Charlie Marsh

my team has spent the entire day processing this number from mr @charliermarsh

Python 3.15 is out now.

Here's our overview of what's new, including lazy imports, a new statistical sampling profiler (Tachyon), sentinel types and more.

astral.sh/blog/python-3.…

Python 3.15 is OUT 🐍🚀

Lazy imports, Tachyon in the stdlib, frame pointers on by default, frozendict, better error messages... this one is LOADED 🔥🔥🔥

Go upgrade 🤘

discuss.python.org/t/python-3-15-…

Doesn't look like much but we work very hard to ensure same-day releases for new Python minor versions (like 3.15) in python-build-standalone and uv.

Python 3.15 is out now and available in uv 0.13.0.

Day 5/

Composer predictions in the desktop app. Often leads to a double take with how on point they are. Included in the Pro plans without consuming usage.

My job with composer predictions

These are truly so, so good

We've reduced the uv binary size by over 40% since July (57% on Windows!).

To put that in context, it'll reduce download bandwidth across public mirrors by almost 4 PB per month.

Python 3.15 is now here for real

Interesting because this is effectively how we test all of Ruff, uv, and ty

hyperfine 2.0 is out now! It includes support for alternative performance metrics like peak memory usage, CPU cycles, and more.

I'd say it is a bit trickier than that.

We most likely don't have a type system that can precisely express all JavaScript idioms. Elixir is not as complex as JavaScript in many respects, and we still had to develop a lot of new theory for its type system!

So when you find an unsupported idiom, your options as a type system developer are:

1. Develop new theory and prove it correct (which I am skeptical models can do autonomously today)

2. Reject the idiom, sacrificing compatibility (which would make adoption harder)

3. Being a gradual type system, fall back to dynamic typing

Option 3 then introduces its own trade-offs because dynamic code has to interact with static code (and the more unsupported idioms, more dynamics!). So you need to choose between:

a. Calling static code from dynamic code returns dynamic(), dropping the opportunity of further static typing checks (this is sound)

b. Calling static code from dynamic code returns the static types, potentially leading to mismatched types at runtime (this is unsound)

c. Calling static code from dynamic code returns the static types with additional runtime checks, unifying both behaviours but potentially impacting performance (this is sound)

d. Use the strong arrows theory we developed for Elixir (elixir-lang.org/blog/2023/09/2…) but it may not be a great for highly polymorphic languages like JS

On the other hand, my understanding is that TypeScript made deliberate trade-offs, especially around developer ergonomics, and a new type system could certainly explore different choices aligned with the current trends.

PS: I'm not a type theorist by training, although I have implemented one used at scale. The above my contain inaccuracies!

We've updated our GitHub math repo with 6 new Lean formalizations, 19 modifications, and 3 withdrawals. The repo now has ~42% top-line results formalized. We will continue to update the repo with new formalizations and with any errata we notice.

github.com/openai/math/bl…

Switching to codegen-units = 1 made the uv build ~50% faster, cut peak memory by 67%, and reduced binary size by 17%.

I was really surprised by this... Increasing the number of codegen units typically _improves_ build times.

I’m testing GPT-6 Sol in Chat right now, and I won’t lie, it’s absolutely incredible.

Not just the model’s personality, but the way it organizes its outputs too. The structure feels extremely refined, clean, and visually polished, with elements that genuinely help me understand the content better.

(OpenAI actually seems to have followed through on the promise to listen to our feedback, which I’ve called them out on like 10 times here.)

We're going beyond text and are releasing a new version of GPT-6 to all users in ChatGPT. Lots of model and infra improvements coming together here to scale it to 1.2B users!

You think I'm joking but the uv workspace is now up to 87 crates

I just called Chungus (my OpenAI Dot) on the phone while picking up mexican food and asked it to transfer some Rex benchmarks from last night on my Mac Studio to my MacBook Pro via Tailscale and I just got to my desk and there they are. What a fucking time to be alive.

GPT-6 and Intelligent UI, now rolling out in ChatGPT for everyone.

Intelligent UI in ChatGPT delivers fast, interactive answers that make everyday questions more visual, complex topics easier to grasp, and interactive tools for your task available on the spot.