Week 21, 2026 · May 18-24: TILs, Clicker Games and Google I/O
I shipped a TIL site, an Odia 2048 game and a Chhena Poda clicker, learned how agent hallucination differs from LLM hallucination, and dug into Google I/O 2026.
Week 21: TILs, Clicker Games and Google I/O
What I learned
Annual Health Checkup
- Did a full body checkup this week. Good to have a baseline and track progress over time.
Agent Hallucination vs Traditional LLM Hallucination
- I added my first real TIL to til.soumendrak.com.
- In traditional LLM hallucination, the model makes up facts. In agent hallucination, the model makes up actions: it claims it called a tool and got result X, but never actually executed the tool call.
- For observability, detecting a fake tool call in a trace is different from detecting a factually wrong response. You need to verify the side effects, not just the output.
A Consultant’s Pricing Framework
- Met a seasoned consultant who shared a practical pricing framework.
- Calculate your hourly rate, then quote a range: the minimum is your actual rate, the maximum is what you charge clients you would rather not work with.
- Minimum 6 month engagements, max 2 clients at a time. The reason for 6 months: the problem is always people, not technology. It takes months to understand organizational dynamics and actually fix things.
First Aid Notes
- I spent an afternoon noting down first aid procedures: cardiac arrest vs heart attack, stroke FAST signs, seizure stages, choking, bleeding, burns, heatstroke, food poisoning, dog bites and fractures.
- Turned it into an AgentWiki page. Hope I never have to use it, but good to have it searchable.
Google I/O 2026
- Gemini 3.5 Flash, Spark (background agents), Omni for video generation.
- The most interesting announcement for me was Gemini Spark: a 24/7 personal AI agent running background tasks across connected apps.
- Directly relevant to the agent observability work I do. If agents run autonomously in the background, tracing what they did and why becomes even more critical.
MCP Dev Summit Bengaluru
- June 9-10, Linux Foundation event. Early bird (Rs 7,900) ends May 26. Co-located with KubeCon India. Worth attending if I can make it.
What I built
TIL Site
- Built a proper Today I Learned micro-blog at til.soumendrak.com.
- First version was static HTML with hardcoded content. My AI assistant Nini refactored it into a proper build pipeline: Markdown posts in
posts/with YAML front matter, a Node.js build script that validates and renders todata.js, GitHub Actions CI that checks every post and auto-deploy to Cloudflare Pages. - The first real TIL about agent hallucination is live. More to come.
Odia 2048 Game
- A standalone HTML/CSS/JS version of 2048 with Odia numerals, live at 2048.openodia.com.
- No build step, no dependencies, one file. Dark and light mode, touch and keyboard support, animations.
Chhena Poda Clicker Game
- An idle clicker game themed around the Odia dessert Chhena Poda.
- Tap the Chhena Poda to earn points, hire Mausis, build Bhatis, expand from a local shop to alien exports. 10 upgrade tiers with compounding costs.
- Vanilla JS, single HTML file, auto-save in localStorage.
Widescope Project Page
- Added a proper project page for widescope on soumendrak.com/projects.
- Logo as cover image, screenshot in the body, linked to the live site. It now appears first on the projects page.
OpenOdia Hub Events Page
- Built an events page for openodia.com tracking Odia AI community conferences and workshops.
- Added a daily crawl-events skill that checks source pages for new events. The page now has 40+ events from GDG chapters, Odisha AI and OdiaGenAI.
Newsletter Digest Pipeline
- The email watcher now processes newsletters on arrival: summarizes, researches action items and appends to a running digest draft.
- Pushed to lili-backup repo for easy reading. 175+ entries logged this week.
What I failed
- Odia transliterator: Tried to build a roman-to-Odia script converter. The syllable mapping needed linguistic knowledge I didn’t have. Gave up and made the repo private. Some problems need a computational linguist, not a weekend hack.
- PDF cover image: The Nemotron workshop PDF went through 6 iterations before the cover looked right. Photo overlays with CSS gradients are hard to preview from a sandbox. Lesson: pure CSS designs are safer than photo composites.
- TIL data.js syntax error: An extra
];broke the entire site. The build looked fine locally but the JSON was invalid. Fixed by using a proper build script with validation. - Cloudflare Pages build config: Setting up Zola builds on Cloudflare required multiple attempts. The build command kept getting lost in the API. Eventually gave up on auto-build and deployed manually.
Random Thoughts & Quotes
- Two months ago I was managing a team. Now I am a Staff IC exploring consulting. The shift is disorienting but right.
- The same week I built a clicker game about Chhena Poda, I also wrote about Buddha and impermanence. That range feels right.
- I have switched from written journaling to dictating notes to my AI assistant. Faster, more immediate. Not sure if I lose something by not writing.
Newsletters That Mattered This Week
AI & Agents
- Thinking Machines Interaction Models: Mira Murati’s lab released their first preview. 276B MoE model with 200ms micro-turns, real-time audio/video/text, background reasoning. The interface model paradigm is different from the chat paradigm. Worth watching.
- Agent Skill Lifecycle (SkillsVote paper): Converts agent trajectories into reusable executable skills. +7.9 pp on Terminal-Bench. Most actionable paper for agent evaluation pipeline design.
- RELEX: RLVR weight trajectories are rank-1 and near-linear. Using 15-20% of training steps, RELEX extrapolates to match full RLVR on Qwen3-8B. Could cut RLVR compute by 80%.
- Code as Agent Harness survey: UIUC/Meta/Stanford paper. Three-layer taxonomy of how code serves as the operational substrate for agent systems.
AI Observability
- Arize: Self-Improving Agent on a Context Graph: Builds a procurement agent that improves from 53.8% to 83.1% human agreement by mining human override patterns from traces. No fine-tuning needed.
- Arize Phoenix v16.0.0: Code Evaluators: sandboxed Python/JS eval functions in the Phoenix UI. Breaking changes around sandboxing.
- OpenTelemetry graduates CNCF: OTel is now a graduated project. Enterprise adoption accelerates from here.
- Langfuse: Skills for Coding Agents: Marc Klingen’s talk on building a skill so Claude Code can reliably instrument applications. Six lessons on making skills work in practice.
- Why Rust is different: Alice Ryhl (Google Android Rust team, Tokio maintainer) on what makes developers stick with Rust.
Industry & Markets
- NVIDIA earnings: Revenue $81.6B (+85%), net income $58.3B (+211%). Dividend hiked 2,400%. Signs of AI capex continuing.
- SpaceX IPO: Largest in history at $1.75T valuation, June 12 Nasdaq debut. Starlink prints $7.2B EBITDA but xAI burns $6.4B. Three companies stapled together.
- Aschenbrenner’s $8.5B short on AI semis: Net short on NVDA, AVGO, AMD, MU. Long on CoreWeave, Applied Digital. Thesis: semi scarcity peaking, value accrues to DC operators.
From the Week’s Read Later List
- Managed Agents (Google I/O): Single API call spins up a complete agent in an ephemeral Linux sandbox. Configurable via AGENTS.md and SKILL.md files. Competes directly with MCP-based runtimes.
- Cohere Command A+: 218B MoE (25B active), Apache 2.0 license. First truly open enterprise-grade agentic model. Lossless quantization at 4-bit. Runs on 2 H100s.
- ByteByteGo: RAGs vs Agents: RAG is retrieve + generate. Agents add tool-use and multi-step reasoning. The distinction is becoming more important as teams choose between the two patterns for production.
- Pi-Bench: New benchmark for proactive personal assistant agents. 100 multi-turn tasks with hidden intents and cross-session continuity. Mentions OpenClaw in the abstract.
- DelTA: Discriminative Token Credit Assignment for RLVR. Fixes the problem where standard RLVR rewards get diluted by formatting tokens. +3.26 pts on Qwen3-8B math.
- RTPurbo: Full attention to sparse in 100 training steps. 9.36x prefill speedup at 1M context. Near-lossless. Only retrieval heads need full context.
- OpenOdia.com verified on Google Search Console: The domain is now indexed. Sitemap submitted with 4 URLs. More coverage needed for individual project pages.