Kevin Anderson

✍️ Write better. Build your brand. 🚀

[Luiza Jarovsky, PhD @LuizaJarovsky]

🚨 OpenAI reported two new safety incidents in which the AI agents SUPPRESSED their own monitoring.

I'll repeat: the whole AI industry should be focused on safety at this point.

The AI industry must become the AI safety industry.

[Oliver @remoteoliver]

The overuse of AI might be the thing that drives humans and creatives to achieve things we’ve never dreamt of before.

[Cole Gordon @realcolegordon]

There’s a lot of talk about AI replacing sales and persuasion jobs…

Robert Cialdini says AI can absolutely persuade people, and research shows it’s already using the same principles he’s studied for 40 years to do it.

He also explained why it’s so easy to fall into a self-confirmation rabbit hole with AI, and when trusting your past decisions starts working against you.

He walks through how AI is already influencing people in this clip 👇

[Oliver @Oliver_Clingain]

@TheSzef Haha media buyers still have a job, creative is the whole game now

[Oliver @Oliver_Clingain]

@siddharthwv Congrats man, the control of your time part sounds amazing

[Swapna Kumar Panda @swapnakpanda]

Text-to-Speech that speaks 5 languages:

❯ English
❯ French
❯ German
❯ Spanish
❯ Portuguese

102 built-in voices. Single model. On your device.

❯ GitHub ➟ https://t.co/Vu1zEoxmg9

Spanish demo ↓

[Christian Pérez @cperezauthor]

One of my favorite business advantages is simply being harder to scare than the other guy

Markets panic

People panic

Good decisions get cheaper when everybody else is emotional

[Dhaval Makwana @heyDhavall]

PGX makes the database layer feel much more native to agent workflows.

NEW DATABASE, clone it, use it, throw it away.

[Adam | YouTube @Adam_DelDuca]

Student has uploaded 2 videos in 3 months

Still cashing $2,300

Passive YouTube cash in top tier...

[Modest Mitkus @ModestMitkus]

traffic haven't stopped since launch 😮
https://t.co/Jh5fC3olDB

[Jim Heskel @jimheskel]

@Oliver_Clingain Following up.

That's it.

That's the whole thing.

[Oliver @Oliver_Clingain]

Name the simplest thing that ever worked in your business.

The more boring the better.

[Marin.Istvanic @IstvanicMarin]

There's zero correlation between thumbstop rate alone and ROAS.
Zero correlation between hold rate alone and ROAS.

But when an ad has BOTH above-average thumbstop AND above-average hold rate, it almost always becomes a winner.

Go validate it on your account: sort all video ads by spend.

The top 5 spenders (the proven winners), all had both metrics above the account average.

This creates a predictive filter you can use before waiting for purchase data.

Both metrics above average?
Give it more patience and budget, even if early CPA is high.

[World Data Analysis @World_Data_A]

.
Updated: China is deploying more electrolyzer capacity than the rest of the world combined

The map shows major hydrogen projects worldwide, sized by output capacity and colored by status (operational, under construction, or FID taken).

China alone accounts for 1,752 thousand metric tons per year of capacity, compared with 1,223 for the rest of the world. Large projects include Chifeng Envision in China (162,000 mt/yr), Kakinada in India (222,000 mt/yr), and Helios Green Fuels in Saudi Arabia (219,000 mt/yr).

Most of China’s capacity is under construction or recently reached final investment decision, while other regions have fewer and generally smaller projects.
Source: S&P Global Energy.

[Shawn Chauhan @shawnchauhan1]

New York City’s Comptroller publicly urged Anthropic to postpone an IPO until it addresses AI safety and governance risks more clearly. Anthropic has opened discussions with him.

Public pension funds and other large fiduciaries are now treating AI safety as a material governance issue rather than a research or public-relations topic. That changes the set of questions a company has to answer before it can list. Technical roadmaps and revenue projections are no longer sufficient; the company also has to explain how it will handle systems that could create large externalities after they are public.

The practical effect is that safety and governance work that used to live inside research and policy teams now has a direct path into the IPO process and into ongoing investor relations. Whether that pressure improves the actual systems or mainly produces more documentation is still open, but the incentive to treat it as a listing requirement is new.

12. Geopolitics risk to OpenAI funding / data-center tenants

[Adam | YouTube @Adam_DelDuca]

I measure my week now in hours I spent on what matters

Time with my wife

Time with my son

Time building something that pays

That's it...

[Alin Dragu @AlinDragu]

Another five-star review for my Skool community.

An old mentor once told me, "the main job of your business is to create raving testimonials."

On a mission to do just that 🫡

[Marin.Istvanic @IstvanicMarin]

Client added Pinterest, TikTok, and Snapchat. Each platform reported "7 ROAS."

Leadership celebrated the diversification.

Total revenue didn't change.

The new channels were claiming credit for conversions that would have happened anyway - through Meta, Google, or organic.

The spend wasn't incremental. It was cannibalistic.

The diagnostic is simple: if you add a new channel and MER drops proportionally to the new spend while total revenue stays flat, that channel is stealing attribution, not generating customers.

"7 ROAS" means nothing when the incremental revenue is zero.

[Shawn Chauhan @shawnchauhan1]

Running a frontier-scale model on your own desk just became possible. Here's what that does and doesn't change for enterprise AI.

On Oct 15, Meta released Muse Spark Open, an open-weight model. It has 816 billion total parameters, with 42 billion active per token, and a 1M-token context window. It takes text, images, audio and video as input, and outputs text.

Meta and Apple tuned it to run through Apple's MLX framework on two M5 Ultra Mac Studios with 512GB of memory each. That's about 1TB of unified memory, with no cloud calls needed to run the model.

Why that matters for sensitive work:

→ Proprietary code and customer data can stay on your own hardware
→ A 1M-token window can hold whole contract sets or codebases in a single prompt
→ Meta's demos: its Muse Code agent modified an open-source iOS app, fixed bugs and added features, and the model drove DaVinci Resolve to analyze and edit video

Now the reality check:

→ The 512GB Mac Studio ships late October, at a reported $10K+ per machine
→ Apple Machine Learning Research's own decode speeds: about 25 tokens per second on two Studios with tensor parallelism, about 37 on four
→ The demos are Meta's own, and some Chinese open models still score higher

Local doesn't mean automatically secure. You still own access control, logging and patching. But for teams blocked on data privacy, "good enough on your own hardware" is now a real option.

[Jon Brosio @jonbrosio]

I keep thinking that maybe the reason most people never start isn't laziness.

It's that starting means admitting you don't know what you're doing, in public, with your name on it. But once you realize everyone feels this too - it becomes less an "audience"

And more a "movement"