If you would like a narcissistic confirmation of what a fantastic person you really, *really* are, just ask ChatGPT to "Please describe me based on all of our previous conversations."

Seriously, I wonder if anybody gets a discouraging response. This has the feel of horoscope-level stuff.

The 'readme.txt' files I recently started using in my current project folders have been lifesavers. After 3 weeks away dealing with other stuff, I would be staring blankly at all my data / results files without these simple 'note to self' virtual post-it stickers. Should have started this long ago.

I opened Claude this morning to find 3.7 has just landed, along with a new reasoning functionality:

I got one of those CO2 meters everyone is talking about and was horrified to learn that the CO2 levels in my home and office are fine and that I cannot easily boost my intellectual abilities by opening a window. Stuck with coffee for now.

An unemotional, insightful and clear-eyed analysis. Highly recommended reading.

www.nybooks.com/online/2025/...

A fail from Claude and a win from OpenAI 4o this morning:

Why is there an "Undo" button for just about everything on your computer *EXCEPT* when you delete a tab in Excel???

Alan Thompson's projected timeline for new models this year:

lifearchitect.ai/timeline/

A new study out of Anthropic using exceptionally high-quality data (that they have made freely available). Shows IMO astoundingly low use of AI outside the "Computer & Mathematical" occupational tasks.

www.anthropic.com/news/the-ant...

I think Nate Silver's G x S x P model for human capital is a good framework to start thinking about the effects of AI on the knowledge economy.

www.natesilver.net/p/sbsq-17-ho...

In a big representative sample of humans against GPT-4: "the creative ideas produced by AI chatbots are rated more creative [by humans ]than those created by humans... Augmenting humans with AI improves human creativity, albeit not as much as ideas created by ChatGPT alone” docs.iza.org/dp17302.pdf

Working paper out of Harvard with evidence that AI-based teaching methods are better than the active-learning classroom. Just one study but we can already see signs that early-stage methods of AI teaching > long-established best practice in traditional approaches.

doi.org/10.21203/rs....

Just appeared on my instance of OpenAI's model family. I have a live ML project under development using Claude that I can now parallel-test on 03:

[Jan 2025] Here are my top 5 AI newsletter subscriptions:

Alan D. Thompson | The Memo
Ethan Mollick | One Useful Thing
Zvi Mowshowitz | Don't Worry About the Vase
Jack Clark | Import AI
Alberto Romero | The Algorithmic Bridge

also
Charlie Guo | Artificial Ignorance

DeepSeek is a really good model, but it is not generally a better model than o1 or Claude.

But since it is both free and getting a ton of attention, I think a lot of people who were using free “mini” models are suddenly being exposed to what a early 2025 reasoner AI can do & are surprised.

The "DeepSeek Effect"👇. IMO, DeepSeek is not the giant-killer that the markets think. In my daily workflow, it is a very distant 3rd behind Claude Sonnet 3.5 and OpenAI 4o and o1. The main use-case for DeepSeek: it's FREE. On that benchmark alone, it is very good. But not Claude or 4o or o1 good.

Yes!! [From Jack Clark | Import AI 397]

Every few months, I write an opinionated guide for general purpose users about which AI to pick, especially for newcomers.

Here is my brand new one, which I actually had to update multiple times in the few days I was writing it. Things are changing fast. open.substack.com/pub/oneusefu...

A pretty negative and depressing report. But useful to recent graduates as a guide on what NOT to do as you search for employment. It's a US-based study, but the behaviours described are unfortunately probably universal.

www.intelligent.com/1-in-4-hirin...