Archive
Every issue of the AlphaSignal newsletter, newest first.
649 issues total
May 2026
28 issues- May 31 Vector search can't find error codes or file paths. Your coding agent needs a terminal, not embeddings.
- May 29 One user? Great. Fifty users? The architecture breaks. Viktor is what enterprise AI actually needs under the hood.
- May 29 Anthropic ships Claude Opus 4.8, raises $65B at $965B valuation, LMMs Lab beats Qwen3-VL by 44 points, Roboflow runs detection on CPU via 30MB weights
- May 28 Cognition raises $1B at $26B, Google DeepMind ships universal embedding model, Qwen3 8B runs agent on old GPU, KAIST optimizer trains 1.5x faster
- May 27 Bearly ships OpenADE, Qwen releases 1M context model, Datacurve drops DeepSWE, Figure robots hit retail, DeepSeek infra papers praised
- May 26 xAI opens Grok Build beta, Perplexity open-sources AI security tool, Pliny jailbreaks Qwen to 4% refusal, coding agents lose 30% on real databases
- May 25 Karpathy's Claude config hits 94% accuracy, DeepMind solves 9 ErdΕs problems, DHH ships 30k GPT-5.5 lines, Qwopus 3.6 27B releases
- May 24 Agentic loops on frontier models drain budgets fast. Specialized models are changing the math.
- May 22 Meta releases SAM3 for video tracking, NVIDIA ships verified agent skills, KAIST 10M model hits 97% Sudoku, Google launches managed agents on Gemini
- May 21 Hallmark stops generic AI UIs, Anthropic ships self-hosted sandboxes, ColBERT retrieval speeds up 3-5x, Karpathy drops free neural net course
- May 20 Google drops Gemini Omni video model, GitHub Spec Kit hits 95K stars, grep beats vector search, ByteDance ships 3B multimodal model
- May 19 Anthropic ships MCP tunnels, Microsoft releases 4B image-to-3D model, Unsloth ships Qwen3 GGUFs at 2x speed, new Transformer rival cuts parameters 30%
- May 19 Cursor ships Composer 2.5, Anthropic doubles token limits, OpenAI Codex gets 10-50x faster Git ops, Harvard red-teams live agents
- May 18 Claude Code gets persistent memory, NVIDIA releases Sana diffusion model, Qwen3.6-35B ships free GGUF, HKUDS turns apps into agent CLIs
- May 17 What if AI reasoned without words? New recursive models do exactly that and it's 100x faster.
- May 15 Musk open-sources X ranking algo, xAI drops Grok Build CLI, Nous Research cuts LLM pretraining 2-3x, Alibaba shrinks Qwen 80B to 23B
- May 14 WiFi router becomes human radar, Anthropic ships free legal tool, DeepSeek runs locally at 27x less energy, AI finds 18-year-old NGINX RCE bug
- May 13 Thinking Machines ships real-time AI, open-source video model runs locally, Rust browser loads in 85ms, Meta halves LLM decoding
- May 12 Anthropic ships Claude Code sessions view, OpenAI launches Daybreak security tool, ByteDance releases 7B GUI model
- May 11 Anthropic cuts Claude blackmail 3x, NVIDIA sparsity boosts H100s 20%, GitHub spec-kit hits 92k stars, CloakHQ ships stealth browser
- May 10 The next shift isn't a better model, it's agents that build their own infrastructure.
- May 8 Anthropic ships Claude to Office apps, Google launches on-device Chrome AI, Zyphra releases 8B reasoning model, Qwen 3.6 adds multi-token prediction
- May 7 Anthropic partners SpaceX for 300MW compute, Claude adds self-learning memory, Stanford theory speeds training 5x, MiniMax M2.7 shrinks to 74GB
- May 6 OpenAI defaults GPT-5.5 Instant, xAI ships Grok 4.3 at $1.25/M tokens, Anthropic cuts unsafe behavior 54% to 7%, Google speeds Gemma 4 3x
- May 5 Voice-Pro dubs YouTube into 100+ languages locally, Ouroboros verifies AI workflows, Perplexity enters Teams, DeepMind drops RL textbook
- May 4 xAI adds voice cloning API, dev burns transformer to FPGA at 50k tokens/sec, Scrapling runs 784x faster than BeautifulSoup
- May 3 Multi-agent systems can amplify errors by 17x. Before you scale up, read why single agents usually win.
- May 1 Anthropic launches Claude Security, Cursor releases agent SDK, Ant Group open-sources 1T-param model, agent optimizer boosts AppWorld score to 89.5