AI · Weekly · May 13, 2026

frontier reasoning just jumped, but not where AGI discourse is staring

AI co‑mathematician systems from Google DeepMind hit 48% on FrontierMath Tier 4, the hardest slice of that benchmark, outperforming all other AI systems tested.[2]

Fields Medalist Timothy Gowers reports GPT‑5.5 Pro produced PhD‑level math research in under two hours and also used GPT‑5.5 to find fatal errors in FrontierMath problems themselves.[3]

OpenAI’s o1 model reached 67% accuracy on ER patient diagnoses, beating physicians on the same cases.[4] Meanwhile, Alan’s AGI countdown has sat at 97% for five months and communities argue current systems are only making incremental progress toward AGI, with definitions of AGI itself still hotly disputed.[5]

The net picture is narrow superhuman reasoning in math and medicine coexisting with skepticism that simple scaling or current agents (which still lack continuity and social awareness) are enough for AGI.[6]

security: LLMs are now both the red team and the worm

Claude Mythos has identified 271 vulnerabilities with low false positives, and Mozilla credits it with helping Firefox fix more bugs in April 2026 than in the prior 15 months, with the US President citing it in AI safety testing discussions.[7]

At the same time, some users say GPT‑5.5 found more critical bugs than Mythos on specific targets, undercutting the narrative that only specialist models matter.[8]

Offense is scaling too: cybercriminals reportedly used AI to find and exploit a zero‑day for the first time, and the Mini Shai‑Hulud npm worm compromised 84 TanStack packages and ~160 npm projects to exfiltrate GitHub and SSH credentials during install.[9]

PyPI is seeing similar supply‑chain attacks with malware quarantined only hours after upload, while OpenClaw and ClawHub were seeded with over 575 malicious skills and thousands of Lovable‑style AI‑generated apps leaked secrets through basic vulnerabilities.[10]

Gray‑market Claude APIs sold at 90% discounts using stolen credentials round out an ecosystem where the same LLM tooling that finds bugs is also embedded in the exploit chain.[11]

throughput hacks are quietly redefining ‘frontier’

DFlash speculative decoding is posting up to 8.5× speedups while matching baseline accuracy, with Gemma‑4‑26B reported at ~600 tokens per second on a single RTX 5090.[12]

Multi‑Token Prediction drafts boost Gemma 4 inference by ~40% and let Qwen 3.6‑27B run about 2.5× faster with contexts up to 262k tokens, hitting 80–135 tok/s on commodity GPUs in llama.cpp.[13]

Blackwell‑class NVFP4 quantization pushes 270 tok/s in the lab while DeepSeek’s FP4‑aware training halves QK selector cost at 99.7% recall, and vLLM shows ~7× token throughput when scaling across multiple B200 8‑GPU machines.[14]

The catch is that DFlash gains fall off beyond ~20k context, MTP often spikes VRAM usage, and some NVFP4 configs plateau at ~50 tok/s on older GPUs, making these wins sharply task‑ and hardware‑dependent.[15]

Local open‑weight workflows are increasingly "good enough" on top of these tricks, so a mid‑sized model plus aggressive decoding/quantization can feel more frontier than a slower giant model over API.[16]

agents: hermes shows the value is moving into orchestration

Hermes Agent has become the most used AI app on OpenRouter, processing about 271B tokens in a day, overtaking OpenClaw and Claude Code and backed by nearly 1,000 contributors.[17]

It runs both against cloud APIs and local GGUF/MLX models, while Gemini‑powered agents like AlphaEvolve are being deployed across genomics and power‑grid optimization, and n8n‑based flows now chain prompts→images→videos→music into full YouTube pipelines.[18]

Underneath, MCP is solidifying as the tool substrate, standardizing how agents see capabilities and providing a security boundary around keys and endpoints, while LangGraph 1.2 adds delta channels and checkpointing so long‑running agents can roll back state.[19]

The practical experience is still rough: builders report debugging memory and state taking more time than prompt design, silent loops blowing budgets, and bloated contexts degrading performance.[20]

OpenClaw’s trajectory—145k stars, then 575+ malicious skills, operational fragility, and a rapid slide in usage as Hermes and Claude Code eclipse it—shows how fast a "general agent IDE" can become a liability surface.[21]

price/perf gravity is drifting toward china and your GPU

Chinese models are now dictating price/perf: DeepSeek V4 Flash is ~90% cheaper than GPT‑5.4 Mini and ~70% cheaper than Gemini 3.1 Flash Lite, and Kimi K2.6 offers a 1T‑parameter MoE at ~$700/month vs GPT‑5.5 at ~$5,500 and Opus 4.7 at ~$4,500.[22]

Kimi quickly took #1 on OpenRouter’s programming leaderboard while Tencent’s Hy3 preview processed 3.66T tokens in a week with ~298% growth, suggesting real usage is following those economics despite Kimi’s reported weakness on hard math/coding relative to Claude and Opus.[23]

At the same time, Qwen 3.6‑27B and Gemma 4 are hitting 2.5× speedups with MTP and DFlash on single RTX 30‑ and 50‑series GPUs, and MiniCPM‑V4.6 now beats Qwen3.5‑0.8B on edge hardware, with local model capability on laptops improving faster than Moore’s Law.[24]

This plays out against a backdrop where AI data centers consume city‑scale power, 69 US jurisdictions have banned new ones, and Maryland residents face a $2B bill to upgrade grids for out‑of‑state AI clusters.[25]

With Starlink‑style mega‑clusters and Colossus‑class 220k‑GPU farms on one side and hyperscale‑class local throughput on a single RTX 5090 plus on‑device SD on Android on the other, compute is polarizing between a few giant clouds and surprisingly capable edge setups.[26]