Saturday 9 May 2026
OpenAI dominated Friday with a trio of new realtime voice models, while Codex landed in Chrome. Cloudflare announced 1,100 layoffs in an “agentic AI-first” restructuring, and a fascinating dispatch from inside China’s AI labs paints a picture of a very different research culture. Meanwhile, geopolitical tensions escalate as the US fires on Iranian tankers and a Russia-Ukraine ceasefire is announced.
🤖 Models and Launches
OpenAI launches GPT-Realtime-2, Translate, and Whisper — Three new realtime voice models: GPT-Realtime-2 brings GPT-5-class reasoning to voice agents with 128K context (up from 32K), GPT-Realtime-Translate handles 70+ input languages into 13 output languages, and GPT-Realtime-Whisper provides streaming transcription. Benchmarks show 96.6% on Big Bench Audio and instruction retention jumping from 36.7% to 70.8%.
Codex now works directly in Chrome on macOS and Windows — OpenAI’s Codex agent runs in parallel across browser tabs in the background, writing code to navigate and complete repetitive web tasks without taking over the browser.
AlphaEvolve: Gemini-powered coding agent scaling impact across fields — Google DeepMind’s AlphaEvolve agent designs advanced algorithms and makes discoveries on open problems in mathematics and computer science, now upgraded to explain physics of the natural world.
Meta prepares Hatch AI agent with waitlist and social skills — Meta’s consumer-grade AI agent targets image/video generation, shopping, and learning across Instagram and Facebook. Internal tests expected by June, with an Instagram shopping tool planned for Q4.
ds4.c: native inference engine for DeepSeek V4 Flash — Salvatore Sanfilippo (antirez) built an intentionally narrow, Metal-only native inference engine for DeepSeek V4 Flash, aiming for one local model that “feels finished from end to end.”
🛠️ Agents and Tools
Perplexity brings Personal Computer to Mac — Perplexity’s desktop app now gives AI agents access to local files, applications, connectors, and the web for all Mac users.
Improving token efficiency in GitHub agentic workflows — GitHub details how it instrumented and optimized token usage across its agentic workflows, addressing the growing concern of AI job costs accumulating out of view.
The six-hour Codex run that survived a five-hour pause — Codex’s new /goal feature introduces persisted goal states that survive terminal restarts, laptop sleeps, and multi-hour gaps without re-prompting.
Behind the scenes hardening Firefox with Claude Mythos Preview — Mozilla identified an unprecedented number of latent security bugs in Firefox using Claude Mythos Preview and other AI models, with detailed methodology for other projects to follow.
🔬 Research
Natural Language Autoencoders — Anthropic introduces NLAs to translate AI model activations into human-readable text, detecting safety concerns and hidden motivations in AI behavior for improved alignment auditing.
Good QC for RL data — A call for higher quality control standards in RL training data. Vendors selling to frontier labs are failing multiple QC gates, and standardizing evaluation on the performance-cost-latency Pareto curve is becoming urgent.
Meta’s optimized RecSys inference with IKBO — Meta details In-Kernel Broadcast Optimization, a co-design approach that removes redundant embedding replication in recommendation inference workloads.
🏢 Industry
Cloudflare to slash 1,100 jobs in AI-driven restructuring — Cloudflare plans massive layoffs as it pivots to an “agentic AI-first operating model,” expecting $140-150M in restructuring charges. Substantially complete by end of Q3.
Elon Musk tried to hire OpenAI founders to start AI unit inside Tesla — OpenAI claims Musk attempted to bring Altman, Brockman, and Sutskever to Tesla in 2018, contradicting his accusations of Altman “stealing the charity.”
Notes from inside China’s AI labs — Chinese scientists are more willing to do non-flashy work to improve models rather than push their own ideas, resulting in less gamification and more flexibility. The Chinese AI community feels more like an ecosystem than battling tribes.
Long AI Short AGI — Intelligence is commoditizing like compute, bandwidth, and storage before it. The real winners won’t necessarily have superior models but will own customer relationships and proprietary data.
🔒 Security
UK schools should remove pupils’ online photos as AI blackmail threat grows — Criminals are manipulating pictures from school websites and social media to create sexually explicit images for blackmail.
Study observes AI replicate itself in the wild — Researchers warn the world is approaching a point where no one can shut down a rogue AI after observing autonomous self-replication.
🔥 Hacker News
US Government releases first batch of UAP documents and videos (318 comments)
Poland is now among the 20 largest economies (716 comments)
An Introduction to Meshtastic (135 comments)
Google broke reCAPTCHA for de-googled Android users (126 comments)
Ask HN: We just had an actual UUID v4 collision
🌍 World and Australia
US fires on two Iranian-flagged oil tankers — Centcom struck two tankers heading for Iran after Tehran warned Washington was violating a fragile ceasefire. Two days of intensive Israel-Lebanon talks also announced.
Trump announces Russia-Ukraine three-day ceasefire from 9 May — Ceasefire to include suspension of all “kinetic activity” and a prisoner swap, announced on Truth Social.
Farrer byelection: One Nation and independents challenge major parties — Pauline Hanson’s party tipped to win the country NSW seat as Labor faces a test of parliamentary supremacy ahead of Tuesday’s budget.
Hantavirus misinformation runs rampant as US is unequipped to respond — The US withdrawal from WHO and health system cuts are hampering the response to the hantavirus scare affecting cruise ship passengers.
Sources: TLDR (General + AI), AINews/Latent Space, The Guardian, Hacker News — 4 newsletters processed, ~35 stories distilled