shanhui dev notes

where i add interesting articles to read that i will “eventually” get to, i promise..

• https://www.brethorsting.com/blog/2026/05/domain-expertise-has-always-been-the-real-moat

• Edge AI, Hardware, Manufacturing

• https://alexlavaee.me/blog/open-source-ai-edge-coding-agents
• https://www.copetti.org/writings/consoles/playstation - useful to learn about optimising edge hardware and chips where there are compute constraints

• https://squeezlabs.github.io/handcrank - no electric, no internet edge voice llm

  • https://squeezlabs.com/team.html

• Agentic harness for 3D CAD Design, automated modeling of articulated, kinetic physical objects for manufacturing

  • Artifact10k project: https://articraft3d.github.io
  • https://github.com/Adam-CAD/CADAM
  • a discontinued CAD oss project https://mattferraro.dev/posts/cadmium - worth reading to understand fundamental building blocks and challenges of building CAD software such as 2d constraint solver, and B-rep kernals
  • great tutorial for learning FreeCAD https://www.youtube.com/watch?v=wxxDahY1U6E

• starting a manufacturing business in the US https://hardwareishard.com/starting-a-manufacturing-business - contains some good tibits about supplying hardware to startups

• Agentic and Harness Engineering

• mario on building pi agent https://mariozechner.at/posts/2025-11-30-pi-coding-agent
• cc prompt change history https://cchistory.mariozechner.at
• openai harness engineering https://openai.com/index/harness-engineering
• https://medium.com/@NMitchem/if-ai-writes-your-code-why-use-python-bf8c4ba1a055
• interesting case examples:
  • https://openai.com/index/building-self-improving-tax-agents-with-codex - how openai created an agent harness and self improvement loop for tax prep workflows - key takeaway is that new automations should only be implemented based on repeatable evidence traces from real operations and why working direct with operators is so important to engineering useful agents, you need a closed loop
• https://arxiv.org/abs/2602.11865

• Cursor 3-way interview with Baseten modal research engineers https://www.youtube.com/watch?v=-jnwTZ789V0

• Voice Agents and Audio Engineering

• https://moq.dev/blog/webrtc-is-the-problem
  • response to https://openai.com/index/delivering-low-latency-voice-ai-at-scale
  • https://datatracker.ietf.org/doc/html/draft-ietf-webtrans-http3
  • author: meself@kixel.me

• open, on device tts / cpu only: https://huggingface.co/Supertone/supertonic-3

  • https://huggingface.co/hexgrad/Kokoro-82M-v1.1-zh
  • https://github.com/OpenMOSS/MOSS-TTS

• Data Engineering

• why is duckdb so fast? https://www.greybeam.ai/blog/duckdb-internals-part-1

• Frontier model labs releases

• Edge models

  • https://blog.google/innovation-and-ai/technology/developers-tools/introducing-gemma-4-12b - opensource weights 12b that can be run on a decent laptop with just 16gb of vram, according to HN post roughly compares to GPT-4.1, although according to consensus, for a 16gp laptop, Qwen-9B seems to be the go to choice.
  • https://huggingface.co/buckets/zshanhui/Fara-7B-bucket - interesting 7b computer use edge sml model

• General machine learning and LLM foundations

• https://opencv.org/opencv-5
• https://www.0xkato.xyz/how-llms-actually-work - no maths!
• https://12gramsofcarbon.com/p/deep-learning-is-applied-topology
  • neural networks can be thought of as topology generators
  • article claims that reasoning is part of this topological space so you can classify good/bad reasoning just like you can classify cats/dogs as points on a manifold
• https://www.generalanalysis.com/blog/supabase-mcp-blog - how an prompt injection attack works in practice against a well known mcp database

• entropy, compression, information theory, and intelligence https://www.youtube.com/watch?v=l6DKRf-fAAM

• CS and ML Challenges/Puzzles

• https://blog.janestreet.com/can-you-reverse-engineer-our-neural-network

• AI and societal changes, how to use AI more effectively

• https://www.dailycal.org/news/campus/academics/failing-grades-soar-as-professors-see-greater-ai-usage-dwindling-math-skills-in-uc-berkeley/article_16fad0bf-02cb-4b8c-8d88-888ffd9f8608.html - Ai use can be a great tool for learning and creation, but can not be a substitute for understanding and judgement

• https://www.nair.sh/guides-and-opinions/communicating-your-expertise/why-senior-developers-fail-to-communicate-their-expertise

• Dev Tools, Evals

• https://github.com/aattaran/deepclaude
• https://cognition.ai/blog/frontier-code

• good nuggets/advice on working with AI coding tools https://news.ycombinator.com/item?id=48553224

• Product design

• https://a16z.com/four-key-product-principles-from-wechats-creator

  • prioritise users first
  • kpis are secondary, be data informed, not data driven
  • good products do not need explanation
  • prioritise organic traffic and sharing, no amount of promotion will turn a bad product good
  • simple core, decentralise ecosystem
  • active search is better than passive discovery

• random curiosities

• https://lr0.org/blog/p/arabic

04/28 - https://www.koshyjohn.com/blog/ai-should-elevate-your-thinking-not-replace-it - https://techtrenches.dev/p/the-human-cost-of-10x-how-ai-is-physically - need to read “Ironies of Automation” 1983 Lisanne Bainbrige

AI Papers worth reading - https://www.youtube.com/watch?v=3rWSvrFahIY YC 5 papers worth reading - SAGE: A Service Agent Graph-guided Evaluation Benchmark https://browse-export.arxiv.org/pdf/2604.09285

04/23 - https://crawshaw.io/blog/building-a-cloud - same founder: https://tailscale.com, vpn identity startup? - LLM-as-a-judge HTTP Proxy, https://www.brex.com/crabtrap - https://wheelfront.com/this-alberta-startup-sells-no-tech-tractors-for-half-price - https://qwen.ai/blog?id=qwen3.6-27b - web as an image stream from llm https://flipbook.page/n/491751647a724297990c7bc2423e27b7 - https://metauto.ai/index.html - https://github.com/OpenHands/OpenHands, opensourced atoms.dev/? - llm/agents as judge https://github.com/metauto-ai/agent-as-a-judge - https://arxiv.org/pdf/2410.10934

04/18 - draft an essay about inputs (reading/listening) and outputs (writing/speaking) - make the argument that reading is increasingly in importantce in the agentic era

04/14 - https://www.quantamagazine.org/the-ai-revolution-in-math-has-arrived-20260413

04/08 - https://github.com/cloudflare/vibesdk - building realtime voice agents on CF https://blog.cloudflare.com/cloudflare-realtime-voice-ai - voice turn detection https://developers.cloudflare.com/workers-ai/models/smart-turn-v2 - https://z.ai/blog/glm-5.1 - https://github.com/KCORES/vector-db-bench

• MiniMax2.7 https://www.minimax.io/news/minimax-m27-en

https://kitchenknif.github.io/SlowChinese/modern-china/204-mei-tu-xiu-xiu.html - rss https://kitchenknif.github.io/SlowChinese/podcast.rss - we should prioritise support for basic podcast v0 on LCN

04/04 - https://nworbmot.org/blog/solar-battery-world.html - https://deepmind.google/models/gemma/gemma-4 - https://github.com/garrytan/gstack

04/01 - https://github.com/chenglou/pretext - https://ccunpacked.dev/

03/31 - https://github.com/drona23/claude-token-efficient/blob/main/CLAUDE.md?plain=1 - https://alexhwoods.com/dont-let-ai-write-for-you - ai desk pet https://www.reddit.com/r/SideProject/comments/1s7ulff/got_tired_of_ai_just_being_a_text_box_so_i_spent

03/09 - https://sytse.com/cancer

03/26 - https://mariozechner.at/posts/2026-03-25-thoughts-on-slowing-the-fuck-down/#toc_1 - AttnRes breakthrough: https://github.com/MoonshotAI/Attention-Residuals - https://epoch.ai/frontiermath/open-problems/ramsey-hypergraphs - interesting projects - https://github.com/karpathy/autoresearch - https://arcprize.org/tasks/ls20 - read it later app, https://hq.getmatter.com/company - hiring for swe remote: https://getmatterapp.notion.site/Matter-Software-Engineer-Backend-42f8d048188a4f31ac6f815d674406df - good fit for our current skillset and interests - yc supply chain startups for ref - Flexport for LATAM, https://nowports.com - manufacturing, https://www.ycombinator.com/companies/terra-2 - food chain/distribution, https://www.ycombinator.com/companies/burnt - industrial sourcing, https://www.ycombinator.com/companies/andustry - might be worth it to reach out about supplying battery and forklift components

• https://blog.skypilot.co/scaling-autoresearch
• look more into OSS Maestro E2E QA agent https://maestro.dev
• Kitten TTS models - smallest TTS less than 25MB https://news.ycombinator.com/item?id=47441546
• automated qa testing agent for web/mobile https://www.runcanary.ai
• andrejk autoresearch https://eu.36kr.com/en/p/3725521482578567

• https://news.ycombinator.com/item?id=47291123

• how i built a sub-500ms latency voice agent from scratch: https://www.ntik.me/posts/voice-agent

• read first pass, the author used 2mb speech detection model for simple turn based state system for a basedline voice audio
• created notes in mlai-notes.md
• https://github.com/zshanhui/shuo-zhongwen - lets remix it to speak fluent Chinese

• chinese tiny llm 2b parameters: https://chinese-tiny-llm.github.io

• andreik microgpt: https://karpathy.github.io/2026/02/12/microgpt

• first pass: really interesting stuff. lets set 4 hours tomorrow morning to dive deep into this and get a high level understanding, Friday 5th Mar

• NVIDIA PersonaPlex 7B, https://blog.ivan.digital/nvidia-personaplex-7b-on-apple-silicon-full-duplex-speech-to-speech-in-native-swift-with-mlx-0aa5276f2e23 (speech2speech model, not asr->llm->tts 3 step)

• worth looking more into… when we have more time

People and personal blogs

• https://bowenbryanwang.github.io

China AI developments (中文)

• articles from 虎嗅 news
• 黄仁勋中国父母 https://www.huxiu.com/article/4867927.html
• https://www.huxiu.com/article/4864204.html
• https://www.huxiu.com/article/4862416.html

• https://www.huxiu.com/article/4864393.html

• 字母AI https://www.huxiu.com/member/2793879.html