Google DeepMind Antigravity accelerates drug discovery

Today in brief

• ● 7 Small Specialist Models Beat Giant LLMs at Specific Tasks KDnuggets
• ● Google Antigravity accelerates drug discovery with agentic cheminformatics Google DeepMind
• ● LangSmith BYOC Now Runs on AWS, Azure, and Google Cloud LangChain
• ● OpenAI shares advances in AI-assisted engineering for physical products OpenAI
• ● Securing the agentic enterprise with data lineage MIT Technology Review AI

Model & Product Releases 3 stories

KDnuggets Nov 4 · 3 min read

7 Small Specialist Models Beat Giant LLMs at Specific Tasks

The article details seven open-weight specialist models under 10 billion parameters that outperform larger LLMs on single tasks like transcription, OCR, routing, and speech. Builders can swap expensive frontier model calls for these task-specific models running locally, but should verify each model's license and benchmark claims against their own data.

MarkTechPost Nov 4 · 2 min read

Tencent Releases WeVisOne-VL Vision Model Unifying Detection, OCR, Segmentation

Tencent's WeChat Vision team released WeVisOne-VL, an Apache 2.0 vision-language model in 2B and 4B sizes that handles detection, OCR, counting, segmentation, and alpha matting with a 9.6M-parameter pixel decoder. Developers can replace separate detector, OCR, and segmentation pipelines with one model, but should test its "None" output behavior for absent objects in their workflows.

Decagon Nov 3 · 3 min read

A better payment experience from reminder to resolution | Decagon

An AI agent for the full payment relationship

Research 4 stories

Google DeepMind Nov 3 · 3 min read

Google Antigravity accelerates drug discovery with agentic cheminformatics

Google Antigravity with new Cheminformatics Skills searches thousands of compounds and proposes novel rigidified macrocycles with synthetic routes in one hour instead of days. Teams building drug-discovery pipelines can evaluate whether agentic workflows reduce iteration time for their specific targets.

Hugging Face Nov 3 · 3 min read

Palette fine-tunes a 20B model to generate slide decks in custom design languages

Researchers fine-tuned a 20B open-weight model on synthetic data and a verification harness to produce slide decks that match a specific design language instead of defaulting to generic styles. Engineers who need on-premise deck generation can test whether this approach beats prompting frontier models for brand-consistent output.

Allen Institute for AI Nov 3 · 3 min read

StateBench benchmarks long-context state tracking across four difficulty levels

StateBench evaluates how well language models track evolving entity states across long documents, revealing that recurrent and hybrid architectures like Gated DeltaNet handle hard dependency chains better than transformers. Developers choosing architectures for long-context tasks should check whether their use case resembles the benchmark's harder chain-tracking variants.

Also covered by Hugging Face.

TechCrunch AI Nov 3 · 2 min read

Anthropic is working on a system to prove AI models’ identity

Called FlashProve, the system would let anyone check that an AI response really came from the model it claims to be from.

Tooling & Infrastructure 4 stories

LangChain Nov 4 · 3 min read

LangSmith BYOC Now Runs on AWS, Azure, and Google Cloud

LangSmith Bring Your Own Cloud (BYOC) is now generally available on AWS and in beta on Azure and Google Cloud, keeping traces, evals, datasets, prompts, and agent deployments in your own cloud account while LangChain manages the control plane. This lets you run agent workloads and store telemetry in your preferred cloud region without operating the LangSmith platform yourself.

LangChain Nov 3 · 3 min read

LangChain Optimizes Browser Use for Deep Agents with Stagehand

LangChain evaluated seven browser automation toolsets for its Deep Agents and selected Stagehand from Browserbase, then used LangSmith traces and evals to iteratively optimize its harness for cost and trajectory. Builders adding web-browsing capabilities to agents should test tool choices against realistic multi-step tasks before committing, since the toolset heavily influences both latency and success rates.

AWS Machine Learning Nov 3 · 3 min read

Bedrock AgentCore Gateway Adds OBO Token Exchange for ServiceNow

Amazon Bedrock AgentCore Gateway now supports on-behalf-of (OBO) token exchange with Microsoft Entra ID, allowing AI agents to call the ServiceNow REST API while preserving the end user's identity without custom Lambda interceptors. Teams building enterprise agents that need audit trails and user-level access controls can configure this at the gateway target and apply it to other OAuth-protected APIs.

Meta Engineering Nov 3 · 3 min read

Meta's F3 Lightning Decouples Feature Releases for Ads ML

Meta's F3 Lightning adds an infrastructure layer atop the Facebook Feature Framework so each of the 20,000-plus features per ad request can be compiled, validated, deployed, and rolled back independently instead of as a single merged graph. The system preserves compute efficiency through stable computation identifiers, a shared execution plan, a shared memory buffer, and dense index mapping for direct array access.

Industry & Policy 3 stories

OpenAI Nov 3 · 2 min read

OpenAI shares advances in AI-assisted engineering for physical products

OpenAI detailed how its models now assist mechanical and electrical engineers with CAD, PCB design, simulation, and cross-discipline coordination, based on internal use and early industry collaboration. Engineers evaluating AI for hardware workflows should note these capabilities are still early and OpenAI is actively seeking feedback on where they add the most value.

OpenAI Nov 3 · 3 min read

Cisco pairs network telemetry with OpenAI models for machine-speed defense

Cisco integrated OpenAI's frontier models with its network telemetry to distribute security controls across microservices, containers, and clusters, leveraging OpenAI's zero data retention and verifiable isolation for enterprise privacy. Security teams considering AI-augmented defense should verify that their data governance requirements align with OpenAI's enterprise privacy architecture before deploying similar integrations.

AWS Machine Learning Nov 3 · 3 min read

Gen Digital launches Norton Revamp on Amazon Bedrock AgentCore for reputation management

Gen Digital built Norton Revamp, a professional reputation management platform, using Amazon Bedrock AgentCore Web Search to ground AI-curated news feeds and content suggestions in real-time, cited web knowledge with zero data egress. Developers building grounded AI applications can reference this implementation of AgentCore Web Search for privacy-preserving, real-time information retrieval in regulated contexts.

Also Worth Reading 3 stories

MIT Technology Review AI Nov 4 · 2 min read

Securing the agentic enterprise with data lineage

The article describes a scenario where AI agents pull pricing data from restricted systems, internal communications, and archives without human oversight, leading to confidential information leaking into a proposal that cannot be traced back to its source. Teams building agentic workflows need to implement data lineage tracking to maintain governance over data that moves, transforms, and replicates across multiple agents and systems.

Microsoft Research Nov 3 · 3 min read

Learn the rules: The hacker mindset

Research Manager Mike Walker reflects on a career spanning hacker culture, cybersecurity, and AI, sharing lessons about trust, teamwork, leadership, and why solving difficult problems requires curiosity, judgment, agility—and an understanding of the rules.

TechCrunch AI Nov 4 · 3 min read

TechCrunch Founder Summit 2026: Doors open at 9 a.m. ET in Boston

Today is the day to join TechCrunch Founder Summit 2026 in Boston's SoWa Power Station. Register now to join this founder-focused conference today.