Authorial takes on the day's AI-engineering news: the conclusion the source itself does not state, grounded in the primary source and the New Runtime archive.
Giotto markets Alto and Solo as sovereign European models, but it publishes no base model, weights or parameter count. The claim that Alto inherits Kimi weights rests on user screenshots and a reported CEO remark that we could not trace to its source.
Gemini 4 Hydrogen matters less as another fast model than as part of a workflow that can assign routine execution and harder reasoning to different models. That makes routing quality—not any single benchmark—the bet builders should test.
Huawei can build large campuses where electricity is abundant, but the harder metric is useful compute from each megawatt. The Horinger case suggests energy is an enabler, not the decisive advantage.
mem0, Zep, and Letta are third-party agent memory layers, unlike Claude/ChatGPT built-in memory which is product-scoped personalization inside those chat apps. mem0 extracts facts into SQL + vector embeddings (+ Platform Graph Memory entity linking), with Platform cloud or OSS self-host (pip/npm Mem
The board has installed an interim CEO but named no deadline or successor. For customers, the key risk is not the title change itself; it is whether Nvidia can sustain execution while infrastructure financing and data-center access are already under pressure.
SemiAnalysis's AgentX comparison shows Vera Rubin NVL72 with a 1.8x to 2.0x lead in tokens per dollar at 100 to 120 tok/s/user. The double-digit multiples appear only at high interactivity, where GB300 falls off. All Rubin numbers are previews on early software.
Meta's post describes a shared model that predicts which ad requests deserve more compute, stacked allocators for region, hardware and host, and AI agents that propose tuning changes. The reported gains are real but coarse: a cumulative ~10% ranking improvement since 2017 and an offline-only 20% benchmark gain, with no compute-savings figure.
Nvidia’s board, Congress, and the president are reacting to a New York Times intelligence story that US agencies will not confirm on the record. Builders should harden executive-access controls without treating a successful secrets transfer as proven.
Six weeks after New Runtime’s earlier coverage, the case for Sonnet 5.5 on bounded everyday work still holds, but “new baseline” is too broad. Anthropic’s cheaper Haiku option and evidence that stale instructions add tool calls make workflow design part of the cost equation.
Prompt caching reuses a stable prompt prefix so later requests pay a discounted cache-read rate instead of full input. Anthropic is opt-in via cachecontrol (automatic top-level or up to 4 explicit block breakpoints); default TTL 5 minutes (optional 1h); writes cost 1.25x/2x base input and reads usua
Open Athena's architecture write-up gives the full recipe for a 535B-total, 23B-active MoE trained on 768 GB200s: a 3:1 local/global attention mix, 8-of-384 latent MoE, MuonH, and a switch to dropless expert parallelism mid-run. The final Paloma loss was 1.982 against a forecast of 2.039, though some of that gain comes from a contaminated eval subset.
AlphaProof Code used Gemini Carbon agents to write about 6.5 million lines of Lean about libsignal's AArch64 binary in four weeks. The proofs and specifications are open. The agent framework itself is not.
agents
security
gemini
No analyses match these filters.
Every analysis, by month
The whole archive on one page, newest first: 340 titles with their dates.