Summary
The evening slot is thin on AI and heavy on everything else, but one item is worth the whole read: Tim Cook is lobbying Washington to let Apple buy memory from China's CXMT and YMTC, accusing Micron of gouging at 80% margins while simultaneously calling the DRAM shortage a "100-year flood." That is the memory-supply crunch escalating from a supplier problem into a national-security fight, and it is the same constraint that sits under every inference cost curve. The second real cluster is open weights: Elie Bakouch of HuggingFace posted three times on the momentum, flagging Kimi K3 open weights landing Monday, a run of competitive releases from smaller labs, and a new signatory on NVIDIA's open-model letter. Tom Yeh contributed the slot's only teaching content, a hand-worked vector database walkthrough and an announced Kimi 3 lecture that will dissect Kimi Delta Attention and its 16-of-896 expert sparsity. Everything else is noise: two accounts produced 34 of the 45 posts, almost entirely French domestic politics, Middle East war coverage, and US fraud-enforcement messaging, with zero curated reposts to anchor the slot.
Posts
Tim Cook lobbies to buy Chinese memory as the DRAM shortage bites (@MarioNawfal). Apple wants clearance to source memory from CXMT and YMTC for non-US devices, both designated Chinese military companies, with YMTC on the Entity List, and Cook is publicly accusing Micron of 80% margin gouging. The internal contradiction is the tell: if the shortage really is a hundred-year flood, then 80% margins are the price signal, not the abuse, and this is the clearest sign yet that memory supply has become the binding constraint on device and inference economics (memory hierarchy, NVIDIA and SK's HBM codevelopment).
HuggingFace on the open-weights moment (cluster of 3: @eliebakouch · @eliebakouch · letter PDF). Bakouch calls out Kimi K3 shipping open weights on Monday plus competitive releases from Thinking Machines, Poolside, Motif, and Upstage, and separately notes a new logo signing NVIDIA's open-model letter. His framing is that open and closed are both accelerating at once rather than one displacing the other, which is a more honest read than the partisan version circulating this afternoon (NVIDIA open-weights letter).
Opus 5 is probably a much smaller model than its scores suggest (@eliebakouch). Bakouch reads Opus 5 as on par with Mythos 5 on AI research work but not clearly better, and infers from the pricing that it sits well below the Mythos and Fable parameter tiers. If that inference holds, the interesting number is not the benchmark parity but the cost per unit of capability, which is the same axis Anthropic is pricing against (Claude Opus 5).
Kimi 3 lecture on Delta Attention and extreme MoE sparsity (@ProfTomYeh · event). Tom Yeh is hand-deriving Kimi 3's architecture on Aug 6, specifically Kimi Delta Attention, attention residuals, and a mixture-of-experts layout that activates just 16 of 896 experts in a 2.8T-parameter model. That activation ratio is under 2%, far sparser than the usual 8-of-64 designs, and the derivation is worth attending if you care about where routing sparsity is actually headed (attention mechanisms).
Vector databases worked by hand, cell by cell (@ProfTomYeh). A ten-step walkthrough that indexes three sentences and answers a query by nearest-neighbour search, with every embedding lookup, encoding step, and distance computation filled in manually on a toy vocabulary of 22 words and four dimensions. It is the retrieval layer under RAG stripped of library abstraction, and it is the kind of thing worth twenty minutes if you have only ever called a client SDK.
A video-native coding agent in preview (@brivael). Described as a Claude Code equivalent built for video, with native avatar primitives, a built-in editor, connections to every model on the market, and user-authored templates, aimed eventually at Netflix-quality series and documentaries. Preview-stage hype with no demo or benchmark attached, but the harness pattern of agent plus domain-native primitives plus model-agnostic backend is the one that keeps recurring.
Grok Imagine ancient Greek fashion promo (@MarioNawfal). Marketing copy for a text-to-video demo with an xAI tag list attached. Skip.
Opus 5 shut down Codex rumor (@ns123abc · @heavypulp). Two joke posts riding the launch cycle, one an all-caps fake breach rumor, the other a Musk repost. Skip.
Non-AI volume (@brivael, @MarioNawfal, @WHFraudTF, @BrettRatner). Thirty-four posts of French domestic politics, Iran and Ukraine war reporting, a fraud task force weekly scorecard, and a book promotion. Two accounts alone carry three quarters of the slot by count. Skip.