"Stack Mac Minis at home" is just the mainframe argument wearing a turtleneck. There was a moment for it — maybe two years ago, when the only models worth running were too large for a single device to load. That moment passed. Machines got better. Models got smaller. @thewebAI

??? Already

Founders win by staying curious. Curiosity rests on humility. Nothing is personal—it’s all information. Take weight out of the room, and direct stress where it actually needs to go.

LLMs are the new disk drives: commodity infrastructure you hot-swap for whoever’s cheapest + best. The fantasy that the model is a moat just expired. ☠️

Modern generative AI is just executing a kinematic equation. It treats intelligence as a ballistic trajectory. The model activates all neurons equally, or worse, uses a central router (MoE) to assign work. This is Central Planning. No central router can effectively manage the

The next wave of AI isn’t about bigger models in the cloud — it’s about smaller, faster, more specialized intelligence running close to the user and the data.

Knowledge Graph RAG running on iPad. Locally. Privately. Offline. Imagine you're an aircraft technician with no internet. You need answers NOW. Query maintenance manuals, technical specs, safety protocols in real-time. No cloud. No waiting. Just intelligence when you need it.

Saying that foundational models are the only viable AI investment is like saying search engines are the only viable internet companies — ignoring the vast ecosystem of apps, deep tech platforms, and services built on top. very wishful thinking

From zero to 100% adoption in one hour. Working with @springshot_com, we deployed a computer vision model across @SpiritAirlines' entire network—validating fire suppression line compliance on 500+ flights per day. The entire deployment took less than an hour. No friction.

Proud of the work @thewebAI is doing with @SpiritAirlines and @springshot_com. "In one hour, we achieved zero to 100% adoption validating fire suppression compliance for more than 500 flights per day across Spirit Airlines' entire network ... webAI brought the ability to run

When the bubble pops, it will expose the inherent limitations of the cloud-centric paradigm. This model flourished in the internet age, where data was largely untethered from immediate user locality and could be centralized without severe real-time penalties. In AI, however,

Today the cloud went down. AWS outage → global ripple. Forbes piece → on-prem is the future with webAI. webAI → built for that world. Resilient. Private. Everywhere. https://t.co/etyXMJFaxi

Where are all the pro-Palestine supporters? Hamas slaughters innocent civilians after a peace deal, and free Palestine actors are quiet. The most intellectually dishonest group of people I have witnessed in recent affairs.

https://t.co/fSbgpbmtkv

New paper from @thewebAI shows a great way to make RAG much faster and more accurate by freezing a small language model, then only training a tiny adapter and a classifier head. Result: 99.95% top-1 with soft embeddings + classifier vs 12.36% for frozen LLM + MIPS, plus 2.62x

We’re growing our public sector leadership with 5 key hires - Scott McKeever, Eric Volmar, Jared Evans, @rdhjr, & Ryan Castonia Together, we’re building AI that government agencies can fully own and operate - secure, mission-ready, and cloud-independent. This is trusted AI where

We just published research that redefines what Retrieval-Augmented Generation (RAG) can be. Instead of patching today’s broken retrieval stack, we rebuilt it: ✅ Retrieval as learning — a classifier replaces heuristic vector search ✅ Soft embeddings — a tiny pre-block adapter

We asked @davidstout (Co-Founder & CEO, webAI) about the advantages of running AI on devices. "There's no free lunch, but there are things cloud players can't do. It's RAM intensive, yes, but by distributing across devices, we can unlock more accuracy and context. But if you're

Morning. On today’s show: – Delian and Everett debate – GPT-5, one week later – China’s AI training woes – Should Trump buy Intel? Guest call-ins: – @zebulgar & @EverettRandle (Founders Fund / Kleiner Perkins) – @niubi (Sinocism) – @ohlennart (RAND) – @davidstout (webAI) –

On the latest TinyLlama 1.1B benchmarks (MacBook Pro M4): Unquantized: ~8× faster than llama.cpp 8-bit & 4-bit quantized: still leading — and getting faster month over month From July → August, every variant of webframe-cpp improved its speed.