Agents held the top of the papers chart for a fourth straight week, with benchmarks close behind; LongStraw's 2M-token RL run is the paper the community ranked first.
Agent-platform hiring rose again across the job boards, led by forward-deployed teams, and hallmark and OpenCut carried star velocity on the code side.
The bank added 110 questions this run, most of them LLM fine-tuning and system design, pulled from four new sources.
TOP MOVERS · CODE
+5.9KNutlope/hallmark
+5.6KOpenCut-app/OpenCut
+3.8KGraphify-Labs/graphify
+2.6KShubhamsaboo/awesome-llm-apps
+2.1K666ghj/MiroFish
TOP MOVERS · PULSE
+312A 7B open model closes the gap on coding evals
+198Agent frameworks converge on a shared tool protocol
+154Serving costs fall as quantized inference
+139Open-weight AI is having its Kubernetes moment
+121Kimi K3 is competitive with Fable on open evals
INTAKE · LAST 30 DAYS
JUN 25JUL 2JUL 9JUL 16JUL 24PULSEJOBSCODEPAPERSQUESTIONS
Papers
UPDATED 4H AGO
PAPERS BY TOPIC
Latest
Past 7 days
Past 30 days
HUGGING FACE DAILY
All sections
HF daily
cs.LG
cs.CL
cs.CV
cs.AI
cs.RO
stat.ML
cs.NE
cs.IR
cs.MA
cs.SD
eess.AS
eess.IV
Code
UPDATED 4H AGO
Latest
Past 7 days
Past 30 days
NEW REPOS & MODELS BY TOPIC
STAR VELOCITY
+5.9KNutlope/hallmark
+5.6KOpenCut-app/OpenCut
+3.8KGraphify-Labs/graphify
+2.6KShubhamsaboo/awesome-llm-apps
+2.1K666ghj/MiroFish
+1.4Kopeninterpreter/openinterpreter
+1.2Kaaif-goose/goose
+0.9Kusestrix/strix
FEED
528K ★codecrafters-io/build-your-own-xgithub.com · 2h ago
119K ★rustdesk/rustdeskgithub.com · 3h ago
114K ★microsoft/generative-ai-for-beginnersgithub.com · 4h ago
The cache stores a key and a value vector per token, per layer, across all KV heads: 2 × layers × kv_heads × head_dim × seq_len × batch elements, times bytes per element.
The visual deep dive, forever: biweekly on Free. PRO steps it up to weekly, opens the full archive, and adds the intelligence layer, the interactive articles, and the roadmaps on top.
How is this different from following arXiv or another newsletter? +
A feed shows you everything and ranks nothing. PRO watches the whole field, shows what actually moved, and connects the papers, repos, postings, and interview questions to each other, then narrows it to what matters for the work you do.
How often do the trackers update? +
Continuously. The pipeline pulls from hundreds of sources every day and re-ranks on each run, so what you see reflects the current state of AI.
How is the data selected? +
We monitor hundreds of sources continuously, then a proprietary ranking algorithm using 50+ factors decides what surfaces. Items are attributed: every interview question carries its company, role, and when it was asked, plus an answer and working code where it applies. Your preferences decide what leads.