Kimi-K3 Technical Report [pdf]
Kimi Delta Attention gives up to 6.3× faster decoding at a million tokens; attention residuals buy ~25% training efficiency for under 2% overhead.
voilr
You can't read forty sources every morning. voilr can.
It reads every source you follow, drops the noise, summarizes what you care about, and writes one cited brief on what actually changed. It works with any source — including paywalled ones like the WSJ, Bloomberg, Substack and X.
- 93 sources tracked
- 10 kinds of source
- 40,025 posts read and scored
- 113 days of searchable archive
It signs in the way you do
A subscription site opens in a real browser carrying a real session, not through an API nobody offered. So what gets read and written up is the article itself — not the teaser above the wall. RSS is the easy half of this job.
It knows when it's the same story
One release reached the feed above as four headlines from two sources over nine hours. voilr worked out they were one story, kept the item worth reading, and threaded the rest underneath it. A repost never costs you a second read.
It finds the sources you're missing
voilr watches which outlets keep showing up in the links you read, pulls their recent articles, scores them with your filter, and offers a subscription only when the result would clear your own bar. No directory, no guessing.
It gets sharper the more you use it
Mark something irrelevant and voilr doesn't just hide it. The keep and discard rules its filter scores against are rewritten from what you keep, skip and follow — so the filter reading your sources in month three is not the one you started with.
Log in
Request access
voilr is in private beta. Leave your email and we'll send you an invite.
The brief · 24 June 2026
What mattered
- The rise of agentic engineering. A dominant theme across the day: the move from chat interfaces to agentic workflows — code as an operational substrate for reasoning, hierarchical orchestration engines for autonomous coding, and agentic loops that automate planning and execution. 34032
- Model efficiency and speed. Multi-token prediction reported at up to 53% faster inference, and surgical optimizations taking GLM-5.2 throughput nearly 20× on high-end hardware. 2053
Must read
- Computer use in Gemini 3.5 Flash. Native computer-use lands in the model — direct browser, mobile and desktop interaction for long-horizon tasks. 30
- LLMs use safety-specific neurons to find code vulnerabilities. Interpretability work finds dedicated detector heads rather than general vulnerability signatures. 5
FYI
- OpenAI Codex infrastructure. Inefficient logging is reported to be driving excessive SSD writes, and the cost with it. 13
Delivered every morning, to the browser, email or Telegram. Every numbered marker opens the post it came from.
The feed underneath it
Everything that cleared your bar, written up rather than linked — the article read in full, the discussion weighed, the claims checked against the archive. One story appears once however many sources carried it, and stays searchable months later.
KIMI K3'S WEIGHTS ARE OUT!