AI Titus's Morning Wire

AI NEWS TODAY

THREE DAYS TO PATCH: A HOLE IN THE SOFTWARE THAT RUNS AI TRAINING IS BEING ATTACKED RIGHT NOW
★ Must-Read / Watch agentic engineering / build-better-agents📚 Learn state of AI coding / useful technique

⚡ AI Frontier · smol.ai

★ Must-ReadCerebras put three wafer-sized chips in one box and says it beats GPUs by 30 times
The CS-4 launched August 19 and is the first machine built on the company's Nexus design. On the GPT-OSS-120B model it produced more than 4,400 words-worth of output per second for a single user. Cerebras claims up to 30 times the speed of GPU systems, twice the speed of its own last generation, and ten times more output per watt of power. It also uses half the parts of the machine before it, which matters because fewer parts means a simpler data center install. It is in early access now and generally available later this quarter.
Anthropic had Claude design proteins from scratch, and a lab confirmed 14 of 15 worked
Claude Opus 4.8 and a preview model produced 1,320 candidate designs against 15 biological targets. Two outside labs, Adaptyv Bio and Twist Bioscience, actually made and tested them: 354 worked, covering 14 of the 15 targets. Hit rates ran 22 to 35 percent against an industry norm of 10 to 15 percent. On one immune receptor, 72 of 90 designs bound. Read it with the limits attached, though. No independent review has been published yet, and a working binder is many years and several trials away from being a drug.
A skeptic with real credentials says the protein result is not the breakthrough it sounds like
Martin Shkreli, whatever else you think of him, has drug-development experience, and he called the work unimpressive. Worth reading next to the Anthropic post rather than instead of it. The disagreement is about what a binder actually proves: making a molecule stick to a target is the easy end of drug discovery, and the hard parts (stability, safety, does it help a patient) are all downstream. This is the shape of most AI-in-science claims right now.
The flaw itself: a browser trick beat Ray's only defense
CVE-2025-62593, scored 9.4 out of 10. Ray's dashboard tried to tell browsers apart from scripts by checking whether the request said it came from Mozilla. Browsers can change that line freely. Pair that with DNS rebinding, a trick that makes a bad website look like it is talking to your own machine, and a malicious page or even an ad can run code on a developer's laptop. Fixed in 2.52.0.
Linear says AI now writes nearly half the issues on its platform, and pull requests have doubled
AI authorship of issues went from 0.1 percent to close to 50 percent, and teams using AI agents tripled their weekly pull requests, with volume up 111 percent since 2024. Take the direction seriously and the precision less so. What it does not measure is whether the extra pull requests are good, and that is the number an engineering lead actually needs.
Pennsylvania just gave local towns a veto over AI data centers
Governor Shapiro signed an order August 18 making the state's GRID standards binding instead of voluntary. In his words, if the local community does not approve a project, the state will not either. Developers must pay their full electricity cost, source a real share from clean energy, hire locally, and sign a community benefits agreement. Nondisclosure agreements on these projects are banned outright, and data centers are out of the fast-track permitting program Amazon used for a $20 billion build last year.

📺 Watch · latest videos

📚 LearnThe State of Model Routing — NVIDIA, Cognition, OpenRouter
Featured by Latent Space.
AI Engineer
📚 LearnVercel accuses Cloudflare of stealing
Featured by Latent Space.
Theo - t3․gg
📚 LearnWF2026: Software Factories & Keynotes ft. Microsoft, OpenAI, OpenClaw, Z.ai (GLM), MiniMax, HF
Featured by Latent Space.
AI Engineer
From Ambient Documentation to Clinical Intelligence — Chaitanya Asawa, Abridge
Clinicians call it pajama time: the roughly two hours a day spent writing visit notes after work has finished. Abridge started there, and within two t
AI Engineer · 37 views · 9 likes
A Week of Grok Bot Lessons in 10 Mins
Latest from Nate Herk.
Nate Herk · 4.7K views · 208 likes
★ Must-WatchHow to Turn a Business Question Into a Strategy Deck With ChatGPT Work | Tutorial
Turn a business question into a clear recommendation—without spending days gathering research and building slides. In this demo, Arvind from OpenAI’s
OpenAI · 4.7K views · 143 likes
6 Open-Source AI Projects Trending NOW
Latest from Matthew Berman.
Matthew Berman · 58.4K views · 2K likes
FIXING Opus 5: PROOF that Prompt Engineering IS NOT DEAD
Opus 5 is one of the BEST models ever released, and one of the WORST to talk to. Verbose, "load-bearing" everywhere, torching your output tokens. So w
IndyDevDan · 22.2K views · 1K likes
The Ultimate Guide to Making Your Entire Development Cycle AI Native
Latest from Cole Medin.
Cole Medin · 7.4K views · 273 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 · 3K views · 93 likes

💻 Builder Desk · Hacker News

The Amazon tax
1268 pts · 648 comments
OpenLogi
1105 pts · 311 comments
Being ambitious and being a dad
695 pts · 512 comments
How does IKEA come up with names for its products?
404 pts · 276 comments
And then the men with guns tell you to do it anyway
370 pts · 283 comments
Scientists stunned by children's lung recovery in ultra low emission zone
345 pts · 291 comments

🗣 Voices & Blogs

★ Must-ReadDagmar Timler On What Actually Changed At The AI Engineer World Fair
The line worth stealing: verification is becoming more valuable than generation. Agent failures usually do not throw an error, because the run returns an answer and looks finished. Teams are now running long-lived reviewer agents that sample real production traces to find the failures nothing alerted on. If that sounds familiar, it is the same problem as a call that logs SUCCESS while the caller was told no.
The Decoder On Anthropic Handing A Language Model The Whole Protein Design Stack
The more careful write-up of the protein result, with the mode-by-mode numbers and the caveats kept in. Notably, the model did better one target at a time (35.1 percent) than juggling several at once (26.7 percent), which is a lesson that generalizes well past biology.
Lovelytics On Governance And Evaluation Being The Thing That Separates Agent Pilots From Production
Their argument: the biggest gains come from better context, not a better model, and specialized agents beat general ones wherever domain rules or compliance are involved. Start with one narrow workflow and a clear threshold for what counts as working.
Google's Developer Team On Five Things That Separated The Winning Agents In Their Bake-Off
A rare writeup where the agents were actually compared against each other rather than described. Useful if you are choosing between architectures instead of reading about them.
An open source rival to Claude Managed Agents just launched
AI infrastructure platform company TrueFoundry has launched its open source agent harness TrueForge. The technology is directly billed as an The post… · The New Stack

🌏 The Wire · Drudge / Breitbart

UNITREE OPENED 629 PERCENT ABOVE ITS IPO PRICE AND BRIEFLY HIT A $66 BILLION VALUATION
The humanoid robot maker priced at 150.8 yuan and opened at 1,100 on Shanghai's STAR Market August 19. It gave back some of that and closed up 460 percent, worth about $48 billion. Retail investors had oversubscribed the offer 8,000 times. It is the first listed humanoid robot maker in mainland China, so its price now sets the benchmark for 30 to 50 peers lining up to list.
CISA ADDED THE RAY FLAW TO ITS KNOWN-EXPLOITED LIST ON MONDAY, DEADLINE THURSDAY
Three days from listing to mandatory fix, under the newer BOD 26-04 rules that shortened patch windows specifically because AI speeds up exploitation. Federal agencies only, but the KEV list is what most private security teams copy their own priorities from.
OPENAI SAYS WATCHING ITS OWN MODELS NOW COSTS IT 20 PERCENT OF ITS COMPUTE
Reported August 19: monitoring overhead on advanced models, including the GPT-5.6 Sol class and Astra, runs about a fifth of inference cost. That is a real number to keep in mind when anyone quotes you a price per million tokens, because safety checking is not free and somebody pays for it.
GOVERNOR SHAPIRO CALLS IT THE NATION'S STRICTEST GUARDRAILS ON AI DATA CENTERS
Pennsylvania's order landed August 18. The clause worth noticing is the ban on nondisclosure agreements, which is what has kept most residents from learning what was being built near them until the concrete was poured.
TECH LAYOFFS IN 2026 HAVE ALREADY PASSED ALL OF 2025, WITH FIVE MONTHS LEFT
Salesforce, LinkedIn, Etsy, Zillow and Rapid7 are among the names, with automation cited as a driver. Treat the causation carefully: companies have an incentive to call a cost cut an AI efficiency gain, because one reads as strategy and the other reads as trouble.
BYTEDANCE AND TENCENT EACH TOOK DELIVERY OF ROUGHLY 10,000 H200 CHIPS, STAGED IN HONG KONG
First major shipments, held in Hong Kong outside mainland customs control. The staging location is the whole story, since it is the part that decides whether this counts as an export.
PEW: 52 PERCENT OF AMERICANS ARE NOW MORE WORRIED THAN EXCITED ABOUT AI
A majority for the first time, and the sharpest move is among young adults, who were the most positive group in 2021. If you sell AI to clients, this is the room you are walking into.
A STUDY SAYS THE USUAL RESEARCH METHODS MISS ABOUT HALF OF HOW PEOPLE REALLY USE AI
Roughly 48 percent of actual ChatGPT use gets filtered out of mainstream usage studies, including health questions, relationship talk and adult content. Every adoption statistic you have read is measuring the half people are willing to be surveyed about.

🤖 Trending Models · Hugging Face

Qwen/Qwen3.8-27B
image-text-to-text · ★11.4K · 1M dl
unsloth/Qwen3.8-27B-GGUF
model · ★1.9K · 4.3M dl
Lightricks/LTX-2.5
image-to-video · ★1.3K · 556K dl
deepseek-ai/DeepSeek-V4-Pro-0813
text-generation · ★619 · 37.6K dl

📈 Markets

NVDA 219.74 ▼2.3%
MSFT 481.63 ▲0.3%
GOOGL 344.20 ▲0.1%
AMZN 259.45 ▼0.7%
META 543.67 ▼4.4%
AMD 484.39 ▼4.3%
AVGO 380.00 ▼3.2%
PLTR 171.54 ▼0.6%
SPCX 143.34 ▼2.0%
TSLA 336.87 ▼0.7%

The BriefRay is the software many teams use to spread AI work across lots of machines. A flaw in it lets an attacker run their own code on your computer. All they need is for you to open a bad web page while Ray is running. The old versions tried to block this by checking if the request came from a browser, but that check is easy to fake. CISA, the US cyber agency, says attackers are using this now. It put the flaw on its must-fix list on Monday and gave federal agencies until Thursday, August 20. That is tomorrow. The fix is Ray version 2.52.0. Anything older is exposed. Here is the part that matters for an MSP: Ray usually is not on anybody's software list. A data scientist installs it on a laptop or a spare server to run one job, and it stays there. So the real question is not "did we patch Ray", it is "do we even know where Ray is running". That is the same question behind most AI tools showing up in your clients' networks right now.

Level UpSpend twenty minutes today finding out whether Ray is running anywhere you are responsible for, including your own machines. Do not start with the patch, start with the inventory. Ray listens on port 8265 for its dashboard, so a scan for that port across your networks will find most of it. Anything you find on a version below 2.52.0 needs updating. Then ask the bigger question this bug is really about: what other AI tools did somebody install for one experiment and never tell you about? Those are the ones with no patching schedule, no owner and no monitoring. Why CISA treated this as a shadow-infrastructure problem, not just a patch