AI lab announcements, 2–8 March 2026

27 announcements from 7 labs, first-party only

OpenAI

OpenAIAvailability change· 6 Mar

Codex Security: now in research preview

OpenAI announces Codex Security, an AI application security agent now available in research preview.

OpenAI· 6 Mar

How Descript engineers multilingual video dubbing at scale

Descript uses OpenAI reasoning models to automatically localize large content libraries while preserving timing and meaning.

OpenAI· 6 Mar

How Balyasny Asset Management built an AI research engine

Balyasny Asset Management built an AI research engine that uses OpenAI's full platform and agent workflows to transform investment research.

OpenAIResearch paper· 5 Mar

Reasoning models struggle to control their chains of thought, and that’s good

OpenAI introduces CoT-Control and finds reasoning models struggle to control their chains of thought, reinforcing monitorability as an AI safety safeguard.

OpenAINew model· 5 Mar

Introducing GPT-5.4

OpenAI announces GPT-5.4, a frontier model with state-of-the-art coding, computer use, tool search, and a 1M-token context window.

OpenAICapability change· 5 Mar

Ensuring AI use in education leads to opportunity

OpenAI announces new tools, certifications, and measurement resources for schools and universities to close AI capability gaps and expand opportunity.

OpenAI· 5 Mar

The five AI value models driving business reinvention

OpenAI outlines five AI value models that guide leaders in sequencing AI adoption from workforce fluency to process reinvention for durable business advantage.

OpenAIProduct launch· 5 Mar

Introducing ChatGPT for Excel and new financial data integrations

OpenAI launches ChatGPT for Excel alongside new financial data integrations, leveraging GPT-5.4 to speed up modeling, research, and analysis in regulated environments.

OpenAI· 5 Mar

Introducing the Adoption news channel

OpenAI introduces a news channel that offers practical insights and frameworks to help businesses turn AI progress into business advantage.

OpenAIResearch paper· 4 Mar

Extending single-minus amplitudes to gravitons

OpenAI announces a new preprint that extends single-minus amplitudes to gravitons, with GPT-5.2 Pro assisting in deriving and verifying nonzero graviton tree amplitudes in quantum gravity.

OpenAIProduct launch· 4 Mar

Understanding AI and learning outcomes

OpenAI introduces the Learning Outcomes Measurement Suite, a tool to assess AI's impact on student learning over time.

OpenAI· 4 Mar

How Axios uses AI to help deliver high-impact local journalism

Axios COO Allison Murphy explains how the company uses AI to support local reporters, streamline newsroom workflows, and deliver high-impact local journalism at scale.

Together AI

Together AIResearch paper· 5 Mar

FlashAttention-4: Algorithm and Kernel Pipelining Co-Design for Asymmetric Hardware Scaling

Together AI presents FlashAttention-4, an algorithm and kernel co-design for Blackwell GPUs that maximizes overlap between matmul and resource bottlenecks like exponential units and shared memory…

Together AI· 5 Mar

Key research and product announcements at the AI Native Conf

Together AI announces seven research and product releases at the AI Native Conf, including FlashAttention-4, the Reinforcement Learning API, ThunderAgent, together.compile, and Megakernel, covering…

Together AIInfrastructure· 4 Mar

Cache-aware prefill–decode disaggregation (CPD) for up to 40% faster long-context LLM serving

Together AI announces cache-aware prefill–decode disaggregation (CPD), a serving architecture that separates cold and warm workloads by cache hit rate to improve throughput and reduce…

Together AIStory of the week· 2 Mar

Introducing Together AI’s new look

Together AI announced a refreshed visual brand identity, including a new logo and design language created with Pentagram, to better reflect its mission of open innovation and its platform connecting…

Google Research

Google ResearchResearch paper· 6 Mar

WAXAL: A large-scale open resource for African language speech technology

Google Research introduces WAXAL, a large-scale open-access speech dataset for 27 Sub-Saharan African languages, featuring approximately 1,846 hours of ASR data and over 565 hours of TTS data,…

Google Research· 6 Mar

Where wild things roam: Identifying wildlife with SpeciesNet

SpeciesNet is an open-source AI model developed by Google Research that identifies wildlife in camera trap images. It was released open-source a year ago.

Google ResearchResearch paper· 4 Mar

Teaching LLMs to reason like Bayesians

Google Research teaches LLMs to reason in a Bayesian manner by training them to mimic the predictions of an optimal Bayesian model, improving performance on a recommendation task and enabling…

Hugging Face

Hugging Face· 5 Mar

Bringing Robotics AI to Embedded Platforms: Dataset Recording, VLA Fine‑Tuning, and On‑Device Optimizations

This guide presents NXP's hands-on best practices for recording reliable robotic datasets, fine-tuning VLA policies (ACT and SmolVLA), and highlights the real-time performance that NXP i.MX 95 SoC…

Hugging FaceDeveloper tool· 5 Mar

Introducing Modular Diffusers - Composable Building Blocks for Diffusion Pipelines

Hugging Face introduces Modular Diffusers, a new composable block-based system for building diffusion pipelines, allowing users to mix and match reusable blocks instead of writing entire pipelines…

Hugging FaceResearch paper· 3 Mar

PRX Part 3 — Training a Text-to-Image Model in 24h!

Photoroom describes a 24-hour training speedrun for a text-to-image model using a combination of techniques such as x-prediction in pixel space, perceptual losses, token routing with TREAD, and…

Cursor

CursorCapability change· 5 Mar

Automations

Cursor introduces automations, allowing users to create always-on agents that run on schedules or trigger from events (Slack, Linear, GitHub, PagerDuty, webhooks).

CursorAvailability change· 4 Mar

Cursor in JetBrains IDEs

Cursor is now available in JetBrains IDEs (IntelliJ IDEA, PyCharm, WebStorm, etc.) through the Agent Client Protocol (ACP), enabling developers to use frontier models from OpenAI, Anthropic, Google,…

LG AI Research

LG AI ResearchResearch paper· 5 Mar

[AAAI 2026] RL-Studio: A System for Multi-Phase Reinforcement Learning Experimentation

LG AI Research presented RL-Studio, a platform for multi-phase reinforcement learning experimentation, at the AAAI 2026 Demonstration Program.

LG AI Research· 4 Mar

[LG Accountability Report on AI Ethics EP.2] Proving Safety, Ensuring Trust: Inside the 2025 Report

LG AI Research published a report detailing its 2025 progress in AI ethics and trustworthiness, including the K-AUT risk taxonomy, hallucination mitigation research, deepfake detection, internal…

Google DeepMind

Google DeepMindNew model· 3 Mar

Gemini 3.1 Flash-Lite: Built for intelligence at scale

Google DeepMind announces Gemini 3.1 Flash-Lite as the fastest and most cost-efficient model in the Gemini 3 series.