{"date":"2026-08-02","title":"What happened in AI — Aug 2, 2026","generated_at":"2026-08-02T21:11:26Z","intro":["China's open-weight labs set the pace again: DeepSeek shipped a 28-cent agent model and a cheaper V4-Flash release, while demand for Moonshot's Kimi K3 is outrunning serving capacity — AMD's MI355X now undercuts Nvidia's B300 on the cost to run it.","Builders kept shipping narrow agent components instead of platforms: a sub-1MB C++ reimplementation of Codex, an `await human()` escalation primitive, and a risk-manager agent that can veto trades. Forbes reported DeepSeek-powered AI already being used to launch attacks, underscoring the open-weights security debate Simon Willison rounded up today."],"highlights":["DeepSeek's 28-cent agent model and new V4-Flash release extend its low-cost model cadence from yesterday's $0.28 pricing floor.","Kimi K3 demand is outpacing serving capacity; AMD's MI355X reportedly undercuts Nvidia's B300 on the cost to run it.","MicroCodex reimplements OpenAI's Codex coding agent in C++ as a sub-1MB binary.","Handoff and Walsh push human-in-the-loop and risk-veto patterns into agent code as first-class primitives, not bolt-ons.","Forbes: DeepSeek-powered AI is already being used to launch attacks, with researchers warning agentic threats are moving past one-off incidents."],"article_count":11,"categories":[{"name":"Builders Ship Narrow Agent Components, Not Platforms","slug":"agent-components-not-platforms","summary":"Today's agent tooling releases were small, single-purpose pieces — a minimal coding-agent binary, a human-escalation primitive, and a risk-veto agent — rather than new all-in-one platforms.","articles":[{"title":"Show HN: MicroCodex Coding Agent – OpenAI/codex reimplemented in C++ <1MB binary","summary":"MicroCodex reimplements OpenAI's Codex coding agent in C++ as a sub-1MB binary, trading the usual Python/Node runtime overhead for a minimal footprint.","source":"hackernews_ai","url":"https://github.com/paoloanzn/microcodex","published":"Sun, 02 Aug 2026 20:11:35 +0000"},{"title":"Show HN: Handoff is await human() for AI agents","summary":"Handoff packages human-in-the-loop escalation as a single `await human()` call in agent code, treating the handoff-to-human pattern as a first-class API primitive.","source":"hackernews_ai","url":"https://github.com/OmegaAgent/handoff","published":"Sun, 02 Aug 2026 12:20:56 +0000"},{"title":"Walsh: Multi-agent research pipeline with risk manager that can veto trades","summary":"Walsh pairs a multi-agent research pipeline with a dedicated risk-manager agent that can veto trades — a concrete example of a guardrail agent constraining an action-taking one.","source":"hackernews_ai","url":"https://github.com/ats4321/walsh","published":"Sun, 02 Aug 2026 00:13:52 +0000"},{"title":"Show HN: Replaybook, an Infrastructure Agent Evaluation Framework","summary":"Replaybook started as an incident-replay trainer and is now doubling as an evaluation framework for infrastructure agents, repurposed eval tooling rather than built for it.","source":"hackernews_ai","url":"https://github.com/ducks/replaybook","published":"Sun, 02 Aug 2026 15:52:25 +0000"}]},{"name":"China's Open Models Keep Undercutting on Cost","slug":"china-open-models-cost","summary":"DeepSeek and Moonshot pushed further on price and scale — a new sub-$0.30 agent model, a cheaper V4-Flash release, and cost-effective inference hardware for Kimi K3 — as demand for Kimi K3 starts to outrun serving capacity.","articles":[{"title":"DeepSeek's new 28-cent agent model","summary":"DeepSeek's newest agent model runs tasks at roughly 28 cents, extending the low-cost pricing push that set an agentic-output floor of $0.28 yesterday.","source":"search_cn_open_weight_labs","url":"https://news.google.com/rss/articles/CBMibkFVX3lxTE9nSjdRT2hZdDU1cmtzOE9Zd0dGc2xxbXhpS1ZIOXVtc1BnYmh0RFhhak5IV19VWXBoNmVlUVFRc0xINk9hWHhrYzMyaGpxeHd1NUN3Vll3cDIxODFWTUN6blVHLUZzdFBIb2dYTFpn?oc=5","published":"Sun, 02 Aug 2026 19:31:10 GMT"},{"title":"DeepSeek Makes a Splash with Small, Affordable V4-Flash Model","summary":"DeepSeek shipped V4-Flash, a smaller and cheaper model, continuing its aggressive low-cost cadence alongside today's 28-cent agent-model release.","source":"search_cn_open_weight_labs","url":"https://news.google.com/rss/articles/CBMimgFBVV95cUxOTjdaMHZvTnZTd1YyUy1zeFNKem9fUUxtemFnb210MkVxZjN4bkNEdVdVT3hFekNDeFZBdU0zWkZKdzhuZVEydzJhZUFhM2IzUzVqbVVqMEVJeWVzNFVIOEozajJkN0xIejVrT0VwQ3lUQUoxemI0STVVMzFlVWFoLXlqR3hkZk1fSE81UXJEZ21rakpCejFPZExn?oc=5","published":"Sun, 02 Aug 2026 04:10:00 GMT"},{"title":"Global interest in China's Kimi K3 model strains capacity and intensifies pressure in the US-China AI race","summary":"Demand for Moonshot's Kimi K3 is outpacing available serving capacity, adding to the competitive pressure between US and Chinese AI labs.","source":"search_cn_open_weight_labs","url":"https://news.google.com/rss/articles/CBMi0wFBVV95cUxQV0VybHVwcTZ4bTJzLUpIcmNxXzZmekZfOURjaGtncmRjdF9CSXJmUzFNbENfdnFYTDl0RUl1ZHN3REZVekRoR181dk9oYW1sTGU3SVF6QTZnQ2J4dm5URUFFMWxkSld1eWszZnEzb1ZuaXQtZ1NCRWhBUlV4eGM1a2dYb2liS25FQkJNRFZwTlMzSTF3aHRXVXFGV3ZGZFRnaGRMT2J5cUlwN0prY3FoV2lwMkNDN2hYeHF3T0drWFlBTXE2MHhRajc0eWZkRHd4OFRn?oc=5","published":"Sun, 02 Aug 2026 10:00:16 GMT"},{"title":"AMD's MI355X Undercuts Nvidia's B300 on Cost to Run China's Kimi K3","summary":"AMD's MI355X reportedly serves Kimi K3 more cheaply than Nvidia's B300, a data point for platform teams weighing inference hardware for open Chinese models.","source":"search_cn_open_weight_labs","url":"https://news.google.com/rss/articles/CBMilwFBVV95cUxPaGpQQnl3a2lPemY4anZYaV9uR3Fmb3VPX3ZJc0JhZHBYVzdHMTM0ZWx2ZE01VHBRWmpOTHI1R2hLNnJCdEx2QlJ3N0xWT2xfQUJLU2xNVFNjQzRIa0l2ZV94NEg1blZVaFNqZ0JxRDFRel9JSEFFTUNOc0xSUnQ4ZDFXTDJfajNfdl9JbDRhbnkyU0pmMXRn?oc=5","published":"Sun, 02 Aug 2026 05:17:39 GMT"},{"title":"Latest open artifacts (#23): Laguna S2.1, Inkling, & Kimi K3 show the utility of open models on the Pareto frontier","summary":"Interconnects AI's latest open-artifacts roundup lines up Laguna S2.1, Inkling, and Kimi K3 as evidence open models are closing in on the cost/quality Pareto frontier.","source":"search_cn_open_weight_labs","url":"https://news.google.com/rss/articles/CBMidEFVX3lxTE9LeWFtSnhpUW9yLXMzRWI5SDM3bms5SE12aEZOUnRlRUtBX3FKUmY1RTVBRXRtOVJTSjZnbzVzZVlJa0NKUE9hcVJDR01vSmlLNXVqNVdHbUs1bjhyS2gxNnA0UF9NZHZqOFdfbDBTOGpmUlpQ?oc=5","published":"Sun, 02 Aug 2026 13:01:40 GMT"}]},{"name":"Security and Policy Catch Up to Agentic Risk","slug":"security-policy-agentic-risk","summary":"As open agent models get cheaper and more capable, reporting turned to how that capability is already being misused and to the policy debate over open weights underneath it.","articles":[{"title":"DeepSeek-Powered AI Used To Launch Attacks — Agentic Threats May Go Beyond One-Offs","summary":"Forbes reports DeepSeek-powered AI has been used to launch attacks, with researchers warning agentic threats are moving past one-off incidents toward repeatable tooling.","source":"search_cn_open_weight_labs","url":"https://news.google.com/rss/articles/CBMi0AFBVV95cUxPandUYVhVbWJLQVN3Vks2b01rOVBhQU1VU3dEd0cxb2QtR0taRm5VUlJncHNaaFkybzk5aXlpdVgxU0ZTUFAxemw0V00ydkkxS2d0YmIwdEVyZEVWdzZSalNoODIxVS1kNjM1eE1BT2NiZ3dJMTRUNjBPamh5d1paRFZ0OXo0V0RfYm8tN0JlR3JmLWoyWEZsaHRHUmhPS0p3Vi1hMWNTWXRnaG9jM0NKWmJaSl9qQVpkT2FOODVyOS1temJqUVlIMWRzcEE3eDdO?oc=5","published":"Sun, 02 Aug 2026 07:58:25 GMT"},{"title":"Open letters about AI development","summary":"Simon Willison rounds up the past few weeks of open letters on AI development, including the 'Open Weights and American AI Leadership' debate over whether open models help or undercut US AI strategy.","source":"simon_willison","url":"https://simonwillison.net/2026/Aug/2/open-letters/#atom-everything","published":"2026-08-02T04:16:52+00:00"}]}]}