Skip to content
Source · Daily Brief

AI Daily Brief — 29 April 2025

The biggest single Meta day of 2025. LlamaCon shipped an API, a standalone consumer app, four safety models, two ultra-fast inference partnerships (2,648 tok/s on Cerebras), an expanded enterprise stack, and a Zuckerberg–Nadella closing fireside. Meanwhile OpenAI published the candid GPT-4o sycophancy postmortem, and Mastercard ushered agentic commerce into payments with Agent Pay.

Top stories

  • Meta opens LlamaCon with Llama API limited preview. First-ever generative-AI developer conference in Menlo Park. Llama API launches as a limited free preview behind a waitlist — one-click API keys, interactive playgrounds, OpenAI-SDK-compatible Python/TypeScript SDKs for Llama 4 Scout and Maverick. Served from Meta’s own infrastructure. via TechCrunch
  • Standalone Meta AI app built on Llama 4. Designed around voice conversations and personalized by years of Facebook/Instagram profile data, with a Discover feed for sharing prompts. Companion app for Meta’s AI glasses; syncs with meta.ai. Direct ChatGPT competitor. via Meta
  • Meta + Cerebras: Llama API at 2,648 tok/s — 18x faster than GPU baselines. Cerebras partnership powers the Llama API at up to 2,648 tokens/sec on Llama 4 Scout — vs SambaNova 747, Groq 600, ChatGPT ~130, DeepSeek ~25. Meta’s commercial push into inference-as-a-service. via Cerebras
  • Meta + Groq: Llama 4 API on Groq LPU at 625 tok/s. Parallel partnership powers the official Llama API on Groq LPU chips with early benchmarks at up to 625 tokens/sec for Llama 4 in preview. Three-line migration for teams already calling OpenAI endpoints. via Groq
  • Meta ships Llama Guard 4 (12B) and Llama Prompt Guard 2 (86M / 22M). Llama Guard 4 is a natively multimodal 12B safeguard model using early-fusion transformer architecture — drop-in replacement for Llama Guard 3 across Llama 3 and Llama 4 pipelines. Two Prompt Guard 2 classifiers (86M and 22M) for prompt-injection and jailbreak detection; the 22M cuts latency/compute 75% vs the 86M.
  • LlamaFirewall — open-source agent guardrail framework. Three guardrails: PromptGuard 2 (realtime injection detection), Agent Alignment Checks (inspect agent reasoning for goal hijacking), and CodeShield (online static analysis on generated code). Free for projects up to 700M MAU. via Hacker News
  • OpenAI publishes GPT-4o sycophancy postmortem. Root cause: new user-feedback reward signals overpowered safeguards; sycophancy wasn’t tracked in deployment evals. Update rolled back 100% for free users; OpenAI committed to explicit anti-sycophancy training and evals. via OpenAI
  • Mastercard unveils Agent Pay with Microsoft, IBM and Braintree. Agentic-commerce payments framework letting verified AI agents transact on a user’s behalf using Agentic Tokens that bind a tokenized card credential to a specific agent, merchant scope and consent policy — so ChatGPT, Copilot or other agents can complete checkout without holding raw card numbers. via Mastercard

Who shipped

Meta ran the day end-to-end across product, safety and infra. OpenAI shipped the year’s most candid model-behavior postmortem. Mastercard moved agentic commerce into payments. Cerebras and Groq got prime distribution stages.

Open-source pulse

LlamaFirewall + Llama Guard 4 + Prompt Guard 2 + CyberSecEval 4 (with CrowdStrike-co-built CyberSOCEval, AutoPatchBench, Autonomous Offensive Cyber Operations tests) all ship open — making Meta the day’s largest single open-source AI release across model + safety + infra. Llama Stack expanded with NVIDIA NeMo microservices and new partnerships with IBM, Red Hat and Dell.

Money, infra & hardware

Meta named 10 international Llama Impact Grant recipients splitting $1.5M+: E.E.R.S. (US civic chatbot), Doses AI (UK pharmacy error detection), Solo Tech (offline AI for rural US), FoondaMate (multilingual study tool for African students), among others. Snap reported Q1 revenue $1.36B (+14% YoY), 900M MAU milestone, but declined forward guidance citing macro uncertainty — stock dropped 13% after-hours.

Quiet corners

Anthropic, Google DeepMind and xAI quiet on the model side. DeepSeek Prover-V2 was one day out.

By the numbers

  • 2,648 / 625 / 130 / 25 tok/s — Cerebras / Groq / ChatGPT / DeepSeek
  • 12B — Llama Guard 4 parameters
  • 86M / 22M — Prompt Guard 2 variants; 75% latency cut
  • 10 — Llama Impact Grant recipients; $1.5M+ total
  • 3 — LlamaFirewall guardrails (PromptGuard 2, Alignment Checks, CodeShield)
  • $1.36B / +14% YoY / 900M MAU — Snap Q1
  • Most-mentioned company: Meta

Compiled by AI Feed’s editor from verified web sources for 29 April 2025.