Meno's paradox, measured: Kimi K3's nervousness index
Minu ChoiMinu Choi
A Malayalam grammar puzzle only one document on earth contains, run through Opus 4.8 and Kimi K3 — and what their traces reveal about search, memory, and trust.Writing
The frontier is open-source today
Hrishi OlickelHrishi Olickel
A detailed analysis and comparison of GLM-5.2 and Opus 4.8, both pushing forward state-of-the-art in timestamp-accurate diarization.Synthetic Hires
Hrishi OlickelHrishi Olickel
Synthetic Hires: High context agents in the critical path - the extended version of Hrishi's keynote at SuperAI 2026 Singapore. No Country for Old Code: Ship outcomes not tools
Hrishi OlickelHrishi Olickel
Extended version, slides and interactive elements from Hrishi's talk at AI Engineer Singapore. Codons, sentinels, budgets, and the case for shipping outcomes instead of tools.How Not to Leave Greenfield
Hrishi OlickelHrishi Olickel
An impromptu talk at SQ Collective.Agentic Programming Patterns: 1
Hrishi OlickelHrishi Olickel
GAN-inspired loops, automated review surfaces, and continuous improvement curves — patterns we've discovered building long-running agents with Hankweave.RISC won: Building towards Data AGI
Hrishi OlickelHrishi Olickel
A deeply technical retrospective on the first year of Southbridge.Systems of Lasting Value
Hrishi OlickelHrishi Olickel
Why does all code become legacy code, and what can we do about it in the age of AI?CCEPL-driven development
Hrishi OlickelHrishi Olickel
Coding agents are the new REPL: A place to test, but not to deploy. Here's the pattern of what we see that comes after.Antibrittle Agents
Hrishi OlickelHrishi Olickel
The bible on how to build reliable long-horizon AI agents that can work for hours or days without breaking. A guide to the architectural principles and practices that make agents antibrittle.Tap the Bed
Minu ChoiMinu Choi
What a Bambu 3D printer's bed-tapping ritual teaches us about building reliable AI agents - and why Hankweave forces agents to forget.We are so back
Hrishi OlickelHrishi Olickel
Agentic systems today feel like raw LLMs did two years ago - unpredictable, and sparking with the same AGI static. The agent (LLM + tools + loop) has become the new unit of computation, and it has dragged back every problem we thought we'd solved.Working with AI as a team
Hrishi OlickelHrishi Olickel
Focusing on intent: how to keep AI from eroding the trust collaboration runs on.Hiring in 2025
Hrishi OlickelHrishi Olickel
Open-sourcing our take-homes and our hiring process, in the face of an increasingly agent-enabled world.Conducting smarter intelligences than me: new orchestras
Hrishi OlickelHrishi Olickel
Everything we learned painstakingly orchestrating subagents made of every flagship model, manually reviewing their output, trying different methods of coordination, and figuring out what works.Everything I'll forget about Evals
Hrishi OlickelHrishi Olickel
The actually-actionable guide to AI evals that nobody else got around to writing, built from months of shipping real datasets and benchmarks. The puzzle at its center: how do you get a dumber intelligence to verify the output of a smarter one?Auto-generating Insanely Good Reports
Hrishi OlickelHrishi Olickel
A repeatable report generation pipeline.Something Weird is Happening in AI
Hrishi OlickelHrishi Olickel
Comparing models across small and big.Southbridge: Scalable data substrates
Hrishi OlickelHrishi Olickel
The fundamental thesis underpinning Southbridge.Making use of the rest of our data
Hrishi OlickelHrishi Olickel
Semistructured data has been the bane of my life - at my last company, half of all engineering hours went to wrestling it into shape. Why it is so stubborn, and how we started to fix it.A Comparative Analysis of Three AI Agent Approaches to a Complex Bug Fix
Hrishi Olickel & Claude Opus 4Hrishi Olickel & Claude Opus 4
Three AI agents, one gnarly state-management bug in gemini-cli, three very different routes to a fix. Comparing Claude Code with and without subagents against Windsurf's Cascade, move by move.The Ghost in the Layout: A Debugging Story of Reactivity and Renders
Hrishi Olickel & Claude Opus 4Hrishi Olickel & Claude Opus 4
A React layout bug that vanished the moment you opened dev tools - the worst kind of ghost. A debugging story about reactivity, render timing, and chasing a non-deterministic gremlin back to its source.Mem0: Technical Analysis Report
Hrishi Olickel & Claude Opus 4Hrishi Olickel & Claude Opus 4
An AI-generated teardown of Mem0, the memory layer for LLMs: its architecture, storage systems, and the novel pieces that make it work. Produced with the unsupervised report-writing workflow from 'new orchestras.'Claude Code: An analysis
Hrishi Olickel & Claude Opus 4 & Gemini 2.5 ProHrishi Olickel & Claude Opus 4 & Gemini 2.5 Pro
A deep, AI-generated dissection of Claude Code's internals: dependencies, data structures, control flow, tools, and file editing. Generated by Opus 4 with help from most of the flagship models; the human write-up on how it was made lives alongside it.Latent space reasoning and the inner monologue
Hrishi Olickel & Gemini 2.0 Pro & OpenAI o1-proHrishi Olickel & Gemini 2.0 Pro & OpenAI o1-pro
What can the way LLMs reason in latent space teach us about our own inner speech? A wander through two papers on recurrent latent reasoning and the parallels to how minds talk to themselves.A review of editing with LLMs
Hrishi Olickel & Claude 3.5 SonnetHrishi Olickel & Claude 3.5 Sonnet
A survey of how LLM systems actually edit code - Tabby, Claude, Aider, and the rest - and the one problem they are all really solving. How do you get a model to change exactly the right lines, reliably?Understanding Code Assistance in Zed
Hrishi Olickel & Claude 3.5 SonnetHrishi Olickel & Claude 3.5 Sonnet
How the Zed editor wires up AI code assistance under the hood - extraction, prompts, structured output, tool calling, and streaming. Written in Rust, explained with TypeScript analogies.Flash vs O1pro Long Context Understanding
Hrishi Olickel & Gemini Flash 2.0 & OpenAI o1-proHrishi Olickel & Gemini Flash 2.0 & OpenAI o1-pro
Gemini 2.0 Flash versus o1-pro on long-context comprehension, using Google's MMoE paper as the exam. The surprise is which model actually explains a complex architecture better.Better OCR with logprobs
Hrishi OlickelHrishi Olickel
Can a model's own confidence scores tell you where its OCR went wrong? Testing log-probabilities as an error detector for multimodal OCR on complex tables, with Gemini Flash and GPT-4o.Approaches for a priori estimation in holey bread
Hrishi Olickel & AI Cameos (Claude 3.5 Sonnet, GPT-4o, Gemini exp-1121, QwQ, OpenAI o1-preview)Hrishi Olickel & AI Cameos (Claude 3.5 Sonnet, GPT-4o, Gemini exp-1121, QwQ, OpenAI o1-preview)
Can a model estimate the calories and macros of a fancy holey bagel from a single photo? A bake-off across Claude Sonnet, GPT-4o, Gemini, and the reasoning models QwQ and o1-preview.Comparative Analysis of Three Songs
Hrishi Olickel & Gemini 1.5 ProHrishi Olickel & Gemini 1.5 Pro
A multimodal model's close reading of three pieces from three traditions - a Carnatic kriti, an Urdu ghazal, and a Western folk-pop track - covering raga, structure, and cultural context. Each analyzed on its own, then against the others.Qwen2.5Coder-32b Code Critique
Hrishi Olickel & OpenAI o1-preview & Claude 3.5 Sonnet & GPT-4o & Qwen 2.5 CoderHrishi Olickel & OpenAI o1-preview & Claude 3.5 Sonnet & GPT-4o & Qwen 2.5 Coder
How well can a 32B specialized coding model critique real code, next to o1-preview, GPT-4o, and Sonnet 3.5? The results say something surprising about specialized coding models.Entropixplained
Hrishi Olickel & Claude 3.5 SonnetHrishi Olickel & Claude 3.5 Sonnet
A guide to Entropix, _xjdr's context-aware sampler that tunes generation on entropy and attention rather than a fixed temperature. What it does, how the code works, and why it can improve coherence and cut hallucinations.OCR with GOT and Sonnet
Hrishi Olickel & Minu ChoiHrishi Olickel & Minu Choi
Complex Indian electoral tables that stumped Document Intelligence and Sonnet on their own, read at 98.79% accuracy. The trick: pair GOT's raw extraction with Claude Sonnet as a corrector.