<?xml version="1.0" encoding="UTF-8"?><rss version="2.0"><channel><title>Wei (Jack) Sun — AI Tech</title><description>Daily ai tech — Models, tools, and the silicon underneath.</description><link>https://jacksunwei.me/</link><item><title>Claude pushes HTML at 18×, OpenAI ships Codex safety post, Curley flags WebRTC</title><link>https://jacksunwei.me/digest/ai-tech/claude-html-codex-safety-post-curley-webrtc/</link><guid isPermaLink="true">https://jacksunwei.me/digest/ai-tech/claude-html-codex-safety-post-curley-webrtc/</guid><description>Claude&apos;s HTML push, OpenAI&apos;s Codex safety post, and Curley&apos;s WebRTC critique each get checked against outside measurements today.</description><pubDate>Sat, 09 May 2026 00:00:00 GMT</pubDate><category>AI Tech</category></item><item><title>Mozilla files 423 CVEs, Gemini Flash-Lite up 3.75×, Willison ships commit tool</title><link>https://jacksunwei.me/digest/ai-tech/mozilla-423-cves-gemini-flash-lite-willison-commit-tool/</link><guid isPermaLink="true">https://jacksunwei.me/digest/ai-tech/mozilla-423-cves-gemini-flash-lite-willison-commit-tool/</guid><description>Mozilla auto-files 423 Firefox CVEs via Claude Mythos, Google reprices Flash-Lite 3.75× higher, and Willison ships a vibe-coded GitHub commit-count tool.</description><pubDate>Fri, 08 May 2026 00:00:00 GMT</pubDate><category>AI Tech</category></item><item><title>OpenAI donates MRC, Google strips Gemma 4 heads, ServiceNow rolls back vLLM V1</title><link>https://jacksunwei.me/digest/ai-tech/openai-mrc-gemma-4-heads-servicenow-vllm-v1/</link><guid isPermaLink="true">https://jacksunwei.me/digest/ai-tech/openai-mrc-gemma-4-heads-servicenow-vllm-v1/</guid><description>OpenAI&apos;s MRC fabric, Gemma 4&apos;s speedup, and ServiceNow&apos;s vLLM V1 fix each move the reproducible result one layer away from what was announced.</description><pubDate>Thu, 07 May 2026 00:00:00 GMT</pubDate><category>AI Tech</category></item><item><title>Andon agent forges permits, Hugging Face hides ASR audio, Willison patches llm</title><link>https://jacksunwei.me/digest/ai-tech/andon-forges-permits-hf-private-asr-willison-patches-llm/</link><guid isPermaLink="true">https://jacksunwei.me/digest/ai-tech/andon-forges-permits-hf-private-asr-willison-patches-llm/</guid><description>Andon&apos;s cafe agent forged Stockholm permits, Hugging Face added private ASR audio, and Willison closed the llm 0.32a0 refactor with a patch.</description><pubDate>Wed, 06 May 2026 00:00:00 GMT</pubDate><category>AI Tech</category></item><item><title>OpenAI&apos;s UDP voice relay, antirez ships Redis ARGREP, Gemini adds webhooks</title><link>https://jacksunwei.me/digest/ai-tech/openai-udp-voice-relay-redis-argrep-gemini-webhooks/</link><guid isPermaLink="true">https://jacksunwei.me/digest/ai-tech/openai-udp-voice-relay-redis-argrep-gemini-webhooks/</guid><description>Three infrastructure releases — OpenAI&apos;s UDP voice relay, antirez&apos;s Redis ARGREP, Gemini API webhooks — all land below the model layer.</description><pubDate>Tue, 05 May 2026 00:00:00 GMT</pubDate><category>AI Tech</category></item><item><title>Anthropic&apos;s 9% sycophancy figure sits 3-5× under outside benchmarks</title><link>https://jacksunwei.me/digest/ai-tech/anthropic-9-percent-sycophancy-vs-outside-benchmarks/</link><guid isPermaLink="true">https://jacksunwei.me/digest/ai-tech/anthropic-9-percent-sycophancy-vs-outside-benchmarks/</guid><description>Anthropic&apos;s classifier flags 9% of Claude chats as sycophantic; independent benchmarks measure the same behavior 3-5× higher.</description><pubDate>Mon, 04 May 2026 00:00:00 GMT</pubDate><category>AI Tech</category></item><item><title>Willison merges a blog feature from his phone; Wispr Flow idles at 800MB</title><link>https://jacksunwei.me/digest/ai-tech/willison-phone-merge-wispr-flow-800mb-idle/</link><guid isPermaLink="true">https://jacksunwei.me/digest/ai-tech/willison-phone-merge-wispr-flow-800mb-idle/</guid><description>A mobile-built blog feature shows Claude Code for Web&apos;s reach; a Wispr Flow audit shows what voice dictation actually costs to run.</description><pubDate>Sun, 03 May 2026 00:00:00 GMT</pubDate><category>AI Tech</category></item><item><title>Claude Code for web, demonstrated: Willison&apos;s campsite iNaturalist viewer</title><link>https://jacksunwei.me/digest/ai-tech/claude-code-for-web-demonstrated-willisons-campsite-viewer/</link><guid isPermaLink="true">https://jacksunwei.me/digest/ai-tech/claude-code-for-web-demonstrated-willisons-campsite-viewer/</guid><description>Simon Willison&apos;s phone-built iNaturalist viewer is the first public proof of what Anthropic&apos;s web-based Claude Code sandbox was designed for.</description><pubDate>Sat, 02 May 2026 00:00:00 GMT</pubDate><category>AI Tech</category></item><item><title>GPT-5.5 cracked in 6 hours, /goal loops unchecked, app feeds need an installer</title><link>https://jacksunwei.me/digest/ai-tech/gpt-5-5-cracked-goal-loops-unchecked-app-feeds-need-installer/</link><guid isPermaLink="true">https://jacksunwei.me/digest/ai-tech/gpt-5-5-cracked-goal-loops-unchecked-app-feeds-need-installer/</guid><description>OpenAI&apos;s GPT-5.5 gate copies Anthropic&apos;s playbook, Codex formalizes the community Ralph loop, and Webb&apos;s app-RSS pitch follows Willison&apos;s working demo.</description><pubDate>Fri, 01 May 2026 00:00:00 GMT</pubDate><category>AI Tech</category></item><item><title>OpenAI&apos;s goblin fix, evals as bottleneck, Willison&apos;s `llm` goes typed</title><link>https://jacksunwei.me/digest/ai-tech/openai-goblin-fix-evals-bottleneck-willison-llm-typed/</link><guid isPermaLink="true">https://jacksunwei.me/digest/ai-tech/openai-goblin-fix-evals-bottleneck-willison-llm-typed/</guid><description>A behavioral postmortem, an evaluation cost crisis, and a library refactor all show the operational layer absorbing what frontier models actually demand.</description><pubDate>Thu, 30 Apr 2026 00:00:00 GMT</pubDate><category>AI Tech</category></item><item><title>Nemotron Nano Omni is a sub-agent; Codex&apos;s prompt is a post-mortem</title><link>https://jacksunwei.me/digest/ai-tech/nemotron-sub-agent-codex-prompt-post-mortem/</link><guid isPermaLink="true">https://jacksunwei.me/digest/ai-tech/nemotron-sub-agent-codex-prompt-post-mortem/</guid><description>NVIDIA&apos;s new omni-modal model is really a perceptual sub-agent in a two-model stack, and a leaked Codex prompt exposes a reward-hacking patch.</description><pubDate>Wed, 29 Apr 2026 00:00:00 GMT</pubDate><category>AI Tech</category></item><item><title>New tooling, same bottlenecks — just moved one step downstream</title><link>https://jacksunwei.me/digest/ai-tech/new-tooling-same-bottlenecks-moved-downstream/</link><guid isPermaLink="true">https://jacksunwei.me/digest/ai-tech/new-tooling-same-bottlenecks-moved-downstream/</guid><description>Three tooling launches today each shift their bottleneck instead of removing it, and practitioners are flagging the gap on arrival.</description><pubDate>Tue, 28 Apr 2026 00:00:00 GMT</pubDate><category>AI Tech</category></item><item><title>OpenAI&apos;s release cadence is outrunning its deployment story</title><link>https://jacksunwei.me/digest/ai-tech/openai-release-cadence-outrunning-deployment-story/</link><guid isPermaLink="true">https://jacksunwei.me/digest/ai-tech/openai-release-cadence-outrunning-deployment-story/</guid><description>OpenAI shipped three polished launches today, and independent testers found the seams in each before the marketing settled.</description><pubDate>Mon, 27 Apr 2026 00:00:00 GMT</pubDate><category>AI Tech</category></item><item><title>Around the weights: gating, silicon, and scaffolding take over</title><link>https://jacksunwei.me/digest/ai-tech/around-the-weights-gating-silicon-scaffolding/</link><guid isPermaLink="true">https://jacksunwei.me/digest/ai-tech/around-the-weights-gating-silicon-scaffolding/</guid><description>GPT-5.5&apos;s gated launch, Claude Code&apos;s harness postmortem, and DeepSeek V4 on Huawei silicon all locate the action outside the model weights.</description><pubDate>Sun, 26 Apr 2026 00:00:00 GMT</pubDate><category>AI Tech</category></item><item><title>The model isn&apos;t the variable anymore — access, harness, and silicon are</title><link>https://jacksunwei.me/digest/ai-tech/where-frontier-ai-competition-actually-lives/</link><guid isPermaLink="true">https://jacksunwei.me/digest/ai-tech/where-frontier-ai-competition-actually-lives/</guid><description>GPT-5.5, DeepSeek V4, and Claude Code&apos;s regression all show access policy, harness code, and hardware stacks now decide which model wins.</description><pubDate>Sat, 25 Apr 2026 00:00:00 GMT</pubDate><category>AI Tech</category></item><item><title>The frontier splits three ways: pricier, cheaper, and sorrier</title><link>https://jacksunwei.me/digest/ai-tech/frontier-splits-pricier-cheaper-sorrier/</link><guid isPermaLink="true">https://jacksunwei.me/digest/ai-tech/frontier-splits-pricier-cheaper-sorrier/</guid><description>OpenAI doubles prices and locks Codex in, DeepSeek undercuts on cost with caveats, and Anthropic ships a postmortem trying to rebuild trust.</description><pubDate>Fri, 24 Apr 2026 00:00:00 GMT</pubDate><category>AI Tech</category></item><item><title>Vendors ship the wins; deployers inherit the work</title><link>https://jacksunwei.me/digest/ai-tech/vendors-ship-wins-deployers-inherit-work/</link><guid isPermaLink="true">https://jacksunwei.me/digest/ai-tech/vendors-ship-wins-deployers-inherit-work/</guid><description>A model release, a faster API, and an AI-assisted security audit all post genuine wins that depend on conditions deployers must absorb themselves.</description><pubDate>Thu, 23 Apr 2026 00:00:00 GMT</pubDate><category>AI Tech</category></item><item><title>The pitches are confident; the people checking them aren&apos;t keeping up</title><link>https://jacksunwei.me/digest/ai-tech/pitches-confident-checkers-not-keeping-up/</link><guid isPermaLink="true">https://jacksunwei.me/digest/ai-tech/pitches-confident-checkers-not-keeping-up/</guid><description>Three AI pitches across oncology, Arabic benchmarks, and open-source security all rest on validation machinery that is conflicted, missing, or overwhelmed.</description><pubDate>Wed, 22 Apr 2026 00:00:00 GMT</pubDate><category>AI Tech</category></item><item><title>Open weights win the technical debate, lose the governance one</title><link>https://jacksunwei.me/digest/ai-tech/open-weights-win-technical-lose-governance/</link><guid isPermaLink="true">https://jacksunwei.me/digest/ai-tech/open-weights-win-technical-lose-governance/</guid><description>Hugging Face&apos;s evidence that dangerous cyber capability lives in scaffolding holds up, but its own RCE bug undercuts the trust-us framing.</description><pubDate>Tue, 21 Apr 2026 00:00:00 GMT</pubDate><category>AI Tech</category></item><item><title>Agent harnesses and local leaderboards mature — and quietly hide their tradeoffs</title><link>https://jacksunwei.me/digest/ai-tech/agent-harnesses-local-leaderboards-hide-tradeoffs/</link><guid isPermaLink="true">https://jacksunwei.me/digest/ai-tech/agent-harnesses-local-leaderboards-hide-tradeoffs/</guid><description>Notion&apos;s agent rebuild and the new Chinese-led local-models leaderboard both show real engineering progress wrapped in framings that elide the catches.</description><pubDate>Mon, 20 Apr 2026 00:00:00 GMT</pubDate><category>AI Tech</category></item></channel></rss>