Tuesday, August 18, 2026

Pulse (last 3d) · Research 58 · Agents 49 · Products 45 · Models 21 · Open Source 19 · Industry 18 · Legal 17 · Policy 8 · Releases 7 · Inference 7

Trending · Claude 15 · Anthropic 9 · Claude Code 8 · ChatGPT 7 · Cursor 7 · OpenAI 7 · Codex 5 · llama.cpp 4 · volcengine 4 · Google 3 · Nvidia 3 · verl 3

---

Improving the matrix multiplication exponent with modern optimization and AlphaEvolve - Evolutionary search via AlphaEvolve tightens matrix multiplication exponent constants. Paper

DumpsterCluster: From Dumpster Diving to Serving LLaMA-70B on $60 GPUs - Oxford/Shumailov team empirically tests whether clusters of retired sub-$100 GPUs can realistically serve LLaMA-70B inference. Paper

ConceptCLIP, concept-supervised vision-language FM with built-in explanations - Concept supervision at CLIP pre-training time delivers zero-shot diagnostic gains across 78 datasets plus clinician-trusted explanations. Nature BME

Policy Iteration with Human Feedback: Bringing Post-Training RL to In-context Learning - Reformulates policy iteration RLHF as in-context learning so post-training updates happen without weight changes. arXiv

NeuronScope - CLI and MCP server for mechanistic interpretability on open-weight models, exposing attention-head and neuron-level drivers for AI agents. Product Hunt

Unmark - Free browser tool that strips invisible Unicode characters from AI text and rewrites wording to disrupt SynthID-style statistical watermarks. Product Hunt

Paitify - MCP-native payment policy engine that lets AI agents get instant approve/deny decisions with a human-in-the-loop audit trail. Product Hunt

Context for Claude - Continuous memory layer that lets Claude agents see everything you see across sessions. Product Hunt

Stripe will reportedly acquire AI gateway startup OpenRouter for $7B+ - Stripe reportedly in talks to acquire LLM routing startup OpenRouter for $7B+. Reddit

Anthropic's annualized revenue surges to $65B - TechCrunch reports Anthropic's annualized revenue has surged to $65B. TechCrunch

Anthropic explains how Claude's invisible text watermarks will work - The Verge covers Anthropic's explanation of how Claude's SynthID invisible text watermarking will work. The Verge

Groq raises $350M to fuel its pivot from AI chips to neocloud - Groq raises $350M as it pivots from selling AI chips to operating as a neocloud. TechCrunch

Teaching Everyone to Fish for Tokens - Nathan Lambert reframes alignment/RLHF progress in terms of who extracts tokens from whom. Interconnects

AI;DR (AI; Didn't Read) - Discussion of an 'AI;DR' browser extension and the broader problem of AI-generated prose polluting codebases and PRs. Rickmanelius

After pushing 1M+ tokens through Qwen 3.8 27B, here is my optimal llama.cpp config for 16GB VRAM - Detailed llama.cpp configuration tuning Qwen 3.8 27B to 73k context on 16GB VRAM for agentic coding workloads. Reddit

llama.cpp adaptive MTP PR#27210 - llama.cpp PR #27210 introduces adaptive multi-token prediction for speculative decoding. Reddit