AINews is a daily newsletter that summarizes top AI discords, reddits, and X/Twitter posts, delivering a comprehensive roundup to AI engineers every weekday.
- Daily AI news roundups
- Curated from top AI communities
- Easy to read and digest
- Focused on practical AI engineering content
- Clone the repository:
git clone https://github.com/smol_ai/ainews-web.git- Install dependencies:
cd ainews-web
pnpm install- Start the development server:
pnpm dev- Build for production:
pnpm build- daily - bash updates
pnpm tsx oneoffs/process-emails.ts --file "$(ls -t src/content/issues/*.md | head -n 1)" && gadmit "latest post" && gpomTo keep Vercel builds under the 8 GB build-memory limit, issues before 2026 are frozen as static HTML instead of being fully rendered through Astro content collections on every build.
- Metadata for pre-2026 posts stays in
src/content/issues/so listings, tags, RSS metadata, and search facets still work. - Full pre-2026 Markdown bodies are preserved in
src/content/frozen-issues/. - Static HTML snapshots are generated into
public/frozen-issues/byscripts/generate-frozen-issues.mjs. pnpm buildruns the frozen-issue generator beforeastro check && astro build.- Pre-2026
/issues/<slug>/routes redirect to/frozen-issues/<slug>.html; 2026+ issues continue to render through Astro normally. src/content/oldissues/was removed because it duplicated archived content and inflated the Astro content store.
If you edit an archived pre-2026 issue, update the preserved file in src/content/frozen-issues/, then run:
node scripts/generate-frozen-issues.mjs
pnpm buildRecent build timing after freezing: frozen HTML generation takes about 45 seconds for 538 archived posts, Astro content sync takes under 1 second, and a full production build completes in about 2–3 minutes locally instead of failing with Vercel out-of-memory errors during server entrypoint bundling.
AINews includes a newsletter subscription functionality using Resend. To set up the subscription feature:
- Create a Resend account at resend.com
- Set up an audience in Resend to collect subscribers
- Add the following environment variables to your
.envfile:
RESEND_API_KEY=re_your_api_key_here
RESEND_AUDIENCE_ID=your_audience_id_hereThe subscription system features:
- A dedicated
/subscribepage (with/signupalias) - An embedded form on the homepage
- Form validation and submission animations
- Success and error state handling
The production site is deployed on Vercel under the smol-ai scope as ainews-web-2025 and serves news.smol.ai.
Use the checked deploy script for normal production deploys:
pnpm deploy:vercelWhy this is preferred:
- It forces the global Vercel CLI on this machine (
/Users/shawnwang/Library/pnpm/vercel) so deploys do not accidentally use the old repo-local Vercel CLI. - It targets the production project explicitly:
smol-ai/ainews-web-2025. - It verifies
.vercel/project.jsonis linked toainews-web-2025before building, so a deploy cannot silently go toswyxs-projects/ainews-web. - It runs
vercel buildlocally and deploys the generated.vercel/outputwithvercel deploy --prebuilt. - It uses
--archive=tgz, which avoids Vercel's per-file upload limit (api-upload-free) when the output contains thousands of static issue/search files. - It validates and repairs the Vercel Build Output route order so static files are served before generated 404 routes.
- It rewrites known Cloudflare-poisoned asset URLs to versioned paths so cached 404s do not keep breaking CSS after a good deploy.
Observed timing from the April 2026 fix:
- Frozen issue generation: about 45 seconds for 538 archived pre-2026 issues.
- Full local Vercel build: about 160–165 seconds.
- Prebuilt upload: about 27–30 seconds for a 162–163 MB tgz upload.
- Vercel output deployment/aliasing: about 70–75 seconds.
- End-to-end
pnpm deploy:vercel: about 236–240 seconds.
The script reports elapsed time for each major step and the total runtime.
Override the Vercel scope if needed, but production should normally stay on smol-ai:
VERCEL_SCOPE=smol-ai pnpm deploy:vercelOverride the project name only for intentional one-off deploys:
VERCEL_PROJECT=ainews-web-2025 pnpm deploy:vercelOverride the Vercel binary if you have already verified your shell resolves to a modern CLI:
VERCEL_CLI=vercel pnpm deploy:vercelIf the output mentions swyxs-projects/ainews-web, stop. That is the wrong Vercel project and will not update news.smol.ai. Relink the repo before deploying:
/Users/shawnwang/Library/pnpm/vercel link --yes --scope smol-ai --project ainews-web-2025Expected .vercel/project.json shape:
{
"projectName": "ainews-web-2025"
}If vercel build already succeeded and you only changed .vercel/output post-processing, redeploy the prebuilt output directly:
/Users/shawnwang/Library/pnpm/vercel deploy --prebuilt --prod --scope smol-ai --target production --archive=tgzExpected timing:
- Upload: about 27–30 seconds for the current 162–163 MB tgz.
- Vercel output deployment/aliasing: about 60–75 seconds.
- Total: about 1.5–2 minutes.
Use this only when .vercel/output is already current. It skips pnpm build, so it will not regenerate frozen issues, Pagefind, Astro output, or route/asset patches from the source scripts unless those patches have already been applied to .vercel/output.
Avoid using raw vercel --prod --scope smol-ai for this repo unless you are intentionally debugging Vercel itself.
Problems with the raw command:
- It can use the locally linked project from
.vercel/project.json; if that link points atswyxs-projects/ainews-web, the deploy will succeed butnews.smol.aiwill not change. - Without
--archive=tgz, it can hit Vercel's upload file-count/rate limit for this large static output. - It can rebuild remotely on the Vercel 8 GB build machine, which previously failed out-of-memory while bundling server entrypoints.
- It does not run this repo's route-order and cache-busting checks before upload.
If you must run a raw deploy, prefer the prebuilt, archived form after a successful local build:
pnpm build
/Users/shawnwang/Library/pnpm/vercel deploy --prebuilt --prod --scope smol-ai --target production --archive=tgzvercel.jsonintentionally does not setoutputDirectory; the@astrojs/verceladapter emits Vercel Build Output under.vercel/output.src/pages/api/og-alt.tsis server-rendered (prerender = false) to avoid build-time network fetches from the OG image renderer.- Vercel project settings should show Node.js 22.x, framework preset Astro, root directory
., and no output directory override. - The build output is large by design because pre-2026 issues are frozen as static HTML and Pagefind indexes thousands of pages. Recent output reports were about 529–530 MB on disk, compressed to about 162–163 MB for tgz upload.
- If production returns a Vercel
NOT_FOUND, confirm the deploy is targetingsmol-ai/ainews-web-2025, not a personalswyxs-projects/ainews-webproject. - If production HTML loads but CSS 404s, check for Cloudflare-cached 404s on immutable asset paths. The deploy script currently rewrites known affected assets to versioned URLs; if new hashed asset paths are poisoned, bump or version those URLs before redeploying.
After deploy, check the live page and the assets it references:
node -e 'const base="https://news.smol.ai"; const html=await (await fetch(base+"/",{headers:{"cache-control":"no-cache"}})).text(); const assets=[...new Set([...html.matchAll(/\/(?:_astro\/[^"'"'"'\s]+\.(?:css|js|woff2?)|pagefind\/[^"'"'"'\s]+\.(?:css|js)(?:\?[^"'"'"'\s]+)?|favicon\.ico(?:\?[^"'"'"'\s]+)?)/g)].map(m=>m[0]))]; for (const path of assets) { const res=await fetch(base+path,{headers:{"cache-control":"no-cache"}}); console.log(path, res.status, res.headers.get("content-type")); }'Expected result: all referenced CSS, JS, font, Pagefind, and favicon assets should return 200.
The site can also be deployed to Vercel or Netlify with a single click:
pnpm tsx oneoffs/process-emails.ts --file "$(ls -t src/content/issues/*.md | head -n 1)" && gadmit "latest post" && gpom
## or
pnpm tsx oneoffs/process-emails.ts --file "$(ls -t src/content/issues/25-08-13*.md | head -n 1)" && gadmit "latest post" && gpom
