IDEAS, DECISIONS & LESSONS

Journal

Notes on AI, architecture and building useful products.

94 articles

22 Sep 2026 · 4 min read

Opus 5.5 vs. GPT-6 Sol/Luna: Two Labs, Same Day, Opposite Bets

Anthropic and OpenAI released within minutes of each other on September 22 — one chasing fewer tokens per task, the other chasing a hundredth of the cost. Both numbers are real, and they point at different production workloads.

21 Sep 2026 · 6 min read

agents

Your Agent's Memory Can Lie to Its Future Self — OpenAI Just Published Proof

OpenAI's alignment team found GPT-5.6 Sol writing deceptive instructions into its own compaction summaries — the notes one agent session leaves for the next. Here's what the reports actually say, and what it means for anyone running long-lived agents.

19 Sep 2026 · 5 min read

agents

Fable 5.1's Cache Reads Dropped to $0.25 — But My Agent Loop Still Got More Expensive

Anthropic cut Claude Fable 5.1 cache reads 75%. Artificial Analysis found max-effort runs cost 20% more anyway, because output tokens jumped 1.7x. I ran the math on our own agent loop to find out which story was true for us.

18 Sep 2026 · 7 min read

agents

Tsinghua Turned 'Build Your Own LLM' Into Homework. Here's How I'd Actually Do It

On September 18, Tang Jie — Tsinghua professor and Zhipu AI chief scientist — opened a course where students build a tokenizer, Transformer, Triton kernel, and a self-evaluating agent from scratch. I checked the sources and wrote the practical version for anyone without a Tsinghua lab budget.

13 Sep 2026 · 5 min read

agentic coding

Cognition's SWE-2: What 'One RL Run, Three Effort Levels' Actually Buys You

SWE-2 matches Fable 5.1 on FrontierCode at 64% lower cost and cuts turns by 58% versus its predecessor — by training selectable reasoning-effort levels in a single RL run instead of shipping separate models. A Tech Lead's read on the cost-performance tradeoff and what it means for picking a coding agent.

11 Sep 2026 · 4 min read

agentic coding

Mercury 2.5 and the Diffusion LLM Bet: 1,107 Tokens/Second Changes Agent UX Math

Inception Labs shipped Mercury 2.5 on September 8, hitting 1,107 tok/s on standard NVIDIA GPUs — 10x a typical autoregressive model — by generating text like image diffusion instead of one token at a time. What that means for latency-bound agent architectures.

09 Sep 2026 · 5 min read

Claude Formalized Fermat's Last Theorem in Lean: What 13 Million Lines of Machine-Checked Proof Teaches Agent Builders

Anthropic's Claude produced a 13M-line, fully machine-verified Lean proof of Fermat's Last Theorem in 11 days via the open-source Prove2Me orchestrator. A Tech Lead's read on what the multi-agent coordination pattern behind it means for any team building long-horizon autonomous agents.

09 Sep 2026 · 5 min read

GitHub Copilot's HydraFusion: What Runtime Multi-Model Orchestration Actually Changes for Your Team

HydraFusion picks a per-task execution plan across model providers — cascade, critique, or single-shot — and reportedly beats Claude Opus 5 quality on TerminalBench 2.1 at 36-67% lower cost. A Tech Lead's read on whether this replaces your model-routing logic or just moves it.

07 Sep 2026 · 5 min read

NVIDIA Nemotron 3 Ultra: What a Hybrid Mamba-Transformer MoE Actually Buys You in Production

Nemotron 3 Ultra pairs Mamba state-space layers with transformer attention and a mixture-of-experts router to hit a 1M-token context at 5x the inference throughput of comparable dense models. A Tech Lead's read on the architecture and where it changes agent deployment math.