AI Daily Digest · November 14, 2026: NVIDIA Loses Its Founder-CEO, ChatGPT Gets a Wallet, llama.cpp Hits 1.0

Saturday's digest covers a founder era ending at NVIDIA, the first serious wallet built for AI agents, a $3B vote of confidence in agent infrastructure, and three releases that matter to people who build: llama.cpp reaching 1.0, an audio codec that encodes 361 times faster than real time, and a genome model that reads DNA the way cells do.

1. NVIDIA announces CEO transition: Jensen Huang resigns

NVIDIA's board announced on November 12 that Jensen Huang has resigned as President, Chief Executive Officer, and a member of the board, effective immediately. The company that defined the hardware layer of the AI boom will be run, for now, by Colette Kress, who has served as EVP and CFO since September 2013. She takes the interim CEO title while keeping her finance role until a permanent successor is named. The board has started a search and says it will consider both internal and external candidates.

Two governance changes stand out. Harvey C. Jones, a semiconductor veteran who previously led Synopsys as CEO, becomes the first chair of the board in NVIDIA's history. The company has operated for 33 years without one, and creating the role now signals how the board wants decision authority structured while the founder is gone. In the announcement, Jones expressed full confidence in Kress and said operations, customer commitments, and the product roadmap are expected to continue without interruption.

The announcement lands a week after unusual turbulence on NVIDIA's own newsroom page: a personal statement from Huang on November 8, a board statement the same day noting an independent investigation led by Joon H. Kim of Wachtell, Lipton, Rosen & Katz, and a follow-up statement from independent directors on November 10. The resignation release itself does not state a reason. Markets reacted before the close, with the stock finishing down over 4 percent on the day, and attention now shifts to the Q3 fiscal 2027 earnings report scheduled for November 17, where Kress will face Wall Street as the company's public face for the first time.

— NVIDIA · CNBC

🔗 NVIDIA Newsroom · SEC filing

2. OpenAI introduces ChatGPT Wallet, a payment rail for agents

OpenAI launched ChatGPT Wallet on Friday, its first product that lets agents complete purchases on a user's behalf. The design keeps the actual card out of the loop entirely: after a user adds a card once, the payment network issues a payment token for each transaction, so neither the agent nor the merchant ever sees card numbers or security codes. Every purchase requires a fresh approval through a device security check such as Face ID or a fingerprint scan, and each approval creates a record of what the user authorized.

Access starts narrow. The wallet is available through dots for ChatGPT Pro subscribers in the US, limited to users 18 and over holding a US-issued Visa, Mastercard, or American Express card. OpenAI built the feature with all three payment networks, and the announcements from each partner stress the same point: this should strengthen the existing trust chain rather than bypass it. Cardholders keep their rewards and their dispute rights, and banks stay responsible for evaluating transactions.

The harder problem is the merchant side, and OpenAI is addressing it with a Commerce Partner Playbook published alongside the wallet. It tells businesses how to claim and verify their presence in ChatGPT, upload product feeds, and configure CDNs and firewalls so signed agent requests using Web Bot Auth are not mistaken for bot traffic. OpenAI reports 1.2 billion people use ChatGPT each week, and the company says it expects to eventually let users pre-approve purchases under parameters like spending limits and purchase categories. For now, every charge keeps a human in the loop.

— OpenAI · CNBC

🔗 OpenAI · CNBC

3. LangChain raises $300M at a $3B valuation to build agent infrastructure

LangChain announced a $300 million Series C on Friday at a $3 billion valuation, led by existing investor CapitalG, Alphabet's independent growth fund. New investors GIC and DFJ Growth joined with significant participation, alongside IVP and Sapphire Ventures, with Salesforce Ventures, Insight Venture Partners, QuantumLight, and MongoDB Ventures also coming in, and Sequoia, Amplify, and Benchmark adding to their positions. Total funding now stands at $460 million, and the valuation more than doubles the $1.25 billion at which the company raised a year ago.

The bet is not on a model. CEO Harrison Chase frames the mission as enabling every company to own its intelligence: choosing models freely, controlling data and infrastructure, and improving agents continuously under centralized governance. The commercial product carrying that pitch is LangSmith, a framework-agnostic platform organized around three pillars: Agent Runtime for deployment, Agent Improvement for testing and monitoring, and Agent Governance for model budgets, data protection, and tool access control.

The customer numbers give the round its substance. More than 7,000 organizations use LangSmith, including NVIDIA, Toyota, and Salesforce at the enterprise end and Harvey, Rippling, and Listen Labs among startups. Clay uses it to monitor hundreds of millions of agent runs per month, and Workday reports cutting the time to deploy a new agent from five months to under a day. A release event on November 20 will show the next set of LangSmith capabilities.

— LangChain · TechCrunch

🔗 LangChain · CapitalG

4. IBM acquires AuthMind to watch what agents do with their credentials

IBM announced the acquisition of AuthMind, an identity security company focused on identity observability and threat detection, and plans to fold its technology into HashiCorp Vault. The gap being addressed is easy to describe: vaults control what happens at the moment a credential is issued, but say nothing about what an AI agent actually does with that credential afterward, across applications, tools, and data. AuthMind's technology builds an identity activity graph that maps relationships and access paths between identities and resources, using network traffic analysis to spot shadow agents operating outside governance.

IBM lays out a five-part model for securing the agent lifecycle: register each agent with a trusted identity and an accountable owner, govern the authority delegated to it, enforce least-privilege access with short-lived credentials, audit what permissions are actually used, and revoke access in real time when behavior turns risky. AuthMind's runtime visibility feeds the audit and response stages, turning static permission lists into a continuous view of what agents are doing in hybrid environments.

The deal deepens an existing relationship. IBM Ventures invested in AuthMind's seed round in 2023, and the two companies had already collaborated through IBM Identity Protection on governance and threat detection for non-human identities. Financial terms were not disclosed. As agent deployments multiply, the buyer here is every security team that has started asking who authorized which agent and what it touched.

— IBM Newsroom · Unite.AI

🔗 IBM Newsroom · Unite.AI

5. llama.cpp and ggml ship v1.0.0, ending three and a half years of 0.x

llama.cpp reached version 1.0.0 on November 13, together with ggml v1.0.0, the tensor library underneath it. The tag ends roughly three and a half years of rolling 0.x builds for the most widely embedded local inference stack, and it converts the public C API from a moving target into a managed contract. The rules are explicit: a major bump means a breaking change to the public C API in the include directory, a minor bump covers backward-compatible additions such as new backends or model support, and a patch leaves the API untouched.

The release is more than a number. It removes the deprecated public API, so code still calling old functions breaks on upgrade, and it tightens llama-server defaults: tools and MCP on a non-loopback address now require an API key or an explicit opt-in flag, CORS handling is fixed with DNS-rebinding protection, and proxies refuse to reach localhost or private addresses. For a server that many hobbyists expose on home networks, these are sensible guardrails shipping in the same package as the version contract.

Distribution spans eight channels, from Homebrew and Winget to Docker, Debian packages, conda-forge, and an XCFramework for Apple platforms, pointing at nightly build b12359. For the large ecosystem of desktop apps, language bindings, and on-device products built on libllama, the practical deliverable is predictability: versions can be pinned, upgrades planned, and any future major bump treated as a visible review gate rather than a surprise.

— ggml-org · GitHub

🔗 Hugging Face blog · GitHub release

6. Liquid AI open-sources LiquidCodec, an audio codec built for on-device voice

Liquid AI released LiquidCodec, the streaming audio codec behind its LFM3-Audio-2B and LFM3-Omni-2B voice models, with open weights on Hugging Face under the LFM Open License v1.0. The codec is small, about 74 million parameters against the 96 million of Kyutai's Mimi, which it replaces in Liquid's stack. It processes 24 kHz mono speech at a 12.5 Hz frame rate, emitting 12 tokens per 80-millisecond frame, roughly 1.65 kbps, and it streams in both directions.

The speed numbers are the story. On a single CPU thread of an Apple M4 Max MacBook Pro, Liquid AI measured encoding at about 361 times real time and decoding at about 87 times real time, with an encode real-time factor of 0.0028 against Mimi's 0.0361. Two design choices drive it: grouped finite scalar quantization, whose 12 codebooks are independent so all tokens in a frame can be predicted in parallel, and a decoder that predicts spectral coefficients directly instead of running a learned upsampling stack.

On quality, LiquidCodec posts the best ViSQOL and lowest mel distance across eight test sets including LibriSpeech and FLEURS, and a same-architecture TTS comparison shows a 3.15 percent word error rate on LibriSpeech against Mimi's 6.27 percent. The weak spot is accuracy in some languages: Japanese comes in at 7.29 percent, the highest of the codecs compared. Training covered 264,000 hours across 107 languages, then 55,000 hours of full-band speech with adversarial critics, plus quantization-aware training that cuts the decoder footprint by 3.5 times for edge deployment. Quantized GGUF builds for llama.cpp are already available.

— Liquid AI · Hugging Face

🔗 Liquid AI · Hugging Face

7. DeepMind announces AlphaGenome 2, a DNA model that thinks in cell types

Google DeepMind announced AlphaGenome 2, the next generation of its DNA sequence model, and the change is architectural as much as it is statistical. The first model took a million-letter DNA sequence and predicted molecular properties across many tissues in parallel. AlphaGenome 2 is a conditional model: alongside the sequence it takes a biological condition, a tissue, cell type, or state, as input at every stage, so it learns how the same stretch of DNA is read differently in a heart muscle cell versus an immune cell. It can even accept a scientist's own single-cell experiment through a cell encoder and generalize to states it has never seen.

The model was trained on more than 100 million individual cell profiles across 68 tissues, and DeepMind reports it identifies nearly twice as many variants that change gene expression as the previous state of the art. On the TenK10K benchmark of 28 primary immune cell types, recall of causal variants at 90 percent precision rose from 16 to 30 percent, and direction-of-effect prediction from 23 to 43 percent. New capabilities include DNA methylation prediction and cross-modality imputation, where one measurement such as single-cell gene activity lets the model predict unmeasured properties like chromatin accessibility.

The practical payoff shows in disease examples. AlphaGenome 2 reproduces, computationally, a Nature-published finding that an enhancer driving inflammatory bowel disease works by over-activating the ETS2 gene in gut macrophages, and it flags variant effects that broad-tissue averages would wash out. The model is available now for non-commercial research through the AlphaGenome API, with a private preview for commercial use, and DeepMind says weights will follow at formal publication.

— Google DeepMind · Nature

🔗 Google DeepMind · Nature

KD Agentic · AI Daily Digest