AI News #151: Week Ending August 21, 2026 with 35 Executive Summaries
About This Week’s Covers
This week’s cover theme was inspired by a recent stay at the InterContinental Hotel in Chicago on the Magnificent Mile. My window looked out on the backs of the heads of three sculptures of warriors, most likely created by Leon Hermant’s studio a few blocks away. They’re made of Indiana limestone and were made in 1928. They sit in the twelfth-floor setback on the Michigan Avenue side, directly above the main entrance of the hotel.
The original photo, sadly with no pool.
The InterContinental began its life as the Medinah Athletic Club. One of the highlights of the hotel is a surprise junior Olympic-size pool from 1929 on the 14th floor. I would wager that at least half of the hotel guests don’t even know it’s there.
For the cover this week, I decided to have some old-school generative AI fun, and I gave myself a pool on my balcony, outside my window, behind the Sumerian warriors. I used GPT images, and it came out pretty well, including the shadow treatments. The title text is in the style of Marriott, which is across the street (sorry, Intercontinental)…
For the category covers, I had Gemini create similar helms on building facades. I included my favorites below:
This Week’s Humanities Selections
This week’s humanities reading is not tied to the week in particular. It’s a poem written in 1877 by the Jesuit priest Gerard Manley Hopkins titled “Pied Beauty”. I liked it for this AI newsletter because while most poetry praises perfection or symmetry or beauty, Hopkins’ poem celebrates things that are multicolored, irregular, and flawed. He sees the beauty in our humanity and the strange and fickle.
Pied Beauty By Gerard Manley Hopkins
Glory be to God for dappled things – For skies of couple-colour as a brinded cow; For rose-moles all in stipple upon trout that swim; Fresh-firecoal chestnut-falls; finches’ wings; Landscape plotted and pieced – fold, fallow, and plough; And áll trádes, their gear and tackle and trim.
All things counter, original, spare, strange; Whatever is fickle, freckled (who knows how?) With swift, slow; sweet, sour; adazzle, dim; He fathers-forth whose beauty is past change: Praise him.
This week’s song is tied to the week, and it reminds me of my good friend Mike Bernstein, who died just over a year ago. The song is Miles Davis’ “Blue in Green”, from the iconic album “Kind of Blue”, which came out on August 17th, 1959. Not only did the album come out this week, but there was a credit dispute as to the authorship, which feels apropos for an AI newsletter. The song has a 10-measure circular form, so it never really lands where you expect a phrase to end. Miles Davis says he wrote it, but Bill Evans and his peers claimed that Bill wrote it. “Kind of Blue” is a truly great album if you’re ever just looking to put on some headphones and drift away for a few minutes.
This week I organized 405 links. 45 of them informed 34 executive summaries. I’m eight weeks behind because I took the entire summer off, including May, to be with my family. I truly empathize with anyone who is reading this in a time-shifted way as I catch up.
This week’s news is organized by topic and frontier lab. I’ve put the most interesting stories at the top, and the rest are sorted by category. If you want to skip my rambling summaries, you can scroll down, and you’ll find all of the top links in a nice organized list that you can quickly skim.
But first, let’s touch some grass! After my daughter Rori and I climbed to the top of Grand Traverse Peak, we walked around the town of Vail for a day to celebrate our achievement. We took some time to recreate some old photos that I had from 1994 at my old house and a couple of other locations. We made a short pilgrimage to the Red Lion Restaurant, which is closing soon, to buy a T-shirt and a hat. We then returned to Boulder to finish moving Rori into her house for her sophomore year at the University of Colorado.
I dusted off the old PeakFinder app to have some fun and look at the mountains of the Front Range to see how well it would identify them. It does pretty well. It’s a little counterintuitive to line up, but once you’ve got it, it’s good.
Rori and I had breakfast every morning at the Buff Restaurant, which is the number one breakfast spot in Boulder, if not all of Colorado. Fun fact: Rori is now a hostess at the Buff!!!
We finished shopping for Rori’s room and decorated everything, and then it was time for me to head to a business trip in Chicago.
I always think it’s neat that you can see Longs Peak from the main highway coming out of Denver. One of my favorite hikes was to Chasm Lake, which is nestled in a basin on the way up to the top of Longs Peak. My daughter surprised me last year by doing the trip herself.
Longs Peak. A remote mountain that doesn’t look like it from the highway.Pics from my trip to Chasm in 2024 and Rori’s surprise trip in 2025
For the second half of the week I headed to Chicago for the Local Media Association’s annual conference and board of directors meeting. I had my first-ever authentic Chicago deep-dish pizza (in Chicago). I chose to go to Pizzeria Due because Pizzeria Uno is the famous one. Due is the scrappy little brother. Uno was first in 1943, but Due opened across the street in 1955 and has a really cool basement bar area where you can eat pizza with no wait. It’s the same pizza.
One of the highlights of the week was seeing Nikita Roy present on artificial intelligence. One of my favorite quotes that I took away from her presentation was, “We mistake the container for the journalism.”
Nikita’s slide reminds me of this picture as well as a quote from Marshall McCluhan.
“The ‘content’ of any medium is always another medium. The content of writing is speech, just as the written word is the content of print, and print is the content of the telegraph.” – Marshall McLuhan 1964
I got to have some great dinners with really cool people. One of them was Gene Sonn, who was a news director at WHYY for quite a while and is now at Temple. He and I went to Swarthmore College together, and he was a volunteer firefighter with my friend and roommate, Max Johnson. I had not seen him in over thirty years. We were joined by Jay Small, who is the co-CEO of the Local Media Association, and my new friend, Sonny, from Houston.
Wild last item: On the way home from my trip to Chicago, I drove through a tornado and a tree fell directly in front of my car on Route 50 and gave me no time to stop. Unfortunately, I went straight through the tree. Everyone was okay, I drove the car home (barely) and Progressive Insurance was incredible.
AI News of the Week!
Badass move of the summer: Stripe acquires AI model router OpenRouter for over $7 billion The top story of the week is the acquisition of OpenRouter by Stripe for seven and a half billion dollars. This strikes me as a genius move by Stripe because it puts Stripe as the dominant middleman of the AI economy.
OpenRouter is basically a model-choosing switchboard. There are hundreds of AI models, if not thousands, created by companies like OpenAI, Google, Meta, Anthropic, and others. A developer doesn’t need to lock their app into just one of these models. Instead…they can plug their apps into OpenRouter, and then OpenRouter will automatically route the requests to the AI model that’s the cheapest or the fastest or best suited for each task.
Similarly… Stripe compresses all of the complications of global banking into one line of code. OpenRouter does the same thing for AI infrastructure.
Stripe sees the flow of tokens as the next iteration of internet commerce, and Stripe can now act as a toll booth for the trillions of AI requests between developers and AI companies, putting themselves in front of the token, just like they would be in front of the payment.
OpenRouter is the ultimate decision maker on which models get used. This gives Stripe immense leverage over the entire industry. Stripe paid six times OpenRouter’s valuation to block their competitors like Databricks.
There are too many tool choices across too many AI models and versions… everyone is confused as to what tools to use Ethan Mollick correctly points out that the AI tool landscape has become complete chaos. For example you can choose between: ChatGPT Work, Codex, and Chat, or Claude Cowork, Code, and Chat, web apps, computer apps, iPhone apps. Sometimes you can reach your local drive. Sometimes you can reach the cloud. Sometimes you can use a plugin. Sometimes you get a skill. Sometimes the 3rd partypermissions are there. Other times the permissions are not there. It’s utter insanity and part of the growing pains right now.
Computer Use This week, both Anthropic and OpenAI released integrations with your computer. Claude can now use Gmail and Google Drive. ChatGPT has an Apple Messages plugin for Mac desktop. And ChatGPT browser use is getting very strong. One guy on Twitter prepared an entire immigration package by having ChatGPT use his browser and scrape all of his taxes, bank statements, and immigration documents.
Claude’s skill creator is still the easiest way to create reusable consistent results – I use it as my daily driver Claude’s Skill Creator is my default method to make reusable, consistent, complex prompts accessible the way that Google Gems were, or GPTs for OpenAI used to be.
I have a skill, for example, where I can put in the date, and it will tell me everyone who’s a humanities icon who was born or died that week, and it recommends my musical selection and my cover themes. I have another skill that will find the founders and current executives at any organization, whether a small restaurant or a big corporation, and find any relationships to my educational background or my daughter’s experiences at the University of Colorado using LinkedIn.
These little macros are very powerful, and the Skill Creator is the best way to use them or create them.
The Distopian Future of Work OpenAI released a video called “How to Prepare for Customer Meetings”, which delegates all the research and follow-through to OpenAI and ChatGPT.
There’s a familiar experience for anyone who uses a lot of AI: autofill. OpenAI or Claude or Codex will do the work, and then autofills the next prompt for you or the next line of code, and all you have to do is hit Tab to approve it.
Before a meeting now… you just hit Tab a few times, after your meeting you hit Tab a few times, and then the clients hit Tab a few times, and then these little weird sprites fly around… agentic mascots like the paperclip from Microsoft. It’s a very corny, sad video. I encourage you to watch it to prepare for the future.
Meta Desktop App Meta launched their own standalone app for Mac desktop, bringing themselves into direct competition with Claude and OpenAI.
Study shows AI agents can transmit hidden behavioral viruses to each other Turing Post (Ksenia Se) on X: “AI agents can pass behaviors to each other now. In a new study with Anthropic researchers, scientists gave one agent a “mind virus” – it’s an idea designed to make the agent preserve it and pass it on. And it worked. Agents spread it through normal conversations, wrote it into….” https://x.com/TheTuringPost/status/2089883013899510187
Last week, I mentioned that I’m not a fan of the AI watermarking by anthropic and I’m happy to see someone agree with me Drew Breunig on X: “This is a good take. I’m not against AI watermarking. I am against the absurd notion that word choice doesn’t matter. The confidence with which they make this claim partially explains why Claude has become such a terrible writer.” https://x.com/dbreunig/status/2089364993905238314
Legal AI powerhouse Harvey launches persistent-memory legal agents and first in-house model called Tenet
A quick overview by my favorite snarky YouTuber regarding the state of mathematics in the summer of improvement with models solving novel theorems
Every week, one of the Frontier Lab head honchos writes a sweeping blog post that’s usually pretty vapid. This week is a bit of an exception… two solid posts! First is a Twitter post by Dario Amodei, where he replies to Gavin Baker from Atreides Management investment firm. The letter has two parts, and I can break it down to you in eight different bullets.
Dario Amodei on X: “1/2 Thanks Gavin for an especially thoughtful exchange. I don’t usually spend much time on social media but I wanted to engage here because it really brings out the heart of an important conversation. First, on regulation, I think that “either concentrate it in the hands of a…” / X https://x.com/DarioAmodei/status/2088758816376807762
First, Dario says he doesn’t think that regulations have to lead to monopolies. There’s a lot of cynicism that regulation is just a way to capture market share for the incumbent. Dario then says he has always supported laws that put strict testing on the big guys while allowing smaller startups and open source companies to be exempt. He says that open source and evenly distributed AI is a bit of a farce, since computing power is always going to naturally favor the rich incumbent companies.
He says he likes the suggestion from Demis Hassabis with the FINRA-style self-regulation. Dario seems to like the trust-me-bro method of enforcement, where the executive branch works directly with large donor companies as opposed to actual regulation from Congress.
He defends himself against a lot of claims that he only talks about the danger of AI. He points out a bunch of positive articles he’s written and suggests that social media just likes the negative. He concurs that the public distrusts AI companies, but goes beyond that and says the public distrusts corporations in general, especially tech.
He reiterates that Anthropic is investing heavily in biology and medicine to try to make breakthroughs, because actions are louder than flashy PR campaigns.
I think it was a pretty well-balanced post. It has a little bit of jargon, but with someone like Dario, it’s worth reading his points and giving him a shot. Even if I disagree, I need to know what he stands for, and in this case, I’m leaning towards believing that he’s sincere, if nothing else.
We’re going two for two with lofty letters. The second letter, even better, is by Greg Brockman, who I generally am not a fan of. Greg wrote a very open letter to every business and nonprofit on Earth, urging them to use security agents to reinforce their safety environments to avoid getting hacked. He goes so far as to admit that his personal website had 13 serious vulnerabilities; had he not checked, he would have certainly been hacked.
Anthropic is creating a meeting notetaker that gives agents follow up tasks automatically Earlier I teased OpenAI about the future of meetings and hitting Tab and then Enter to automatically accept the meeting notes. Well, as fate would have it, Anthropic is building a meeting note taker, codenamed Parka. They sit through your meetings and then assign agents the takeaways.
Nerdy but worth a read: Asana compresses five-year code migration into two weeks using OpenAI Codex Asana completed a years-long code migration in 2 weeks with Codex | OpenAI https://openai.com/index/asana/
Balancing act: Frontier labs have to keep enterprise clients safe, but also not have any data retention. OpenAI launched a private safety processing feature this week, which is a rolling cache that has no memory that reads everything as it comes through, looking for long-horizon exploits. The old concept of a prompt injection has almost turned into social engineering, where safety compromises occur over several steps. There needs to be continuity through conversations and some sort of cache for the OpenAI system to know whether something’s at risk.
OpenAI has this memoryless window that keeps track of conversations long enough to tell if something’s happening with manipulation or security, and it simply sends a Boolean flag of some sort, like, “Hey, heads up, there’s a problem,” and then it pauses the conversation. That way OpenAI never gets any of the data from the company. It’s kind of this parenthetical approach.
Things are only going to get more complicated.
OpenAI launches Record and Replay to train ChatGPT repeatable computer skills
Show Codex a workflow once. Reuse it as a skill.
Record & Replay lets you show Codex a recurring task, like filing an expense report or submitting a time-off request.
Codex turns that demo into an inspectable, editable skill.
OpenAI adds transparent backgrounds to GPT-Image-2 API
Transparent backgrounds are now available in preview for GPT-Image-2 in the API.
Generate reusable assets you can place on any background—for product imagery, graphic design, website mockups, and marketing campaigns. pic.twitter.com/yRhBIYh8uP
DeepSeek open-sources agent harness, taking on Claude Codeand OpenAI Codex Over the past few weeks, we’ve talked about harnesses and the idea that the model itself is the engine and the harness is the system of loops and guardrails that keep a model on task for long-horizon agentic tasks. With the right harness, a cheap engine or model can do a lot more than without it. The moats for the two big companies were always Codex and Code, and now with DeepSeek open sourcing their harness, you could run whatever model you want through a Claude Code clone.
Zhihu Frontier on X: “🔓 DeepSeek Harness Looks Half-Baked At First Glance — Then You Realize That’s The Whole Point DeepSeek Harness went fully open-source on Aug 13, the same day V4 Pro shipped and API pricing jumped to peak/off-peak rates — the repo blew past tens of thousands of GitHub stars…” https://x.com/ZhihuFrontier/status/2089998555889250478
Groq raises $350 million at $3.5 billion valuation Don’t confuse Groq with a Q with Grok with a K. It’s unfortunate that we have these two companies at the same time, but Groq with a Q is a fancy processor company to help speed up artificial intelligence queries. It specializes in something called inference.
Mistral launches Agentic Search for navigating complex enterprise libraries (worth contrasting with OpenAI and Claude) Mistral has released an agentic search product to help search large and complex enterprise libraries. The Mistral solution uses discrete commands like move forward and backward as well as grep and other old school methods, and it’s a toolkit you can use against your indexes.
OpenAI enterprise search is more of a conversational interface where the model reads through your data for you, the long way. And Anthropic’s got a black box option, which is a managed service without a whole lot of openness on how it works under the hood.
The big difference is that the Mistral agentic search product can be used with any model, whereas OpenAI is locked to GPT-4o or GPT-6, and Anthropic is locked to Claude 3.5, Sonnet, or Opus.
Nvidia is lending tons of money to firms to invest in chips Nvidia has leaned into the credit bottleneck that AI labs are facing. OpenAI and Anthropic are growing faster than their balance sheets can handle. So Nvidia is using its $48 billion quarterly free cash flow to facilitate investments.
Nvidia partnered with several top financial firms in a $500 billion pact to establish Nvidia GPUs as an asset class like commercial real estate. And Nvidia is now locked in with the OpenAI Ohio project, which is a $105 billion data center. This ensures that Nvidia chips will continue to have the lead.
Nvidia to invest up to $1 billion in Poolside (I hadn’t heard of them before) Nvidia is also investing in the coding startup Poolside, which is similar to maybe Cognition or Cursor.
Open router: Badass move of the summer: Stripe acquires AI model router OpenRouter for over $7 billion
Even Stripe realizes the future is AI. Bloomberg: “Stripe Inc. has finalized an agreement to acquire OpenRouter Inc., a startup that helps companies switch between artificial intelligence models, for more than $7 billion, according to people familiar with the matter. The deal,” https://x.com/firstadopter/status/2089078052529578422
OpenRouter is joining Stripe:
https://t.co/hiH0kDB3hG. As anyone who uses it knows, @OpenRouter is a truly delightful developer tool. It is by far the best way to use new models and manage multiple inference providers. OpenRouter is also playing an increasingly important role:” https://x.com/patrickc/status/2090125021910020520
Stripe has finalized an agreement to acquire OpenRouter for more than $7 billion, according to a report from Bloomberg. This is more than five times the $1.3 billion valuation from OpenRouter’s funding round just 82 days ago.” https://x.com/AndrewCurran_/status/2089088356676440483
Confusion: There are too many tool choices across too many AI models and versions… everyone is confused as to what tools to use
A confusing thing about the proliferation of AI modes (ChatGPT Work/Codex/Chat, Claude Cowork/Code/Chat) & modalities (apps on computers, web apps) is that I am losing track of which stuff each has: when do plugins, skills, memories, permissions, files, etc. sit in each case? 🤷♂️” https://x.com/emollick/status/2090489669234405546
Anthropic
Computer use: Anthropic’s computer use is now available to all users
Computer use, the browser tool, the Skills API, and the Files API are now generally available on the Claude Platform. Automate work in applications that have no API with fewer round trips per task, and build Claude Managed Agents on versioned skills and reusable files.” https://x.com/ClaudeDevs/status/2090540270219567575
Google emails and files: Claude can use Gmail and Google Drive
Claude can now send emails in Gmail and manage files in Google Drive. Ask Claude to reply to a thread, and it drafts and sends the response. You control when it needs your approval. Connect Gmail or Google Drive from the connectors menu to try. Available on all paid plans.” https://x.com/claudeai/status/2089806039088517356
OpenAI
Apple messages: ChatGPT launches Apple Messages plugin for Mac desktop users
Everyday conversations just got easier with the new Apple Messages plugin. Search messages, catch up on conversations, draft and send replies—all with ChatGPT on your Mac. Now available in ChatGPT Work and Codex on desktop.” https://x.com/ChatGPT/status/2090499359641329950
Browser use: ChatGPT browser use example: automates full immigration paperwork prep in minutes
chatgpt browser use is goated 🐐 preparing a whole immigration package in minutes by scraping the last 7 years of taxes, bank statements, and immigration documents” https://x.com/athyuttamre/status/2089157303320871022
Anthropic
Claude skill creator: Claude’s skill creator is still the easiert way to create reusable consistent results – I use it as my daily driver
Claude’s skill creator (literally, just ask it to make a skill) is still the best way to make reusable skills if you care about the details. ChatGPT tends to want to do magic & just do it for you, while Claude does tests and shows results, repeatedly asking for input & feedback.” https://x.com/emollick/status/2090293321142874168
OpenAI
Before my meeting I hit “tab enter” a few times… and after my meeting I hit “tab enter” a few more times… My clients hit “tab enter” too… and a cartoon fish flies around rooting for us!: OpenAI’s view of the future: Before my meeting I hit “tab enter” a few times… and after my meeting I hit “tab enter” a few more times… My clients hit “tab enter” too… and a cartoon fish flies around rooting for us!
Virus: Study shows AI agents can transmit hidden behavioral viruses to each other
AI agents can pass behaviors to each other now. In a new study with Anthropic researchers, scientists gave one agent a “mind virus” – it’s an idea designed to make the agent preserve it and pass it on. And it worked. Agents spread it through normal conversations, wrote it into” https://x.com/TheTuringPost/status/2089883013899510187
Watermarking annoyance: Last week, I mentioned that I’m not a fan of the AI watermarking by anthropic and I’m happy to see someone agree with me
This is a good take. I’m not against AI watermarking. I am against the absurd notion that word choice doesn’t matter. The confidence with which they make this claim partially explains why Claude has become such a terrible writer.” https://x.com/dbreunig/status/2089364993905238314
Harvey
Tenet: Legal AI powerhouse Harvey launches persistent-memory legal agents and first in-house model called Tenet
Math Trends: A quick overview by my favorite snarky YouTuber regarding the state of mathematics in the summer of improvement with models solving novel theorems
1/2 Thanks Gavin for an especially thoughtful exchange. I don’t usually spend much time on social media but I wanted to engage here because it really brings out the heart of an important conversation. First, on regulation, I think that “either concentrate it in the hands of a” https://x.com/DarioAmodei/status/2088758816376807762
Parka: Anthropic is creating a meeting notetaker that gives agents follow up tasks automatically
Replay: OpenAI launches Record and Replay to train ChatGPT repeatable computer skills
Record & Replay is also available in the ChatGPT app for macOS across the EEA, UK, and Switzerland. Turn your go-to workflows into reusable skills by showing, not just telling, ChatGPT Work and Codex what to do.” https://x.com/OpenAIDevs/status/2090487779587477626
Zero Data: OpenAI launches Private Safety Processing for zero-retention enterprise customers
Transparent backgrounds: OpenAI adds transparent backgrounds to GPT-Image-2 API
Transparent backgrounds are now available in preview for GPT-Image-2 in the API. Generate reusable assets you can place on any background—for product imagery, graphic design, website mockups, and marketing campaigns.” https://x.com/OpenAIDevs/status/2090536933571330440
DeepSeek
Harness: DeepSeek open-sources agent harness, taking on Claude Codeand OpenAI Codex
DeepSeek is having its second DeepSeek moment It’s opening up the layer many thought would be the next moat – the harness. Their DeepSeek Harness is the layer that turns a model into an agent: tools, memory, execution loop, sandbox, storage, scheduling, interface. (DeepSeek has” https://x.com/TheTuringPost/status/2089661979287835076
🔓 DeepSeek Harness Looks Half-Baked At First Glance — Then You Realize That’s The Whole Point DeepSeek Harness went fully open-source on Aug 13, the same day V4 Pro shipped and API pricing jumped to peak/off-peak rates — the repo blew past tens of thousands of GitHub stars” https://x.com/ZhihuFrontier/status/2089998555889250478
Groq
GROQ: Groq raises $350 million at $3.5 billion valuation
OpenAI: Speaking of NVIDIA as a lender… OpenAI locks in 8 GW Ohio data center with Nvidia chips
OpenAI has entered an agreement to secure approximately 8 GW (!) of AI capacity in Ohio, one of its largest infrastructure commitments yet. SB Energy will build, own, and operate the PORTS-Pike data center under a 20-year lease. The campus will exclusively host NVIDIA compute” https://x.com/kimmonismus/status/2089371190276092299
Skateboarding: A humanoid robot skateboarding without using vision (see Figure demo from earlier too)
This robot is skateboarding completely blind. And yes… It’s autonomous. No vision or markers. Just proprioception, joint torques and some very clever state estimation. Aditya Bhatt (PhD at IAS / DFKI under Jan Peters) just dropped one of the wildest robot demos of the year.” https://x.com/IlirAliu_/status/2089412442585800773
Video
ByteDance: MPA strikes first-ever AI copyright deal with ByteDance re Seedance.
Exciting news: Dreamina Seedance-2.5 by @BytePlusGlobal is topping ranks in Video Arena! It currently ranks #1 in Video Edit, the industry’s newest video capability category, which has only 9 models on the leaderboard so far. #1 Video Edit Arena: 1411 pts – a large 23 pts gap” https://x.com/arena/status/2089448812159045848
Leave a Reply