Pulse (last 3d) · Research 95 · Models 93 · Agents 58 · Industry 40 · Products 29 · Legal 26 · Open Source 24 · Media 22 · Infra 16 · Releases 14

Trending · OpenAI 18 · Anthropic 11 · Claude 10 · Apple 7 · ChatGPT 5 · Bonsai 27B 4 · GPT-5.6 4 · Google 4 · Prime Intellect 4 · Claude Code 3 · GPT-5.6 Sol 3 · Sam Altman 3

Daily AI Brief, July 15, 2026

Top story

Last 30 Days: The AI Scientists. Four autonomous discovery systems cleared peer review in four months, yet every one was already a year old; the piece highlights the reliability research that's being overlooked. Source

Tools

PrismML Bonsai 27B. First 27B-class multimodal model claimed to run on a phone, based on Qwen3.6-27B with a compression method that retains ~90% capability under 10GB. Source

ExLlamaV3 v1.0.0. Major release delivering significant performance upgrades for local LLM inference. Source

Agnost AI (YC S26). New launch that extracts user feedback from agent conversation logs. Source

FrontierFinance Benchmark. New evaluation suite for frontier models with initial leaderboard including GPT 5.6 Sol. Source

Research

Blind-Spots-Bench. New HuggingFace benchmark designed to probe blind spots in multimodal model behavior. Source

llama.cpp SYCL/Intel Updates. Recent PRs add Flash Attention via oneDNN on Xe2 and fix ops, yielding notable Qwen prefill speedups. Source

Industry

DeepSeek Nears $500M ARR, Eyes IPO. Report positions DeepSeek alongside OpenAI and Anthropic as the next potential AI IPO at a $71B valuation. Source

Meta Muse Spark 1.1 vs. OpenAI/Anthropic. Analysis argues Spark 1.1 wins on API price but loses on consumer subscription positioning vs. Claude and ChatGPT. Source

Community

Closing-Line Edge Leakage Question. Backtest shows edge vs. closing lines but inference used incomplete line-movement features; classic leakage debate. Source

Claude Memory "Heist". HN discussion on extracting sensitive user data from Claude via memory features, mostly social engineering rather than a fundamental breach. Source