AI Titus's Morning Wire

AI NEWS TODAY

AN AI AGENT JUST BEAT ALL 183 LEVELS OF ARC-AGI-3, THE PUZZLE TEST BUILT TO STUMP MACHINES. NVIDIA'S AVO SHOWS THE SCAFFOLDING AROUND THE MODEL NOW MATTERS AS MUCH AS THE MODEL
★ Must-Read / Watch agentic engineering / build-better-agents📚 Learn state of AI coding / useful technique

⚡ AI Frontier · smol.ai

★ Must-ReadClaude designed working protein binders against 14 of 15 drug targets, and outside labs proved it
Anthropic had Claude run an entire protein design pipeline on its own, from researching each target to picking tools and ranking candidates. Two independent labs, Adaptyv Bio and Twist Bioscience, then built and tested the designs. Hit rates ran 22 to 35 percent, against the 10 to 15 percent typical for human-led campaigns. AI agents doing real, checkable science matters more than any benchmark score.
★ Must-ReadOpenAI cut GPT-5.6 Sol prices to $4 in and $20 out per million tokens, through at least November 21
Posted August 21: input pricing fell 20 percent and output 33 percent for at least three months. If you run anything on GPT-5.6 Sol, your bill just shrank with no code change. Frontier prices keep sliding, so it pays to re-check your model choices every quarter.
DeepSeek's cheap Flash model can now see images, and DeepSeek says its agent scores come close to Opus 4.8
Released August 21 as deepseek-v4-flash-vision-exp on the API. It reads documents, charts and screenshots at Flash prices, and on multimodal agent benchmarks DeepSeek reports a big jump over its last version. Cheap vision changes the math on what is worth automating, like reading invoices or screenshots in a ticket.
★ Must-ReadNevada cleared up to 8,000 robotaxis from Tesla, Waymo and Uber over the next 12 months
Approved August 20 for Clark County, home of Las Vegas: 5,000 permits for Tesla, 1,000 for Waymo and 1,000 for Uber's partners. Tesla's own engineer expects to field about half its permit in year one. Either way, self-driving cabs are about to feel normal in a major US tourist city.
📚 LearnPew: 1 in 10 webpages shows clear signs of AI writing, and it is over 1 in 3 for pages made after ChatGPT launched
Pew ran nearly 500,000 English webpages through AI detection. The tells include em dashes and pet AI vocabulary. Worth reading twice: once before you trust a random article, and once before you publish text that sounds like a machine wrote it.
Starcloud raised $250 million to put AI data centers in orbit
The startup, now valued at $2.3 billion, builds satellites that run AI inference in space and plans up to 88,000 spacecraft riding SpaceX's Starship. Power and land for data centers are so tight that orbit is now a serious answer, not science fiction.

📺 Watch · latest videos

📚 LearnAI Research Legend’s Honest Assessment of Where We Are
Featured by Latent Space.
Unsupervised Learning: With Jacob Effron
📚 LearnWF2026: Autoresearch & Keynotes ft. Anthropic, Google DeepMind, Amazon AGI, Sonar, Arena, Recursive
Featured by Latent Space.
AI Engineer
📚 LearnThe State of Model Routing — NVIDIA, Cognition, OpenRouter
Featured by Latent Space.
AI Engineer
Everything Goldman Sachs Taught Me About AI (In 10 minutes)
Latest from Nate Herk.
Nate Herk · 440 views · 42 likes
Intelligence EXPLOSION: Harness Engineering with Pi Agent, Deepseek, and Gemini
The Intelligence EXPLOSION is HERE. 🔥 Five plus model releases in five days, LLM pricing wars in full effect, and most engineers are still locked insi
IndyDevDan · 876 views · 62 likes
Grok Bot saves so much time and is so easy
Latest from Matthew Berman.
Matthew Berman · 1.6K views · 22 likes
★ Must-WatchWhat Codex Unlocks for NTT Data
After introducing Codex, adoption grew to more than 10,000 active users and became one of the largest internal community in just a few months. The NTT
OpenAI · 3.8K views · 111 likes
CODEX FOR BUSINESS FULL COURSE (4.5 HOURS)
Latest from Nick Saraev.
Nick Saraev · 10.4K views · 551 likes
The Agent Behind the Curtain: Building the Oz Cloud Agent Platform — Safia Abdalla, Warp
Warp open sourced about three months ago and went from roughly 20,000 GitHub stars to over 60,000, with thousands of pull requests and hundreds of con
AI Engineer · 1.9K views · 35 likes
DeepSeek is back... and Silicon Valley is terrified
Take BlueDot's Future of AI course for free - https://bluedot.org/fireship OpenAI just paused its biggest training run in history just as things were
Fireship · 929.7K views · 29.2K likes

💻 Builder Desk · Hacker News

ElevenLabs, TwelveLabs, ThirteenLabs
438 pts · 132 comments
Hister – A private, full content search index that you control
423 pts · 94 comments
Why your local LLM feels dumber than it is
417 pts · 171 comments
Scrap (2006)
386 pts · 212 comments
typ.ing
321 pts · 102 comments
To become a better writer, read as much as you can
316 pts · 208 comments
Thinking in Python
251 pts · 50 comments

🗣 Voices & Blogs

★ Must-ReadSimon Willison: Linus Torvalds credits AI help on a stubborn kernel debugging problem
August 22. The Linux creator, long a loud AI skeptic, says AI assistance helped him crack a persistent debugging problem. When the most famous code reviewer on earth finds the tool useful, the claim that it is all hype gets harder to hold.
📚 LearnAdaptyv Bio's wet-lab case study, the independent half of the Claude protein story
August 19. The lab that actually built and tested Claude's 1,320 protein designs walks through the numbers, including a 26.8 percent hit rate that beat human-led design competitions. Read this one because it is the outside check, not the vendor's own telling.
Zvi Mowshowitz: AI text watermarking is free and good
August 21. He walks through Scott Aaronson's watermarking work and argues invisible marks on AI text cost almost nothing and help everyone. Pairs neatly with Pew finding a third of new webpages showing AI fingerprints.
Dr. Alex Wissner-Gross: Welcome to August 23, 2026
A weekend sweep of the whole frontier: US versus Chinese models on cost, chip and energy crunches, robotics and where AI autonomy is heading. A good single read if you skipped the news all week.
Thomson Reuters trained its own AI model. Then it kept using Anthropic’s anyway.
Thomson Reuters has developed its own AI model for legal, tax and compliance work, trained on the company’s proprietary professional The post Thomson… · The New Stack

🌏 The Wire · Drudge / Breitbart

ALIBABA LAUNCHES A $10.2 BILLION SHARE SALE TO FUND CHIPS, DATA CENTERS AND AI MODELS
Announced as the week opens, and it could be the largest primary follow-on offering ever by a Hong Kong listed company. Also on deck this week: Nvidia reports earnings Wednesday, with about $92 billion in quarterly revenue expected.
NVIDIA MAKES FOUR AI BETS IN A SINGLE DAY: POWER LAND, ORBITAL COMPUTE, TRAINING DATA AND A $6 BILLION MODEL LICENSE
In one Thursday: a stake in power developer Cloverleaf, a piece of an orbital data center round, funding talks with training-data firm Mercor at a reported $20 billion, and a $6 billion license to Poolside's Model Factory plus a $1 billion investment. Nvidia is buying into the whole supply chain, not just selling chips.
MICRON COMMITS $10 BILLION TO A US RESEARCH HUB FOR THE MEMORY THAT AI RUNS ON
Announced August 20, planned over the next decade, based in Boise. Memory, not just GPUs, is a choke point for AI compute, and Micron wants the next generation of it invented in America.
WAYMO REVEALS A CUSTOM 5NM CHIP THAT FUSES LIDAR, RADAR AND CAMERAS IN REAL TIME
Over 1,000 TOPS of machine learning performance, engineered only for driving. Like Google before it, Waymo decided the chip it needed did not exist yet, so it built its own.
AI TRAINING-DATA STARTUP MICRO1 HITS A $500 MILLION RUN RATE, UP 5X IN EIGHT MONTHS
The labs' hunger for human-made training data is minting whole companies: rival Mercor is at $2 billion annualized. Teaching AI is quietly one of the fastest-growing businesses in tech.
DUTCH REGULATOR FINES UBER 825 MILLION EUROS FOR FIRING DRIVERS BY ALGORITHM
August 21. Uber deactivated driver accounts automatically, without real human review or a fair way to contest it, which breaks EU data rules. A loud warning for anyone letting software make decisions about people unsupervised.

🤖 Trending Models · Hugging Face

Qwen/Qwen3.8-27B
image-text-to-text · ★12.4K · 2.6M dl
unsloth/Qwen3.8-27B-GGUF
model · ★2.8K · 7M dl
OBLITERATUS/Qwen3.8-27B-OBLITERATED
text-generation · ★673 · 312.6K dl
HauhauCS/Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-MTP-GGUF
image-text-to-text · ★562 · 762K dl

📈 Markets

NVDA 214.72 ▼1.0%
MSFT 483.24 ▲0.4%
GOOGL 344.82 ▲1.2%
AMZN 258.63 ▼0.6%
META 549.90 ▲0.7%
AMD 473.25 ▲0.8%
AVGO 368.45 ▲1.2%
PLTR 179.94 ▲3.4%
SPCX 136.97 ▲2.2%
TSLA 362.86 ▲5.1%

The BriefNVIDIA built an agent system called AVO that just scored 100 percent on ARC-AGI-3, a set of video-game puzzles designed to stump AI. The agent ran for hours on its own, using Claude Opus 5 as its brain plus memory, tools and self-checks wrapped around it, and it even used fewer moves than earlier systems. The lesson for anyone putting AI to work: the harness around the model, the memory, the supervision, the tools, now decides success as much as the model itself.

Level UpAnthropic just opened Claude Academy, a free learning site with more than 20 courses on Claude, Claude Code, the API, MCP and building agents. You do not need an Anthropic account to use it. Pick one course this week, the agent skills or sub-agents ones are a strong start, and actually finish it. Thirty minutes a day beats another week of feeling behind. Claude Academy, the free official course catalog

Caught Up?
  • Attackers are actively exploiting a critical hole in MLflow, a common AI model tool. CISA says patch to 3.15.0 or newer now.
  • Anthropic moved Computer Use plus the Skills and Files APIs out of beta, so click-and-type agents are now production plumbing.
  • ChatGPT on Mac can now read, search and send your text messages once you grant it access. Pause before a whole office enables that.
  • Z.ai shipped GLM-5.3, a coding model so good at finding security bugs that its open weights were held back for a safety review.
  • An independent nonprofit graded the big AI labs on their safety controls. Nobody scored above a C+.