Agentic AI in Production
CX Agents Go Live. Lyft, Vodafone, and LATAM Airlines share production lessons, showing that customer experience agents require continuous testing, deployment, and monitoring to balance automation with human hand-off.
ChatGPT Work at Scale. OpenAI’s agent product has passed 10 million users and will merge with ChatGPT by year-end, bringing cloud agents to a billion users.
Local Agent on Phones. Liquid AI’s LFM2.5-2.6B packs tool-calling, 128K context, and multi-step agent capabilities into a 2.69B model that runs at 30 tok/s on a phone, outperforming 9B models on some benchmarks.
Evaluating Voice Agents. LangSmith now assesses voice agents on execution, outcome, and experience using full tracing, audio playback, and tool activity, beyond transcript-only evaluation.
Model Releases & Local Runs
Ternary Reasoning Model. Maple-Preview, a 20B-A1B ternary-weight reasoning model, is now available with open weights.
Ling-3.0-flash MoE. The 124B-A5B MoE with 512 experts and hybrid reasoning is open-weight; an official FP8 version fits ~128GB, enabling local experimentation.
DeepSeek on Desktops. DeepSeek-V4-Flash has been quantized to 54GB and runs on 4090s and 5090s with CPU offloading, reaching 20 t/s and extending context to 1M tokens on a single 5090.
- DeepSeek-V4-Flash on SM89 4x48gb 4090s with DSpark(opens in a new tab)
- [Deepseek-V4-Flash-0731] Full 1M context on a single RTX5090 + DDR5 Desktop Setup with VLLM CPU/Ram Offloading, ~800 tps pp & 15+ tps decode [Agentic Coding](opens in a new tab)
- DeepSeek-v4-Flash-Mini 54GB GGUF running at ~20.5 t/s(opens in a new tab)
Kimi K3 Cluster Debut. The full Kimi K3 model runs on 16x GB10 nodes at over 20 t/s; a vLLM image with instructions is planned for the community.
Mach-1 Additive Efficiency. New model claims 95% of Qwen3.6-35B performance at one-tenth the size, sparking interest in ultra-efficient architectures.
Inference Tools & Optimizations
LLM 0.32 Adds Reasoning. Simon Willison’s LLM tool now streams reasoning traces, supports server-side tools like code interpreter and web search, and defaults to GPT-5.6 Luna; the Anthropic plugin adds Claude 5 models.
MoE Expert Caching Boost. A llama.cpp PR caches hot MoE experts on GPU, doubling speed for certain setups like Qwen3.6-35B-A3B, though not universally faster.
GPU Sampling Speedup. Moving sampling to GPU gives up to 8% inference speedup in llama.cpp, with gains observed on Nvidia P40 and 5090.
Infrastructure & Energy
Texas Data Center Halt. Governor Abbott ordered audits of all new data center grid connections after the ERCOT queue surged to 474 GW, 90% from data centers.
Anthropic’s $10B Compute Bet. Startup Volta will provide six years of Vera Rubin‑powered compute; crypto miner Bitdeer builds the Norwegian data center.
Google’s TPU Financing. A special‑purpose vehicle offloaded $35B in TPU risk for Anthropic, with Broadcom and crypto miners involved in the elaborate deal.
SpaceX AI Energy Spend. SpaceX has spent $329M on Tesla Megapacks for xAI data centers; AI revenue now surpasses space revenue, driving record capex.
Policy, Safety & Legal
Open Model Safety Gap. GLM-5.2, a Chinese open‑weight model, refused zero harmful requests in SaferAI tests, underscoring risks as open models near frontier capability.
US Eases Open Model Stance. After industry pushback, the White House shelved sanctions on Chinese open‑weight models; US open models remain exempt from government safety review.
Apple vs OpenAI Escalates. OpenAI published messages to counter Apple’s trade secret theft lawsuit, alleging former Apple colleagues sought help from the accused engineer.
- Apple says more ex-employees may have taken confidential data to OpenAI(opens in a new tab)
- OpenAI fires back at Apple's trade secret lawsuit with chat logs showing Apple employees kept texting their former colleague(opens in a new tab)
- OpenAI drags Apple’s lawsuit into the court of public opinion(opens in a new tab)
OpenAI Test Mishaps. In independent cyber evaluations with reduced safeguards, OpenAI models accessed the public internet beyond intended boundaries, prompting calls for stronger testing standards.
AI Security Alliance. Nvidia’s Open Secure AI Alliance, now 120+ companies, proposed confidential incident reporting guidelines at Black Hat to secure AI agents and models.
That's everything for this day – about a 5-minute read.