The October 22 engineering account uses Kimi K3, with roughly three trillion parameters, as an illustration. This is newly selected technical coverage, rather than a new model release.
Each WSE-3T wafer has 44 GB of on-chip memory. The example assumes four-bit weights, an eight-bit KV cache, the model’s stored working memory for a prompt, and 256,000-token average requests.
Its simplified sizing assigns about 16 GB of weights per layer. Two linear-attention layers fit per wafer because their request state stays fixed in size as prompts grow.
Full-attention layers need more memory as prompts grow, so each gets its own wafer. The example’s 69 linear-attention and 24 full-attention layers require 35 plus 24 wafers, or 59 altogether.
Weights stay on their assigned chips while intermediate results travel between them, with a claimed roughly two-microsecond network hop. Different requests occupy different stages concurrently to keep the pipeline busy.
Cerebras claims 1,000–2,000-plus output tokens per second for large models. Its account recommends comparing total output per unit of power at matched answer speed, rather than treating concurrency alone as efficiency.
Caveat: This vendor illustration supplies neither a reproducible end-to-end comparison nor a public Kimi service commitment, so the speed claim is not a deployment guarantee.
(HN 185 points · 62 comments)
An agent is a model running a multi-step task with tools. Scrawl’s MIT-licensed source packages marked screenshots, a transcript, interaction events and a structured manifest for one to inspect.
Its v0.3.13 release, published October 22 UTC, adds macOS 15 support and a custom-command transcriber. Apple transcription stays on-device, while a custom program controls its own data handling.
The desktop app supports Apple Silicon Macs on macOS 15 or later and an unsigned Windows preview. The site still lists Linux as forthcoming.
ScrawlKit records inside an iOS 18-or-later app, while cabled capture provides another iPhone route. The assistant can request more frames around a narrated moment instead of receiving every frame upfront.
Default cleanup moves raw media to Trash after seven days and whole bundles after thirty. Retained screenshots and transcripts remain sensitive, and automatic OpenCode delivery can be disabled.
Caveat: On-device transcription does not keep the eventual handoff private, because the receiving coding assistant may send those screenshots and words to its model provider.
(lobste.rs discovery · checked release, repository documentation and platform limits)
Custom developer tools need tests that outlast the demo#
Luu’s essay identifies code search, Rust compilation and filesystem operations as bottlenecks once models write code faster. His examples target those specific waits.
He reports roughly 50-times faster cold builds and twice-as-fast warm builds from a Rust interpreter experiment. He explicitly describes quick comparisons without rigorous performance testing.
He points to pgrust’s source and its author’s upstream bug reports as inspectable evidence of testing. One August 1 PostgreSQL report includes a query sequence exposing history-dependent date parsing.
These are older examples supporting a newly circulated argument, rather than newly discovered PostgreSQL defects. The essay also acknowledges pgrust’s premature preview submission to a performance comparison.
Its counterargument is operational: a narrowly optimized tool can overfit one workload and become hard to maintain. Luu likewise separates personal-use quality from the work required for a reusable product.
Caveat: This is a practitioner’s argument with selected examples and informal timings, so it establishes neither a general productivity multiplier nor production readiness for the linked database.
(HN 451 points · 397 comments · lobste.rs)
Kimi’s influencer-promotion controversy → UNVERIFIABLE: the report explicitly leaves the purported campaign document unauthenticated and its origin unresolved. It supports neither proven company misconduct nor replaying the release scores.
The smart-TV reverse-engineering thread → UNVERIFIABLE: the inspected threads supply account allegations and screenshots, without establishing the reason for suspension or whether detected audio was transmitted.
Apollo’s no-payoff claim → HYPE: the same aggregate comparison was already assessed yesterday. It cannot establish that every non-tech business received zero benefit.
The private Replit discount report → LOW_UTILITY: the unchanged report offers neither a generally available discount nor an inspected integration change.