The Pulse, Thursday 3 September 2026
Meta's Muse Spark 1.3 and Google's Gemini 3.8 Flash landed within a day of each other, both pitched at long-running agentic work. On arXiv, two papers put numbers on things we lean on daily: code-embedding retrieval and LLM-as-a-judge. And Nvidia says it is acquiring Hugging Face.
Models
Meta: Muse Spark 1.3
Meta's multimodal reasoning model for long-running agentic, multi-agent, and coding workflows, listed on OpenRouter with a 1M-token context at $1.25 in and $4.25 out per million tokens, plus a Contributor tier at $0.10 and $0.20.
A 1M-context agentic model with a near-free contributor tier is worth a run through the wehshi review panel and the SalesKick grading evals before we decide it is just another launch.
Introducing Gemini 3.8 Flash and 3.8 Flash Cyber
Google's next Flash generation, pitched at agentic workflows, with a cyber-defence variant offered to governments and partners through a limited-access program (vendor claims).
Flash-class models are what we run for high-volume calls; a new generation is a candidate for the EverRaise call-disposition classifier and the cheaper lanes of the review panel.
Research
ExecRetrieval: Measuring the Functional-Correctness Gap in Code-Embedding Retrieval
A benchmark that plants controlled functional bugs to test whether code-embedding retrieval returns correct code rather than lexically similar code.
Our coding agents and the review panel's full-file context both rest on retrieval; this is a way to check whether an embedding model is fooled by near-duplicates.
EarlyEval: Cheaper Agent Evaluation via Early Outcome Prediction
Predicts an agent's final outcome from its early trajectory to cut evaluation cost, which the authors put at hundreds to thousands of dollars per pass over an agentic benchmark.
Agent evals are where our OpenRouter spend goes; early-outcome prediction is a cheap trick to try on the review-panel calibration runs.
Hearing the Whispers: Black-Box Membership Inference Attacks on Finetuned TTS Models
Black-box membership inference against fine-tuned text-to-speech models, showing that private voice data used for personalisation can be exposed.
EverRaise voices are built from customer recordings; any TTS fine-tuning we adopt needs this threat model in the vendor conversation.
CHIME: Credit-Aware Hierarchical Memory Evolution for Long-Horizon Agentic Planning
A self-evolving memory for agents that accumulates reusable experience from interaction outcomes and credits the memories that actually helped planning, avoiding the cost of test-time search or retraining.
Long-horizon agent memory is our problem too — Botterfly's recall and any SalesKick agent that learns across calls face the same question of which stored experience to trust.
Releases
langchain-anthropic 1.7.1
Adds Claude Fable 5.1 support and stops sending middleware trace inputs, a small performance fix.
The services that reach Anthropic through LangChain should bump to this before anyone tries Fable 5.1 through them.
Langfuse v4.28.0
Evals now track judge prompt metadata, ClickHouse query outcomes are counted for read SLIs, and experiments get a search bar.
Judge prompt metadata per eval run is what the grading calibration work wanted to see; worth upgrading the Langfuse instance before the next round.
Hacker News
Nvidia to Acquire Hugging Face
Nvidia's own announcement that it is acquiring Hugging Face, the hub most open-weight models are published through.
Half the models we evaluate arrive via Hugging Face; the questions in the thread about hosting, licensing, and neutrality are ours too.
Pre-Release of Polars 2.0
The Polars team's pre-release notes for 2.0, the first major since the dataframe library took over a good share of pandas workloads.
Our data-prep scripts for grading and call analytics are pandas today; a 2.0 pre-release is the moment to test a Polars port on one of them.
Claude outage – Resolved
Anthropic's status page for yesterday's Claude outage, which hit Hacker News the same day as a ChatGPT outage.
Both major providers dropped within hours of each other; the review panel's OpenRouter routing and the cron jobs need their fallbacks exercised, not assumed.
Audacity 4.0
The first major Audacity release in years, and the day's top story by points.
Worth knowing, no direct use for us yet.
First issue. Python.org and Z.ai had nothing new in the window; everything else ran clean.