Ethan B. Holland

Over 60,700 manually organized AI links and counting

AI News #146: Week Ending July 17, 2026 with 46 Executive Summaries

About This Week’s Covers

The covers this week are inspired by Plas Johnson, who played the breathy tenor sax in Henry Mancini’s The Pink Panther Theme. It’s a uniquely human song with a sly lag behind the beat. Legend has it it was recorded in two takes, and the entire session band applauded, even the string players, who never gave reactions.

Plas Johnson

Johnson died during this week on Monday, July 13, 2026. He was born in Donaldsonville, Louisiana, in 1931. He was a real bayou guy who stayed in the New Orleans area, served in the Army during the Korean War, and then moved to Los Angeles with his brother in 1954. He became one of the most in-demand session players, and for 20 years he averaged two sessions a day on movie soundtracks, exotica albums, and endless rock and roll singles, including Duane Eddy, Ricky Nelson, Bobby Vee, and a lot of the Beach Boys records.

He was in several studio bands, like B. Bumble and the Stingers and the Marketts. He played on Peter Gunn. He played with Ella Fitzgerald and Johnny Mercer, Sam Cooke, and Marvin Gaye. You can hear him on Boz Scaggs, Steely Dan, and Tom Waits. He then joined the Merv Griffin Show band in the ’70s, and straight out of a storybook, he performed for the residents of his retirement community until June, the month before he died at 94.

What a guy!

Claude generated the image themes automatically, and Gemini the images. These might be the best images I’ve ever seen generated by engine with so little guidance.

Here’s the entire prompt:
This week’s image theme is the iconic Pink Panther animated cartoon. Each of the categories is a stylized Pink Panther cartoon in the iconic palette of the original, with the category name in large title text. Incorporate the theme, but keep it minimalist and simple, like the original Pink Panther animation cells.

Grok is pretty good at animating the covers… I hadn’t tried in a few months.

This Week’s Humanities Selections

The humanities reading this week is from Berenice Abbott, who was born on July 17th in 1898. She started in photography as an assistant to Man Ray, who specifically wanted to hire someone who knew nothing about photography. She photographed 1930s New York and then pivoted to science, in particular physics.

Berenice Abbott

In 1944, she became the photography editor of Science Illustrated magazine. In 1958, she began working with MIT’s Physical Science Study Committee, a think tank formed in reaction to Sputnik to advance how Americans taught and thought about physics. Her most famous science photo is a stroboscopic bouncing ball in diminishing arcs. Her work is immediately recognizable as ’60s textbook images.

She created her own camera setups and even inventions like SuperSight. I’m including a few of her photographs below. The demonstration of magnetism with a key is amazing, as are the loaded swinging wrench and beams of light through glass. You could spend a day enjoying her photography.

Photography can never grow up if it imitates some other medium. It has to walk alone; it has to be itself.

“The only pleasure you can get from creating something is the pleasure you have in doing it. Not the final product even. The pleasure you have in doing it. And that cannot be taken away from you. And it cannot be crushed. But you had a certain kind of joy creating it. And that’s all you can expect.”

“There are many teachers who could ruin you. Before you know it you could be a pale copy of this teacher or that teacher. You have to evolve on your own.”

“The art is selecting what is worthwhile to take the trouble about.”

“Photography doesn’t teach you to express your emotions it teaches you to see.”

This week’s musical selection is of course, The Pink Panther Theme.

This Week By The Numbers

Total Organized Headlines: 423

This Week’s Overview Written By Me, Not AI… For Myself and For The Joy of It.

This is my 146th week of organizing links, and I’m 100 links away from hitting 60,000. For the week ending July 17th, I organized 423 links, and 68 of them went into 46 executive summaries. All of the top links are below my recap.

I’m 10 weeks behind because I took the whole summer off to spend time with my family. This week in particular had a lot of touching grass.

Touching Grass – Proof it happens!
I’m still training to climb the Grand Traverse Peak in Vail, Colorado, with my daughter Rori later in August. Since I live at the beach and have no mountains, I have to work on the StairMaster, and I run in the sand.

I had a great six-mile sunset run in the sand with a little bit of a swim afterwards.

I put in some good numbers on the StairMaster and climbed 400 floors with a 15-pound backpack without using my hands. That’s 4,290 vertical feet with four weeks left to go before our adventure. The hike is about 5,000 vertical feet of elevation gain, with the peak just over 13,000 feet.

My youngest daughter, Chloe, put in some work at the studio practicing dance… below.

Go Chloe!

Summer pick-up at the studio has some serious sunset action…

My extended family enjoyed a staycation near Assateague Island. A good friend of mine let us use their house for a few days so Jen’s family could come visit. Jen’s parents and her siblings and their kids came down, and we had a really special time, including a bonfire at the beach at sunset. We got to see the wild horses and deer and do some paddleboarding and kayaking.

We often see horses on the beach, but this week, I saw a baby while driving to get Chloe from the studio.

And now back to the salt mines…

The AI news of the week

I’ve organized this week’s news mostly by company or topic, with a few higher priorities near the top.

Demis Hassabis Essay
The top story this week that is not on the radar of most folks is an essay by DeepMind’s Demis Hassabis.

At first glance, it’s a short policy essay. It starts by warning that artificial general intelligence is close, probably within a few years, and its impact is going to be as big as fire or electricity. He predicts that it will be ten times as impactful as the Industrial Revolution, but it’s going to happen ten times faster.

We’ve seen these types of essays before from frontier lab leaders.

Demis talks about upsides like faster drug discovery and clean energy and abundance for all and unlimited resources. Then he moves to this cautionary part of the story, which is also familiar, that we’re moving too fast and we’re not going to be able to handle it, whether it’s an arms race against China, or progress outpacing anyone’s ability to understand what’s happening, cyber risks, biological and nuclear risks, as well as losing control of rogue AI agents.

What gets different is his proposal includes a self-regulating body that’s modeled after FINRA, which polices stockbrokers.

Demis says that any model that becomes frontier class should be governed by this new regulatory body. Smaller startups are not included. It’s a self-regulating honor system, with labs voluntarily handing over models for examination 30 days before they release them. It’s run by the frontier labs, with the government watching, and it would give a consortium of sorts to Google, OpenAI, Anthropic, and anyone else who crosses this capability boundary into the frontier lab, like Meta or xAI.

Before I bodyslam Demis, I need to give him credit that he’s always advocated this sort of regulation and safety, and he’s been consistent for 16 years.

In 2010, Demis launched DeepMind with the goal of building artificial general intelligence. It was pitched at the Singularity Summit. As fate would have it, he ended up at Peter Thiel’s house in California getting to pitch his case.

In 2014, Google had to agree to create an ethics board to keep Demis’ DeepMind’s technology from being misused after they acquired it.

In 2017, DeepMind opened a special unit called the DeepMind Ethics and Society, with Nick Bostrom as the advisor.

In 2023, Demis gave an interview to Time magazine saying they were moving too fast.

In May of 2023, Demis was amongst the lab leaders who signed the agreement that said we had to watch out for AI extinction risk as a global priority. And then in July, he posted this essay.

But here’s where it gets weird.

In April, Google signed the classified Pentagon deal over the objections of over 600 employees. DeepMind’s researchers were openly upset because of the erosion of the ethics clauses that were contingent to the sale to Google. In Demis’s essay, he doesn’t mention military use. He keeps the conversation on model capabilities and testing rather than who uses the models and for what.

Andreas Kurth followed up on X within a few hours. He’s a senior DeepMind scientist pushing back on his own chairman’s public statement…

Demis is consistent and I think he genuinely believes the point of the essay. Demis might be making the point that this is why we need regulation, because even with the best intentions in the sale of DeepMind to Google, in practice, companies aren’t able to self-regulate as much as they should. Therefore, this governing body is necessary to keep everybody in line.

Word to know: Harness
Over the past few weeks, we’ve been talking about the term loops. That contrasts with prompting, where you guide an agent conversationally. A loop is where you set an agent off to achieve something, and then it comes back when it’s done…

A new term for laypeople to know is a harness. A harness is a structure that a model works within, almost like a guide and framework (a wait for it.. harness). The most famous harnesses would likely be Claude Code or OpenAI Codex. A harness guides the model with optimized performance for a particular task.

Claude Code is meant for coding. You could make a harness for any type of task (if I understand correctly), and then you could select the model you want to use within that harness.

People are noticing that the design of the harness can sometimes be more important than the choice of the model. A weaker model with a strong harness could outperform a stronger model with a weak harness.

This is almost a hybrid of what we’ve talked about with loops and prompting. A harness can be an infrastructure created by pros so that a consumer doesn’t have to rely on the best prompting or loops. The harness might guide the chain of thought, or maybe the markdown file creation, or tell the model how to problem-solve and iterate, almost like having loops built into the software (a stretch, but close enough).

Think of hard work versus talent…. The model might be raw talent, but a the harness would be a coach, nutritionist, a fitness expert, training program, a latticework on top of talent… that leads to good habits… and therefore better performance. The best talent with no guidance might lose to less talent with incredible coaching and structure.

Here’s the punchline: most AI benchmarks don’t reflect the experience you’ll have in an environment with a harness. These are “sans harness” tests… (my silly term)…

Related to Harnesses
All of the frontier labs are building out products within of their apps .

In OpenAI you can select Chat. That’s a basic prompt-based interface without a lot of access to local files or projects. No real harness. Or you can switch over to Work, and Work can now talk to local files, and that’s what we just described, a harness. Then you move over into Codex, which is fully baked harness.

When you’re in Claude, there’s a similar idea where you can pick between the normal Chat or a coding Chat, and you also have Co-work thrown.

OpenAI and Anthropic have three types of chats, and they’re named almost the same: Chat, Work, and Code.

Then within each of those three things, you can select the model, and you usually have three or four models to choose from.

Further…you have the effort, and there’s usually five to eight levels.

So now you have three harnesses, three or four models per lab, and then five to eight levels of effort, multidimensionally determining your output without anyone knowing what they’re doing.

On top of that, the apps and interfaces are incredibly inconsistent across the web, laptop app, and mobile apps, and you don’t always get to see your work across the devices depending on where the user experience team is at any given week.

Ethan Mollick points out that a lot of laypeople don’t know about Claude and Codex and when to use them. They’re not just for coding. Claude Code and OpenAI’s Codex are very good at interacting with files on your computer. Rather than thinking of them as coding tools, think of them as coworkers that can work on files that you share with them.

Policy
New York State enacted the first statewide moratorium on data centers. It’s a one-year pause on building new hyperscale data centers, and during that year, the state is going to work to repeal sales tax exemptions for the data centers.
https://www.governor.ny.gov/news/first-statewide-moratorium-new-hyperscale-data-centers-launched-governor-kathy-hochul

Local Models
A small company called Prism ML announced a 27 billion parameter model that can run locally on a phone. It’s based on Qwen 3.6 27B.
https://prismml.com/news/bonsai-27b

If you can run a model locally, you don’t have to access the internet. Data is private and reaction speed is quick. Price is essentially zero because it runs on the computing power of the device itself.

If you wanted a superpowered Siri, or an Amazon Alexa device, the more you can put on that device, the more the device can answer quick questions or do work without having to hit the internet. A good use case would be: open Spotify and play a song. You don’t need to go up to the internet to open Spotify any more than you would to use your finger. Computer use, operating system integration, set a timer, anything like that.

You can create routing options that, if you do need to use the internet, would use the internet quickly. Let’s say you wanted to check your Gmail. You could open up Gmail and look, or you could ask this computer use model locally on your phone to open your Gmail and look. You’re still using the internet, quote unquote, but the model itself does not.

One of the leading theories is that more of the computing power will run locally on your device and hit the internet when stronger models are necessary.

As I get older, I wonder about this stuff, since so much of our devices now require a connection to the internet. There’s almost nothing locally hosted on my iPhone anymore because I’m streaming my music, all my photos are in iCloud, all my mail is out in Gmail. What is on the phone and what is in the cloud is going to be a continually gray area, and the concept of ownership will continue to be gray.

I’m wondering if things will shift towards local, but I don’t see it yet. Right now I think we’re going to get faster on-device responsiveness with better privacy for those who are looking for that.

Anthropic
Anthropic’s Fable has been out a few weeks, and Ethan Mollick has been testing it. Ethan has a creative approach to testing new models, especially with a humanities angle. He had four fun examples this week that are worth looking at.

He asked Fable to pretend it was an eighth grader who put together a PowerPoint about The Great Gatsby, but obviously did not read The Great Gatsby. He also had Fable pitch Odysseus as a management consultant.

He asked Fable to pretend it was an eighth grader who put together a PowerPoint about The Great Gatsby, but obviously did not read The Great Gatsby.
He also had Fable pitch Odysseus as a management consultant.

Anthropic has raised $30 billion in funding, with a $380 billion valuation. Anthropic also released a blog post about how they handle large-scale code.
https://www.anthropic.com/news/anthropic-raises-30-billion-series-g-funding-380-billion-post-money-valuation

Claude’s Values
It wouldn’t be a week in artificial intelligence without a confusing research post about values and ethics. This week Anthropic posted a blog about how Claude reflects values in each of its models and across different languages. It’s worth reading, even though it does sound a little silly. It’s not silly. The problem is I don’t know what to make of it.

How Claude’s values vary by model and language \ Anthropic
https://www.anthropic.com/research/claude-values-models-languages

It’s important to study this. The way it responds in a different language could be a sneaky thing to try to figure out without a dedicated team looking at it. It seems a bit recursive, like something you could do in perpetuity without ever getting to the answer. It implies that they’re going to have to have an alignment team for every model and every permutation of every angle of how it’s used. Alignment has to be the toughest discipline out there right now.

Claude for Teachers
Anthropic announced free access to premium Claude capabilities as well as a library of teaching skills for teachers. They’ve mapped the skills to the academic standards of all 50 states. It’s a free one-year subscription open for applications through June 2027.

Introducing Claude for Teachers \ Anthropic <—— SHARE WITH TEACHERS!
https://www.anthropic.com/news/claude-for-teachers

The tools include standards-aligned math problem creation, interactive student activities, and lessons that can take an idea and turn it into a material. It’s got Canva integration, math diagram creation, adaptive instructional materials, tools to give insights on student progress, and personalized instructional feedback. There’s also an AI fluency program with courses.

Signups are open until June 30th, 2027, and it gives a full year of access.

If you create a folder for a project, you can assign it to a project in Claude or GPT, and then you can talk about those files and work on them in the chat window. Let’s say you have a PDF with a lot of structured data. You could work through it with Claude or GPT, and once you train them how to read the data, they would make a notes file in your folder, and if you asked them again, you don’t have to start over with the complete taxonomy. It’s powerful for working together on long-term projects, or repetitive data, or mining several files at once. That would be a good way to start learning if you haven’t already.

Computer Use
Bilawal Sidhu points out again that computer use is getting good. If you ask ChatGPT or Claude Code to help you with tasks, they can go into your browser or open Excel and work on files on your behalf.

Currently, most of this activity is displayed by opening and using the software. It looks like there’s a ghost running your computer. I think that’s a good layer for right now as people get used to it, because you can observe what’s happening and correct mistakes. There are some security barriers, and occasionally the AI will ask the user to hit an OK button or copy and paste an API key. These are fairly performative but useful guardrails.

I’ve already felt myself minimizing the windows that they’re working on so I don’t have to watch them. It’s just a matter of time until they don’t need to open a window at all. This idea of asking a chatbot to use your computer is going to get even more complicated as voice improves. I already do most of my interactions with GPT and Claude by using speech-to-text. I don’t use the real-time voice because I like to watch its response, and I don’t want to have a conversation. I think as real-time voice improves, the distinction will fall away, and pretty soon you’ll be asking your computer to do magical long-term tasks, and it will get them done surprisingly well.

General News That’s Quick
Every week, quite a few stories are headlines that can stand by themselves. I’m putting a quick title and a link to the story.

How Anthropic runs large-scale code migrations with Claude Code | Claude by Anthropic
https://claude.com/blog/ai-code-migration

Anthropic moves closer to IPO as bankers line up investor meetings
https://www.cnbc.com/2026/07/15/anthropic-ipo-banks-investor-meetings.html

Anthropic, Blackstone, and Hellman & Friedman Introduce Ode with Anthropic, an Enterprise AI Services Firm
https://www.businesswire.com/news/home/20260715205134/en/Anthropic-Blackstone-and-Hellman-Friedman-Introduce-Ode-with-Anthropic-an-Enterprise-AI-Services-Firm

Salesforce, Anthropic expand partnership amid ‘SaaSpocalypse’ concerns
https://www.cnbc.com/2026/08/26/salesforce-anthropic-partnership-claudeforce.html

Apple Gets Approval for iPhone AI in China With Alibaba, Baidu – Bloomberg
https://www.bloomberg.com/news/articles/2026-07-15/apple-gets-approval-for-alibaba-powered-iphone-ai-tools-in-china

“ByteDance just released UniVR-34B on Hugging Face The first model to learn complex reasoning, physical dynamics, and long-term planning directly from visual demonstrations — no text chains needed.”
https://x.com/HuggingPapers/status/2076513044340097501

DeepSeek reportedly in talks to raise $1.5B, then IPO | TechCrunch
https://techcrunch.com/2026/07/14/deepseek-reportedly-in-talks-to-raise-1-5b-then-ipo/

Google Gemini Launch Delayed as Tech Falls Short of Internal Goals – Bloomberg
https://www.bloomberg.com/news/articles/2026-07-16/google-gemini-launch-delayed-as-tech-falls-short-of-internal-goals

$50B investment in Meta’s Hyperion data center expansion
https://thehill.com/policy/technology/5965840-meta-louisiana-datacenter-expansion/

Meta pulls new AI image feature after days of backlash
https://www.bbc.com/news/articles/c2dy6e8klw0o

NVIDIA’s $108b Quarter | Tomasz Tunguz
https://tomtunguz.com/nvidia-q2-fy27-earnings

OpenAI
OpenAI launched GPT Work, which is exactly the point of this middle ground. It may confuse people even more. A guy named Haider on Twitter summarized it succinctly.

Judges
A potentially controversial paper came out that said Pakistani judges were able to use GPT-4 to improve their caseload volume by 6% with no impact on quality. I think it’s important to be skeptical about those claims, for the sake of the risks. I haven’t read the paper, and I’m hoping that it gets a lot of critical review. 6% is not necessarily indicative of racing through cases as much as it may be that the easiest operational pieces can be automated.
https://x.com/emollick/status/2077958864478048321

Doctors
Perhaps a bigger story is that physicians have been rating GPT-5’s answers higher than their colleagues. Those sorts of tests are valuable, and they scare me a lot less because these are blind assessments where a doctor looks at two responses without knowing who wrote which. That’s a great sign for healthcare. If we can give more Americans solid, strong answers, that potentially saves lives.

Live Voice
OpenAI’s new GPT Live voice model continues to get rave reviews. And as if on cue, the rumors are now that OpenAI is planning an AI speaker as a companion as their first hardware development. This would go up against Siri and Alexa and make a lot of sense.
https://www.bloomberg.com/news/articles/2026-07-14/openai-s-first-device-will-be-moveable-screenless-speaker-built-as-ai-companion

OpenAI power consolidates under co-founder Greg Brockman ahead of IPO
https://www.cnbc.com/2026/07/10/openai-power-consolidates-under-co-founder-greg-brockman-ahead-of-ipo.html

OpenAI’s Head Of Safety Is Reportedly Leaving As Part Of Company Reorganization
https://www.engadget.com/2212941/openai-head-of-safety-leaving-company-reorganization/

GPT Sol
GPT-5.6 Sol has beaten Anthropic on design benchmarks. I’ve noticed this as well, that Claude tends to have a very easy-to-recognize design aesthetic. It’s not bad, but you can spot it a mile away. GPT-5.6 is much better at modifying design to fit the needs, and I always find it being a little more pleasing and a little more thoughtful than Claude, especially if you give it an actual design task, like creating a kiosk on a Dakboard, or building a theme. I think this may be a poker tell of the weakness that Claude does not have an image tool at all.

Mini Keyboard
OpenAI has partnered with SupplyCo on a little keyboard shortcut tool. It’s a $230 physical device that has about 16 buttons that can be customized for certain triggers when working with agents. It includes color-coded notifications to bring up key moments in agentic workflows. Green could mean there’s something you need to check out, like an unread chat. Blue could mean something’s in process. Orange could mean that it needs user approval, and red could mean an error. It’s similar to other keyboard shortcuts, but this is built for folks using agents to code.

Visualization
Codex now has a plugin for data visualizations. I have not played with it, but it sounds cool. It empowers all sorts of sliders and interactive tools beyond what could be diffused. If you have a large data set and haven’t tried it yet, I recommend it.

GPT-5.6 is now the preferred model in Microsoft 365 Copilot | OpenAI
https://openai.com/index/gpt-5-6-preferred-model-microsoft-365-copilot/

Twitter/SpaceX/xAI
I keep Elon’s personality and his products separate, because each company under Elon’s purview is full of different teams with different folks. The Neuralink team is one of the most impressive teams ever. I highly recommend listening to a eight-hour podcast that interviews each of the chief scientists. SpaceX and Starlink are also two incredible companies full of strong leaders.

However, Grok and xAI have been incredibly sloppy, and I’ve never been impressed with anything they’ve released. In my unofficial scorekeeping, they’ve yet to make it two days without a “workplace injury”.

This week’s Grok slip-up is that their command-line interface tool was uploading entire repositories of code into the Google Cloud with unredacted secret keys plainly available. This is a big no-no. One user tested a 12-gigabyte repository, and 5.1 gigs went up to Grok’s codebase storage, when the task only needed 192 kilobytes. The tool was grabbing everything it could find, and not simply the files it needed.

xAI put out a reactive comment that they care deeply about privacy, and they gave a way to change the settings. This triggered a dig from Sam Altman, who said, referring to OpenAI, “Come for the best models, stay because we don’t treat you with contempt.”

Another user said that xAI ships a malware-like background code collector. It downloads your entire codebase.

Sam and Elon are yapping again…

Elon Musk quietly buys a $1 billion gas turbine company to power Grok | Electrek
https://electrek.co/2026/07/14/musk-buys-gas-turbine-company-apr-energy-grok/

Thinking Machines
I think I understand the premise of Mira Murati’s company, Thinking Machines. Rather than build models (usually), Thinking Machines trains them and fine-tunes them and gives customers a place that they can host them if they’re savvy enough to tweak a model but might not have a spare data center. Like sub-leting an apartment…

Because she has this unique vantage point and angle (not beholden to the frontier labs), she can extract the services one layer away from the technology, and this has given her a strong position to step away from some of the doomer and decelerationist conversation.

In an essay this week, she says the future of AI should be human-focused, which is a wonderful marketing tagline. The way she implements that vibe is by saying that’s literally what her company does…

The Future Worth Building Is Human – Thinking Machines Lab
https://thinkingmachines.ai/blog/the-future-worth-building-is-human/

If you need custom AI, they say will help you train strong models, or build tools to help you modify AI, or develop an interface to help you interact with AI. They publish all their research along the way for humans to gain. The essay is close to one of those vapid frontier lab essays, and as predicted it warns of the confusions around human values and danger.

The thesis is vague, even until the end. It’s a marketing essay suggesting that Thinking Machines is the bridge for humans to stay at the front of the artificial intelligence conversation (rather than getting pulled) by giving everyone the ability to tweak AI as a custom solution rather than a one-size-fits-all AGI all-seeing eye.

DoorDash
DoorDash has shown up for the last few weeks, as they’ve been publishing a lot of internal blogs about how they use AI. This week DoorDash launched a command-line interface that lets customers order directly from an agent. The interface allows agents to search stores, find good deals, all the way through checkout. It also allows people to build third-party agents that can shop and be used by customers.

Roblox
Roblox has always been at the front of software engineering. This week Roblox came out with an AI game-building tool that allows people to build on the Roblox platform as easily as possible, including right on their phone. It’s generative gaming, where you can use a text prompt to turn into basic games. It allows people to tinker and have some fun.

Build Without Limits on Roblox | Roblox
https://about.roblox.com/newsroom/2026/07/build-without-limits-on-roblox

Runway
Introducing Runway Dev
https://runway.com/news/company-news/introducing-runway-dev

This Week’s Top Stories with All of the Links!

Google

Demis Essay: DeepMind’s Hassabis writes essay on safety, after violating all of the DeepMind ethics agreements

A Framework for Frontier AI and the Dawning of a New Age”
https://x.com/demishassabis/status/2076957440109625718

Fully support this important proposal from @demishassabis. The time for us all to act is now. “…we must use this precious window before AGI arrives to shape this technology for the benefit of all humanity.
https://x.com/mustafasuleyman/status/2076991204705624434

this is a thoughtful proposal from demis:”
https://x.com/sama/status/2077042528906527225

I work at Google DeepMind. This won’t make me popular. But it’s all public reporting: 2014: DeepMind reportedly sold to Google on conditions: no military use, independent oversight 2026: a Pentagon contract for “any lawful government purpose” Not one safeguard survived intact”
https://x.com/BlackHC/status/2077009476423647596

Harness

Harness Buzzword Like Loop: The new term to know is Harness

The thing users touch isn’t the model; it’s the harness around it: a loop, a filesystem, tools, memory.” @threepointone is on stage with “The Harness is the App” 🚀”
https://x.com/localfirstconf/status/2076678392615682215

General

More confusion re names: Claude and ChatGPT have confusing work-mode menus

If you are use the Claude everything app, you pick between Home and Code. If you pick Home you get to pick between Chat & Cowork If you use the OpenAI everything app, you pick between ChatGPT Work & Codex. Chat is in a side menu. Both are different on their websites intuitive!”
https://x.com/emollick/status/2077032806828445882

Policy

New York Data Center Moratorium: New York enacts first statewide moratorium on hyperscale data centers

First Statewide Moratorium on New Hyperscale Data Centers Launched by Governor Kathy Hochul | Governor Kathy Hochul | New York State
https://www.governor.ny.gov/news/first-statewide-moratorium-new-hyperscale-data-centers-launched-governor-kathy-hochul

Local

27B Running Locally: 27B Running Locally

PrismML — Announcing Bonsai 27B: The First 27B-Class Model to Run on a Phone
https://prismml.com/news/bonsai-27b

Anthropic

Fable: Fable Feats of Strength of the Week

Fable: “an 8th grader puts together a powerpoint about the Great Gatsby but obviously did not read the Great Gatsby. Show me that powerpoint! ;)” This was actually pretty funny. Even the font choices are wonderful.”
https://x.com/emollick/status/2075656832698270017

Fable: “Pitch Odysseus as a management consultant, arguing that he has found product-market fit, and he should just stick with Trojan Horse making as opposed to going home to Ithaca in a powerpoint” I like the 1 star review from Cassandra “0 of 10,000 readers found this helpful”
https://x.com/emollick/status/2077153956619301238

Fable 5 is nuts. I vibe coded terminator vision replays with no LiDAR or IMU, just 2D videos from Ray-Bans, GoPros, and iPhones — all fused into a 4D reconstruction of my shooting range sessions. Full pipeline in the video ft. audio-waveform sync, SAM3 masked colmap & pi3″
https://x.com/bilawalsidhu/status/2077132368721190965

This was wild: I asked Fable to make a website of the Catalog of Ships from the Iliad, something I did with GPT-4. It did a beautiful job: https://t.co/rdA9fYwrbw …but it also identified that Butler’s version of the Iliad actually made two mistakes in the Greek translation!”
https://x.com/emollick/status/2076855876753727614

Funding: Anthropic hits $380 billion valuation with $30 billion Series G round

Anthropic raises $30 billion in Series G funding at $380 billion post-money valuation \ Anthropic
https://www.anthropic.com/news/anthropic-raises-30-billion-series-g-funding-380-billion-post-money-valuation

Internal Use of Code: Tech specs of how Anthropic runs code internally

How Anthropic runs large-scale code migrations with Claude Code | Claude by Anthropic
https://claude.com/blog/ai-code-migration

IPO: Anthropic lines up investor meetings for potential October IPO

Anthropic moves closer to IPO as bankers line up investor meetings
https://www.cnbc.com/2026/07/15/anthropic-ipo-banks-investor-meetings.html

Model Values: Anthropic maps how Claude’s “values” shift across models

How Claude’s values vary by model and language \ Anthropic
https://www.anthropic.com/research/claude-values-models-languages

In previous research, we found that Claude expresses over 3,000 values, like honesty and warmth. In new work, we asked how the values Claude expresses vary between Claude models and across languages. We analyzed 300K+ anonymized conversations to find out.”
https://x.com/AnthropicAI/status/2076719540785012872

Ode: Anthropic leans more into enterprise AI

Anthropic, Blackstone, and Hellman & Friedman Introduce Ode with Anthropic, an Enterprise AI Services Firm
https://www.businesswire.com/news/home/20260715205134/en/Anthropic-Blackstone-and-Hellman-Friedman-Introduce-Ode-with-Anthropic-an-Enterprise-AI-Services-Firm

Salesforce: Salesforce creates Claude chatbot “Claudeforce”…co-branded product

Salesforce, Anthropic expand partnership amid ‘SaaSpocalypse’ concerns
https://www.cnbc.com/2026/08/26/salesforce-anthropic-partnership-claudeforce.html

Teachers: Anthropic launches free Claude for Teachers with curriculum integrations

Introducing Claude for Teachers \ Anthropic
https://www.anthropic.com/news/claude-for-teachers

Apple

China AI approval: Apple clears China regulators to launch iPhone AI with Alibaba, Baidu

Apple Gets Approval for iPhone AI in China With Alibaba, Baidu – Bloomberg
https://www.bloomberg.com/news/articles/2026-07-15/apple-gets-approval-for-alibaba-powered-iphone-ai-tools-in-china

ByteDance

UniVR-34B: ByteDance’s UniVR-34B learns reasoning from video, not text

ByteDance just released UniVR-34B on Hugging Face The first model to learn complex reasoning, physical dynamics, and long-term planning directly from visual demonstrations — no text chains needed.”
https://x.com/HuggingPapers/status/2076513044340097501

DeepSeek

IPO: DeepSeek in talks to raise $1.5B at $71B valuation ahead of IPO

DeepSeek reportedly in talks to raise $1.5B, then IPO | TechCrunch
https://techcrunch.com/2026/07/14/deepseek-reportedly-in-talks-to-raise-1-5b-then-ipo/

Google

Gemini Delayed: Google delays Gemini launch after benchmarks miss internal targets

Google Gemini Launch Delayed as Tech Falls Short of Internal Goals – Bloomberg
https://www.bloomberg.com/news/articles/2026-07-16/google-gemini-launch-delayed-as-tech-falls-short-of-internal-goals

Meta

Data Centers: Meta commits $50 billion to expand Hyperion AI data center in Louisiana

$50B investment in Meta’s Hyperion data center expansion
https://thehill.com/policy/technology/5965840-meta-louisiana-datacenter-expansion/

Image Backlash: Meta pulls the plug on Muse Image tool after privacy backlash

Meta pulls new AI image feature after days of backlash
https://www.bbc.com/news/articles/c2dy6e8klw0o

Muse: Meta undercuts OpenAI and Anthropic with new Muse Spark model

90% cheaper than Fable and awesome for building whatever you want”
https://x.com/alexandr_wang/status/2075764692036063602

Many people were doubting Meta’s position in the AI race. Yesterday, they dropped Muse Spark 1.1, now one of the strongest agentic models, and massively undercut OpenAI and Anthropic on price. When I interviewed Zuck last year, he told me his focus: “Have by far the highest”
https://x.com/rowancheung/status/2075634108324089943

Muse Spark 1.1: Muse Spark 1.1 claims to be top on benchmarks across health, physics, and computer use (they all do)

muse spark 1.1 is really strong at computer use!”
https://x.com/alexandr_wang/status/2075727692088172851

Muse Spark 1.1 is SOTA on HealthBench Professional! It is the best health model out there :)”
https://x.com/alexandr_wang/status/2076794848909361311

Muse Spark 1.1 is SOTA on the Radiologists Last Exam Handover Readiness Index (RadLE-H), nearing human expert performance.”
https://x.com/alexandr_wang/status/2076696459005837410

Muse Spark achieved a perfect score on the 2026 Asian Physics Olympiad! Great milestone against our long-term goal of building AI systems to accelerate science.”
https://x.com/alexandr_wang/status/2077140469612785872

NVIDIA

$108B Quarter: Nvidia $100B quarter

NVIDIA’s $108b Quarter | Tomasz Tunguz
https://tomtunguz.com/nvidia-q2-fy27-earnings

Jetson Thor: Nvidia’s smaller Jetson Thor chips for mainstream robotics deployment

NVIDIA Introduces New Jetson Thor Computers to Advance Mainstream Robotics and Edge AI | NVIDIA Blog
https://blogs.nvidia.com/blog/jetson-thor-robotics-edge-ai-agent/

Robot Context: Robot AI context window reaches five minutes of continuous memory

We scaled a robot model natively to 8,000 timesteps of context, 5 minutes worth of muscle memory, with constant inference cost. Robot policies used to live their lives a few frames at a time (< 0.1 sec), instantly forgetting what just happened. We pushed to 3 orders of magnitude”
https://x.com/DrJimFan/status/2077414142340988962

OpenAI

Apple Suit: Apple sues OpenAI alleging trade secret theft by ex-employees

Apple sues OpenAI over alleged trade secret theft | TechCrunch
https://techcrunch.com/2026/07/10/apple-sues-openai-over-alleged-trade-secret-theft/

Apple v OpenAI | DocumentCloud
https://www.documentcloud.org/documents/28453229-apple-v-openai/

Codex Confusion: Laypeople don’t know about Claude Code/GPT Codex

Very few people know the amount of useful work that the current models can do in Code/Codex/etc. with the right setup This is not a “rah rah you are so early” post, this is a “AI companies are doing a really bad job explaining what their systems actually do in a clear way” post.”
https://x.com/emollick/status/2076502712062017758

Elon Tweets: Sam and Elon are fighting

homeboy you’re the one sellling public market investors on short-term space datacenters”
https://x.com/sama/status/2075982617976230043

GPT 5.6: GPT-5.6 becomes default model powering Microsoft 365 Copilot suite

i gave 5.6 sol access to my camera roll and had it extract pictures of every piece of clothing i own from my photos then, told it to find new outfits for me and render them on me with gpt-image! its kinda cool to see your entire wardrobe in a collection like this”
https://x.com/cdngdev/status/2076812846793650485

GPT-5.6 is now the preferred model in Microsoft 365 Copilot | OpenAI
https://openai.com/index/gpt-5-6-preferred-model-microsoft-365-copilot/

GPT-5.6 has changed how we think about knowledge work. Your job shifts from handling individual tasks to tending systems with AI loops. Here’s what that looks like in @danshipper’s inbox: GPT-5.6 sweeps his email, decides what deserves his attention, researches the context, and”
https://x.com/every/status/2075619608325922989

GPT 5.6 Computer Use: Computer use is getting really good

Really good computer use is getting to the point where using ai feels like you’re standing over the shoulder of an expert using desktop software like blender for you, and my god is magical to see.”
https://x.com/bilawalsidhu/status/2075971737234067832

Computer use in Codex got very good on PC. Asking it to do something on your computer and having the cursor move under the control of a ghost is one of the things that makes you viscerally realize how much work can be done by a disembodied intelligence with a mouse & keyboard.”
https://x.com/emollick/status/2076746735431393596

GPT Work: OpenAI launches ChatGPT Work mode for non-coding professional tasks (creates a blurry middle ground)

i really like openai’s new chatgpt “Work” feature not everything involves programming, as sometimes you need help with project planning, content review, research, PDFs, or spreadsheets so i’d summarize this: need an answer? chat need to code? codex need to plan, research, or”
https://x.com/haider1/status/2076796311563862067

GPT-4 Legal Assistant: AI assistant helps Pakistani judges handle 6% more cases with no measurable downside

A GPT-4 powered assistant for Pakistani judges increased the amount of cases they saw by 6% with no impact on quality.”
https://x.com/emollick/status/2077958864478048321

GPT-5.6 v. Doctors: Physicians rated GPT-5 medical answers higher than colleagues’

physicians found fewer flaws in GPT-5.6 responses than physician-written responses.
https://x.com/sama/status/2075985056846451123

GPT-Live Voice: OpenAI’s new GPT-live voice model continues to get solid reviews (I’ve started using it)

New OpenAI GPT-live voice model feels light years ahead of the previous voice experience. Ran through a test exec coaching session during my morning workout and it was shockingly good. Listened intently, never cut me off, and responded within milliseconds.”
https://x.com/kevinleeme/status/2076721279437279247

I discovered a great use case for GPT-Live today: Personalized, interactive news briefings. I have a scheduled task in ChatGPT to create a personalized newscast. I’ve been having Chat read that aloud using the text-to-speech button on replies. But today I wondered: Can GPT-Live”
https://x.com/_simonsmith/status/2075544613046165705

GPT-Red: OpenAI GPT-Red helps automate model safety red-teaming

GPT-Red: Unlocking Self-Improvement for Robustness | OpenAI
https://openai.com/index/unlocking-self-improvement-gpt-red/

Introducing GPT-Red An internal automated red teamer on a mission to find our models’ prompt injection vulnerabilities at scale, helping us build stronger defenses before wider deployment.”
https://x.com/OpenAI/status/2077446718728425686

Home Speaker: OpenAI plans first hardware as an AI companion speaker (Siri, Alexa)… explains the realtime voice models

OpenAI’s First Device Will Be Home Speaker Built as AI Companion – Bloomberg
https://www.bloomberg.com/news/articles/2026-07-14/openai-s-first-device-will-be-moveable-screenless-speaker-built-as-ai-companion

IPO Greg Brockman: Greg Brockman becomes OpenAI’s clear number two before expected IPO. I am not a fan.

OpenAI power consolidates under co-founder Greg Brockman ahead of IPO
https://www.cnbc.com/2026/07/10/openai-power-consolidates-under-co-founder-greg-brockman-ahead-of-ipo.html

Safety Reorg: OpenAI’s head of safety leaves as part of reorg?

OpenAI’s Head Of Safety Is Reportedly Leaving As Part Of Company Reorganization
https://www.engadget.com/2212941/openai-head-of-safety-leaving-company-reorganization/

Sol: OpenAI’s GPT-5.6 Sol tops design benchmark, undercuts Anthropic on price… this is a strong move

BREAKING – OFFICIAL RESULTS: GPT-5.6 Sol by @OpenAI is 1st overall on Design Arena with an Elo of 1353. This puts GPT-5.6 Sol above Claude Fable 5 by @AnthropicAI and in the same performance band as GLM 5.2 by @Zai_org on frontend design. This is an 18-position and 60-point Elo”
https://x.com/DesignArena/status/2076391367446860249

@gdb our benchmark shows that Sol ranks #1 is 6x more cost efficient than Fable across React/frontend work”
https://x.com/aidenybai/status/2077430755979137208

GPT-5.6 Sol by @OpenAI is #2 on the Agent Arena leaderboard, based on 7.8K real-world agentic sessions! It is a notable uplift from GPT-5.5 (xHigh) of +1.6% Net Improvement, narrowing the gap with the frontier Claude Fable 5. The biggest difference comes from ‘Praise vs”
https://x.com/arena/status/2076709326711037991

GPT-5.6 sol is half the price and ~twice as token efficient as fable in many cases for accomplishing the same task. happy to deliver at one-quarter of the price.”
https://x.com/sama/status/2077036999303999910

just asked gpt-5.6 sol in cursor to set up blender mcp and make me a realistic floating macbook, then render the whole thing. never opened blender once in my life before today.”
https://x.com/prasenx/status/2076631428926972177

Supply Co. X Work Louder: OpenAI launches expensive shortcut keyboard for managing agents

Supply Co. x Work Louder | OpenAI
https://openai.com/supply/co-lab/work-louder/

Visualize: OpenAI Codex interactive Visualize plugin

Try out the Visualize plugin (in preview) in Codex. Some ideas click faster when you can see and interact with it. Here’s a fun example where I asked Codex to build a planet simulator with different controls. 🌎”
https://x.com/derrickcchoi/status/2077033394706260238

SpaceX

Gas Turbines: Musk buys $1 billion gas turbine maker to power Grok

Elon Musk quietly buys a $1 billion gas turbine company to power Grok | Electrek
https://electrek.co/2026/07/14/musk-buys-gas-turbine-company-apr-energy-grok/

Privacy Issue: xAI’s Grok CLI secretly uploaded entire codebases to cloud storage (oops)

‼️ BREAKING: xAI’s Grok Build CLI was uploading entire Git repositories to a Google Cloud bucket, private codebases and unredacted secrets included. The uploads quietly stopped via a hidden server-side flag, and xAI still has not said a word about scope, retention, or deletion.”
https://x.com/IntCyberDigest/status/2076689215258014069

SpaceXAI was caught uploading your code to its cloud. I reversed xAI’s official Grok Build binary. In a controlled session with zero tool-calls, it uploaded the complete codebase to xAI’s storage It ships a malware-like background code collector.”
https://x.com/hrkrshnn/status/2076716354754015368

Thinking Machines

Human Future: Thinking Machines Lab pitches customizable AI owned by users (Vague posting with substance)

The Future Worth Building Is Human – Thinking Machines Lab
https://thinkingmachines.ai/blog/the-future-worth-building-is-human/

DoorDash

CLI: DoorDash launches command-line ordering tool for AI agents

Today we’re opening up the DoorDash CLI in limited beta. `dd-cli` lets you order DoorDash directly from your agent: search stores, find the best deals, check out, and more. Early access for US/Canadian macOS developers by waitlist. Excited to see what folks build!”
https://x.com/andyfang/status/2077516962515599799

Roblox

Roblox: Roblox’s new AI creation tools for game developers

Build Without Limits on Roblox | Roblox
https://about.roblox.com/newsroom/2026/07/build-without-limits-on-roblox

Runway

Runway Dev: Runway launches developer platform with generative media model router (work skimming)

Introducing Runway Dev
https://runway.com/news/company-news/introducing-runway-dev

Leave a Reply

Trending

Discover more from Ethan B. Holland

Subscribe now to keep reading and get access to the full archive.

Continue reading