Justin Wright

CrossFit Games → AI startup. Solutions Engineer @mavenagi. Writing about tech. Podcast @MostlyHumansPod

[Justin Wright @justinwrightai]

https://t.co/fPjFhWJUaH

[Justin Wright @justinwrightai]

Last week, Anthropic published a paper calling for "the option to slow or temporarily pause frontier AI development."

This week, the US government took them up on it.

On Monday, Anthropic launched Claude Fable 5. State-of-the-art on nearly every benchmark. The most capable AI model publicly available.

On Friday at 5:21 PM, Fable 5 and Mythos 5 were subject to export controls.

The company that asked for guardrails just learned they cut both ways.

Full analysis 👇

[Justin Wright @justinwrightai]

https://t.co/E2m5N9Gacz

[Justin Wright @justinwrightai]

80% of the code merged into Anthropic's production codebase in May was written, tested, merged by Claude.

Before Claude Code launched in February 2025, that number was in the low single digits.

The typical Anthropic engineer now merges 8x as much code per day as they did in 2024.

Code quality? "Somewhat worse" than humans in late 2025, at parity today, and expected to be "strictly better" within the year.

More in this week's Monday Momentum 👇

[Justin Wright @justinwrightai]

https://t.co/Pf8CoF0gyE

[Justin Wright @justinwrightai]

Anthropic just became the most valuable AI company on earth. The number ($965B) matters less than the order

May 6: Rented all of SpaceX's supercomputer
May 14: $200M Gates Foundation partnership for healthcare + education
May 28: Released Opus 4.8 and raised $65B

Compute->Mission->Product->Capital. 22 days

Compare:
OpenAI launched ads in ChatGPT
Google embeds Gemini to protect its ad empire

The company that built better products, spent less, and led with mission just won

Full breakdown 👇

[Justin Wright @justinwrightai]

Big week for Anthropic as they announce Opus 4.8 and $65 billion in fresh funding.

This comes off the heels of a $200 million partnership with the Gates foundation for vaccine research, AI tutoring for students, and support for impoverished farmers.

Full newsletter coming Monday at https://t.co/LDWUKRRVBQ!

[Justin Wright @justinwrightai]

https://t.co/Yxa1DDO7Hk

[Justin Wright @justinwrightai]

Google just gave its AI agent an email address.

At I/O on Tuesday, Sundar Pichai unveiled Gemini Spark: a personal AI agent that runs 24/7 on a dedicated Google Cloud virtual machine. It does not stop when you close your laptop or pause when you lock your phone.

Price: $100/month (cut from $250). Available next week.

Last week Google killed the app. This week it is coming for the 9-to-5.

Full analysis in this week's Monday Momentum 👇

[Justin Wright @justinwrightai]

The biggest mistake people make with AI workflows is rubber stamping the output.

LLMs are verbose by default. They generate content that's dense, convoluted, and buries the actual useful stuff under paragraphs of filler. If you're using AI to create follow-up docs from call transcripts (a super common use case), ask yourself this before you send it: if someone got this with zero context, would it help them make a decision? Or would they have to do more work just to figure out what's going on?

Here's the thing most people skip: you need to actually edit the output. Cut it down and find where the real insights are hiding. That human review step isn't optional, especially when you're first building these workflows.

Once you make those edits, feed them back into your instructions. Tell the AI: "Those last five outputs were too long. Cut this section entirely. The overview was mostly fluff. The stuff that matters was buried in this other section."

Now you're using your human review to create better instructions for the next run. Every edit becomes a training signal. Every correction makes the next output tighter.

Stop rubber stamping AI output. Start treating your edits as the instruction layer for your next automation.

[Justin Wright @justinwrightai]

https://t.co/95VYQml8wX

[Justin Wright @justinwrightai]

Starting this summer on Samsung and Pixel phones, Gemini can read what is on your screen, move across apps, and complete multi-step tasks for you.

When an AI agent handles tasks across apps on your behalf, the user never enters the app directly. The engagement layer that funds mobile software becomes invisible.

The company that built the app store is now making it obsolete. Google is cannibalizing its own Play Store to control the AI layer above it.

Full analysis 👇

[Justin Wright @justinwrightai]

https://t.co/RVwFziPvcs

[Justin Wright @justinwrightai]

As I have gone down the rabbit hole of building AI tools and optimizing my stack, one of the more interesting projects has been leveraging LLM’s to run traditional ML experiments in bulk.

I recently built an investing model which combines multiple strategies into one unified portfolio. I’m using machine learning models to train and build the strategy, but I created specialized agents to run the experiments and argue over the results.

The final steps of these experiments always go through human review, but it has allowed me to compress months of experiments into weeks. I was originally inspired by Karpathy’s auto-training repo, and the specialized agents add another layer of process and review.

For a deeper dive into the setup, and how you can apply the same to any domain, check out our latest episode of Mostly Humans on YouTube, Spotify, or anywhere you stream your favorite podcasts today!

[Justin Wright @justinwrightai]

https://t.co/qbtlfnc39f

[Justin Wright @justinwrightai]

3 months ago Musk called Anthropic "evil"

Last week he leased them his supercomputer

Anthropic rented ALL of Colossus 1: 220K Nvidia GPUs, 300+ MW

The deal generates $3-4B/year for SpaceX with $2.5B in profit. And it positions SpaceX as a hyperscaler ahead of the June IPO

3 weeks, one pattern:

Wk 1: Paid $10B for Cursor (Grok can't code)
Wk 2: OpenAI missed targets, Anthropic passed at $30B ARR
Wk 3: Musk becomes the landlord of the company beating him

Full breakdown in Monday Momentum 👇

[Justin Wright @justinwrightai]

Space X lets Anthropic use colossus for additional compute, Codex with GPT 5.5 is more powerful, and OpenAI launches new realtime voice model.

[Justin Wright @justinwrightai]

https://t.co/KbfRcXgnlC

[Justin Wright @justinwrightai]

OpenAI has 900 million weekly users

Anthropic just passed it in revenue

WSJ reported OpenAI missed its own revenue and user growth targets. The CFO privately warned $600B in compute commitments may outpace revenue

Why the gap?

Claude Code holds over half the AI coding market. 1,000+ Anthropic customers spend $1M+/year.

OpenAI built the largest audience in AI history. Anthropic built the products enterprises pay for. The revenue followed the product

Full breakdown in Monday Momentum 👇

[Justin Wright @justinwrightai]

You can find us on YouTube or all major podcast platforms: https://t.co/RVwFziPvcs

[Justin Wright @justinwrightai]

You don't need a startup idea to start building with AI.

You need one thing that annoys you.

It could be a process at work that's obnoxious, a piece of software that doesn't check all the boxes, or something small you do every day that makes you think "there has to be a better way."

Here's my challenge: build your own website. If you're a small business owner using Wix or Squarespace, those tools are powerful, but they're limited. Every site ends up looking kind of the same.

So try this as a thought experiment: Think about what your website could look like if there were no constraints.

Then open an AI tool and start building a prototype. It's a well-defined problem. AI tools are really good at it. And it's a low-stakes way to see just how simple building software has become.

If you want more ways to break down the barriers to using AI in your life, check out my podcast 👇