March 11, 2026 is a packed day: Anthropic creates an interdisciplinary institute for public debate on AI, Perplexity rolls out its “Computer” vision with four simultaneous APIs, OpenAI releases the new agentic primitives of the Responses API, and Meta details four generations of custom AI chips developed in two years.
Anthropic Institute: Jack Clark at the helm of public benefit
March 11, 2026 — Anthropic launches The Anthropic Institute, a new effort to advance public debate on the challenges posed by powerful AI. The initiative is led by co-founder Jack Clark, who takes on a new role as “Head of Public Benefit” at Anthropic.
The Institute will bring together an interdisciplinary team — researchers, economists, legal scholars, public policy experts — with unique access to Anthropic’s frontier models. Its mission: analyze and communicate the societal, economic, and safety impacts of AI as they unfold.
Three initial hires were announced:
| Hire | Background |
|---|---|
| Matt Botvinick | Resident Fellow, Yale Law School; former Senior Director of Research, Google DeepMind |
| Anton Korinek | Professor of Economics (on sabbatical), Economic Research team |
| Zoë Hitzig | Former OpenAI, specializing in the social and economic impacts of AI |
The Institute builds on Anthropic’s existing teams: Frontier Red Team, Societal Impacts, Economic Research. In parallel, Anthropic is announcing the expansion of its Public Policy team, focused on model transparency, energy consumer protection, export controls, and global AI governance.
The creation of this Institute marks a notable step: Anthropic is formally structuring its public engagement and giving it a face with one of its co-founders.
🔗 Introducing The Anthropic Institute
Perplexity: coordinated launch of the “Everything is Computer” vision
March 11, 2026 — Perplexity is publishing four announcements simultaneously, forming a coordinated launch around its “Computer” vision: AI as a personal and professional computer.
Personal Computer and Enterprise
The Personal Computer is a dedicated Mac mini that runs 24/7, connected to local applications and Perplexity servers. It acts as the user’s digital proxy — sensitive actions require explicit approval. Waitlist open.
The Computer for Enterprise connects to Snowflake, Salesforce, HubSpot, and hundreds of platforms. Skills are customizable, and Slack integration makes it possible to work in DMs or shared channels. It is based on SOC 2 Type II, SAML SSO, and audit logs. Perplexity puts forward a figure from an internal study of 16,000 queries: $1.6 million saved in labor costs and 3.25 years of work completed in 4 weeks.
Comet Enterprise is a native AI browser with admin controls (domain-level permissions, action logs, MDM) and a CrowdStrike partnership for browser-level protections.
Perplexity Finance gets 40+ live finance tools (SEC filings, FactSet, S&P Global, Coinbase, LSEG, Quartr, Polymarket), with broker connection via Plaid to analyze a real portfolio.
Premium Sources offers access to Statista, CB Insights, and PitchBook directly in search flows — paid sources are cited automatically.
Agent API: full orchestration runtime
The Agent API is a managed runtime for building agentic workflows with integrated search, tool execution, and multi-model orchestration. In a single integration point, it replaces a model router, a search layer, an embeddings provider, a sandbox service, and a monitoring stack.
| Integrated tool | Capabilities |
|---|---|
web_search | Filtering by domain, recency, date range, language, content budget |
fetch_url | Direct URL retrieval |
| Custom functions | Full support |
Four optimized presets cover use cases: fast factual search, balanced research, in-depth multi-source analysis, and institutional research. Deep Research 2.0 is available via the advanced-deep-research profile — it launches dozens of searches per query and processes hundreds of documents.
The Agent API is model-agnostic, supports multi-model fallback chains for near-100% availability, and is available today on docs.perplexity.ai.
Search API: improved snippets and SEAL benchmark
The Search API update focuses on snippet quality and evaluation infrastructure.
The new span-level labeling pipeline identifies which segments of a source document are relevant to a query. Result: smaller and more precise snippets, reducing token costs and improving context management for downstream models.
The SEAL benchmark tests whether a retrieval system can answer questions whose answer changes over time. Perplexity is improving on SEAL-Hard while other providers decline. The search_evals framework is updated as open source on GitHub.
Other improvements: multi-query support (up to 5 in a single API request), filtering by language (ISO 639-1 code) and country, and Python SDK (pip install perplexityai) with native support for all three APIs.
Sandbox API: isolated code execution for agents
Perplexity is opening its internal code execution environment as a standalone service. Each session runs in an isolated Kubernetes pod, with a mounted persistent file system. Supported languages: Python, JavaScript, SQL. Package installation at runtime is possible.
Sessions are stateful: files created at one step are available in later steps, and long workflows can pause and then resume hours later. Security relies on a zero-trust model: no direct network access, outbound traffic via proxy, and code never has access to raw API keys.
The Sandbox API will be integrable into the Agent API with the same API key and credits. Status: private beta coming soon.
OpenAI: the Responses API gets a computing environment for agents
March 11, 2026 — OpenAI publishes an engineering article detailing the new Responses API primitives for building reliable autonomous agents: Unix shell tool, hosted containers, native context compaction, and reusable agent skills.
The shell tool
The shell tool allows the model to interact with a computer via the command line, with access to classic Unix utilities (grep, curl, awk). Unlike the existing code interpreter, which only ran Python, the shell tool supports Go, Java, Node.js, and other environments. GPT-5.2 and later models are trained to suggest shell commands.
The Responses API can execute multiple shell commands in parallel via separate container sessions, and applies an output limit per command to avoid saturating the context window.
Hosted containers
The container is the model’s workspace:
| Component | Description |
|---|---|
| File system | Uploading, organizing, and managing resources via container and file APIs |
| Databases | Structured storage (SQLite) — the model queries tables rather than loading all content |
| Network access | Centralized egress proxy with allowlist, domain-level secret injection |
Native context compaction
For long-running tasks, the Responses API includes a native compaction mechanism: models are trained to produce a compact, encrypted representation of the conversational state. Available server-side (configurable threshold) or via a /compact endpoint. Codex uses this mechanism to maintain long-running coding sessions without degradation.
Agent skills
Skills condition recurring workflow patterns into reusable bundles: a folder with a SKILL.md file and associated resources. The Responses API automatically loads the skill into the context before sending the prompt to the model. Skills are managed via API and versioned.
In parallel, a developer blog post celebrates one year of the Responses API with five customer case studies. The two publications form a coherent picture of the platform’s evolution toward agentics.
🔗 Responses API + computing environment 🔗 1 year of the Responses API
OpenAI: defense strategy against prompt injection
March 11, 2026 — OpenAI publishes a security article on the resistance of AI agents to prompt injection attacks.
Early attacks consisted of inserting direct instructions into external content (Wikipedia pages, emails). As models improved, these attacks evolved toward social engineering: convincing professional context, simulated urgency, alleged authorization. A 2025 example described in the article showed a successful attack in 50% of cases on an older version of ChatGPT.
OpenAI approaches the problem through the lens of the three-party system (employer / agent / malicious third party), analogous to a human customer service agent exposed to manipulation attempts. The goal is not to perfectly identify every attack, but to limit the impact of successful manipulation.
| Countermeasure | Description |
|---|---|
| Source-sink analysis | Detection of combinations of untrusted content + dangerous action |
| Safe Url | Detects whether information from the conversation would be transmitted to a third party — asks for confirmation or blocks |
| Application sandbox | Canvas and ChatGPT Apps detect unexpected communications and ask for consent |
Safe Url also applies to browsing in Atlas as well as searching and browsing in Deep Research.
🔗 Designing AI agents to resist prompt injection
Meta MTIA: four generations of AI chips in two years
March 11, 2026 — Meta publishes a detailed technical article on its family of custom AI chips, MTIA (Meta Training and Inference Accelerator). In two years, Meta has developed four successive generations to serve billions of users at lower cost.
“AI models are evolving faster than traditional chip development cycles.”
| Generation | Main innovation |
|---|---|
| MTIA 300 | First chip optimized for ranking and recommendation models, reusable modular foundation |
| MTIA 400 | Evolution toward GenAI workloads, 72-chip rack in a single scale-up domain |
| MTIA 450 | Doubling of HBM bandwidth, +75% MX4 FLOPS, hardware acceleration for attention and FFN |
| MTIA 500 | +50% HBM bandwidth vs MTIA 450, focus on GenAI inference |
The progression from MTIA 300 to MTIA 500: HBM bandwidth multiplied by 4.5 and FLOPS multiplied by 25. Meta’s strategy is based on high-velocity development (one new chip per year), a focus on inference rather than pre-training, and native PyTorch integration.
“Mainstream GPUs are typically built for the most demanding workload — large-scale GenAI pre-training — while Meta’s primary need is inference.”
The Processing Element architecture combines two vector RISC-V cores, a Dot Product Engine, a Special Function Unit, a reduction engine, and a DMA. The software stack is based on PyTorch, vLLM, Triton, and dedicated MTIA compilers, with vLLM integration via plugin architecture.
🔗 Meta MTIA: scale AI chips for billions
Gemini CLI v0.33.0: Enhanced Plan Mode and A2A Authentication
March 11, 2026 — Gemini CLI releases version v0.33.0, two weeks after v0.32.0, which introduced Plan Mode.
| Category | New features |
|---|---|
| Agent architecture | HTTP authentication for remote A2A agents, discovery of authenticated A2A agent cards |
| Plan Mode | Integrated research sub-agents, annotation support for user feedback, new copy subcommand |
| CLI interface | Compact header with ASCII icon, reversed context window display, 30-day default retention for chat history |
The addition of HTTP authentication for the A2A (Agent-to-Agent) protocol is the main technical novelty: Gemini CLI can now discover and authenticate with remote agents, laying the groundwork for secure multi-agent orchestration.
Gemini in Chrome: expansion to India, New Zealand, and Canada
March 11, 2026 — Google is expanding Chrome’s AI features to three new markets: India, New Zealand, and Canada.
Gemini in Chrome — the AI assistant in the side panel, based on Gemini 3.1 — is now available on Mac, Windows, and Chromebook Plus in these regions. Features being rolled out include: access to Gmail, Maps, Calendar, and YouTube from Chrome, cross-analysis of multiple open tabs, and image transformation directly in the browser via Nano Banana 2. The update adds 50+ additional languages, including Hindi, French, and Spanish.
🔗 Chrome expands to India, New Zealand, and Canada
AlphaEvolve: new bounds for 5 Ramsey numbers
March 11, 2026 — Pushmeet Kohli (Google DeepMind) announces that AlphaEvolve has established new lower bounds for 5 classical Ramsey numbers in extremal combinatorics — problems so difficult that Erdős himself had commented on their complexity, and whose best previous results dated back at least a decade.
AlphaEvolve acts as a meta-algorithm that automatically discovers the necessary search procedures, where historically it was necessary to manually design specific algorithms. This result illustrates AlphaEvolve’s capabilities beyond the Google kernel optimizations for which it was already known.
Gemini Embedding 2: Google’s first multimodal embedding
March 10, 2026 — Google announces Gemini Embedding 2, described as “our most capable and first fully multimodal embedding model”. It is Google’s first natively multimodal embedding model, available to developers via Gemini API and AI Studio.
GitHub Copilot: code review from the terminal and JetBrains advances
March 11, 2026 — Two notable updates for GitHub Copilot.
Code Review from GitHub CLI v2.88.0
It is now possible to request a Copilot code review directly from the terminal. The gh pr edit --add-reviewer @copilot (non-interactive mode) and gh pr create (interactive mode) commands integrate Copilot alongside teammates. Reviewer selection benefits from dynamic search, improving performance in large organizations and fixing accessibility issues. Available on all plans that include Copilot code review — update to GitHub CLI v2.88.0 required.
Agentic improvements for JetBrains IDEs
The JetBrains plugin update brings to general availability (GA): custom agents, sub-agents, and plan agent, as well as automatic model selection for all plans. In public preview: agent hooks (userPromptSubmitted, preToolUse, postToolUse, errorOccurred) via a hooks.json file in .github/hooks/, and support for AGENTS.md and CLAUDE.md files.
Other improvements round out the update: configurable auto-approve by MCP, thought panel for extended reasoning models, context window usage indicator, and deprecation of Edit mode in the menu.
Explore a repository in Copilot on the web
In public preview: it is possible to explore the tree structure of a GitHub repository directly from Copilot’s web interface. Selected files are automatically added as temporary references to the chat and can be made permanent.
🔗 Repository exploration in Copilot
Anthropic Sydney and Claude for Office
Sydney, 4th Asia-Pacific office
March 10, 2026 — Anthropic announces the upcoming opening of an office in Sydney, its fourth in Asia-Pacific (after Tokyo, Seoul, and Singapore). Australia ranks 4th globally for Claude.ai usage per capita; New Zealand ranks 8th by the same indicator (Anthropic Economic Index). The office will initially focus on enterprise customers, startups, and research.
🔗 Sydney, 4th Asia-Pacific office
Claude for Excel and PowerPoint: shared context and Office Skills
March 11, 2026 — The Claude for Excel and Claude for PowerPoint add-ins receive two important updates: context sharing between the two applications (an analyst can extract data from an Excel workbook and use it in a PowerPoint presentation in a single conversation), and the arrival of Skills (one-click reusable workflows) in both add-ins.
Pre-loaded Skills cover the most common use cases: formula auditing, building LBO/DCF models, competitive landscape decks, updating presentations with new data, and reviewing investment banking decks. The add-ins are now available via Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry for compliant deployments. Availability: Mac and Windows on paid plans (Pro, Max, Team, Enterprise).
🔗 Claude for Excel and PowerPoint
NVIDIA Nemotron 3 Super, ComfyUI, and GTC 2026
March 11, 2026 — NVIDIA is very active around GTC 2026.
Nemotron 3 Super offers 5x higher throughput for agentic AI compared with the previous generation. It is a 120-billion-parameter open source MoE (Mixture of Experts) model, optimized for high-frequency inference workloads.
NVIDIA and ComfyUI announced at GDC 2026 (Game Developers Conference) an integration that simplifies local AI video generation for game developers and creators, with support for FLUX and LTX-Video models.
The NVIDIA GTC 2026 Live Updates blog gathers live announcements from the conference in San Jose — Mistral AI is also presenting its frontier models there in this context.
🔗 Nemotron 3 Super 🔗 ComfyUI GDC 🔗 GTC 2026 Live
In brief
Runway Labs — Runway launches an internal incubator led by Alejandro Matamala Ortiz (co-founder, chief design officer). Runway Labs will prototype radically new applications for generative video and world simulation models (General World Models) across various sectors: film, healthcare, education, video games, advertising, real estate. Hiring is open.
Claude Code /btw — A new /btw command enables side chain conversations while a task is running, without interrupting the work in progress.
NotebookLM Flashcards — Quiz and flashcard update: resume where you left off, track passed or failed flashcards, ability to delete or shuffle flashcards.
Meta Canopy Height Maps v2 — Meta and the World Resources Institute release CHMv2, a new version of global forest canopy height maps. The DINOv3 model (Meta’s self-supervised vision model) improves accuracy and global coverage. Applications: climate migration, forest restoration, urban planning. Models available in open source.
Z.ai GLM-5 — GLM-5 is now accessible to Lite users (free tier), after being reserved for Pro users since its launch in February 2026.
What this means
March 11, 2026 illustrates two deep trends that are converging.
The first is the platformization of agentics: OpenAI, Perplexity, and GitHub are releasing complementary primitives on the same day (shell tools, sandboxes, agent hooks, code review). The ecosystem is being structured around reusable blocks — skills, containers, sub-agents — that allow developers to build reliable agents without reinventing the infrastructure.
The second is the race for custom silicon: the detail Meta publishes on its four MTIA generations in two years reveals a clear strategy of independence from consumer GPUs, calibrated for large-scale inference. The same logic is pushing NVIDIA to release Nemotron 3 Super on the very day of GTC, where announcements are piling up.
The creation of the Anthropic Institute is part of a more discreet but durable movement: as capabilities progress, major AI companies are structuring their public impact teams — no longer as superficial communications, but as a full-fledged research effort.
Sources
- The Anthropic Institute
- Sydney, 4th Asia-Pacific office
- Claude for Excel and PowerPoint
- @bcherny tweet (/btw)
- Everything is Computer — Perplexity
- Agent API — Perplexity
- Search API — Perplexity
- Sandbox API — Perplexity
- Responses API + computing environment — OpenAI
- 1 year of the Responses API — OpenAI
- Prompt injection defense — OpenAI
- Meta MTIA chips
- Meta Canopy Height Maps v2
- Gemini CLI v0.33.0
- Gemini in Chrome — expansion
- AlphaEvolve — Ramsey numbers
- Gemini Embedding 2
- Copilot Code Review from CLI
- Copilot JetBrains — agentic improvements
- Copilot — repository exploration on the web
- Nemotron 3 Super — NVIDIA
- ComfyUI + NVIDIA at GDC
- GTC 2026 Live Updates — NVIDIA
- Introducing Runway Labs
- NotebookLM Flashcards update
- Z.ai GLM-5 Lite rollout
This document was translated from the fr version into en using the gpt-5.5 model. For more information about the translation process, see https://gitlab.com/jls42/ai-powered-markdown-translator