A 14MB model just shipped that runs on phones with no business doing inference. And it works. Not "works with caveats," works. That's the story this week: the floor keeps dropping while the frontier labs quietly cut prices because someone cheaper showed up. Both things are true at once, and both change what you build next.
The AIgent stays free for everyone, and reader support is what keeps it that way. Show your support for The AIgent so we can keep providing for and building the community together.
|
Sponsored
|
One message missed. One customer is gone.
Your customers are on Instagram, Messenger, and WhatsApp — and they leave when no one answers. Wati connects those channels into one inbox and helps you reply instantly with AI-powered automations, so no customer slips away.
The Drops
Reponeedle
5.5k stars · cactus-compute/needle A 14MB foundation model built to run on phones, wearables, smart home devices, and robots.
Here is the real problem it solves. every "edge AI" pitch this year still wants a GPU somewhere upstream. This one doesn't. It fits in the RAM budget of hardware that shipped a decade ago.
Quick start: clone it, point it at any ARM device you've got lying around, run the sample inference script before you touch a config file.
Catch: 14MB buys you narrow capability. Don't expect chatbot-grade reasoning, expect a classifier that never phones home.
AffiliateDify
When a 14MB model runs on a decade-old phone, the case for renting every layer of your stack gets thinner. Dify is an open-source platform for building LLM apps: visual workflows, retrieval pipelines, and agents in one place.
Use it for: standing up an LLM app without wiring the orchestration yourself, self-hosted or on their cloud.
We may earn a commission.
Repostable-diffusion-webui-forge
13k stars · lllyasviel/stable-diffusion-webui-forge A rebuilt, faster fork of the Stable Diffusion WebUI, optimized for lower VRAM and quicker iteration.
Use it for: local image generation on a single consumer GPU without the memory-thrashing the original UI is known for.
Catch: still a fork to maintain yourself, expect breaking changes when upstream Diffusion updates land.
MCPcodex-image-context-runtime
shixinnt/codex-image-context-runtime A Codex plugin and local MCP runtime for context-bounded image generation and inspection.
Use it for: giving Codex the ability to actually look at and generate images inside a bounded local context instead of blind text-only reasoning.
Catch: early and narrow, this is a runtime primitive, not a polished product.
RepoForeman
21 stars · Turki-Sh/Foreman A build playbook that turns your coding agent into the foreman of a website build.
Use it for: forcing design and scope decisions upfront, locking a visual system, then writing the build plan before your agent touches code.
Catch: 21 stars means unproven at scale. I'm flagging it because the pattern (interview before code) is the fix for the thing every agent-built site gets wrong first.
Start Here
Ask it to ask you questions before it answers.
Most people type a request, get something generic back, and decide the tool is the problem. Usually it is just missing two or three facts that only you have. So make it ask for them.
1. Type your request the way you normally would, then add one line at the end: "Before you answer, ask me any questions you need."
2. Answer what it asks, even the parts that feel obvious, like who this is for or how long it should be.
3. Ask it to say back what it now understands, in one short paragraph, before it writes anything.
4. Then tell it to go ahead.
TRY THIS: take the last thing you asked an AI that came back disappointing. Paste it in again, add "ask me questions first," and read the two answers side by side.
|
Recommended
Once you are getting sharper answers out of it, the next question is what to point it at. Latestly AI's free PDF, 100 Ways to Make Money with AI, is a solid idea bank: a five-minute weekly briefing on AI income opportunities, with the PDF landing in your inbox instantly. We may earn a commission. |
Frontier Signals
OpenAI and Anthropic cut prices this week. Chinese model providers are undercutting both on cost, and it's working well enough that the two labs with the most pricing power just moved first. If you're on a per-token budget, check your provider's current rate card before you renew anything. (Ars Technica AI)
Google will let you strip the visible watermark off AI-generated images. The invisible detection watermark stays, so this changes nothing for provenance tracking, it just removes the visual tell for anyone who wants a clean export. (TechCrunch AI)
Vercel's CDN now supports Encrypted Client Hello. If you're running anything behind Vercel DNS, this closes a real privacy gap: the domain you're connecting to is no longer visible in plaintext during the TLS handshake. Worth flipping on for anything client-facing. (Vercel Blog)
Meta is handing out big retention equity grants to keep engineers from leaving, and in some cases it is speeding up the exits. If you're hiring against that market, the signal isn't the dollar figure, it's that money alone isn't holding people at a lab that size right now. (Pragmatic Engineer)
Simon Willison shipped a patch release for sqlite-utils fixing a crash from a bad typing-extensions import. A small fix, but if you're running sqlite-utils in any pipeline, this is a drop-in update worth pulling now before it bites you at 2am. (Simon Willison)
|
Recommended reading
If you like The AIgent, a small group of operator-tier publications worth your inbox: see the shortlist. |
Community Pulse
This week on the Builders Board: what should we build next? Drop the one AI tool, skill, or workflow you wish existed. The most-wanted pick each week is the one we try to actually build.
Join the communityBefore You Go
We're heads-down building out The AIgent's engine this week, more coming soon on that front. Reply and tell me what you shipped with what's in today's Drops, I read every one.
See you Tuesday.
|
Want to reach builders shipping with AI every weekday? Advertise in The AIgent. |
