class support in Omarchy, DHH's Arch-based agentic Linux distribution, with desktop app and CLI.
Hermes is our agent framework; seeing it ship in a distro validates the approach and grows the ecosystem we're building on.