Agent & Model Releases
LangChain GTM Agent. LangChain's internal sales agent generated 1,540 artifacts like account plans and proposals in 12 weeks. Deals where reps sent one of these documents moved to a later stage 47% of the time, compared with 16% for matched deals without one.
Databricks Genie update. The updated Genie agent can handle multistep, unsupervised jobs using an ontology built from customer data. CEO Ali Ghodsi demonstrated it analyzing employee surveys and preparing for interviews.
Inflect TTS v3. A 34M-parameter local TTS model with six voices, vocal-event tags like [laugh] and [sigh], and reference-voice tuning. It runs on CPU and CUDA under Apache 2.0.
Interactive world model. A new local model turns any image into an interactive video you control in real time, running on RTX 3060 and above. It uses a transformer with learned summaries and is available on Hugging Face.
Alpamayo 2 Edge. NVIDIA released an open-source self-driving model trained in an RL loop that takes six cameras and recent motion. Embedl optimized it to run over 10x faster than PyTorch on edge hardware like AGX Thor.
KRAFTON Raon-1-nano. A 1.5B-effective-parameter dense model for Korean and English with tool calling and up to 262K context. Released under Apache 2.0 in GGUF, MLX, and LiteRT-LM formats.
Tools & Frameworks
Gemini Decision 1. Google's new Decisions API scores up to 1,000 closed-set questions against one multimodal input in 150-250ms for $0.10 per million input tokens, with no output charge. It targets routing, policy checks, and tool selection inside agents.
- Gemini Decisions API: Fast, calibrated multimodal decisions(opens in a new tab)
- Google Launches Gemini Decisions API With Gemini Decision 1 at $0.10/M Tokens(opens in a new tab)
- Gemini Decisions API(opens in a new tab)
- Google joins the decision-model wave with Gemini Decision 1, scoring 1,000 questions in a single pass(opens in a new tab)
Quantization advances. Unsloth Dynamic V3.1 and Huawei's MACaQ both improve GGUF quality at similar file sizes. A 2.9-bit DeepSeek V4.1 Flash quant on 6x3090s matches unquantized performance on APEX 2025.
- Open-source method (MACaQ) beats Unsloth GGUFs!(opens in a new tab)
- DeepSeek V4.1 Flash 1017 exl3 2.9bpw APEX 2025 benchmark results(opens in a new tab)
- Unsloth Dynamic V3.1 released for Qwen3.8-27B, DeepSeek-V4.1-Flash-1017 and Qwen3.8-Flash-Next(opens in a new tab)
- New Qwen3.8-27B Unsloth Dynamic V3.1 GGUFs(opens in a new tab)
AWS emulator. A Rust-based AWS emulator built by a swarm of AI agents accumulated 29,684 commits in three weeks. It currently requires Docker on macOS/Windows and could become a single-binary test tool.
Code retrieval benchmark. JioHotstar evaluated six approaches to help coding agents navigate large codebases, finding structural search best for concrete tasks and semantic retrieval for ambiguous requirements. Adding retrieval layers did not consistently improve task completion or token use.
Research Highlights
AlphaGenome 2. DeepMind's DNA model predicts variant effects in specific cell types by reading 1 Mb of DNA plus a cell state. It beats fine-tuned AlphaGenome 1 and baselines on single-cell benchmarks; API is non-commercial, weights later.
Signal verification. DeepMind's AlphaProof Code mathematically verified Signal's cryptographic core in four weeks, writing over six million lines of Lean proofs and flagging 1% for human review. The approach could extend to power grids and medical devices.
Attention architecture shift. Open LLM roundup shows no full-attention models remain in late 2026; models use hybrid or local attention. This makes long contexts and agent tool calls cheaper to run.
Industry & Business
Nvidia CEO scandal. Jensen Huang admitted to an affair with a woman U.S. intelligence alleges is a Chinese spy, and the New York Times reported a second affair with an Nvidia executive. Nvidia's board is reportedly preparing to oust him, Congress may subpoena him, and there is no clear succession plan.
- Who could replace Jensen Huang? A highly unscientific list of possible successors(opens in a new tab)
- The unfolding Jensen Huang scandal: a timeline(opens in a new tab)
- Nvidia’s stock drops as board reportedly preps to oust CEO Jensen Huang(opens in a new tab)
- Jensen Huang accused of having another affair — with an Nvidia exec(opens in a new tab)
- Congress to call Jensen Huang in to answer for alleged Chinese spy affair(opens in a new tab)
- Is Jensen Huang’s bad behavior more commonplace than it seems?(opens in a new tab)
- Nvidia is reportedly preparing to fire Jensen Huang(opens in a new tab)
- Nvidia is Preparing to Oust Nvidia CEO Jensen Huang(opens in a new tab)
Anthropic's multi-chip bet. In its IPO roadshow, Anthropic says it trains and serves Claude across AWS Trainium, Google TPUs, NVIDIA GPUs, and AMD GPUs. Its model Mythos ported kernels to AMD in three days, eroding NVIDIA's CUDA lock-in.
OpenAI hits $10B quarter. Quarterly revenue crossed $10 billion for the first time, up from $6.7 billion in Q2, with adjusted operating margin improving to -109%. Growth is strong but still trails Anthropic's enterprise traction.
TSMC invests $1B. TSMC's board approved up to $500 million each in OpenAI preferred stock and Anthropic common shares at IPO. The chipmaker gains direct stakes in leading AI developers.
Fractile raises $720M. British AI chip startup Fractile raised a $720 million Series C at a valuation over £5 billion. It combines custom hardware with a machine learning compiler to cut inference costs.
CoreWeave funding shift. CoreWeave reported Q3 revenue up 168% to $3.66 billion but a $709 million net loss. It is shifting to customer prepayments, contract-backed loans, and shorter deals to fund AI infrastructure.
Data center hype check. Many announced data center projects may never be built due to local opposition, permitting, and power constraints, according to an Equity podcast with Ascend Analytics. The pipeline is wildly inflated.
Policy & Security
RL data exports. At least nine US firms are selling roughly $500 million in reinforcement learning environments to Chinese AI labs, according to a ChinaTalk investigation. The data lets Chinese labs reuse environments built for OpenAI and Anthropic, prompting calls for export controls.
Helsing Clarity. Defense company Helsing unveiled Clarity, an agentic platform that aggregates sensor data to recommend military targets and weapons. It requires human authorization for every strike and emphasizes European sovereignty.
United blocks AI booking. United Airlines says personal AI assistants cannot complete ticket purchases, citing privacy, regulatory, and safety attestation concerns. Agents can still display flights but not transact.
That's everything for this day – about a 5-minute read.