Saturday morning · six from the feeds
11July
2026
Today the tools get sharper: reviews that cost less to run, inference that answers sooner, and an image model that edits like a designer.
Better tools made the code review worse.
GitHub handed Copilot's reviewer the same Unix-style grep/glob/view tools its CLI uses — and quality dropped, because generic instructions sent the agent wandering the whole repo instead of reading the diff. The fix was the prompt, not the toolbox: start from the diff, narrow the evidence, open files only when needed. Reviews came out about 20% cheaper at the same quality.
$ copilot review --pr 4821
▸ before: glob **/* pulled 1,904 files into context
▸ after: start from the diff → 41 files, read on demand
▸ tokens / review ......... ▼ ~20%
✓ review posted · quality held
Session replays that file their own pull requests.
PostHog's Replay Vision watches recordings, spots where users struggle, works out the root cause, and opens a pull request. In one run it went from recording to opened PR — writeup included — in about ninety minutes. The pile of sessions nobody watches becomes a queue of fixes you can merge.
Recording → opened pull request, root cause attached
- Watches
- every session
- Finds
- the struggle
- Files
- the fix
An image model that edits like a designer.
ByteDance's Seedream 5.0 Pro brings point-, lasso-, and material-level control to text-to-image: mark a region and it changes only that region, leaving the rest coherent. It also splits a scene into ten-plus transparent layers you can drag and recompose, and renders text across a dozen languages — Spanish accents, Arabic script, and all. The output stops being a flat picture and starts being a file you can open.
Same answers out — up to 85% sooner.
DeepSeek open-sourced DSpark, a speculative-decoding framework that drafts tokens with a small parallel model and verifies them in a single pass on the big one — 60–85% faster per user on V4, with a scheduler that verifies more tokens when GPUs sit idle. The acceptance rule preserves the target model's distribution, so the output is statistically identical to plain decoding. The training code, DeepSpec, ships MIT-licensed.
// per-user generation, matched throughput · lossless vs. target
Chase the tangent without dropping the thread.
Cursor 3.11 adds side chats — a parallel conversation next to your main one, so you can test an idea or ask a throwaway question without derailing the agent already at work. It also searches past agent transcripts by what was actually said, not just their titles, and scopes the project and repo pickers so you land in the right place faster.
"The main thread keeps working while you think out loud in the margin."
- Side chats running alongside the main agent
- Search transcripts by content, not titles
- Scoped project & repo pickers
Cache that parks next to your origin.
Cloudflare's Smart Tiered Cache now takes a hint about where your origin actually lives — an AWS, GCP, Azure, or Oracle region — and picks an upper-tier data center close to it. Fewer long hauls back to origin on a miss means lower latency and less egress on the requests that get through. It's a one-setting change for anyone serving from a public cloud.
| Origin region | Upper tier picked | Miss path |
|---|---|---|
| aws · us-east-1 | IAD | short hop |
| gcp · europe-west4 | AMS | short hop |
| azure · japaneast | NRT | short hop |
| oracle · ap-sydney-1 | SYD | short hop |
That's Saturday.
Six picks the morning the tooling tightened — cheaper reviews, replays that file fixes, region-level image edits, faster-but-identical inference, side chats, and cache that stays close to home.