AI Daily · 2026-10-19

The main thread today is AI-native development pushing both tooling and memory layers to change. Anthropic says its monorepo merges grew from about 2,500 to more than 30,000 per day, making source control the biggest bottleneck, so it built and migrated to Scion, an internal version control system designed for always-on agents, replacing git within five months. Tencent Hunyuan released GGUF quantizations of Hy-Memory-30B-A3B for local llama.cpp inference, with Q6_K, Q8_0, and Q4_K_M options, aiming to replace general LLMs in external memory frameworks such as Mem0 and MemOS. Numan adds an evaluation perspective: static datasets and scripted multi-turn conversations fail for coaching-style dialogue, so it now uses topic-specific splits and per-category LLM judging. Separately, Modal reports that NVIDIA's Arm-compatible Vera CPUs double throughput in its tests, making them a candidate for bursty multi-tenant sandbox workloads.

North America · First-hand

Anthropic

⭐⭐ [Other] Redesigning source control at Anthropic

Merges to Anthropic's monorepo grew from about 2,500 to more than 30,000 per day, making source control the biggest development bottleneck. The team built an in-house version control system called Scion, designed for always-online agents. According to the post, the team migrated off git in five months.

East Asia · First-hand

Tencent Hunyuan

⭐⭐ [Model Release] tencent/Hy-Memory-30B-A3B-GGUF

Tencent Hunyuan published GGUF quantizations of Hy-Memory-30B-A3B on Hugging Face for efficient local inference with llama.cpp, in three variants: Q6_K imatrix (24.71 GB), Q8_0 (31.99 GB) and Q4_K_M imatrix (18.24 GB). All three files passed llama.cpp --check-tensors, real no-think generation and the full quantization evaluation; on a fixed 20% LoCoMo test split (2 users, 262 questions, 59 memory chunks, top_k=10) they scored 58.78%, 57.63% and 56.49% on LoCoMo Judge, with Q6_K imatrix giving the highest measured quality and Q4_K_M the smallest footprint. Hy-Memory-30B-A3B is a compact long-term-memory writer model (~3B active parameters) built by Tencent Hunyuan to execute the write path inside external memory frameworks, replacing the general-purpose LLM in Mem0, MemOS and Hy-Memory by deciding what to retain, how to represent it, and how new evidence should update the persistent memory bank. It is trained in two stages, Memory On-Policy Distillation followed by Attributed Memory Reinforcement Learning, and transfers directly to unseen frameworks such as EverMemOS.

Ecosystem & Beyond (Products / Agents / Tools / Opinions)

Research

⭐⭐ [Research] How Numan Built an Eval-First AI Health Coach with LangGraph and LangSmith

UK digital health company Numan describes how it evolved the evaluation setup for its AI health coach Nu, built as a LangGraph graph where a supervisor model orchestrates handoffs to domain-specialist sub-agents, each with its own prompt and tools, while staying clear of medical diagnosis, prescriptions, and clinical decisions. The team started with the standard approach — a dataset of inputs and expected outputs scored by an LLM judge — but found it broke down for coaching conversations: single-turn evals on static fixtures assume one right answer, safety depends on surrounding context, and averaging one score across the dataset hides localized failures; one cluster of low scores turned out to be caused by an overly narrow rubric rather than by Nu. Scripted multi-turn conversations failed the same way, because pre-written patient turns ignore what Nu actually said, so once Nu deviates every later turn responds to a conversation that never happened. The current framework runs themed datasets covering coaching quality, safety, medication queries, routing, and summarisation for things decidable from a single reply, scores each category separately with an LLM judge instead of one average, and writes each rubric to ask only what a single reply can show.

Other

⭐⭐ [Other] NVIDIA Vera CPUs double throughput for Modal Sandboxes

NVIDIA hand-delivered the first Vera CPU server to Modal's SF office last week, and it is now running in a rack. Vera is NVIDIA's new Arm-compatible CPU, built for the Vera Rubin GPU platform and also sold as a standalone CPU server; supporting Arm would require changes across Modal's infrastructure and image build system, which has run exclusively on x86. Modal says its benchmarks show a 2x increase in throughput, making Vera a compelling option for powering Modal Sandboxes. The post explains the motivation: in multi-tenant sandbox clusters, workload shapes are not statically known and resource usage is spiky (a 10x surge from an npm install or an expensive grep), and Modal's burst model, which replaces fixed reservations, increases exposure to noisy-neighbor effects; CPU time also drives idle GPU cost in background agents and reinforcement learning rollouts. To test Vera, Modal built Agentless Agent Bench, 50 Terminal-Bench oracle solutions and verifiers with all network calls stripped, prewarmed in concurrent Docker containers.