AI Titus's Morning Wire

AI NEWS TODAY

A FREE MODEL THAT READS CODE, IMAGES AND VIDEO NOW FITS ON ONE GRAPHICS CARD
★ Must-Read / Watch agentic engineering / build-better-agents📚 Learn state of AI coding / useful technique

⚡ AI Frontier · smol.ai

★ Must-ReadGoogle open sourced a compiler that lets AI run on data it never gets to see
HEIR turns a normal model into one that computes on encrypted input, so the service returns an answer without ever holding the plain data. Google demoed it on fraud detection and intrusion detection. Until now this needed a cryptographer on staff.
★ Must-ReadAnthropic published exactly how the invisible watermark in Claude's text works
It steers the random choices the model makes between equally good words, leaving a pattern only a key holder can read. No extra characters, no quality hit, and nothing in it identifies you or your company. The EU AI Act's marking rule took effect August 2.
📚 Learn1Password released a free benchmark that measures whether your AI agent falls for a phishing page
Thirty scenarios across nine attack types, MIT licensed, scoring irreversible mistakes like typing credentials into a fake login form. One model did exactly that within ten seconds. Adding a 1,200 word security block to the system prompt took weaker models from 38 percent to 96 percent.
📚 LearnA small search agent did the digging so the expensive model only has to think
Mixedbread's Toast 1 breaks a question into sub-questions, checks sources, and hands back only the useful context. They report matching Opus 5 and GPT-5.6 Sol on search quality at up to 10 times cheaper and 12 times faster. The pattern is worth stealing even if you never use their API.
★ Must-ReadWriter rebuilt the layer that runs its agents and cut cost 41 percent on everyone's models, not just its own
The harness plans tasks, batches work, and hands pieces to sub-agents. Writer says the rewrite alone made tasks 44 percent faster across Anthropic and OpenAI models too. Paired with its new Palmyra X6 at $2 in and $8 out per million tokens, it claims 52 percent lower cost overall.
📚 LearnNetflix explained how it replaced a stack of recommendation models with one that speaks language
GenRec treats what you watched as a sequence to continue rather than a list to score. A rare look at a large company moving a core money-making system onto a language model.

📺 Watch · latest videos

AI News: ChatGPT Ultrafast, Grok 4.6, 3 New Open-Source Models, and more!
Latest from Matthew Berman.
Matthew Berman · 33.3K views · 1.1K likes
How to Build the Most Powerful System for AI Coding (Full Breakdown)
Latest from Cole Medin.
Cole Medin · 7.9K views · 318 likes
★ Must-WatchPreviewing Ultrafast mode: GPT‑5.6 Sol at up to 14X the speed
Today we’re previewing Ultrafast mode, a new service tier in the OpenAI API for GPT‑5.6 Sol. Powered by Cerebras, Ultrafast runs up to 14x faster than
OpenAI · 16.8K views · 480 likes
How Web Data Infrastructure Powers the Next Generation of AI — Patricija Žemaitytė, Oxylabs
Minutes into a call to demo a search API rebuilt to answer in under a second, the system got blocked, badly, in front of the client. Patricija Žemaity
AI Engineer · 877 views · 15 likes
Never Miss a Customer Call Again (AI Receptionist)
Latest from Nate Herk.
Nate Herk · 2.4K views · 52 likes
★ Must-WatchWhat does AI actually know about you?
The information you share with AI only travels as far as you let it. Zoe from the Anthropic education team explains what happens to your information w
Claude / Anthropic · 1.3K views · 46 likes
Meta's new model wants "deep access" to your personal life...
Latest from Fireship.
Fireship · 435.9K views · 12.1K likes

💻 Builder Desk · Hacker News

Qwen 3.8 27B
1223 pts · 724 comments
AI by Hand
335 pts · 24 comments

🗣 Voices & Blogs

★ Must-ReadMun Logadan On Why A Better Model Can Feel Worse To Work With
The argument is that training rewards confident decisions in ambiguous situations, which wins benchmarks and loses trust. A model that guessed less and asked more would score lower and help more. Worth reading before you blame yourself for the friction.
📚 LearnSimon Willison On Letting The Model Make Up Tags On Purpose
Doug Turnbull's trick, and it inverts the usual advice. Let the model invent whatever tag it wants, then match that invented tag to your real list with embeddings. No fixed vocabulary to maintain.
explainx On What Claude's Watermark Actually Proves
The mark says the text passed through Claude, not that Claude wrote it. Proofreading and translating leave the same mark as generating. If your company is about to write policy around detection, read this first.
📚 LearnYotta Labs On Running Today's Lead Model On One Machine
Ollama, GGUF files and single GPU setup, with the memory math laid out. The practical caution: a 17.9GB file does not fit in a 17.9GB budget once the cache and the operating system take their share.
Grok 4.6 matched Fable 5 Max at an 85% discount. Downloadable models set that price.
I’m Matt Burns, Chief Content Officer at Insight Media Group. Each week, I round up the most important AI developments, The post Grok 4.6 matched… · The New Stack

🌏 The Wire · Drudge / Breitbart

DATABRICKS TAKES $5 BILLION, INVESTORS TRIED TO HAND IT $15 BILLION
Valued at $190 billion, past a $7 billion revenue run rate, growing more than 80 percent year over year.
IBM PUTS THOUSANDS OF CONSULTANTS BEHIND OPENAI MODELS
A dedicated OpenAI practice inside IBM Consulting, aimed at banks, government, telecom and retail. Certified engineers sent to sit with the customer.
UBER AND PONY.AI PLAN 2,000 ROBOTAXIS ACROSS FIVE EUROPEAN CITIES
Zagreb already runs commercially and moves into the Uber app. Four more cities unnamed, with the Middle East next.
CHINA'S BIGGEST CHIPMAKER RAISES WAFER PRICES, FABS RUNNING AT 93.7 PERCENT
SMIC says incoming orders are far past its own forecast. If you buy anything with a chip in it, this is the quarter the AI boom starts showing up on your invoice.
ALIBABA'S BIGGEST MODEL SHIPS UNDER ITS OWN LICENSE, NOT APACHE
The 27B is genuinely Apache 2.0. The 2.4 trillion parameter Max carries a custom license that adds requirements once you get big. Read it before you build a product on it.
META PATENTS GLASSES THAT PUT A NAME TO EVERY FACE IN THE ROOM
Recognize the people in view, then cut a highlight reel of the dinner party. A patent is not a product, but it is a statement of intent.
ANTHROPIC LINES UP OUTSIDE MONEY TO BUILD ITS OWN DATA CENTERS
Macquarie Asset Management and GIC fund and own the sites, Anthropic leases them long term. Anthropic says it will cover the electricity price increases neighbors would otherwise absorb.

🤖 Trending Models · Hugging Face

unsloth/Qwen3.8-27B-GGUF
model · ★1K · 868K dl
Lightricks/LTX-2.5
image-to-video · ★905 · 378.4K dl

📈 Markets

NVDA 225.30 ▲0.5%
MSFT 496.88 ▲0.9%
GOOGL 346.36 ▲0.8%
AMZN 265.13 ▼0.8%
META 594.97 ▲2.8%
AMD 483.01 ▲0.0%
AVGO 417.82 ▲0.4%
PLTR 179.01 ▲4.7%
SPCX 141.29 ▼3.3%
TSLA 339.96 ▲3.8%

The BriefAlibaba's Qwen team put the weights for Qwen3.8-27B up for free download under the Apache 2.0 license, which is the permissive kind you can use commercially without asking anyone. It reads text, images and video, handles 262,000 tokens of context natively, and scores 61.7 percent on SWE-Bench Pro, a test of fixing real bugs in real repositories. The part that matters for a working person is where it runs: quantized to 4-bit it needs roughly 14 to 17GB of video memory, so a single 24GB card holds it. That means a capable coding and document model on hardware you own, with no per-token bill and no customer data leaving your network.

Level UpSpend twenty minutes learning where your AI coding spend actually goes. Anthropic published six concrete habits on Friday: clear the session between unrelated tasks, pick your model and effort level before you start instead of switching mid-conversation, point at files directly instead of making the agent hunt for them, quiet noisy commands, check what is already loaded in a fresh session, and compact before you walk away. Each one is a setting or a keystroke, not a project. Maximizing the value of your Claude Code sessions