April 2026
- Research Kronos — open foundation model for financial markets — 12B K-lines from 45 exchanges, +93% RankIC over best time-series baseline, AAAI 2026 accepted impact 7.8
- Research Google Research's Simula generates 512K synthetic training samples — mechanism-design framework yields 10% math-reasoning gain with Gemma-3 4B student impact 8
- AI Archon — first open-source harness builder for AI coding — hits 18.8k stars, ships 17 default workflows, v0.3.6 released April 12 impact 7.2
- AI OpenAI Agents Python SDK hits 22.6k stars — 1351 commits, 268 contributors, 100+ LLM providers supported impact 7.2
- AI Meta Superintelligence Labs debuts Muse Spark — 58% on Humanity's Last Exam in Contemplating mode, 10x less compute than Llama 4 Maverick impact 8.7
- Research Anthropic's Automated Alignment Researchers: 9 Opus 4.6 copies hit 0.94 PGR on math alignment, 0.47 on coding impact 8.2
- AI Anthropic-Trump relationship thaws — Treasury + Chief of Staff meet Amodei, 'every agency except DoD wants Anthropic' impact 7.7
- AI Google ships Gemini 3.1 Flash Live — 90.8% on ComplexFuncBench Audio, 2x longer conversation context impact 7.4
- AI OpenAI ships GPT-5.4-Cyber — cyber-permissive model that reverse-engineers binaries, gated to vetted defenders impact 8
- Startups Cursor in talks to raise $2B+ at a $50B valuation — $6B ARR forecast, Anthropic-subsidy tailwind impact 8
- Research Berkeley SPEX: GPT-4o mini fails 92% of trolley problems — replacing 4 words reduces failure to near zero impact 7.6
- Research Vidoc reproduces Anthropic Mythos vulnerability-finding with GPT-5.4 + Opus 4.6 — 3/3 on FreeBSD and Botan impact 8
- Research DeepMind's Gemini Robotics-ER 1.6: 93% instrument-reading accuracy, up from 23% — Boston Dynamics Spot ships with it impact 8.5
- AI xAI opens Grok 3 and Grok 3-mini via API — $0.50/1M output tokens, 1/7th the price of Gemini 2.5 Flash thinking impact 7.4
- AI OpenAI sheds side quests — Sora shut down after losing $1M a day, Weil / Peebles / Narayanan out impact 7.6
- Startups OpenAI commits $20B to Cerebras over 3 years — up from $10B, equity warrants for up to 10% stake impact 8.4
- Research Claude 4.7's new tokenizer inflates text 1.325x on average — 80-turn Code sessions now cost 20-30% more impact 7.3
- Research OpenBMB releases VoxCPM2 — 2B-param tokenizer-free TTS, 1.84% WER in English, 30 languages, Apache 2.0 impact 7.7
- AI Addy Osmani's agent-skills hits 17K stars — 20 engineering skills, 7 slash commands, 3 personas for AI coding impact 7.5
- AI Perplexity ships Personal Computer for Mac — $200/mo, orchestrates 19 AI models, runs on a $599 Mac mini impact 7.8
- Startups Converge Bio raises $25M Series A — 4.5x protein yields, Bessemer + Meta/OpenAI/Wiz execs back it impact 7.1
- Research NVIDIA's Nemotron OCR v2: 28x faster than PaddleOCR, trained on 12.2M synthetic images, 34.7 pages/sec impact 7.6
- Research MirrorCode: Claude Opus 4.6 reimplemented a 16000-line Go bioinformatics toolkit that would take humans 2-17 weeks impact 7.9
- AI Mistral opens Studio to MCPs — 20+ enterprise connectors (Databricks, Snowflake, Stripe, Zapier) and custom servers impact 7.3
- AI Google ships Gemini 3.1 Flash TTS — 70+ languages, Elo 1211 on Artificial Analysis leaderboard impact 7.6
- AI Anthropic ships Claude Design — Brilliant cut 20 prompts to 2, Figma drops 4.26% impact 8.3
- Research Tencent Robotics X releases HY-Embodied-0.5: 2B + 32B open-source embodied AI foundation models, leads 16 of 22 benchmarks impact 7.76
- AI Stanford HAI 2026 Index: China narrowed Arena score gap from 1300 pts to 39 — and AI talent flow to US fell 89% impact 7.85
- Research SkillClaw: multi-user LLM agent ecosystems where skills evolve across all users — 276 HuggingFace upvotes in one week impact 7.29
- AI NousResearch hermes-agent hits 96k GitHub stars — open-source self-improving agent with built-in learning loop impact 7.11
- Research Nature: LLMs transmit behavioural traits through hidden signals embedded in training data impact 7.31
- AI White House moves to deploy Anthropic Mythos across US federal agencies impact 7.61
- AI OpenAI launches GPT-Rosalind — specialized frontier AI for life sciences research impact 7.17
- AI OpenAI expands Codex to 'almost everything' — 874 HN pts, 449 comments on agentic coding agent upgrade impact 7.54
- AI €54,000 billing shock in 13 hours: unrestricted Firebase browser key drained by Gemini API abuse impact 6.9
- AI Alibaba open-sources Qwen3.6-35B-A3B: 35B MoE, 3B active params, runs on a laptop, outdraws Claude Opus 4.7 impact 7.28
- AI Anthropic releases Claude Opus 4.7 — most capable Opus yet, 1,752 HN pts, extended thinking + higher rate limits impact 7.96
- Research Squishy photonic switches: IEEE Spectrum reports fast, low-power soft photonic logic impact 6.95
- AI Claude for Word enters beta with cross-document context across Excel and PowerPoint impact 6.93
- Research Quanta Magazine: The AI revolution in math has arrived impact 7.49
- Research "Neural Computer": video-generation architecture trains a world model of a real computer impact 6.82
- AI Taking on CUDA with ROCm: 'One Step After Another' impact 7.15
- AI Anthropic brings "advisor strategy" to Claude Platform: Opus advises Sonnet/Haiku at inference impact 7.06
- Research Biological Computing Company: living neurons power new AI chips and algorithms impact 6.8
- Cybersecurity Anthropic unveils Project Glasswing — Claude Mythos already found "thousands" of zero-days in major software impact 8.5
- AI Anthropic ships Claude Managed Agents — production agents without the infra work impact 7.8
- Cybersecurity Hacker uses Claude and ChatGPT as assistant-in-the-loop to breach multiple government agencies impact 7.8
- AI Gemma 4 crosses 10M downloads in one week; Gemma family at 500M total impact 7.4
- Top Tweets Karpathy: "a growing gap in understanding AI capability" — 19.4K likes impact 7.5
- AI Karpathy's nanochat hits 51.7K stars — ChatGPT clone trainable end-to-end for $100 impact 8.2
- Top Tweets Karpathy's LLM-coding pitfalls compiled into viral CLAUDE.md — #2 on GitHub weekly impact 7.1
- AI Google ships Gemini 3.1 Ultra — 2M tokens, native multimodal, sandboxed code impact 8.8
- Startups OpenAI closes $122B at $852B — most valuable private company in history impact 7.5
- Research DeepMind's TurboQuant: 6.2× KV-cache compression, no perplexity loss impact 7.8
- AI MCP crosses 97M installs; Linux Foundation takes governance at KubeCon impact 8.6
- Research Physics-informed transformer: 34% better RMSE, 12× faster than PINN baselines impact 7.6
- AI Meta's Muse Spark: first flagship since the $14.3B Scale AI deal impact 8.2
- Startups Q1 2026 is the biggest venture quarter ever: $300B, 80% of it AI impact 7.8
- Quantum AI sparks a quantum breakthrough — 'the world is not ready' impact 7.1
- Top Tweets Karpathy's personal-LLM-plus-Obsidian thread hits 18K likes impact 6.7
- Research Tufts neurosymbolic model: 100× less energy, 7 pts better on reasoning impact 8.3
- Cybersecurity Claude Code source leaks; 140 fake repos seed Vidar within hours impact 8.3
- AI OpenAI ships GPT-5.4 — 75% on OSWorld-V, above the 72.4% human baseline impact 9