AI Daily Dev — September 22, 2026

xAI Ships Grok 4.7 With 2.1T Params and Multi-Hour Agent Training

Sept 21: 2.1T-param base (up from 1.5T), 500K context, priced identically to 4.6 at $2/$6 per M tokens.
Trained on a longer RL run weighted toward multi-hour tasks; xAI calls it their 'agent stamina' model.
Benchmarks: 71.0% DeepSWE v1.1 (up from 65.2%), 46.3% CursorBench 4.0, 38.0% Terminal-Bench — still trails Fable 5.1 on the coding suites.
HackerBench v0.3 safety score: allows 3.3% of risky dual-use prompts, xAI's tightest guardrails yet.
HN thread: 490 points, 410 comments — debate is 'meaningful bump' vs 'still behind the frontier.'

models x.ai

Sept 21: Amazon started returning a Conditions-of-Use popup to any Muse-driven checkout after Meta refused a removal request.
Amazon says Muse hides its agent identity and retains customer credentials without disclosure.
Meta counters that Muse runs in a per-user VM and uses Stripe Link single-use cards — passwords stay in secure storage.
First public block of a frontier-lab shopping agent by a major retailer; sets the template for TOS-vs-agent fights.
HN + Bloomberg + TechCrunch all picked it up within hours; 140 pts, 148 comments on HN.

industry techcrunch.com

Open-source protocol where agents declare intent — symbol, api, schema, config, migration, contract — before touching code.
Catches semantic clashes Git misses, like two agents renaming the same symbol in different files.
0.4.0 ships a CLI, JSON API, MCP server, SQLite store, and deterministic conflict detector.
Landed the same week Linear's engineering blog argued CI is now the bottleneck for AI-coding-heavy teams.

tools github.com

Engineering write-up on how Cursor/Claude/Copilot volume broke Linear's CI economics: queue times up, feedback loops slower than the agents writing the code.
Their fix: per-PR test selection, cached graph builds, and a merge queue that admits by predicted risk, not FIFO.
129 pts, 125 comments on HN — the 'we all have this problem' thread of the day.
Reinforces the emerging thesis behind Foremerge and lossless-memory: the surrounding tooling, not the models, is the next bottleneck.

tools linear.app

John Hartnup's June post explodes on HN this weekend after making the point that ChatGPT will produce Bauhaus, Risograph, or cut-paper posters if you ask precisely.
The counter-argument in the comments: the skill floor is now taste and prompt discipline, not the model.
1,856 points and 942 comments — the #1 story across all of HN today, not just AI.
Serves as the counterexample the 'AI slop' discourse has been missing for six months.

community john.hartnup.uk

Personal AI memory layer that stores every conversation verbatim behind a vector index — no rolling summaries, no compression.
Pitched at audit trails, reproducibility, and debugging agent runs; storage cost is the main open question in the thread.
57 pts, 21 comments on HN in its first day; part of a small wave of 'context-window amnesia' tools this month.

open-source github.com